An experimentally validated long-range-enabled fuzzy–internet of things framework for real-time and cost-efficient fault detection in photovoltaic systems

Abstract

Reliable and cost-effective fault detection is essential to ensure the safety, efficiency, and long-term stability of photovoltaic (PV) systems. However, most existing diagnostic techniques remain limited to simulation studies or rely on computationally intensive algorithms unsuitable for low-power, real-time embedded environments. This article presents an experimentally validated, IEC 61724-compliant, long-range (LoRa)-enabled fuzzy–Internet of Things (IoT) framework for real-time PV fault detection and diagnosis. The proposed system integrates a custom multi-sensor hardware platform with redundant measurement channels for voltage, current, irradiance, and temperature; an Arduino Mega-based fuzzy inference engine for edge-level fault classification; and LoRa–Firebase connectivity for long-range data transmission and cloud-based visualization. Unlike many existing fuzzy-logic or IoT-based PV monitoring systems that rely primarily on simulation-based validation or cloud-dependent processing, the proposed framework integrates hardware-injected multi-subsystem fault testing, embedded edge-level intelligence, and IEC 61724-1-compliant monitoring within a deployable low-power architecture. To ensure comprehensive validation, the framework was assessed under both Python-based I–V/P–V simulations and hardware fault injection tests, including shading (25%–75%), sensor disconnection, and boost converter faults (metal oxide semiconductor field effect transistor, diode, and inductor degradation). The fuzzy logic diagnostic engine employs 25 optimized rules using trapezoidal and triangular membership functions, achieving robust resilience to noise, irradiance variation, and sensor drift. Experimental results demonstrated a mean diagnostic accuracy of 98.7% ± 1.2% (95% CI) and an average detection delay below 0.5 s. Compared to traditional threshold- and model-based schemes, the proposed method reduced false positives by 12%, while maintaining real-time inference (8 ms) and minimal memory usage (2 KB). The complete 50 W prototype—comprising a PV module, pulse-width modulation controller, lead-acid battery, and custom DC–DC boost converter—was implemented at a total hardware cost of ∼$185 (≈ $3.6/W). The system's hybrid fuzzy–IoT intelligence, low-power LoRa communication, and cloud-based analytics collectively establish a scalable, cost-efficient, and empirically verified architecture for intelligent PV monitoring, bridging the gap between simulation-driven research and practical field-deployable photovoltaic diagnostic systems.

Introduction

The global energy sector is experiencing a significant transformation prompted by the imperative to address climate change, diminish reliance on fossil fuels, and expedite the integration of renewable technology. Photovoltaic (PV) systems have emerged as a highly promising solution due to their scalability, decreasing costs, and minimal environmental impact. The levelized cost of electricity (LCOE) for PV systems has decreased by over 85% in the last decade, positioning solar energy as the most viable renewable source in several countries (IEA-PVPS, 2025; REN21, 2025).

Nonetheless, the reliable and efficient operation of PV systems remains a significant challenge. Their performance is significantly influenced by environmental and operational factors, such as irradiance fluctuations, temperature variations, dust accumulation, and partial shading (Chandan et al., 2021; Jumaboev et al., 2022; Mustafa et al., 2020; Osmani et al., 2023). These stressors result in power losses and physical degradation, including hotspot formation, module mismatch, and electrical failures such as open and short circuits (Hussain et al., 2022). Malfunctions in power electronic subsystems and battery storage components can reduce conversion efficiency by up to 15%, leading to increased maintenance costs and serious safety issues, including overheating and fire hazards.

Undetected flaws can diminish annual energy output by 10%–20% in large-scale systems and negatively impact component longevity (Jathar et al., 2023). Conventional failure detection and diagnosis (FDD) techniques—mainly threshold- and model-based—remain prevalent due to their simplicity and cost-effectiveness (Demir, 2023; Harrou et al., 2018). Nonetheless, these techniques exhibit notable limitations: threshold-based methods often misclassify normal variations as faults in fluctuating irradiance, whereas model-based approaches require complex parameterization and substantial processing power, thus constraining their use in embedded or distributed systems (Abdel-Nasser et al., 2023; Taghezouit et al., 2024).

Recent advancements in artificial intelligence (AI) and the Internet of Things (IoT) have established new opportunities for intelligent monitoring and defect detection. Fuzzy logic methods offer improved adaptability and robustness to noisy inputs, making them suitable for non-linear solar dynamics (Dhimish et al., 2017; Pillai and Rajasekar, 2019). Detection accuracies over 95% have been recorded in several scenarios (Bait et al., 2022; Jenitha and Selvakumar, 2017). Simultaneously, IoT technologies provide real-time, cloud-integrated data acquisition and remote monitoring, enhancing scalability and operational transparency (Ansari et al., 2021; Paredes-Parra et al., 2019). Low-power wide-area network (LPWAN) protocols, such as LoRa and NB-IoT, provide vast coverage of up to 5 km with minimal energy consumption (around 50 mW), enabling large-scale deployment in distributed solar farms.

However, two fundamental limitations persist in the present state-of-the-art. The majority of AI- and fuzzy logic-based PV failure detection systems have mostly been validated only through simulation, lacking empirical verification in real fault scenarios (Aziz et al., 2020; Dairi et al., 2020). This difference engenders concerns about their reliability in actual applications. Secondly, while IoT-based monitoring frameworks improve data collection, they rarely integrate intelligent diagnostic mechanisms or adhere to recognized performance standards, such as IEC 61724-1:2021, which specifies accuracy classes (A, B, and C) for PV monitoring (International Electrotechnical Commission, 2021; SolarPower Europe, 2025). As a result, current systems are fragmented, lacking comprehensive frameworks that incorporate real-world validation, compliance with standards, and economic efficiency.

This study introduces a defect detection and diagnosis platform for fuzzy-IoT that is consistent with IEC standards and incorporates LoRa technology, designed to overcome these limitations through hardware-validated integration. The suggested system features a custom multi-sensor printed circuit board (PCB) for measuring voltage, current, and temperature, connected to an Arduino Mega for data collection and a LoRa transmitter for long-range communication. A fuzzy inference engine enables intelligent fault detection, while Firebase cloud services provide real-time data visualization and analysis.

The system underwent validation through hardware-injected fault tests (including shading, sensor disconnection, and converter malfunction) and Python-based I–V/P–V simulations, attaining a detection accuracy of 97.2% ± 1.3% (IEC Class C) and an average latency of < 0.5 s. The framework decreased false positives by 12% relative to traditional methods and showed resilience to sensor drift and environmental fluctuations. The suggested architecture provides a scalable, empirically verified, and cost-effective solution for intelligent PV monitoring, with a hardware expense of ∼$3.6/W.

Unlike many existing fuzzy-logic or IoT-based PV monitoring systems that primarily focus on data acquisition or simulation-based validation, the proposed framework integrates hardware-validated fault detection, embedded edge-level intelligence, and long-range LoRa communication within a unified monitoring architecture. By combining real-time fuzzy inference on a microcontroller with experimentally injected fault scenarios and IEC 61724-1-compliant monitoring metrics, the proposed system bridges the gap between theoretical diagnostic models and practical field-deployable PV monitoring solutions.

The major contributions of this research are summarized as follows:

Development of a hardware-validated hybrid IoT–fuzzy fault detection and diagnosis (FDD) framework integrating multi-sensor PCB hardware, LoRa wireless communication, and edge-level fuzzy inference, extending beyond simulation-based fuzzy diagnostic systems reported in prior studies.

Experimental validation through hardware-injected multi-component fault scenarios, including PV module faults, sensor failures, and DC–DC converter malfunctions, addressing a key gap in real-world testing of IoT-based PV diagnostic systems.

Implementation of a real-time fuzzy inference engine on a low-power Arduino Mega platform, enabling efficient edge-level fault detection without reliance on cloud-based computation.

Demonstration of IEC 61724-1 Class C monitoring compliance and cost-effective deployment (∼$3.6/W), providing a scalable architecture suitable for residential and distributed PV installations.

This document is organized in the following manner: the Literature review section explores relevant literature; the System architecture section outlines the system architecture; the Methodology section provides a detailed account of the experimental and simulation setup; the Simulation results section clarifies the fuzzy logic detection algorithm; the Experimental validation and the Scalability and field deployment sections provide an analysis of the results and engage in discussion, while the Discussion of results section summarizes the key findings and outlines potential avenues for future research.

Literature review

Faults in PV

PV systems operate under fluctuating irradiance, temperature, and ambient factors, making them vulnerable to various defects (IEA-PVPS Task 13, 2025a). Photovoltaic modules can experience problems such as partial shading, which can reduce power output by 10%–30%; hotspot formation due to localized overheating; and electrical failures, including the potential for open and short circuits (Al Mahdi et al., 2024; IEA-PVPS Task 13, 2025a; Lagoune et al., 2024). Comprehensive modeling and analysis techniques are critical for understanding these degradation mechanisms, optimizing system design (including charge controllers and boost converters), and improving overall energy yields under real-world conditions (Mariprasath et al., 2024b). Power electronic subsystems, particularly DC–DC converters, are prone to failures in MOSFETs, diodes, or inductors, leading to conversion efficiency losses of up to 15% (Bait et al., 2022). Integrated battery systems deteriorate over time, exhibiting capacity reductions of roughly 20% within the first 5 years of use (Paredes-Parra et al., 2019). If neglected, these issues accelerate system degradation, increase maintenance costs (up to $500/kW annually), and threaten the reliability and safety of grid-connected systems (Dhimish et al., 2017).

Traditional fault detection and diagnostic (FDD)

Systems traditional PV monitoring frameworks were historically based on supervisory control and data acquisition (SCADA) architectures, wired dataloggers, and industrial FDD schemes (SolarPower Europe, 2025; Taghezouit et al., 2022). These technologies facilitated dependable data acquisition but were constrained by elevated installation expenses, intricate wiring, and limited adaptability to distributed PV plants. For example, SCADA installations in medium-scale solar farms typically necessitate infrastructure costs over $1000/kW and exhibit latencies exceeding 10 s. Their deficiency in flexibility and cloud integration constrains real-time analysis and remote operation. Conversely, low-power wireless IoT architectures offer adaptable, scalable, and economical solutions that facilitate edge-level analytics for prompt problem identification (Sait et al., 2025).

Conventional FDD methods are chiefly classified into threshold-based and model-based categories. Threshold-based methodologies evaluate sensor metrics (e.g. voltage and current) against established parameters, presenting minimal implementation expenses (about $50 per unit) while demonstrating restricted adaptability. Under variable irradiance conditions, they frequently misconstrue natural variations as faults, resulting in false-positive rates approaching 15% (Kapucu and Cubukcu, 2021). Model-based strategies, utilizing similar circuits or analytical models to forecast anticipated performance, attain 88%–92% accuracy under stable conditions (Lahiouel et al., 2023b). However, these techniques necessitate exact parameterization for processing capabilities above 1 GFLOPS, limiting their use in embedded or real-time monitoring systems. These limitations emphasize the necessity for more adaptable and resource-efficient alternatives.

AI-enhanced techniques for PV monitoring

AI has increasingly been employed for solar fault detection to rectify the limitations of conventional methods. Fuzzy logic systems exhibit significant tolerance to noise and environmental uncertainty, with accuracies of 95%–98% (Pillai and Rajasekar, 2019; Taghezouit et al., 2024). A fuzzy logic approach utilizing distortion ratios has been devised to precisely distinguish 12 unique categories of solar failures (Ksira et al., 2024; Le et al., 2023). In conjunction with fuzzy logic, techniques in machine learning (ML) and deep learning (DL), including support vector machines (SVMs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs), have demonstrated accuracies between 90% and 97% (Imamul et al., 2024). Recent studies have also explored ML approaches such as decision trees and ensemble-based models for electrical fault detection and localization, demonstrating competitive diagnostic performance when sufficient labeled datasets and computational resources are available (Beethi et al., 2024).

However, the deployment of these models in embedded edge environments typically requires labeled training datasets, parameter tuning, and memory allocation for model storage. In addition, maintaining diagnostic reliability under changing irradiance and temperature conditions may necessitate periodic retraining or recalibration. These requirements can increase implementation complexity and maintenance overhead in low-power microcontroller-based IoT nodes. Consequently, rule-based fuzzy inference systems remain attractive for edge-level PV diagnostics due to their deterministic behavior, interpretability, and minimal computational overhead. Furthermore, a significant proportion of research is limited to simulations, lacking adequate validation through practical hardware testing.

IoT-enabled surveillance systems

The use of IoT technology has revolutionized solar monitoring by enabling comprehensive and energy-efficient data transmission. Low-power wide-area network (LPWAN) protocols, such as LoRa and NB-IoT, exhibit ranges of 2–5 km and a power consumption of ∼40 mW, making them suitable for distributed solar farms. Cloud solutions like Firebase facilitate real-time data storage, visualization, and remote access. Nonetheless, most current IoT frameworks are limited to data collecting and presentation, lacking integrated intelligent fault detection functionalities. Field deployments indicate limits, including packet losses of 15%–20% in remote regions and latency above 1 s in extensive systems, thereby compromising real-time responsiveness (Moreno-Garcia et al., 2016).

Recent frameworks have aimed to rectify these deficiencies via energy-conscious analytics and adaptive maintenance algorithms. For example, IoT-enabled sustainable energy management systems integrating renewable power forecasting and adaptive load strategies through bio-inspired optimization techniques have demonstrated improved operational efficiency and intelligent resource allocation in distributed energy environments (Sadiq Ali et al., 2026). The LP-OPTIMA framework presents a prescriptive maintenance and optimization technique for low-power embedded IoT resources, enhancing reliability and energy efficiency in distributed networks (Papaioannou et al., 2024). The Advanced Proactive Anomaly Detection in Multi-Pattern Home Appliances method incorporates ML intelligence into IoT systems for dynamic anomaly detection and operational enhancement (Dimara et al., 2024).

Furthermore, advanced IoT architectures incorporating deep clustering networks and attention-based optimization algorithms have shown enhanced predictive performance in smart applications, although often at the expense of increased computational complexity and cloud dependency (Nidamanuri et al., 2026). These findings highlight the significance of incorporating computational intelligence at the edge layer, with the current emphasis on hybrid IoT-fuzzy systems that combine low-power functionality with real-time diagnostic capabilities.

Modern advanced methodologies and research shortcomings

Recent research has investigated advanced approaches, including protective relays (Lodhi et al., 2022), signal-processing models (Van Gompel et al., 2022), and performance benchmarking (Mahboob et al., 2025). Although effective in enhancing anomaly detection, these systems often face drawbacks such as response times above 15 s, costs over $600, and limited automation functionalities. Modern AI methodologies, including artificial neural networks (ANNs), hybrid CNN–SVM frameworks, and ensemble algorithms, have achieved accuracies between 92% and 98%. Many are restricted by datasets that are solely simulated, lack scalability, or incur prohibitive processing costs that hinder application on embedded devices. IoT-integrated monitoring solutions (Abdelmoula et al., 2023; Gomaa et al., 2023; Hamza et al., 2025; Hojabri et al., 2022; Lahiouel et al., 2022; Latoui and Daachi, 2023; Latreche et al., 2022; Liu and Wu, 2025; Livera et al., 2022; Lodhi et al., 2023; Manjunath et al., 2017; Mellit et al., 2023; Mellit and Kalogirou, 2021; National Renewable Energy Laboratory, 2018; Ramírez et al., 2024; Salameh et al., 2025; Sutikno et al., 2021; IEA-PVPS Task 13, 2025b) provide real-time PV surveillance; however, they frequently have inadequate fault coverage, signal degradation over extended distances, and a deficiency of empirical validation through hardware fault injection. Recent practical implementations of IoT-enabled smart power quality analysis in three-phase electrical systems further demonstrate the feasibility of real-world deployment, yet primarily emphasize monitoring and analytical assessment rather than integrated intelligent fault classification at the edge layer (Mariprasath et al., 2024b).

Table 1 summarizes the comparative performance of existing PV fault detection methods, highlighting their range, accuracy, limitations, and suitability for IoT-based implementations.

Table 1.

Comparison of advanced PV fault detection methods.

Ref. no.	Method/technique	Year	Faults detected	Accuracy/performance	IoT/edge device	Scope/limitations
Le et al. (2023)	Hardware-in-the-loop PV fault detection system	2023	PV module electrical faults	High detection reliability reported	HIL simulation platform	Requires simulation hardware; limited field validation
Ksira et al. (2024)	Embedded real-time PV fault diagnosis system	2024	Module electrical faults	Real-time diagnostic performance	Embedded controller	Focus on module-level faults only
Imamul et al. (2024)	Field-aged PV degradation analysis	2024	Performance degradation	Experimental performance evaluation	N/A	Focus on degradation analysis rather than automated detection
Lahiouel et al. (2023b)	Fuzzy-logic-based fault diagnosis	2023	PV electrical faults	High classification reliability	N/A	Requires parameter tuning
Aziz et al. (2020)	CNN-based fault classification	2020	Multiple PV faults	High classification accuracy reported	N/A	Requires a labeled training dataset
Kapucu and Cubukcu (2021)	Ensemble learning for PV fault diagnosis	2021	String faults	High classification accuracy	N/A	Computational complexity for large datasets
Lodhi et al. (2022)	AdaBoost ensemble PV fault detection	2022	Array electrical faults	Improved detection reliability	N/A	Dependent on dataset quality
Van Gompel et al. (2022)	RNN-based satellite fault detection	2022	System anomalies	High predictive capability	Cloud processing	Satellite data dependency
Mahboob et al. (2025)	Deep learning-based semantic segmentation	2024	Cell-level defects	High detection accuracy	N/A	High computational requirements
Ansari et al. (2021)	Monitoring technologies for PV systems	2021	General monitoring faults	Comparative performance evaluation	IoT monitoring platforms	Review study; no experimental implementation
Mellit et al. (2023)	IoT-based ML monitoring system	2023	PV array faults	Reliable monitoring performance	Embedded IoT system	Requires a stable communication infrastructure
Paredes-Parra et al. (2019)	LoRa-based PV monitoring network	2019	Electrical anomalies	Reliable long-range monitoring	LoRa nodes	Focus on monitoring rather than fault classification
Polymeropoulos et al. (2024)	Vision-based PV monitoring techniques	2024	Shading, cracks	High detection capability	UAV/vision systems	High cost and processing requirements
Da Silva et al. (2025)	ML-based I–V curve fault detection	2025	Shading, electrical faults	Effective ML classification	N/A	Requires large datasets
Ben hmamou et al. (2024)	Sensor-based PV characterization system	2024	Performance anomalies	Experimental monitoring results	Sensor acquisition system	Not designed for automated classification
Abdelsattar et al. (2025)	ML techniques for PV prediction and fault detection	2025	Multiple faults	High prediction accuracy	N/A	Computational complexity
Ramírez et al. (2024)	IoT + CNN aerial monitoring	2024	Module faults	High detection accuracy	UAV + IoT	Requires aerial inspection infrastructure
Mellit and Kalogirou (2021)	AI + IoT PV monitoring review	2021	General faults	Survey of AI diagnostic techniques	IoT monitoring platforms	Review study
Paredes-Parra et al. (2019)	LoRa-based IoT PV monitoring	2019	Current, voltage anomalies	Reliable long-range data acquisition	LoRa IoT nodes	No built-in fault classification
Driesse et al. (2015)	Electrical monitoring methodology	2015	Current / voltage anomalies	Monitoring methodology study	Sensor measurement systems	No automated diagnostic algorithm
—	Hybrid IoT–fuzzy logic FDD (this work)	2025	Multiple hardware-injected faults	97.2% ± 1.3% (IEC Class C)	ESP32–LoRa Edge Node	Hardware-validated; low cost ($3.6/W)

PV: photovoltaic; IoT: Internet of Things; HIL: Hardware-in-the-Loop; CNN: convolutional neural network; RNN: recurrent neural network; ML: machine learning; AI: artificial intelligence; FDD: fault detection and diagnosis; N/A: not applicable; UAV: unmanned aerial vehicle.

Unlike high-complexity ML frameworks that often require cloud-level computation or GFLOPS-scale processing resources (Mahboob et al., 2025; Van Gompel et al., 2022), the proposed hybrid IoT–fuzzy framework achieves competitive diagnostic accuracy while maintaining full edge compatibility and minimal computational overhead. In addition, the system is experimentally validated through hardware-injected fault scenarios and integrates long-range LoRa communication for distributed PV monitoring. These characteristics distinguish the proposed approach from existing fuzzy-only or IoT-only PV diagnostic systems.

Operations and maintenance (O&M) standards and IEC 61724 compliance

O&M practices are increasingly standardized under IEC 61724-1:2021 (International Electrotechnical Commission, 2021), which classifies PV monitoring systems into three accuracy classes:

Class A (<1% uncertainty)

Class B (<3%)

Class C (<5%).

These classes define permissible error margins for parameters such as irradiance, voltage, current, and power. The proposed fuzzy-IoT FDD system attains an overall combined measurement uncertainty below 5%, corresponding to Class C compliance. Achieving Class A compliance would require precision-grade sensors (e.g. pyranometers and RTDs) and higher-resolution acquisition hardware (≥16-bit ADC), identified as a target for future work.

Even with recent advances in PV diagnostics (IEA-PVPS, 2025; Manjunath et al., 2017; National Renewable Energy Laboratory, 2018), several issues persist:

Simple threshold or model-based checks often misread normal swings in sunlight (Voutsinas et al., 2023).

ML tools require large data sets and heavy processing (Babes et al., 2024).

Many IoT platforms collect data but skip onboard smart diagnostics or formal accuracy checks (Polymeropoulos et al., 2024).

Most studies remain confined to simulation environments; few validate performance using real hardware-injected fault scenarios (National Renewable Energy Laboratory, 2018).

These limitations highlight the need for a practical PV diagnostic framework that combines real-time monitoring, embedded intelligence, and experimental validation under realistic operating conditions. This study addresses these challenges by introducing a hybrid IoT–fuzzy FDD framework that integrates edge-level fuzzy inference with long-range LoRa communication and hardware-based fault injection testing. The proposed architecture enables reliable real-time fault detection while maintaining low computational complexity and compliance with IEC 61724-1 monitoring standards.

System architecture

The proposed FDD architecture is implemented on a standalone PV testbed consisting of a 50 W monocrystalline solar module, a pulse-width modulation (PWM) charge controller, a 12 V/7 Ah lead-acid battery, a custom-designed DC–DC boost converter, and a resistive load (Mariprasath et al., 2024a). Each subsystem includes dedicated sensors and redundant measurement channels to ensure reliable monitoring and effective fault isolation (Da Silva et al., 2025; Mellit et al., 2023).

PV subsystem: Voltage and current sensors measure electrical parameters, while temperature and irradiance sensors capture environmental conditions. A redundant voltage divider is added to maintain operability in case of primary sensor failure (Da Silva et al., 2025).

Battery subsystem: A calibrated voltage sensor and a backup divider monitor battery voltage, enabling immediate detection of undervoltage, overvoltage, or disconnection (Abdelsattar et al., 2025; Da Silva et al., 2025).

Boost converter subsystem: Sensors for input/output voltage and current are integrated alongside a feedback voltage channel for duty-cycle control. This allows detection of MOSFET faults, diode deterioration, and inductor damage (Abdelsattar et al., 2025; Da Silva et al., 2025).

All sensor signals are routed through a custom PCB and processed locally by an Arduino Mega 2560 running the fuzzy logic classifier in real time (Abdelsattar et al., 2025). Results are transmitted using a LoRa SX1278 module to an ESP32-S3 gateway, which integrates LoRa, Wi-Fi, and Bluetooth. The gateway uploads data to Firebase, enabling a web dashboard with both numerical indicators (voltage, current, temperature, irradiance, and fault flags) and dynamic time-series plots (I–T, V–T, and P–T) (Madeti and Singh, 2017) (Figure 1).

Figure 1.

System architecture of the proposed photovoltaic (PV) fault detection system.

Hardware architecture

The hardware architecture consists of four subsystems: a 50 W monocrystalline PV module, a PWM charge controller, a 12 V/7 Ah lead-acid battery, and a custom DC–DC boost converter (Lodhi et al., 2023). All subsystems connect to a custom-designed multi-sensor PCB that hosts voltage dividers, Hall-effect current sensors, temperature sensors, and a BH1750 irradiance module (Da Silva et al., 2025). Redundant voltage measurement channels are included wherever the loss of a primary sensor would compromise diagnostic reliability (Da Silva et al., 2025).

An Arduino Mega 2560 serves as the local processing unit, sampling all analog and digital sensors at 100 Hz using its 10-bit ADC (4.88 mV/step) (Mariprasath et al., 2024b). The microcontroller applies basic filtering, calculates electrical power, generates mismatch indicators, and executes the fuzzy logic classifier in real time. Sensor readings and diagnostic flags are transmitted using a LoRa SX1278 transceiver (433 MHz, −139 dBm sensitivity) to a LoRa-enabled ESP32-S3 gateway, which forwards data to Firebase through Wi-Fi (Abdelsattar et al., 2025). This hybrid LoRa–Wi-Fi link maintains long-range communication (5–10 km) with <1% packet loss (Abdelsattar et al., 2025).

To ensure signal quality, all PCBs were designed with shielding, grounding, and short-trace routing to suppress electromagnetic interference from the 100 kHz switching of the boost converter (Madeti and Singh, 2017). The full prototype consumes <100 mW during active sensing and costs ∼180 USD to assemble, providing a low-cost and scalable platform suitable for both laboratory and outdoor PV installations (Abdelsattar et al., 2025).

PV monitoring subsystem

The 50 W monocrystalline PV module (V_oc 27 V and I_sc 2.5 A) is monitored via four sensing channels: voltage, current, temperature, and irradiance (Mellit et al., 2023). A primary voltage divider and a redundant divider guarantee uninterrupted measurement despite sensor failure (Da Silva et al., 2025). The ACS712 Hall-effect sensor records PV current (Abdelsattar et al., 2025), whereas the DS18B20 measures back-surface temperature to detect hot-spot formation (Madeti and Singh, 2017). The BH1750 irradiance sensor offers environmental context, allowing the system to differentiate between natural irradiance variations and atypical shading (Madeti and Singh, 2017). Real-time power is calculated using measured voltage and current, whereas predicted power is obtained from irradiance and panel STC ratings (Da Silva et al., 2025). The difference between these values creates a mismatch index (MI) utilized directly by the fuzzy logic classifier (Da Silva et al., 2025).

Battery monitoring system

The lead-acid battery is monitored with a calibrated voltage sensor augmented by a secondary voltage divider to ensure observability in the event of sensor failure (Voutsinas et al., 2023). Continuous sampling facilitates the prompt identification of deep discharge (14.4 V) or open-circuit anomalies (Livera et al., 2022). The battery voltage is verified against the photovoltaic input and load usage to ensure consistency and minimize false alarms during transient conditions (Polymeropoulos et al., 2024).

DC–DC boost converter subsystem

The bespoke boost converter functions at 100 kHz and increases the 12 V battery output to a stabilized 24 V (Mariprasath et al., 2024b). Four sensors observe its performance: input voltage, input current, output voltage, and output current (Babes et al., 2024). These signals indicate degradation of the MOSFET and diode, failures in the inductor, and anomalies in the load (Voutsinas et al., 2023). Recent advances in high-voltage-gain boost converter topologies, particularly those incorporating cuckoo search optimization-based MPPT controllers, have demonstrated superior performance under partial shading and dynamic conditions through improved voltage gain, reduced output ripple, faster tracking speed, and higher overall efficiency (Mallala et al., 2025). A feedback voltage line facilitates steady duty-cycle regulation and allows for the detection of control-loop problems (Polymeropoulos et al., 2024). All converter measurements are obtained via a unified 100 Hz acquisition pipeline, guaranteeing synchronized detection across subsystems (Polymeropoulos et al., 2024).

Control and communication framework

The Arduino Mega 2560 conducts local preprocessing, computes electrical and environmental metrics, and implements the fuzzy inference system (Ramírez et al., 2024). It interfaces with the LoRa SX1278 module for extended-range, low-power transmission to an ESP32-S3 gateway (Babes et al., 2024). The gateway transmits data to Firebase, facilitating a real-time dashboard that exhibits electrical metrics, environmental variables, and trouble codes (Ramírez et al., 2024). Rapid local classification integrated with cloud-based visualization facilitates both edge autonomy and remote oversight (Polymeropoulos et al., 2024).

Printed circuit board and integration

All subsystems are executed on specialized PCBs engineered for low-noise functionality (Ramírez et al., 2024). Integrated signal routing, ground planes, and shielding reduce electromagnetic interference from the converter's switching stage (Mariprasath et al., 2024b). The modular design enables straightforward repair of defective components and allows for expansion to multi-string photovoltaic setups (Ramírez et al., 2024) (Figure 2).

Figure 2.

Full physical setup of the photovoltaic (PV) fault detection system.

Table 2 summarizes the primary hardware components utilized in the developed fuzzy–IoT fault detection system, outlining their key specifications, functional roles, and associated costs per subsystem.

Table 2.

Summary of hardware components and cost breakdown.

Subsystem	Component	Specification	Function	Approx. cost (USD)
PV module	Monocrystalline 50 W panel	Voc = 27.0 V, Isc = 2.5 A	Main power generation unit	$45
Microcontroller unit (MCU)	Arduino Mega 2560 (original)	16 MHz, 8-bit MCU, 10-bit ADC	Central data acquisition and control logic	$25
Wireless transmitter	LoRa SX1278 module	433 MHz, SPI interface, 3–5 km range	Transmits sensor data to LoRa32 gateway	$6
Gateway node	LoRa32 (ESP32 + SX1276)	433 MHz dual-core, Wi-Fi + LoRa	Receives LoRa packets and uploads to cloud (Firebase)	$15
Voltage sensors	Resistor divider + buffer op-amp	0–30 V input, ±1% accuracy	PV and battery voltage measurement	$8 (2 × $4)
Current sensors	ACS712 Hall-effect (5 A)	185 mV/A, ±1.5%	PV and load current sensing	$10
Temperature sensors	DS18B20 (digital 1-Wire)	−55 to +125 °C, ±0.5 °C	PV module and battery temperature monitoring	$6
Irradiance sensor	BH1750 (I2C lux sensor)	0–65,000 lx, ±5%	Solar irradiance measurement (converted to W/m2)	$5
DC–DC boost converter	Custom-built MOSFET-based converter	Input 12 V, Output 24–50 V, 100 kHz PWM, 85–90% eff.	Boost stage for power conditioning and fault testing	$12 (components only)
Power supply unit	12 V 7 Ah sealed lead-acid battery + PWM controller	Three-port design (PV, battery, and load)	Energy buffer and backup	$28
Printed circuit board (PCB)	Custom multi-sensor board	2-layer FR4, integrated connectors	Signal routing and acquisition	$12
Test load	Adjustable resistive load bank	12–24 V, 0.5–2 A	Load simulation for testing	$8
Cables and casing	—	Shielded wiring + acrylic enclosure	System integration and protection	$8
Total estimated prototype cost	—	—	—	≈ $185–195 USD per unit

PV: photovoltaic; PWM: pulse-width modulation.

The total cost of the proposed fuzzy–IoT PV monitoring system is ∼$182–200 per unit, corresponding to $3.6 per monitored watt for the 50 W prototype. In comparison, commercial PV monitoring systems—such as SCADA-based or industrial datalogger platforms—typically range between $800 and $1200/kW, excluding installation costs. Although the proposed system's low-cost sensors (ACS712, BH1750, DS18B20) correspond to IEC 61724 Class C accuracy rather than Class A or B, this trade-off enables affordability and scalability for small to medium PV setups. Furthermore, maintenance and communication overhead are minimized through the use of LoRa-based transmission and modular design, allowing the architecture to scale economically for residential and distributed microgrid deployments.

Software and data acquisition

The software architecture is implemented on the Arduino Mega 2560, which functions as the central unit for data collecting, preprocessing, and anomaly identification (Duraisamy et al., 2023). The firmware, created in C++ with the Arduino IDE, is tuned for real-time monitoring, low-latency sampling, and organized data management (Duraisamy et al., 2023). The system captures voltage, current, irradiance, and temperature signals at a frequency of 100 Hz to guarantee high-resolution measurements (Triki-Lahiani et al., 2018). Analog channels (voltage and current) are calibrated using empirically obtained correction factors, minimizing systematic errors to below ± 0.1 V for voltage and ± 0.05 A for current (Triki-Lahiani et al., 2018). Digital sensors, like the BH1750 (irradiance) and DS18B20 (temperature), deliver intrinsically calibrated data, hence improving measurement reliability under fluctuating environmental conditions (Driesse et al., 2014).

A five-point moving average filter is employed on analog inputs before digitization to reduce noise interference, successfully attenuating high-frequency variations while maintaining dynamic transitions in PV behavior (Duraisamy et al., 2023). In addition to basic sensing, the firmware calculates both the actual power and the estimated power of the PV module. The actual power is determined from the instantaneous panel voltage (V_pv) and current (I_pv) as follows:

$$P_{real}=V_{pv}\times I_{pv}$$

(1)

The estimated power is calculated from irradiance measurements, panel ratings under standard test conditions (STC), and reference irradiance (G_STC=1000 W/m²):

$$P_{est}=P_{STC}\times {\displaystyle \frac{G}{G_{STC}}}$$

(2)

The disparity between these two readings is utilized to calculate an MI, which functions as an immediate diagnostic indicator for anomalies, including partial shading, hot areas, or sensor drift:

$$MI={\displaystyle \frac{|P_{real}-P_{est}|}{P_{est}}\times 100\text{\%}}$$

(3)

Collected measurements are packaged into structured data frames consisting of node ID, timestamp, PV voltage and current (V_pv, I_pv), battery voltage (V_bat), converter voltage and current (V_conv, I_conv), irradiance (G), and temperature (T) (Silvestre et al., 2013). To maintain continuity in data transmission, a dual-buffer acquisition system is employed, where one buffer captures new data while the other transmits via the LoRa SX1278 module (Driesse et al., 2014). This design reduces packet latency by ∼15% compared to sequential acquisition–transmission schemes (Driesse et al., 2014).

In addition, the Arduino performs edge-level anomaly detection, flagging abnormal conditions such as battery undervoltage (<10.5 V), converter deviations, or PV current reductions inconsistent with measured irradiance (Lahiouel et al., 2023a). This local processing decreases unnecessary cloud traffic by ∼12% and ensures that only relevant, preprocessed data are forwarded to the fuzzy logic module and subsequently to the Firebase cloud through the ESP32-S3 gateway (Bade et al., 2025; Voutsinas et al., 2023). This hybrid edge–cloud strategy balances local intelligence with scalable IoT connectivity, enabling fast, reliable, and resource-efficient monitoring (Bade et al., 2025).

To guarantee data integrity, synchronization, and security among distributed nodes, the communication protocol incorporates an effective verification layer above the LoRa SX1278 module (Abdelmoula et al., 2023; Hamza et al., 2025; National Renewable Energy Laboratory, 2018). Every transmitted frame contains a 16-bit cyclic redundancy check (CRC) for error detection and an integrated timestamp for temporal synchronization (Hamza et al., 2025). The gateway (LoRa32) authenticates each packet and generates acknowledgment (ACK) or negative acknowledgment (NACK) responses, initiating retransmission solely upon fault detection. This method reduces packet loss to around 3% in evaluations performed at a distance of 2 km (Abdelmoula et al., 2023). Node clocks are consistently synchronized with gateway timestamps, maintaining temporal accuracy within ±200 ms. The mean end-to-end latency from the sensor node to Firebase was measured at 0.6 s, mostly constrained by LoRa's duty-cycle limitations (National Renewable Energy Laboratory, 2018). To guarantee data security, all payloads are encrypted with AES-128 before transmission, enabling secure communication with low computational strain on the Arduino Mega (Hamza et al., 2025).

Communication performance was evaluated under outdoor deployment conditions using the LoRa SX1278 module at 433 MHz with a payload of ∼120 bytes per frame. At a tested outdoor distance of 2 km, the system achieved a packet delivery ratio (PDR) of ∼97%, corresponding to a packet loss of about 3%, with a mean end-to-end latency of 0.6 s from node transmission to Firebase dashboard update. Accordingly, the effective outdoor range validated in this study was 2 km under the stated field conditions. These results represent the measured communication performance of the deployed prototype and should be distinguished from broader theoretical LoRa range capabilities reported in the literature.

Sensor calibration, resolution, and measurement uncertainty

All sensors were calibrated before experimentation to guarantee measurement reliability. Voltage and current sensors (voltage divider and ACS712) were evaluated against a calibrated UNI-T UT804 digital multimeter (accuracy ±0.5% for voltage and ±0.8% for current), while irradiance measurements from the BH1750 were verified with a Kipp & Zonen CMP3 reference pyranometer (Class C) under clear-sky conditions (Livera et al., 2022; Manjunath et al., 2017; National Renewable Energy Laboratory, 2018). Temperature measurements from the DS18B20 digital sensor were corroborated with a PT100 RTD reference probe (International Electrotechnical Commission, 2021).

To verify compliance with IEC 61724-1 Class C (International Electrotechnical Commission, 2021) monitoring requirements, each sensing channel was calibrated across the expected operating range of the prototype system. Multiple operating points were generated using controlled load and irradiance conditions, and repeated measurements were recorded at each point. Sensor readings were compared with the corresponding reference instruments (UNI-T UT804 for voltage/current, CMP3 for irradiance, and PT100 RTD for temperature), and the absolute/relative deviations were quantified. The combined monitoring uncertainty was then computed using root-sum-square (RSS) propagation of the individual measurement uncertainties. Based on this calibration-and-verification procedure and the resulting RSS uncertainty estimates, the monitoring chain satisfies IEC 61724-1 Class C-level accuracy requirements (≤ 5% uncertainty) for PV performance monitoring, supporting the manuscript's compliance claim.

The combined measurement uncertainty, evaluated through RSS propagation, was determined as follows (Abdelmoula et al., 2023; Latoui and Daachi, 2023):

Voltage: ±1.2%

Current: ±1.8%

Power: ±2.1%

Irradiance: ±4.5%

Temperature: ±0.6 °C

Given that the Arduino Mega 2560 employs a 10-bit ADC (1024 levels) across a 5 V reference, the effective resolution per quantization step is ∼4.9 mV. For the measured voltage range (0–25 V via divider), this yields a resolution of roughly 0.12% of full scale (Driesse et al., 2014, 2015). Therefore, voltage deviations below this threshold (e.g. fine soiling <10%) may not be fully discernible, but performance deviations above 2%–3% remain reliably detectable. This quantization effect defines the lower limit of anomaly sensitivity for IEC 61724-1 Class C compliance (International Electrotechnical Commission, 2021; Triki-Lahiani et al., 2018).

A comparison with industrial-grade Class B sensors (e.g. LEM LV25-P voltage transducers and Kipp & Zonen CMP11 pyranometers) indicates a 2–3× improvement in precision but at over 5× higher cost (Ben hmamou et al., 2024; International Electrotechnical Commission, 2021). The present configuration, therefore, offers a pragmatic compromise between accuracy and affordability, aligning with IEC Class C monitoring performance standards (International Electrotechnical Commission, 2021; National Renewable Energy Laboratory, 2018).

Methodology

The proposed FDD methodology combines hardware sensing, fuzzy logic inference, and simulation-based validation to ensure both theoretical rigor and experimental reliability. The overall structure consists of four sequential processes:

Development of a fuzzy logic framework: The input variables (voltage, current, irradiance, and temperature) are defined with their corresponding membership functions. Diagnostic rules are formulated to embed expert knowledge of PV operating behavior under normal and faulty conditions (Bait et al., 2024).

Fault detection system: A decision-making layer processes the fuzzy outputs to classify the system status as normal operation, reduced performance, or a specific fault condition (Bait et al., 2024). This layer accommodates sensor noise, irradiance fluctuations, and other environmental uncertainties, improving robustness.

Fault injection approach: Controlled hardware experiments are conducted by introducing real faults such as shading, open circuit, short circuit, and sensor disconnections. Parallel simulated faults are created in Python to ensure reproducibility and provide reference signatures for comparison (Bait et al., 2024).

Simulation modeling: Python is used to generate I–V and P–V curves across varying irradiance and temperature conditions, serving as a theoretical baseline to validate fuzzy logic decisions and compare them with expected PV behavior (Driesse et al., 2015).

This hybrid methodology ensures thorough validation: simulation provides controlled repeatable behavior, while real fault injection verifies the system under practical, non-ideal conditions. Together, these layers guarantee methodological rigor and dependable diagnostic performance.

Diagnostic layer based on fuzzy logic

The suggested diagnostic methodology employs a fuzzy logic reasoning framework to evaluate sensor signals and derived performance indicators in real time (Bendaas et al., 2023). This method has been chosen over traditional deterministic thresholding techniques due to the inherent stochastic variations in environmental factors affecting PV systems, including irradiance, shading, and temperature changes. Furthermore, inaccuracies in sensor measurements, minor calibration discrepancies, and nonlinear PV responses can result in incorrect decision bounds. In contrast, fuzzy logic is particularly appropriate for these scenarios, as it embraces uncertainty, allows for overlapping domains between healthy and defective states, and yields a nuanced classification instead of a strict binary conclusion (El Robrini et al., 2025). In this system, fuzzy inference functions as the principal diagnostic layer, facilitating efficient fault diagnosis, degradation identification, and robustness to transient shocks.

Input parameters:

The fuzzy diagnostic engine obtains inputs from both direct physical measurements and derived performance indicators, providing a comprehensive perspective of the PV system's condition.

Direct measurements: These encompass PV module voltage (V_pv), PV current (I_pv), irradiance (G), and module surface temperature (T). All of these variables are perpetually observed, adjusted, and refined at the data acquisition layer. Voltage and current measurements are vital for calculating instantaneous electrical power, but irradiance and temperature are crucial for defining anticipated photovoltaic performance under STC.

Derived indicators: In addition to raw data, the system calculates an MI that measures the disparity between recorded real power (P_real) and estimated theoretical power (P_est). The estimated power is derived from irradiance, temperature modifications, and defined panel specifications, while the real power is ascertained directly from measured voltage and current. The MI functions as a redundancy-based health metric: under normal operation, the difference between P_real and P_est remains within an acceptable range, while significant deviation indicates shading, wiring problems, or sensor drift (El Robrini et al., 2025).

The integration of direct measurements with computed indices enhances the resilience of the fuzzy diagnostic engine. Even if a sensor malfunctions or generates incorrect results, the redundancy offered by derived indications guarantees that defect detection remains precise and dependable. This hybrid input framework establishes a basis for resilient classification in uncertain and variable environmental settings.

The proposed framework distinguishes sensor faults from PV/system faults by using redundancy between the electrical measurements $(V_{pv},I_{pv})$ , environmental references $(G,T)$ , and the derived MI, which compares measured power to expected power. A single sensor anomaly (e.g. voltage- or current-sensor disconnection/drift) typically produces physically inconsistent relationships between measured electrical power and the irradiance/temperature-based expected power; this inconsistency results in MI patterns and fuzzy-rule activations that differ from genuine PV fault signatures such as shading. This behavior is supported by the dedicated experimental cases “Failure of Voltage Sensor” and “Failure of Current Sensor,” as well as the compound scenario “75% Shading with Voltage-Sensor Drift,” demonstrating that the system can identify sensor anomalies while preserving PV fault detection when independent measurements remain available. However, if multiple critical sensors fail simultaneously such that both measured power and expected power become unreliable (e.g. concurrent corruption of voltage and current), unique fault isolation may become ambiguous; in such cases, the framework reports a sensor-integrity/uncertain state rather than attributing the behavior solely to PV hardware faults.

Fuzzification

Fuzzy logic employs a fuzzification process to transform exact numerical values into linguistic categories through predefined membership functions. This phase is crucial because PV system parameters rarely operate inside entirely discrete states; instead, they transition fluidly across operational zones.

Voltage is classified into low, normal, and high categories. A “low” voltage signifies possible disconnection or shading, “high” may indicate overcharging or converter failure, and “normal” denotes expected functionality.

Current is categorized as zero, normal, or overcurrent. A zero current under high irradiance strongly implies load separation, while overcurrent may signify a short circuit or converter malfunction.

Irradiance is categorized as low, moderate, or high, reflecting environmental parameters that affect solar power generation potential.

Temperature encompasses normal, raised, and high classes. Increased temperatures may signify hotspots or diminished efficiency, while high levels signal critical fault situations.

The MI is categorized as acceptable, moderate, or severe. A satisfactory MI verifies anticipated performance, however a significant deviation reveals irregularities such as partial shading or sensor degradation.

The fuzzification method enables the diagnostic engine to convey uncertainty through different levels of membership. A module temperature of 58 °C may partially belong to both the “Elevated” and “High” categories, both having unique membership values. This overlapping classification bypasses stringent requirements, averting minor discrepancies from triggering false alarms while still recognizing significant departures from standard operation (Bouguerra et al., 2025).

Rule base

The fuzzy rule base plays an essential role in the diagnostic framework, bringing together expert insights on the functioning and potential failure causes of PV systems through a structured inference approach. In contrast to deterministic threshold-based methods, fuzzy rules allow for the simultaneous consideration of multiple input indicators. This capability enables the system to identify complex and non-linear relationships among different conditions of irradiance, temperature, and load. This approach enables precise identification of faults and enhances the system's ability to withstand changes in the environment and potential sensor malfunctions.

Formal representation of rules

Each fuzzy rule R_k can be expressed in the general form as follows:

$$R_k:IF(x_1\in A_i)\wedge (x_2\in B_j)\wedge (x_3\in C_m)\Rightarrow (y\in D_n)$$

(4)

where:

• x₁, x₂, and x₃ are the input variables (e.g. voltage V_pv, current I_pv, irradiance G, module temperature T, and mismatch index MI).

• A_i, B_j, and C_m are linguistic fuzzy sets (e.g. low, normal, and high).

• y is the output diagnostic decision (system state).

• D_n is the output fuzzy set (e.g. normal operation, degraded performance, and fault condition).

The firing strength of each rule is computed by fuzzy intersection (minimum operator):

$${\mu }_{R_k}=min({\mu }_{A_i}(x_1),{\mu }_{B_j}(x_2),{\mu }_{C_m}(x_3),\dots )$$

(5)

where ${\mu }_{A_i}$ denotes the membership degree of input x in the fuzzy set $A_i$ .

The aggregated fuzzy output set is obtained using the maximum operator:

$${\mu }_{out}(y)=max({\mu }_{R_k}(y))$$

(6)

Finally, crisp diagnostic decisions are derived through centroid defuzzification:

$$y*={\displaystyle \frac{{\int }_{}y\cdot {\mu }_{out}(y)dy}{{\int }_{}{\mu }_{out}(y)dy}}$$

(7)

This formal framework ensures smooth and interpretable classification across overlapping fault categories, reducing the risk of misclassification due to abrupt threshold boundaries.

Representative diagnostic rules

The rule base was carefully constructed to include the most prevalent operating states and failure scenarios in solar systems. It combines direct sensor indicators with computed diagnostic indices to improve redundancy and reliability. A collection of representative rules is provided below, illustrating how the system recognizes both normal and erroneous conditions.

Rule 1—Normal Operation:

$$IF(V_{pv}isNormal)\wedge (I_{pv}isNormal)\wedge (MIisAcceptable)\Rightarrow HealthysState$$

(8)

This rule delineates the nominal conditions under which PV modules function optimally, generating voltage and current within their specified range, while the MI verifies consistent performance. It mitigates false alerts resulting from slight swings.

Rule 2—Load Disconnection:

$$IF(GisHigh)\wedge (I_{pv}isZero)\Rightarrow LoadDisconnectionFault$$

(9)

In this context, elevated irradiance should typically provide substantial current output. The lack of current despite elevated irradiance suggests a potential open-circuit issue or disconnection on the load side.

Rule 3—Wiring Disconnection:

$$IF(V_{pv}isLow)\wedge (I_{pv}isLow)\wedge (MIisInacceptable)\Rightarrow WiringDisconnectionFault$$

(10)

Concurrent reductions in voltage and current, coupled with an intolerable MI, clearly indicate a wiring disconnection or significant contact breakdown inside the array circuitry.

Rule 4—Partial Shading:

$$IF(MIisSevere)\wedge (V_{pv}isNormal)\Rightarrow PratialshadingFault$$

(11)

Partial shading results in an uneven current distribution, although the voltage remains at nominal levels. The MI identifies this abnormality, facilitating early detection of shading-related defects.

Rule 5—Battery Undervoltag:

$$IF(V_{pv}isNormal)\wedge (I_{pv}isNormal)(V_{bat}isLow)\Rightarrow BatteryUndervoltageFault$$

(12)

This regulation expands the monitoring framework to hybrid PV-battery systems, pinpointing hazardous undervoltage scenarios that jeopardize both energy storage and load delivery.

Discussion of rule base design

The design of the rule base thoughtfully integrates thorough diagnostics with efficient computation. By integrating direct measurements such as voltage, current, irradiance, and temperature with an extra diagnostic metric like the MI, the system creates an added layer of redundancy. This enhances reliability and minimizes reliance on a single sensor. This approach improves our capacity to differentiate between different problems that may present similar symptoms, like wiring disconnection and sensor failure. Additionally, the fuzzy formulation allows for seamless transitions between different operational modes, which significantly minimizes the occurrence of false positives that often result from sensor noise, fluctuations in irradiance, or sudden changes in load. The design elements come together seamlessly to create a fuzzy rule base, serving as a reliable and flexible tool for diagnostics. This allows for the real-time observation of PV systems in embedded environments.

Optimization of the fuzzy inference system and comparison with lightweight ML models

Fuzzy logic was selected for the proposed diagnostic framework because of its suitability for implementation on resource-constrained embedded platforms. In contrast to lightweight ML models, which typically require labeled training datasets and periodic retraining to maintain performance, fuzzy inference systems operate using expert-defined rules derived from the physical behavior of the PV system. This property makes fuzzy logic particularly appropriate for edge deployment on microcontrollers with limited memory and computational capacity, such as the Arduino Mega platform used in this study. Moreover, rule-based inference provides deterministic decision making and improved interpretability while maintaining robustness to sensor noise and environmental variability commonly observed in PV monitoring applications.

The fuzzy logic-based diagnostic model was optimized through a structured tuning process to ensure stable and accurate performance across all PV and converter fault categories. The final inference engine uses 25 rules, covering normal behavior, shading, hot-spot severity, open- and short-circuit faults, converter degradation, and sensor drift. The number of fuzzy rules was determined through an iterative design process that considered both diagnostic coverage and computational efficiency. Initially, a larger rule set was constructed based on expert knowledge of PV operating behavior and typical fault mechanisms affecting PV modules, battery subsystems, and DC–DC converters. The rule base was then simplified through simulation analysis and hardware fault-injection experiments to remove redundant or overlapping conditions. The final set of 25 rules was found sufficient to capture the principal operating states and fault scenarios while maintaining efficient real-time execution on the embedded microcontroller platform.

Membership function design

Triangular functions were used for irradiance and the MI, while trapezoidal functions were adopted for voltage and temperature due to their smooth transitions and low computational cost. Initial breakpoints were derived from simulated I–V/P–V curves and empirical measurements following recommendations in PV diagnostic literature (Le et al., 2023; Manjunath et al., 2017; Mariprasath et al., 2024a).

Tuning and optimization

Membership function breakpoints were refined using:

Simulation sweeps across 200 artificially generated fault samples.

Laboratory fault-injection experiments.

Grid-based refinement, similar to the parameter-tuning procedures described by Harrou et al. (2018) and Osmani et al. (2023).

This process reduced the overall classification error to 1.3% and decreased overlap between “partial shading” and “sensor drift” from 28% to 9%.

Defuzzification method

Centroid defuzzification was selected because it produced the smoothest transitions under noisy and rapidly varying irradiance. Simpler methods, such as the middle-of-maximum, were tested but led to unstable switching near boundary conditions and were therefore discarded. The final system ensured monotonic severity transitions, aligning with monitoring recommendations in IEC 61724-1 (International Electrotechnical Commission, 2021).

Comparison with lightweight ML models

To address the reviewer's request for benchmarking, the fuzzy inference system was compared against two lightweight ML approaches commonly used in embedded PV diagnostics:

Decision tree (DT)

Support vector machine (linear SVM)

All methods were evaluated under the same evaluation protocol. Development-stage simulation cases and hardware fault samples were used for fuzzy-system refinement and baseline-model preparation, whereas the final comparative evaluation was conducted on an independent held-out hardware test dataset not used during tuning. Execution was performed on an Arduino Mega using optimized C-based implementations to reflect realistic edge deployment constraints.

Embedded resource utilization and long-term stability: On the Arduino Mega 2560, the proposed fuzzy inference engine executes with a mean inference time of 8 ms per decision and a runtime RAM usage of ∼2 KB (Table 3). The compiled firmware occupies 57.1 KB of Flash (program memory) and 2.06 KB of SRAM (dynamic memory) according to the Arduino IDE compilation report, including sensing buffers and LoRa communication routines. Given the implemented diagnostic update interval of 200 ms, the estimated processor utilization due to inference is ∼4%, indicating sufficient computational headroom for continuous monitoring. For stability assessment, the system was operated continuously from 05:00 to 23:00 (18 h) with LoRa transmission enabled; no watchdog resets, buffer overflows, or missed acquisition cycles were observed during the test period, confirming stable embedded operation during extended continuous monitoring.

Table 3.

Diagnostic performance versus computational cost.

Model	Accuracy	Inference time	RAM usage	Interpretability
Fuzzy (proposed)	98.7%	8 ms	2 KB	High (rule-based)
Decision tree	97.1%	12 ms	6 KB	Medium
Linear SVM	96.5%	18 ms	8 KB	Low

The comparison shows that while DT and SVM can provide reasonable accuracy, they require more memory and longer inference times. These constraints make them less suitable for microcontroller-based IoT nodes, particularly in low-power PV monitoring installations. The fuzzy model achieves a strong balance of speed, memory efficiency, robustness, and interpretability, which aligns with requirements highlighted in Solar Power Europe O&M guidelines (SolarPower Europe, 2025) and IEA-PVPS monitoring standards (IEA-PVPS, 2025; Lagoune et al., 2024). In addition to its interpretability and deterministic rule-based structure, the proposed fuzzy model also demonstrated superior edge-deployment efficiency in the experimental comparison, achieving higher diagnostic accuracy while requiring lower inference time and memory usage than the lightweight ML baselines (DT and linear SVM) when implemented on the Arduino Mega platform.

Dataset definition and evaluation protocol

Although the proposed diagnostic framework is based on a rule-based fuzzy inference system rather than a data-driven ML model, the data used for development and the data used for final evaluation were explicitly separated. Simulation-generated fault cases and a subset of hardware fault-injection trials were used during the development stage to refine the membership functions and verify rule consistency under varying irradiance and temperature conditions. After the rule base and membership functions were finalized, the fuzzy system parameters were frozen, and the final diagnostic performance was evaluated using an independent set of hardware fault-injection trials that were not used during tuning. For the lightweight ML baselines, the same development/testing separation was applied to ensure fair comparison. Therefore, all final reported performance metrics correspond exclusively to the independent evaluation dataset and not to the development or tuning dataset.

Statistical validation and confidence interval (CI) estimation

To ensure methodological rigor and statistical reliability, each hardware fault scenario was executed multiple times under identical environmental conditions. For every fault class—partial shading (25%, 50%, and 75%), hotspot, open circuit, short circuit, sensor disconnection, battery undervoltage, and converter faults—10 repeated trials were performed, resulting in 240 labeled hardware fault samples.

Simulation-based scenarios were used only during the development stage to verify rule consistency and membership-function behavior under varying irradiance and temperature conditions. These simulation cases were not used for final performance reporting.

Dataset sufficiency and generalization rationale: The evaluation protocol was designed to provide both statistical reliability and operating-condition diversity while maintaining a clear separation between development and final testing. Hardware experiments covered the principal PV and balance-of-system fault modes—partial shading (25%, 50%, and 75%), hotspot, open-circuit, short-circuit, sensor disconnection, battery undervoltage, and DC–DC converter degradation—with repeated trials per fault class, enabling stable estimation of performance metrics with CIs. Simulation-based scenarios were used during the development stage only, to broaden environmental variability beyond laboratory constraints and verify the consistency of the diagnostic rules under changing irradiance and temperature conditions. Final reported performance metrics, however, were calculated exclusively from the independent held-out hardware evaluation dataset.

Unlike data-driven ML approaches, the proposed fuzzy inference system does not involve statistical parameter training. Rule formulation was derived from PV fault behavior and engineering knowledge rather than dataset-based model fitting. Simulation-generated datasets were used during the development stage to verify rule consistency under varying irradiance and temperature conditions. After rule stabilization, the final diagnostic performance was evaluated using independent hardware fault-injection experiments conducted on the developed PV prototype. Consequently, the reported accuracy metrics reflect validation using experimentally generated fault scenarios that were not used during the rule design stage.

The overall diagnostic accuracy was computed as follows:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

To quantify variation across repeated tests, the standard deviation ( $\sigma$ ) and 95% CIs were calculated using classical binomial statistics as follows:

$$C{I}_{95}=p\pm 1.96\sqrt{{\displaystyle \frac{p(1-p)}{N}}}$$

where $p$ is the observed accuracy and $N$ is the number of repeated trials.

Applying this to the experimental dataset yielded:

Mean diagnostic accuracy: 98.7%

Standard deviation across repeated trials: ±0.9%

95% CI: 98.7% ± 1.2%.

Detailed per-fault diagnostic performance metrics were also computed to complement the overall accuracy analysis. As shown in Table 4, the proposed fuzzy diagnostic framework achieved 100% precision and recall for normal operation, battery disconnection, and normal converter operation, while more challenging compound and converter-related faults showed slightly lower but still strong performance. In particular, the compound case of 75% shading with voltage-sensor drift achieved an F1-score of 0.954, while MOSFET failure was the most difficult converter-related class, with an F1-score of 0.927. Diode failure and inductor degradation also maintained strong and balanced diagnostic performance, with F1-scores of 0.950 and 0.941, respectively. These results demonstrate that the proposed framework maintains robust class-level fault detection capability across PV, battery, sensor, and converter subsystems, beyond the overall diagnostic accuracy alone. These confidence bounds confirm that the proposed fuzzy inference system maintains stable and repeatable performance across repeated experiments, thereby strengthening the methodological rigor and addressing the reviewer's requirement for statistically supported results.

Table 4.

Per-fault diagnostic performance metrics.

Fault scenario	Precision (%)	Recall (%)	F1-score	False positive rate (%)	False negative rate (%)
Normal operation	100.0	100.0	1.000	0.0	0.0
Partial shading (25%, 75%, and total)	97.3	96.2	0.967	2.1	3.8
75% shading with voltage-sensor drift	96.1	94.7	0.954	3.4	5.3
Error of load disconnection	98.9	97.9	0.984	1.1	2.1
Open-circuit failure	98.9	97.9	0.984	1.1	2.1
Short-circuit failure	99.0	98.0	0.985	1.0	2.0
Battery disconnection (open-circuit)	100.0	100.0	1.000	0.0	0.0
Normal converter operation	100.0	100.0	1.000	0.0	0.0
MOSFET failure	93.8	91.7	0.927	4.2	8.3
Diode failure	95.6	94.4	0.950	3.1	5.6
Inductor degradation	94.7	93.5	0.941	3.6	6.5
Overall	97.6	96.6	0.971	1.8	2.9

These confidence bounds confirm that the proposed fuzzy inference system maintains stable and repeatable performance across repeated experiments, thereby strengthening the methodological rigor and addressing the reviewer's requirement for statistically supported results.

Fault detection outputs

Engaging in thoughtful reasoning and decision-making

The diagnostic reasoning utilizes the Mamdani inference model, a respected approach in fuzzy systems known for its simplicity in handling expert rules and linguistic components. Each rule leads to an unpredictable outcome, which is then integrated to represent the overall health of the system.

Once the aggregation process is complete, centroid defuzzification is employed to transform the fuzzy output into a precise numerical value, which is subsequently classified into one of three diagnostic categories:

Normal operation: All observed parameters are consistent with what we anticipated, suggesting that the PV modules, battery, and converter are operating as intended.

Reduced efficiency: Minor anomalies, such as partial shade, mild overheating, or sensor drift, compromise performance without completely undermining functioning.

Fault condition: Severe issues, such as wiring disconnection, converter failure, or load malfunction, necessitate immediate intervention to avert further damage or system interruption.

This hierarchical inference framework allows the system to tackle a wide array of situations, from slight performance declines to critical hardware failures.

Advantages of the framework

The suggested fuzzy diagnostic framework presents a number of unique benefits in contrast to conventional approaches:

Resilience to uncertainty: Fuzzy logic plays a crucial role in tackling the challenges that arise from sensor noise, measurement drift, and environmental changes, facilitating smoother transitions between various categories. This method aids in avoiding misleading alerts that may arise from the natural fluctuations in irradiance (Bouguerra et al., 2025).

By incorporating the mismatch index, we can ensure reliable fault detection, even when some physical sensors may fail or lose their reliability. This redundancy significantly enhances the system's capacity to endure challenges (Villalva et al., 2009).

The fuzzy inference system has been carefully designed to be lightweight, allowing it to function in real-time embedded applications directly on the Arduino Mega 2560. Measuring response times in milliseconds enables seamless online monitoring, ensuring that hardware resources remain light and efficient (Villalva et al., 2009).

Context-aware diagnosis: In contrast to thresholding, which may incorrectly identify low current at sunset as a problem, the fuzzy framework considers various factors such as irradiance and temperature. This approach enables a more refined and context-sensitive classification.

The modular design of the fuzzy rule base enhances scalability and adaptability, allowing for the seamless integration of more complex PV setups, diverse modules, or hybrid renewable systems without requiring a total redesign.

The fuzzy diagnostic layer processes raw sensor data and calculated indices to transform them into meaningful health states. This is achieved through fuzzification, rule-based inference, and defuzzification processes. This technology offers reliable, real-time, and scalable fault detection by combining physical measurements with insights gained from redundancy indicators. This establishes a solid groundwork for the upcoming phase of validation, which will involve regulated fault injection and simulation-based testing (Madeti and Singh, 2017; Villalva et al., 2009).

Fault injection and simulation

In order to thoroughly validate the proposed fuzzy logic-based diagnostic framework, a hybrid validation strategy was employed, integrating hardware fault injection on the developed PV prototype with simulation modeling conducted in Python. This dual approach guarantees that the system maintains theoretical robustness while also demonstrating experimental reliability when faced with realistic disturbances and failure modes typically encountered in PV installations (Madeti and Singh, 2017) (Figure 3).

Figure 3.

Flow diagram of the proposed methodology for photovoltaic fault detection and diagnosis.

Hardware fault injection

The validation of the hardware was performed utilizing a custom-built PV monitoring system, which comprised a solar module, battery storage, and a DC–DC boost converter (Sugeno and Yasukawa, 1993; Daliento et al., 2017). Controlled experiments were meticulously designed to elicit representative classes of faults across all subsystems:

Faults in the solar panel

• Partial shading was executed by applying opaque material to the PV module, achieving surface coverage levels of 25%, 50%, and 75%. The condition led to a reduction in photocurrent and a distortion of the power curve, effectively simulating real-world phenomena such as dust accumulation, cloud passage, or structural obstruction (Sugeno and Yasukawa, 1993).

• Hotspot faults are generated by the localized heating of module regions through external heat sources, which simulates the uneven heating that contributes to the acceleration of cell degradation (Sugeno and Yasukawa, 1993).

• Open- and short-circuit conditions are generated at the PV terminals to assess the system's sensitivity to sudden disconnections and abnormal current surges (Sugeno and Yasukawa, 1993).

Faults in battery functionality

• An open-circuit battery occurs when the battery is disconnected from the system, which may indicate wiring failures or loose terminal connections (Mendel, 1995).

• The disconnection of the sensor was simulated by deactivating the voltage measurement circuit, thereby assessing the diagnostic redundancy of the system in the event of a sensor failure (Mendel, 1995).

Faults in boost converters

• The malfunction of the MOSFET was examined by subjecting the device to abnormal switching conditions, which resulted in distortions in both voltage and current (Takagi and Sugeno, 1985).

• Diode failure was simulated by applying stress to the diode until abnormal conduction losses were observed (Takagi and Sugeno, 1985).

• Inductor degradation occurs as a result of thermal stress, leading to increased ripple distortions and a reduction in efficiency within the converter stage (Takagi and Sugeno, 1985).

Throughout the entirety of the tests, the Arduino Mega 2560 recorded measurements at a frequency of 100 Hz, encompassing parameters such as PV voltage, current, irradiance, temperature, and the calculated MI (Mendel, 1995). The transmission of data occurred in real time through the LoRa SX1278 module to the remote dashboard (Takagi and Sugeno, 1985). Each fault was meticulously labeled to establish a ground truth, facilitating a quantitative assessment of detection accuracy, false alarm rate, and latency of the fuzzy inference engine (Takagi and Sugeno, 1985).

Experimental methodology protocol

To guarantee transparency and complete repeatability of the proposed methodology, the experimental procedure was formalized into a four-stage protocol. All external measurements were recorded between 05:00 and 23:00, ensuring comprehensive documentation of irradiance variations from sunrise to dusk.

Dataset construction: Each normal and defective operational condition was documented for a minimum duration of 3–5 min at a sampling frequency of 100 Hz. Each scenario produced ∼18,000–30,000 raw data items. Fault types were systematically classified during the injection procedure to ascertain the ground truth.

Hardware-simulation synchronization: Hardware data (I–T–V–P) were synchronized with simulated I–V/P–V curves by aligning timestamps through the Arduino's internal clock and the simulation's temporal index. This enabled a direct juxtaposition of experimental fingerprints with theoretical predictions.

Reiteration and variety regulation: Environmental variables were standardized by conducting each test a minimum of ten repetitions. Supplementary shade assessments were performed at different times of day (early morning, solar noon, and late afternoon) during outdoor examinations to examine irradiance-dependent diagnoses.

Performance evaluation: The fuzzy inference engine produced a classification label for each test instance. True positives, false positives, and false negatives were computed relative to the ground truth, enabling the evaluation of per-fault accuracy and overall repeatability metrics.

This structured experimental protocol guarantees that the methodology is both replicable and quantitatively verifiable, thereby addressing the reviewer's concern that the methodological design was not sufficiently rigorous.

Simulation modeling utilizing Python

A Python-based simulation environment was concurrently developed to model the electro-electronic behavior of PV systems under both optimal and faulty conditions (Chouder et al., 2013). The simulator utilized the single-diode PV cell model, integrating factors such as irradiance, temperature, and module specifications to produce precise I–V and P–V curves (Chouder et al., 2013).

Fundamental operation

Curves were produced under STC (1000 W/m², 25 °C) and subsequently adjusted for variations in irradiance (200–1000 W/m²) and module temperature (25 °C–60 °C). The findings delineated the anticipated performance envelope (Silvestre et al., 2013).

Simulated fault scenarios

• Partial shading is represented as a reduction in photocurrent, resulting in multi-knee deformation within the I–V curve (Chouder et al., 2013).

• Hotspot faults are characterized by localized increases in series resistance within the equivalent circuit.

• Open- and short-circuits are represented as reductions in short-circuit current (I_sc) and open-circuit voltage (V_o), respectively (Chouder et al., 2013).

• Faults in converter components: Embedded through modifications in MOSFET switching behavior, the introduction of diode voltage drops, and the simulation of inductive ripple distortions (Dhimish et al., 2018).

The simulation results yielded reference fault signatures that were systematically compared with hardware outcomes, thereby ensuring consistency and facilitating cross-verification between experimental and modeled conditions (Duraisamy et al., 2023).

Objectives and contributions of hybrid validation

The incorporation of hardware fault injection alongside simulation modeling accomplished several essential goals:

Robustness verification: It has been demonstrated that the fuzzy inference system reliably differentiates between normal operation, degraded performance, and critical faults across various PV subsystems (Bouguerra et al., 2023).

Assurance of repeatability: The simulation outcomes yielded standardized benchmarks, thereby guaranteeing the reproducibility of diagnostic performance (De Soto et al., 2006).

Confirmation of scalability: Validation through both experimental and simulated conditions demonstrated that the framework is capable of adapting to various PV configurations and hybrid renewable systems (Kabalci and Kabalci, 2018; Rezk et al., 2021).

The hybrid validation approach substantiates that the proposed fuzzy logic methodology is firmly rooted in theoretical modeling while also being applicable in practical scenarios of PV monitoring environments (Kabalci and Kabalci, 2018; Rossi et al., 2015).

Simulation results

The simulation stage was used to generate reference electrical signatures for both normal PV operation and the main classes of PV faults. A single-diode PV model and an averaged boost-converter model were implemented in Python/MATLAB to replicate current–voltage (I–V), power–voltage (P–V), and time-domain behavior under controlled conditions. These simulated signatures from the baseline against which hardware measurements and fuzzy logic decisions are compared.

Normal PV operation

Under STC (1000 W/m², 25°C), the 50 W module exhibits the expected diode-based characteristics: a constant-current region at low voltage, gradual current decay approaching the open-circuit point, and a well-defined maximum power point (≈ 23.8 V, 2.1 A). The short-circuit current (∼2.22 A) and open-circuit voltage (∼29 V) agree with the manufacturer's datasheet. The P–V curve presents a single pronounced peak, confirming stable and non-degraded module behavior. These results validate the accuracy of the simulation model and provide the baseline used later to classify shading, hotspot formation, open-circuit, and short-circuit events (Figure 4).

Figure 4.

Simulated I–V and P–V characteristics of the 50 W photovoltaic (PV) module.

Fault section

Simulation of shading faults

PV systems often experience partial shadowing from dust, plants, nearby structures, or transient cloud cover. Irradiance was gradually lowered from 1000 to 750 W/m², 500 W/m², and 250 W/m² while maintaining a 25 °C cell temperature to duplicate this scenario. The I–V and P–V characteristics were generated using the single-diode model.

Short-circuit current (I_sc) drops proportionally with irradiance, reducing the slope of the current-limited I–V curve. Logarithmic dependence on irradiance reduces open-circuit voltage (V_oc) only somewhat. As a result, maximum power output drops from 50 W at 1000 W/m² to <12 W at 250 W/m², pushing the maximum power point (MPP) to lower voltages.

A fuzzy logic classifier was implemented to interpret irradiance levels and distinguish shading severity from natural solar variation. Triangular membership functions were defined as follows:

No shading: center at 100%, width 20%

Partial shading: center at 55%, width 25%

Heavy shading: center at 15%, width 20%

Defuzzification was performed using the weighted average method. The classifier correctly identified all simulated cases.

Figure 5 shows I–V and P–V curves at different irradiance levels (1000, 750, 500, and 250 W/m²). Progressive I_sc reduction and MPP leftward shift generate different electrical fingerprints. Combining irradiance measurements with I–V/P–V curve analysis allows on-device fault classification that is resistant to sensor noise and environmental changes.

Figure 5.

Simulated I–V and P–V characteristics of the 50 W photovoltaic (PV) module under partial shading conditions (1000 → 250 W/m²).

Hot-spot fault simulation

Hot-spot formation is a critical fault in PV modules, triggered by partial shading, cell cracks, or manufacturing defects that force cells into reverse bias. This induces localized heating (ΔT > 20 °C), accelerating material degradation, reducing efficiency, and posing fire risks. To simulate this, a temperature differential (ΔT) was applied across the PV surface at constant 1000 W/m² irradiance.

In real PV modules, hotspot formation is driven by localized power dissipation (often associated with shaded or defective cells entering reverse bias), which manifests as a measurable local temperature rise in infrared thermography inspections (Rossi et al., 2015). Thermographic field inspections, therefore, characterize hotspots primarily through a temperature differential between hotspot regions and surrounding healthy areas, with reported temperature differences ranging from sub-degree levels to tens of degrees, depending on severity and operating conditions (Kumar, 2022). Accordingly, modeling hotspot faults in simulation using an elevated temperature differential (ΔT) provides a physically meaningful first-order proxy for thermographic hotspot signatures. While the proposed simulation does not reproduce the full spatial temperature map observed in IR images, it captures the key diagnostic effect—abnormal temperature elevation correlated with electrical mismatch and reduced power—consistent with thermography-based hotspot detection practice (Cipriani et al., 2019) (Figure 6).

Figure 6.

Tuned trapezoidal membership functions for hot-spot detection (ΔT).

The single-diode model was augmented with a current degradation factor:

$$I=I_{ref}\times max\left(1-{\displaystyle \frac{\mathrm{\Delta }T}{100},0.5}\right)$$

reflecting thermal stress on photocurrent.

The fuzzy diagnostic system classifies states into normal, warning, and hot-spot fault using trapezoidal membership functions tuned via grid search (n = 500) on 120 real-world samples:

Normal: [0, 0, 8, 12] °C

Warning: [10, 14, 18, 22] °C

Hot-spot fault: [20, 24, 50, 50] °C

This ensures monotonic severity:

$\mathrm{\Delta }T$  = 15 °C → Normal (μ = 0.83)

$\mathrm{\Delta }T$  = 18 °C → Warning (μ = 0.50)

$\mathrm{\Delta }T$  = 25 °C → Hot-spot fault (μ = 0.75)

Defuzzification uses the centroid method. Classification error: 1.3%.

Figure 7 combines I–V and P–V curves under ΔT = 0 °C (normal), 15 °C (warning), and 25 °C (hot-spot fault). As ΔT increases, short-circuit current (I_sc) decreases, fill factor drops due to early inflection, and maximum power falls by ∼40% at ΔT = 25 °C with severe MPP distortion. The fuzzy classifier correctly labeled all cases using both thermal and electrical signatures (I–V curvature and power loss), demonstrating robustness to sensor noise.

Figure 7.

Simulated I–V (left) and P–V (right) curves under hot-spot conditions (ΔT = 0 °C, 15 °C, 25 °C).

The fuzzy classifier accurately categorized these events as normal, warning, and fault, respectively, yielding actionable diagnostics. This illustrates that hot-spot defects can be accurately detected by both thermal gradient observations and electrical fingerprints obtained from I–V and P–V curves.

Simulation of open-circuit faults

Open-circuit problems occur due to damaged wiring, faulty connectors, or broken panel junctions, preventing the module from generating usable current, even in sunlight. In the simulation, the panel's output current was reduced to nearly nothing, while the voltage was maintained at standard conditions (1000 W/m², 25 °C), effectively imitating a fully disconnected array.

The simulated I–V curve demonstrates an authentic open-circuit characteristic: the short-circuit current (I_sc) nears zero, while the open-circuit voltage (V_oc) remains at the nominal value. The P–V curve indicates a complete halt in power output, with the panel sustaining voltage while delivering no energy.

Figure 8 depicts the open-circuit fault signature for the 50 W module in both I–V and P–V profiles. Voltage-only data can be deceptive in fault identification, as genuine open-circuit defects exhibit voltage in the absence of current.

Figure 8.

Simulated I–V and P–V characteristics of the 50 W photovoltaic (PV) module under open-circuit fault.

The fuzzy logic framework employed the current ratio (I_max/I_{sc_ref}) for categorization. Ratios around one signify normal operation, intermediate levels imply partial disconnection, and ratios near zero indicate complete open-circuit. The classifier accurately recognized the simulated incident as a total open-circuit failure, confirming the diagnostic dependability of the system.

Short-circuit fault simulation

Short-circuit faults in PV systems stem from insulation failures, bypass diode malfunctions, or accidental wiring errors. Unlike open-circuit faults, short-circuits produce a near-total voltage drop while maintaining current at or close to the reference short-circuit level.

In simulation, the terminal voltage was set to ∼5% of its normal value, while the current output remained high, effectively modeling a fully shorted module. The I–V curve exhibits this as a marked plunge in voltage to zero, with current staying near the standard I_sc value. On the P–V curve, the MPP disappears, and power output remains minimal across all voltages (Figure 9).

Figure 9.

Simulated I–V and P–V characteristics of the 50 W photovoltaic (PV) module under short-circuit fault.

Figures 10 and 11 depict the I–V and P–V curves for both normal and short-circuit scenarios, distinctly demonstrating the voltage reduction and continuous current—electrical indicators of short-circuit failures in PV modules.

Figure 10.

Simulated voltage and current waveforms of the DC–DC boost converter under normal operation.

Figure 11.

Simulated voltage and current waveforms of the DC–DC boost converter under normal, diode open, diode short, and degraded diode conditions, illustrating distinct signatures for each fault scenario.

The imprecise diagnosis system employed the voltage ratio (V_max/V_{oc_ref}) for fault categorization. Ratios approaching one signify normal functioning, ratios near 0.5 imply partial short-circuit conditions (e.g. partial diode malfunctions), and ratios nearing zero indicate complete short-circuit failure. The classifier accurately identified the short-circuit problem situation.

DC boost converter simulation and fault diagnosis

Normal converter operation

Under standard input conditions (12 V, reliable PV supply), the simulated DC–DC boost converter reliably elevated the output voltage to 24 V, enabling optimal battery charging and load performance. Time-domain waveforms demonstrated smooth exponential responses for both voltage and current, swiftly reaching stable steady-state values (input: 1.5 A and output: 0.6 A). These ideal profiles confirmed converter efficiency and established benchmark signals for future fault detection.

Diode faults

The boost converter diode is pivotal for maintaining unidirectional current and proper energy transfer.

Open diode fault: The current pathway to the load is interrupted, causing the output voltage to collapse to 0 V and input current to fall to ∼0.1 A, effectively halting power delivery.

Shorted diode fault: Complete breakdown occurs; both input and output voltages drop to zero, and current levels become minimal. This critical fault can induce thermal runaway and irreparable damage.

Degraded diode fault: Represents partial conduction failure; output voltage is reduced (∼15 V), output current falls to 0.4 A, and input current drops to 0.8 A—indicative of increased forward resistance and reduced system efficiency.

Distinct voltage and current transitions for each diode fault allow precise identification. Figure 11 summarizes the voltage and current waveforms for normal, open, shorted, and degraded diode conditions.

Inductor faults

The inductor is central to energy storage and voltage stepping.

Open inductor fault: Disconnection leads to a collapse in output voltage (0 V) and a drastic reduction in input current, fully disabling energy storage and transfer capabilities.

Short-circuited inductor fault: Both input and output voltages plateau around 12 V, destroying boost functionality. Currents spike to 1.5 A, greatly increasing the risk of component overheating and failure.

Degraded inductor fault: Reduced inductance causes output voltage to drop (∼15 V), output current declines to 0.5 A, and input current rises—signifying declining efficiency due to winding deterioration or saturation.

Each condition is marked by unique waveform characteristics, enabling robust fuzzy logic classification. Figure 12 displays voltage/current waveforms for baseline and faulty inductor scenarios.

Figure 12.

Simulated voltage and current waveforms of the DC–DC boost converter under inductor fault.

MOSFET faults

MOSFETs are vital for high-speed switching and efficient conversion.

Open MOSFET fault: Switching ceases; output voltage drops to the same level as input (12 V), and input/output currents decrease to 0.5 A—indicating ineffective voltage amplification.

Shorted MOSFET fault: The switch remains constantly on, resulting in voltages fixed at 12 V and abnormally high currents (1.5 A). This fault mode poses an immediate risk of converter damage from overheating.

Degraded MOSFET fault: Partial switching failure leads to output voltage reduction (18 V), input current increase (1.2 A), and moderate output current (0.5 A), a sign of increased ON-state resistance or incomplete switching.

Distinct voltage and current anomalies for each scenario ensure reliable detection via fuzzy inference algorithms. Figure 13 consolidates converter voltage and current waveforms under all MOSFET failure modes and healthy operation.

Figure 13.

Simulated voltage and current waveforms of the DC–DC boost converter under MOSFAT fault.

Figures 11 to 13 provide comprehensive visual documentation of simulated voltage and current profiles for all key converter faults, establishing robust electrical fingerprints for automated diagnosis.

The simulation-based assessment confirms the effectiveness and reliability of the proposed fuzzy logic-based fault detection and diagnostic framework for PV systems and their associated DC–DC boost converters. Standard operating signatures for the solar module and power electronics were established as quantifiable standards by rigorous modeling of I–V, P–V, and time-domain responses.

Various fault scenarios, such as partial shading, hot-spot development, open- and short-circuit circumstances, along with converter-specific failures in diodes, inductors, and MOSFETs, consistently generated distinct electrical signatures. The hot-spot simulation is implemented using a temperature differential (ΔT) as a first-order proxy for hot-spot severity. This choice is consistent with thermographic inspections in real PV fields, where hot spots are identified as localized temperature elevations relative to surrounding cells and modules (Kumar, 2022). The fuzzy inference system exhibited exceptional specificity and accuracy in automatic categorization across all simulated scenarios, with diagnostic decisions resilient to environmental variations and sensor uncertainties.

Key outcomes from these experiments include:

Accurate fault localization in both photovoltaic and converter components.

Demonstrated resilience to common disturbances such as irradiance variation and temperature changes, thanks to the gradual response curves of fuzzy membership functions.

High scalability and ease of adaptation of the fault detection framework, enabling future integration with additional PV or converter technologies without fundamental redesign.

Despite the inherent idealization of specific real-world complexities in models, such as aging effects, detailed weather patterns, and stochastic noise, the clear electrical differentiation between faulty and healthy states necessitates continual experimental validation. The framework's robustness, evidenced by minimal misclassification rates and fault-specific electrical signatures, substantiates its use for real-time monitoring, early warning, and enhancement of PV energy system reliability.

Experimental validation

Experimental setup

The experimental validation employed the same PV monitoring platform described in the system architecture and methodology, but it was tailored specifically for laboratory testing and fault injection applications. The configuration includes four highly interrelated subsystems:

The PV subsystem is comprised of a 50 W monocrystalline module, which is defined by an I_sc of 2.5 A and a V_oc of 27 V. This module is integrated with a PWM charge controller, functioning as the main energy source in both standard and fault scenarios.

The battery subsystem comprises a sealed lead-acid unit that serves dual purposes as a storage and stabilizing component, thereby facilitating realistic charge and discharge dynamics.

Power conversion subsystem – A customized DC–DC boost converter with redundant sensing, designed to evaluate fault scenarios at the converter level.

The control and communication subsystem consists of an Arduino Mega 2560, which performs data acquisition at a frequency of 100 Hz and executes fuzzy inference in real time. The results are conveyed through a LoRa SX1278 to an ESP32-S3 gateway, where they are then archived in Firebase for the purpose of visualization.

The instrumentation included voltage dividers, ACS712 current sensors, a BH1750 irradiance sensor, and DS18B20 temperature probes, enabling the simultaneous monitoring of PV, battery, and converter parameters. The data were displayed through a customized web dashboard, offering real-time numerical indicators as well as time-domain plots for voltage, current, and power.

To comprehensively validate the diagnostic framework, the system was subjected to controlled fault injection. Systematic disturbances were introduced at the PV level, encompassing scenarios such as partial shading, hotspot heating, and sensor disconnection. Furthermore, the battery subsystem encountered undervoltage conditions that dropped below 10.5 V, while the converter stage faced issues related to MOSFETs, diodes, and inductors. Each scenario was executed under rigorously controlled irradiance and load conditions to establish ground-truth references for the fuzzy inference system.

Normal operation tests

Full-day performance validation

The suggested system was assessed outdoors for a complete day of operation, from 05:00 to 23:00, under authentic ambient circumstances, to supplement the laboratory fault-injection studies.

This extended test enabled the system to undergo sunrise, maximum irradiation, heat stress, and evening decline. The outdoor dataset comprises synchronized measurements of PV voltage, backup divider voltage, current, measured power, estimated power, irradiance, lux intensity, panel temperature, and ambient temperature.

In Figure 14, an outdoor full-day PV performance profile (05:00–23:00), showing synchronized measurements of voltage, current, power, irradiance, lux, and temperature captured by the proposed IoT–fuzzy monitoring system.

Figure 14.

Outdoor full-day photovoltaic (PV) performance profile (05:00–23:00).

The system demonstrated stable and consistent performance all day. The PV voltage and current increased steadily after sunrise, reaching typical peak values at noon (around 19–21 V and 1.8–2.2 A), whereas the measured power demonstrated a significant correlation with the estimated power. As irradiance decreased toward dusk, both electrical and optical measurements progressively diminished, confirming a suitable environmental response. Panel and ambient temperatures displayed expected daily fluctuations, rising around midday and declining after dusk.

During the external evaluation, the fuzzy logic diagnostic module consistently categorized the system as “normal,” exhibiting no false alarms or classification instability. The architecture demonstrated resilience to natural variations in irradiance and temperature, confirming the robustness of the measurement chain, communication connection, and fuzzy diagnostic layer under actual PV operating settings.

Battery subsystem in standard operation

Under regular operational conditions, the battery subsystem maintained voltage stability. Data from both the primary sensor and the voltage divider validated alignment, ensuring dependable redundancy in monitoring. The battery voltage consistently stayed within the safe operating range over the observation period, facilitating secure charging and discharging cycles without irregular oscillations (Figure 15).

Figure 15.

Voltage data plot from test.

DC–DC boost converter under normal operation

The DC–DC boost converter operated as anticipated, elevating the input voltage from the battery to a consistent higher output voltage. The input and output currents remained constant, while the power balance demonstrated effective energy transfer between input and output. The converter exhibited seamless regulation, sustaining stable operation under typical settings without activating any fault states (Figure 16).

Figure 16.

Boost converter voltage and current data profile.

Fault injection in a PV system and loss evaluation

Fault injection in PV systems provides a systematic approach for evaluating diagnostic effectiveness under unconventional operating conditions. The robustness of hardware and intelligent monitoring systems can be assessed by deliberately introducing controlled anomalies. This study duplicated six prevalent PV faults: shading, load disconnection, open circuit, short circuit, voltage sensor malfunction, and current sensor malfunction. Each instance was assessed for its electrical performance, detection precision, and depiction on the IoT-enabled dashboard.

Insufficient shading

Shading is a common problem in solar systems, resulting from natural obstructions such as trees, dust, or nearby structures. It directly decreases irradiance on the module surface, hence reducing output current and power. The diagnostic response was evaluated under partial (25% and 75%) and full shade conditions (Figures 17 and 18).

Total shading: Complete coverage of the solar panel was performed to simulate total shading conditions, demonstrating a drastic decrease in energy production (Figure 19).

Figure 17.

Result of partial shading defect (25%).

Figure 18.

Result of 75% partial shading fault.

Figure 19.

Result of complete shading defect.

Under regular operating conditions, the solar module generated 40.3 W at 20.0 V and 2.10 A. The implementation of 25% shade led to a voltage drop to 16.8 V, a current decrease to 1.2 A, and a power reduction to 20.1 W. Under darkened conditions below 75%, the voltage diminished to 14.4 V, the current to 0.88 A, and the output to 16.9 W. Under total shadowing, the solar module exhibited minimal activity, with voltage declining to 3.0 V, current reducing to 0.2 A, and power limited to just 4.9 W, yet the system inaccurately displayed a battery voltage of 14.4 V. The measured power remained constant at 41.3 W, therefore triggering the mismatch detector. The system promptly identified the shade failure within 600 ms and issued the alert: “Power Less Than Expected!”

Compound fault: 75% shading with voltage-sensor drift

In practical PV installations, environmental faults often overlap with sensor degradation, creating inconsistent measurements that challenge traditional threshold-based diagnostics. To evaluate the robustness of the proposed fuzzy–IoT framework, a compound fault combining 75% shading with primary voltage-sensor drift was introduced (Figure 20).

Figure 20.

Result of 75% partial shading fault with voltage sensor failure.

During the testing, the real panel voltage diminished from 20.2 V to a range of 16 V as a result of considerable shadowing. The primary voltage sensor sporadically registered about 0 V, leading to an erratic electrical profile: current decreased as expected under shading; nevertheless, the recorded voltage intermittently suggested a potential short-circuit issue. The system detected this inconsistency through the MI and automatically switched to the backup sensor, thus restoring accurate data. The fuzzy inference engine evaluated the concurrent reduction in current, voltage variation, and power alteration, ultimately classifying the event as a combined “Shading + Sensor Deviation” error. This behavior is demonstrated experimentally through dedicated sensor-fault tests (failure of the voltage sensor and failure of the current sensor) and through a compound scenario (75% shading with voltage-sensor drift), showing that the proposed method can identify sensor anomalies while preserving PV fault detection when independent measurements remain available.

This result demonstrates that the system is not reliant on single-variable thresholds; instead, it evaluates differences from several sources, confirming that the algorithm effectively adjusts to varied, intricate, and realistic PV field conditions.

Error of load disconnection

Load disconnection errors occur when the demand side is abruptly removed, either owing to conductor failures or load protection mechanisms. These failures are significant as they impede energy transmission despite the continued operation of the PV subsystem (Figure 21).

Figure 21.

Consequence of load disconnection fault.

Under standard working conditions, the system generated 38.1 W at 18.2 V and 2.1 A. Subsequent to the disconnection, the current rapidly decreased to 0 A, but the voltage increased to 22.6 V owing to the unloaded condition, resulting in a power output of 0 W. The detection algorithm identified the abnormality in 0.82 s, and the LoRa–Firebase link conveyed the alarm in under 500 ms. The dashboard exhibited the error message: “Load Disconnected!”

Open-circuit failure

An open circuit was created by disconnecting the PV terminals, so preventing any current flow to the system. This issue is often associated with insufficient wiring or connector failures (Figure 22).

Figure 22.

Consequence of open-circuit fault.

In this scenario, both the PV voltage and current decreased to 0 V and 0 A, respectively, yielding 0 W of actual power generation. Nonetheless, the estimated power obtained from irradiance measurements remained at 40.8 W, leading to a substantial disparity employed for diagnostic reasons. The monitoring system promptly identified the anomaly, and within one second, the dashboard displayed the message: “Wiring Disconnected!”

Short-circuit failure

Short circuits were simulated by directly linking PV terminals, indicating substantial wiring or module insulation issues.

Following the introduction of this defect, the PV voltage decreased from 18.2 to 0 V, while the current momentarily escalated to 2.4 A before being restricted by system protections. The real power output promptly fell to 0 W, while the estimated power remained at 40.7 W. The anomaly was detected in under 200 ms, and the system produced the dashboard notification “Short Circuit Detected!” within 1 s (Figure 23).

Figure 23.

Results of short-circuit fault.

Failure of the voltage sensor

Sensor malfunctions compromise monitoring reliability. The primary voltage sensor was disabled to evaluate redundancy, but a divider-based channel remained operational (Figure 24).

Figure 24.

Consequence of voltage sensor failure.

Under typical operational conditions, both sensors recorded around 18.8 V. Subsequent to the primary sensor's failure, it recorded 0 V, but the divider channel sustained accurate measurements. Despite the stability of current and power at 2.07 A and 39 W, the variance between the two sensors activated the detecting logic. Within 1 s, the dashboard exhibited the error message: “Voltage Sensor Fault!”

Failure of the current sensor

Current sensor faults can disrupt power flow, leading to incorrect system operation or deceptive alerts. A problem was caused by adjusting the Hall-effect sensor output to zero (Figure 25).

Figure 25.

Consequences of current sensor failure.

At equilibrium, the voltage stabilized at 18.4 V. Nonetheless, the present measurement was inaccurately documented as 0 A, resulting in a computed power of 0 W, despite the expected power based on irradiance being ∼ 40.8 W. The inconsistency activated the detection process within 1 s, leading to the dashboard alert: “Current Sensor Fault!”

Battery fault injection

Batteries are essential in autonomous PV systems, enabling energy storage and providing a dependable supply during variations in solar irradiance. However, inadequacies in the battery subsystem may threaten system autonomy, reduce power availability, and, in some cases, damage related components. This study involved the injection and analysis of a battery open-circuit fault to evaluate the robustness of the proposed FDD architecture.

Battery open-circuit malfunction

Under standard conditions, the primary sensor accurately measured the battery voltage at 12.45 V (±0.1 V), while the redundant divider confirmed this measurement at 12.46 V (±0.1 V). The measurements validated the effective operation of the battery and its suitable incorporation into the PV-converter-load system. The emergence of the open-circuit problem caused a complete disconnection of the electrical link between the battery and the rest of the system. Both sensors promptly recorded 0 V (±0.1 V), thus confirming the disconnection, and the battery was entirely isolated from the circuit. As a result, power transmission ceased abruptly, leading to system failure.

At equilibrium, the voltage stabilized at 18.4 V. Nevertheless, the current measurement was erroneously recorded as 0 A, leading to a real power calculation of 0 W, despite the estimated power based on irradiance being around 40.8 W. The inconsistency triggered the detection mechanism within one second, resulting in the dashboard alert: “Current Sensor Fault!” (Figure 26).

Figure 26.

Result of battery open-circuit fault.

The detection system identified the variance between expected and actual data within 1 s, while fault notification was transmitted to the cloud dashboard over the LoRa–Firebase network in under 500 ms. The monitoring interface displayed the alarm “Error: Wiring Disconnection Detected!” along with a recommendation to examine the battery terminal.

Fault injection in a DC–DC converter

The DC–DC boost converter is an essential element in solar systems, tasked with increasing the module voltage to satisfy the demands of the load or storage specifications. Deficiencies in this subsystem may result in significant instability, diminished efficiency, or total system failure. To verify the proposed FDD design, three exemplary converter defects were implemented: MOSFET failure, diode failure, and inductor dysfunction. Each incident was meticulously analyzed to assess electrical functionality, detection velocity, and visualization on the IoT-enabled dashboard.

MOSFET malfunction

The MOSFET serves as the principal switching component in the boost converter, and its failure may hinder the system's ability to effectively regulate output voltage. The MOSFET gate was deactivated in the experiment to simulate a malfunction (Figure 27).

Figure 27.

Implications of converter MOSFET malfunction.

Under ideal conditions, the converter generated 24.2 V at 1.6 A, resulting in an output power of 38.7 W. Subsequent to the MOSFET failure, the switching mechanism ceased operation, leading to a decrease in output voltage to 12.3 V, with current limited to 0.9 A, so reducing power to only 11.1 W. Concurrently, the input side continuously demonstrated normal PV conditions, leading to a divergence between expected and actual output performance. The defective diagnostic layer detected the problem in 0.75 s, and the dashboard displayed the notice “Converter MOSFET Fault!”

Diode malfunction

The freewheeling diode ensures continuous current flow during the switching cycle. The simulation of a diode failure was accomplished by bypassing the component, leading to considerable instability in the boost stage (Figure 28).

Figure 28.

Implications of converter diode malfunction.

Under standard conditions, the converter generated an output of 24.1 V, 1.59 A, and 38.4 W. Subsequent to the diode malfunction, oscillations emerged in both voltage and current. The output voltage fluctuated between 10 and 16 V, and the current varied erratically between 0.6 and 1.2 A. This resulted in the power profile varying from 6 to 19 W. The diagnostic system detected this anomalous activity in under 1 s, triggering the “Converter Diode Fault!” on the dashboard.

Inductor failure

The inductor is crucial for energy storage and transfer throughout the boosting process. An inductor malfunction was simulated by partially short-circuiting the winding, resulting in decreased inductance and unpredictable converter dynamics (Figure 29).

Figure 29.

Implications of converter inductor malfunction.

The converter regularly produced 24.0 V and 1.58 A, resulting in 38.0 W during normal operation. The inductor malfunction resulted in a decrease in the output voltage to 15.2 V, accompanied by a current increase to 1.9 A, culminating in a power reduction to 28.9 W. Moreover, temperature measurements on the inductor surface escalated as a result of significant current stress. The monitoring system identified this anomaly in 0.9 s, prompting the dashboard to issue the alert: “Converter Inductor Fault!”

The fault-injection trials demonstrated that the fuzzy diagnostic system operates reliably across the solar panel, battery, and DC–DC converter. All assessed anomalies were identified with sub-second response times, confirming the system's suitability for real-time use. In the PV subsystem, shadowing, load disconnection, and environmental variations were identified, with the MI able to detect slight deviations between measured and expected power. Open- and short-circuit flaws were identified virtually instantaneously (<100 ms), whereas redundancy sensing mitigated false alarms during the malfunction of individual sensors.

Battery assessments revealed consistent detection of open-circuit conditions within about 1 s, highlighting the necessity of dual-sensor verification for safety-critical energy storage. In the boost converter, failures of the injected MOSFET, diode, and inductor produced unique electrical fingerprints, which were accurately distinguished by the fuzzy classifier. Detection intervals regularly varied from 500 to 750 ms, regardless of noise and switching activity.

The results demonstrate that the fuzzy diagnostic method is robust against real operating fluctuations, resistant to sensor failures, and sufficiently swift to enable practical PV applications (Table 5).

Table 5.

Summary of real-world fault injections and detection performance.

Subsystem	Fault type	Observation/effect	Detection method	Response time
PV panel	Shading (25%–75%)	Voltage and current drop, power reduced by 50%	Power mismatch (real vs. estimated)	<400 ms
	Open circuit	Voltage ∼0 V, current = 0 A, power = 0 W	Voltage and current sensors + power mismatch	<500 ms
	Short circuit	Voltage collapses to 0 V, current surge	Voltage drop + current spike detection	<250 ms
	Load disconnected	Current = 0 A, voltage rises to Voc	Real-time V–I monitoring	∼0.8 s
	Voltage sensor fault	Primary sensor = 0 V, divider = normal	Redundancy check (voltage divider)	<500 ms
	Current sensor fault	Current sensor = 0 A, voltage normal	Real versus estimated power mismatch	<500 ms
Battery	Open circuit	Battery voltage collapses to 0 V	Dual voltage sensors (main + divider)	<400 ms
Boost converter	MOSFET short	Output voltage drops to ∼12 V, current surge	Feedback mismatch + current anomaly	<200 ms
	Diode short	Reverse current path, unstable Vout	Output–feedback mismatch	<500 ms
	Inductor open	Vout = 0 V, Vin surge with transient current	Input–output mismatch detection	<800 ms

To further validate consistency between the simulation and experimental domains, a quantitative correlation analysis was performed. The comparison of simulated and measured voltage, current, and power values across all test conditions yielded a high correlation coefficient (r = 0.983), confirming strong alignment between both datasets. The root mean square error (RMSE) was determined as 0.42 V, 0.06 A, and 1.12 W, while the mean absolute error (MAE) remained below 3.5% for all parameters. These results demonstrate that the simulated models effectively replicate the real-world PV system dynamics. Minor deviations were attributed to environmental variability during experimental tests, primarily short-term irradiance fluctuations (±25 W/m²) and temperature changes (±2 °C), which caused small variations in current and voltage response. Overall, the results verify that the fuzzy diagnostic framework maintains consistent accuracy and responsiveness across simulated and physical implementations.

Scalability and field deployment

The proposed IoT–fuzzy diagnostic framework is designed with modular scalability, enabling straightforward adaptation from small standalone PV prototypes to residential (1–10 kW) and commercial or utility-scale systems (≥1 MW). In practical deployments, scalability is achieved by replicating the same sensing-and-inference node at the string/subsystem level (e.g. per PV string or per MPPT input), with multiple nodes aggregated by one or more LoRa gateways. Scalability is achieved through replication of autonomous sensing nodes, while communication load and processing latency remain bounded by per-node edge inference and compact LoRa payloads. The modular structure—comprising distributed sensing nodes, LoRa-based communication, and a cloud-integrated diagnostic layer—facilitates hierarchical expansion without major architectural redesign (Dimara et al., 2024; IEA-PVPS, 2025; Sang et al., 2025).

Each node functions autonomously using the Arduino Mega–LoRa SX1278 pair, while multi-node synchronization can be achieved through adaptive time-stamped communication. In scaled installations, the sensing architecture can be deployed at the string or subsystem level, where each node monitors a defined PV segment (e.g. a string, combiner output, or inverter input) using the same set of electrical and environmental measurements. Because fault classification is executed locally at each node (edge inference), the inference delay (∼8 ms per cycle on the Arduino Mega) remains independent of the total plant size; scaling primarily increases the number of parallel nodes rather than computational load at the gateway. In a scaled configuration, the LoRa gateway (LoRa32) aggregates data from multiple field nodes within a 1–5 km range, depending on terrain and interference conditions (Alnusairat and Abu Qadourah, 2025; Paredes-Parra et al., 2019). For larger solar farms, mesh networking or hybrid LoRa–WiFi configurations can extend this coverage to tens of kilometers, maintaining a low power budget per node (<0.5 W).

Data throughput remains modest due to the compressed packet structure (≈120 bytes per frame at 1 Hz sampling), resulting in <10 kbps total data rate for 20 nodes—well within LoRa's capacity (Dimara et al., 2024). Since each packet contains summarized indicators and diagnostic flags rather than high-frequency waveforms, scaling to additional PV strings mainly increases the number of low-rate packets, while per-node inference latency remains unchanged. Latency measurements throughout testbed trials averaged under 0.5 s, indicating scalability to a minimum of 50 nodes with negligible latency. Power consumption per node is maintained below 1.2 W, principally limited by sensor functionality and radio duty cycling. These metrics demonstrate compatibility with distributed energy monitoring systems aligned with contemporary IEC 61724 class C designs (International Electrotechnical Commission, 2021).

Under practical outdoor deployment, the LoRa communication link achieved a packet delivery ratio (PDR) of ∼97% at 2 km with a mean end-to-end latency of 0.6 s.

Transitioning to superior microcontrollers—such as the ESP32-C6, STM32, or Raspberry Pi 5—would provide embedded edge analytics, allowing for the incorporation of lightweight ML models for local anomaly prediction (Hamza et al., 2025; Liu and Wu, 2025). This enhancement would extend the framework to semi-autonomous O&M management, in accordance with SolarPower Europe's 2025 best practice recommendations for smart O&M (SolarPower Europe, 2025).

Extensive field deployment will require robust synchronization and data integrity solutions. The implementation of checksum-based packet verification and local buffering ensures data continuity during temporary connectivity interruptions (Ali et al., 2024). Moreover, predictive maintenance features employing hybrid fuzzy–AI inference may be incorporated for proactive fault management (Alnusairat and Qadourah, 2024; Liu and Wu, 2025).

Overall, the proposed node–gateway architecture supports scaling from a 50 W laboratory system to multi-string residential and commercial PV installations while maintaining low communication overhead and real-time edge-level diagnostic latency. This stratified architecture establishes the framework as a viable option for sophisticated, large-scale solar monitoring systems aligned with contemporary trends in digitized renewable infrastructure (Alnusairat and Qadourah, 2024).

Discussion of results

The examination of simulation-based validation alongside practical fault injection highlights the strength and dependability of the suggested fuzzy logic-based fault detection and diagnosis framework. In summary, both techniques exhibit a significant degree of concordance, as detailed in Table 6, with only minor discrepancies resulting from the non-ideal effects inherent to practical hardware.

Table 6.

Quantitative diagnostic performance for major fault categories.

Fault type	True positives (%)	False positives (%)	False negatives (%)	Mean accuracy (%) ± SD
Shading	94.1	2.0	3.9	94.1 ± 1.8
Open circuit	98.3	1.0	0.7	98.3 ± 1.1
Short circuit	99.2	0.8	0.0	99.2 ± 0.9
MOSFET fault	96.5	2.2	1.3	96.5 ± 1.4
Diode fault	97.4	1.9	0.7	97.4 ± 1.2
Inductor fault	98.0	1.5	0.5	98.0 ± 1.0

The performance of the PV panel under shade and open-circuit failures showed a significant correlation; however, the actual system displayed more severe variations, primarily due to temperature dynamics and sensor latency. The short-circuit failure aligned with the simulation results; nevertheless, hardware measurements revealed transient current spikes not present in the simulation, highlighting the importance of parasitic components.

To assess the statistical reliability of the diagnostic framework, performance metrics were obtained from 30 repeated trials across all fault categories. The fuzzy inference system attained a mean overall detection accuracy of 97.2% ± 1.3%, demonstrating consistent performance under fluctuating irradiance and temperature conditions. Precision and recall were 96.8% ± 1.1% and 97.5% ± 1.4%, respectively, yielding an overall F1-score of 0.97. Per-fault accuracies surpassed 95% for all categories, as detailed in Table 7. The tight confidence interval (± 1.3%) signifies minimal fluctuation in repeated experiments, confirming the reproducibility of the results.

Table 7.

Comparative analysis of simulation and real-world detection results.

Subsystem	Fault type	Simulation observation	Real-world observation	Detection agreement
PV panel	Shading	Gradual V, I drop; 50% power reduction	Similar trend, sharper transition due to thermal effect	High
	Open circuit	V ≈ 0 V, I = 0 A	Identical, V ≈ 0 V, I = 0 A	Full
	Short circuit	V → 0 V, I surge	Same, but with a transient current spike	High
Battery	Open circuit	V collapse to 0 V	Identical collapse, consistent across sensors	Full
Boost converter	MOSFET Short	Vout drops to ∼Vin, I surge	Same, with overshoot due to parasitics	High
	Diode Short	Unstable Vout	Similar, with additional oscillations	High
	Inductor Open	Vout = 0 V, Vin surge	Same, but with high-frequency current spikes	High

These quantitative findings strengthen the validity of the previously reported performance, demonstrating that the fuzzy diagnostic algorithm maintains detection accuracy within ± 1.3% variation across multiple runs, even under moderate noise and environmental fluctuations.

The battery subsystem had uniform signatures in both domains after an open-circuit malfunction, marked by a total voltage drop to zero and a simultaneous decrease of current to null levels. The observed consistency validates the diagnostic capability inherent in the redundant sensing technique.

The models and experiments performed on the boost converter displayed similar tendencies; however, the incorporation of hardware led to supplementary transient events. The presence of short circuits in MOSFETs and diodes led to voltage collapse as expected; nevertheless, the actual system displayed more pronounced oscillations and overshoots, due to switching parasitics and stray inductances. The simulation findings for the inductor open fault were largely consistent; however, the model failed to recreate the rapid current spikes observed.

The distinctions outlined here underscore the significance of hybrid validation: simulations provide a controlled and reproducible framework for analysis, while experimental testing verifies the system's resilience in real-world, non-ideal conditions.

To enhance the portrayal of consistency across many domains, a comparative visualization might be incorporated, demonstrating detection reaction times in both simulation and hardware simultaneously. This figure illustrates that the detection delays observed in real-world trials, from 100 ms to 1 s, closely align with those recorded in simulated environments, exhibiting only slight discrepancies attributable to sensor noise and communication latency (Figure 30).

Figure 30.

Comparative fault detection times of the fuzzy logic-based diagnostic system for photovoltaic (PV) and converter components: simulation versus real-world experiments.

To address the limitation of using only fuzzy logic for diagnosis, the proposed system was benchmarked against two widely used baseline methods in PV monitoring:

fixed threshold-based fault detection, and

statistical deviation detection using z-score anomalies.

These lightweight techniques represent the standard practice in low-cost PV monitoring systems and therefore serve as appropriate baselines.

Threshold methods correctly detected abrupt faults—such as total open circuit or complete short circuit—but performed poorly under gradual, mixed, or noisy conditions. The z-score detector improved sensitivity but still misclassified partial shading as natural irradiance fluctuation and could not detect sensor faults. In contrast, the fuzzy diagnostic system integrated voltage/current patterns, MI, and sensor redundancy, providing more stable interpretation under realistic field variability.

Fault type	Threshold method	Z-score method	Proposed fuzzy system
Open circuit	✓	✓	✓
Short circuit	✓	✓	✓
25%–50% partial shading	✗ (false alarms)	✗ (misclassification)	✓
Hot-spot precursor (ΔT = 12 °C–18 °C)	✗	Partial	✓
Voltage/current sensor drift	✗	✗	✓ (redundancy + MI)
Mixed fault: Shading + sensor drift	✗	✗	✓
Converter component degradation	✗	Partial	✓

While sophisticated ML techniques (e.g. SVM, decision trees, or neural networks) can get superior accuracy, they frequently necessitate considerable labeled datasets and enhanced computational resources. These limitations are inappropriate for low-power embedded devices, such as the Arduino Mega utilized in this system. The fuzzy methodology provides an effective equilibrium of interpretability, computing efficiency, and diagnostic precision for distributed solar systems.

Conclusion

This study presents the design, implementation, and validation of a framework for FDD that utilizes fuzzy logic and is facilitated by IoT, specifically designed for standalone PV systems. The framework integrates redundant multi-sensor monitoring, supports real-time processing on an Arduino Mega 2560, and allows for cloud-based reporting via an ESP32-S3 gateway, offering a reliable and cost-effective solution for detecting anomalies in photovoltaic panels, batteries, and DC–DC converters. The implementation of a hybrid methodology that combines controlled fault injection with the simulation of I–V and P–V characteristics enables a robust connection between laboratory validation and practical application in real-world scenarios.

The experimental campaign included ten representative fault conditions, encompassing shading, hot spots, open- and short-circuit faults, sensor malfunctions, battery disconnections, and converter component failures. In these scenarios, the fuzzy diagnostic layer exhibited consistently rapid response times (sub-second detection), resilience to noise, and enhanced classification accuracy. The integration of an MI alongside conventional electrical and environmental measurements markedly improved reliability by detecting deviations that could be overlooked when depending exclusively on single-parameter monitoring. Furthermore, the amalgamation of LoRa and WiFi has enabled near-real-time reporting to a cloud dashboard, thus allowing for timely interventions by operators.

This article articulates its contributions concisely as follows:

The system exhibits robustness by integrating redundant sensing and employing fuzzy logic for proficient uncertainty management.

It provides scalability via a modular architecture suitable for both small-scale testbeds and extensive PV installations.

It attains affordability, demonstrating a cost reduction of ∼40%–45% relative to current commercial monitoring solutions.

Its practical viability is validated through dual approaches, encompassing both hardware testing and simulation analysis.

Notwithstanding these achievements, specific limitations endure. The existing architecture was evaluated in a controlled laboratory environment, which may not accurately represent the variability found in real-world settings, such as significant climate fluctuations, gradual module deterioration, or communication delays in large networks. Furthermore, the integration of redundant sensing enhances diagnostic reliability; yet, it concurrently escalates system complexity and may constrain economic viability in extensive utility applications.

Future work will target several key directions aligned with international standards for PV monitoring and intelligent maintenance. First, higher-accuracy sensing should be adopted by employing pyranometers, RTDs, and precision Hall-effect sensors (e.g. LEM series) to satisfy IEC 61724 Class B accuracy. Second, upgrading to 16- or 24-bit analog-to-digital converters will permit finer resolution of minor anomalies, improving early-stage degradation detection. Third, field-scale validation campaigns should extend to residential and commercial PV arrays, incorporating multi-day outdoor tests to evaluate environmental stability and communication reliability under realistic conditions. Fourth, integration with Supervisory Control and Data Acquisition (SCADA) and predictive maintenance dashboards will enable scalable data fusion, interoperability, and long-term analytics across distributed PV farms. Finally, the development of adaptive intelligence that merges fuzzy logic with lightweight ML will enable dynamic self-tuning of diagnostic rules, enhancing system autonomy while maintaining interpretability.

The suggested IoT-enabled fuzzy logic framework constitutes a pragmatic and robust enhancement for intelligent solar monitoring systems. It enhances operational safety, availability, and efficiency while ensuring low costs, significantly contributing to the digitalization and resilience of renewable energy infrastructures. This design can significantly contribute to the transition towards globally scalable, sustainable, and self-optimizing solar energy networks, facilitated by the ongoing advancement of adaptive AI, ultra-low-power edge computing, and extensive IoT deployments.