Data preprocessing-based bearing fault diagnosis in PHM: A systematic review

Abstract

Preprocessing, a component of Prognostics Health Management, involves data collection, cleaning, etc., from sensor measurements to accurately assess bearing health. This step is critical for detecting incipient faults and performing effective diagnostics. By systematically preprocessing data from various sensors, engineers can extract meaningful features that reflect bearing degradation, enabling the generation of an accurate Health Index. Prognostics Health Management utilizes these indicators to predict the Remaining Useful Life of bearings, allowing for proactive maintenance and minimizing downtime. This paper outlines multiple methods for analyzing bearing faults. It is divided into two main sections. The first section addresses preprocessing methods, while the second section focuses on dataset collection, which helps researchers compare the results obtained using similar preprocessing methods. In summary, we create a preprocessing graph of the methods used.

Graphical abstract

Introduction

Bearings are critical components in rotating machinery, and their failure can cause significant problems, leading to operational disruptions. Bearings come in a range of types, each designed for specific functions and uses. ¹ For example, cylindrical roller bearings are well-suited for supporting heavy radial loads and are commonly used in gearboxes and mills. Deep groove ball bearings, on the other hand, are widely employed in motors and household appliances, where they manage radial loads and some axial loads. ² Angular contact ball bearings, designed to handle combined loads, are typically paired to enhance their capacity, making them ideal for use in machine tools and wheel hubs. ³ Lastly, self-aligning ball bearings are engineered to operate smoothly even when there is angular misalignment between the shaft and housing, ensuring reliable performance despite imperfect alignment. ⁴ Figure 1 illustrates the different types of bearings and their respective uses.

Figure 1.

Different types of bearings and their respective uses.

Despite their importance, there are several factors that can lead to bearing failures. Common causes of bearing failure include: lubrication deficiencies, where elevated temperatures cause lubricants to degrade; contamination from external particles; misalignment due to bent shafts or improper installation; and improper fitting when bearings are mounted on incorrectly sized shafts. ¹ Notably, inadequate lubrication is responsible for around 80% of bearing failures, highlighting the critical role of proper maintenance in ensuring bearing reliability. ⁵ In the case of wind turbines, bearings are a leading cause of failure, with the MB accounting for 30% of failures in the powertrain components, which is particularly significant given its position between the rotor and gearbox. The MB absorbs non-torque loads and supports the rotor, preventing these forces from reaching the gearbox. As turbine sizes and output continue to grow, MB failure has become a primary cause of downtime. ⁶ Figure 2 illustrates the various causes of failure and the components involved.

Figure 2.

(a) Common bearing faults (adapted from Gbashi et al. ⁶ ) and (b) major sources leading to bearing failure (adapted from Khan et al. ⁷ ).

Bearing failure is closely associated with progressive mechanical degradation ⁸ corresponding to four common bearing faults: BPFO, BPFI, BSF, and FTF, ⁹ which can compromise the reliability of rotating machinery. ¹⁰ To prevent operational disruptions, early detection of bearing faults is crucial. Vibration signal analysis, based on data acquired from different sensor technologies such as acoustic, temperature, and pressure sensors, as well as visual imaging, is widely used to monitor this degradation ¹¹ and has significantly improved early fault detection by providing precise measurements close to the bearing. ¹² Each fault type (BPFO, BPFI, BSF, and FTF) generates characteristic frequency components in the vibration spectrum. However, fault detection depends ultimately on the quality of the acquired data and this is where an important limitation arises.

Data acquisition, which is a critical step in monitoring and diagnostics, introduces measurement errors that stem from factors such as sensor accuracy and instrument precision. These errors contribute to uncertainty, which can directly impact the dependability of the acquired data and, by extension, the performance of PHM systems. ¹³ Addressing this uncertainty is key to enhancing the effectiveness of fault detection and ensuring the smooth operation of industrial processes. Additionally, parameter estimation in ML propagates uncertainty to model predictions, which is crucial for reliable prognostics. PHM faces three significant dilemmas related to uncertainty. ¹⁴ First, data quality is limited due to sparse, noisy observations, insufficient failure data, and the high costs of manual labeling. Second, ML models are often perceived as “black boxes” with unobservable processes that hinder trust and interpretability. Lastly, ML solutions may not always align with physical laws, as they prioritize fitting training data over adhering to physics constraints.

Among the various sources of uncertainty in PHM, data-related issues are particularly critical for bearing applications, for the reasons outlined above (sparse and noisy observations, limited run-to-failure experiments, and costly manual labeling). These challenges introduce both epistemic and aleatory uncertainties, which directly affect the performance of data-driven diagnostics and prognostics. While model interpretability and physics-based consistency also matter, this review emphasizes data quality and preprocessing, as they underpin reliable bearing fault detection and RUL prediction.

By addressing these two types of uncertainties, PHM systems can provide more accurate and reliable assessments, enabling industries to make informed decisions and maintain optimal system health. To determine epistemic uncertainty, Chen et al. ¹⁵ used uncertain random accelerated degradation modeling to address the deterministic degradation ¹⁶ of product performance over time. Makeev et al. ¹⁷ characterized fatigue crack models with uncertainty. Aleatory uncertainties, in contrast, arise from ML that has been embedded with incomplete physics knowledge and based on sparse data. ¹⁸ The relation between the two types of uncertainty and PHM is described in Figure 3.

Figure 3.

Categories of uncertainties encountered in systems (adapted from Rocchetta et al. ¹⁹ ).

To reduce uncertainty in data, preprocessing is employed. It plays a critical role in predicting bearing health degradation, ²⁰ significantly enhancing the accuracy and reliability of the prognostic process. These essential preprocessing steps include noise reduction, normalization, outlier detection, feature selection, feature reduction, handling missing values, and feature extraction. ²¹ The preprocessing steps in fault prediction and their relation with PHM cycles are described in Figure 4. ²² The data collected from these sensors creates different types of datasets, which are analyzed using two main approaches: traditional methods and DL. DL techniques offer the advantage of automatically extracting hierarchical features from raw data, but they require extensive datasets for training and high computational power. On the other hand, traditional methods rely on manual preprocessing, such as feature extraction based on domain knowledge, and have lower data and computational requirements, making them more practical in resource-limited environments.^23,24

Figure 4.

PHM cycles and detail of the preprocessing steps (adapted from Wardhana et al. ²⁵ ).

These methods play a crucial role in the broader framework of PHM, which aims to optimize maintenance strategies and ensure the reliability and performance of industrial machines by detecting potential failures early and predicting the RUL of bearings. PHM involves monitoring bearing conditions, diagnosing faults, and estimating RUL to avoid unexpected downtime and reduce maintenance costs. Condition monitoring methods, such as vibration and acoustic analysis, help detect anomalies that indicate bearing wear or failure. A reliable HI is used to quantify the bearing’s condition, reflecting wear trends over time. The RUL prediction helps plan maintenance activities, ensuring that bearings are replaced or repaired before complete failure occurs. Advanced approaches such as LSTM and CNN are used for analyzing large datasets, identifying complex patterns, and predicting bearing health. Combining these data-driven techniques with physics-based models offers a more comprehensive view of bearing behavior and enhances the accuracy of failure predictions. By integrating PHM systems, ²⁶ industries can significantly reduce unplanned downtime, extend machinery life, and lower operational costs, while improving safety and process efficiency. This approach hinges on the ability to effectively monitor and detect faults, ensuring that components such as bearings operate smoothly and reliably in industrial applications. A simplified representation of the process from data acquisition to preprocessing is illustrated in Figure 5, providing the foundation for subsequent condition monitoring and fault analysis.

Figure 5.

Overview of bearing degradation progression and the corresponding frequency bands associated with different bearing fault conditions.

In this review, we examine the data preprocessing stage of vibration-based bearing fault diagnosis within the broader PHM framework. We conducted a SLR of recent data-driven studies, with a particular focus on how preprocessing techniques are applied to widely used academic bearing datasets. The objectives are to (i) categorize and summarize the main preprocessing methods, (ii) relate these methods to the characteristics of the datasets on which they are evaluated, and (iii) highlight strengths, limitations, and open challenges relevant to researchers and practitioners working on bearing PHM.

Despite the rapid growth of data-driven PHM, existing evidence on data preprocessing practices for bearing applications remains fragmented. This study addresses that gap by characterizing, analyzing, and synthesizing the preprocessing methods reported in recent work, and by relating them to widely recognized bearing datasets from multiple academic institutions. In doing so, it compiles and documents available academic bearing datasets to broaden the range of options for model development and training, and it identifies promising directions for future research. The results provide a clearer picture of current preprocessing practices, including their applications and limitations, and establish a foundation for subsequent technical investigations into innovative preprocessing methods capable of efficiently handling large-scale data.

The article is organized as follows: Section 2 presents the systematic literature review, methodology, detailing the research questions, search strategy, screening criteria, and analysis steps used to identify relevant studies on data preprocessing for bearing PHM. From an initial set of 3358 articles, a rigorous filtering process led to a final selection of 89 papers, mainly centered on the CWRU and PRONOSTIA benchmark datasets. Section 3 compiles and describes the most widely used bearing fault datasets, detailing their sensor setups, sampling conditions, fault types, and application domains to support consistent benchmarking in PHM research. It highlights key differences across datasets—such as operating conditions, data volume, and fault complexity—helping researchers select appropriate data sources for diagnosis and prognostics studies. Section 4 provides a comprehensive overview of preprocessing techniques used in bearing fault diagnosis, covering the full pipeline from outlier detection and noise reduction to handling missing data, normalization, class imbalance correction, feature selection, extraction, and dimensionality reduction. It highlights how each method improves data quality and model performance, linking specific techniques to the characteristics of widely used bearing datasets. Section 5 identifies key research gaps in current preprocessing practices, emphasizing the need for more realistic industrial datasets, improved real-time data handling, multi-sensor fusion, and better integration of uncertainty quantification. It also highlights opportunities for developing more interpretable, scalable, and adaptive preprocessing methods to strengthen PHM applications in real-world environments. Finally, Section 6 summarizes the main findings of the review, showing that effective preprocessing—such as noise removal, outlier detection, feature extraction, and dimensionality reduction—is essential for reliable vibration-based bearing fault diagnosis. It also highlights key research gaps, including limited real-world datasets, insufficient multi-sensor integration, and the need for more explainable and industrially robust PHM approaches.

Systematic literature review

To gain detailed insights into the preprocessing methods applied to bearing health in PHM over the last 5 years, specifically in the CWRU and PRONOSTIA datasets, we conducted a SLR.

We focused only on these two datasets because they are the most widely used and well-established benchmarks in recent bearing diagnostics research. Our search results show that these two datasets consistently generated the highest number of relevant publications, while other public datasets were mentioned only occasionally and lacked sufficient research volume for meaningful comparison. Moreover, CWRU supports fault diagnosis, whereas PRONOSTIA provides run-to-failure data for prognostics, offering complementary coverage of key PHM tasks. Including rarely used datasets would have reduced comparability and weakened the consistency of the review.

This study follows a SLR approach guided by the five-step process outlined by Han et al. ²⁷ The steps in this process are as follows:

Formulating the research question(s)

The first step in conducting a SLR is to carefully define the scope of the study. This sets clear research boundaries and defines key questions, establishing a strong foundation for a structured SLR. Based on the objectives outlined in the previous section, the primary aim of this study is to find the preprocessing methods that are commonly used in PHM using bearing datasets (CWRU and PRONOSTIA). The research questions guiding this evaluation are:

• Q1: What key levels of analysis and decision-making are explored in the literature?

• Q2: What methodologies have been used to increase the quality of bearing health prognostics?

Searching and identifying core studies

To ensure comprehensive coverage, this study used three major databases: Science Direct, Scopus, and IEEE. These databases were used to retrieve articles published in international peer-reviewed journals over the last 5 years. To identify relevant studies, an initial set of keywords was developed by analyzing highly cited papers in the field. The main keywords had to be in the area of preprocessing, PHM, and bearings. These keywords ultimately created a comprehensive query that was used to search the titles, abstracts, and keywords of publications. The final search query employed Boolean operators to combine the related terms that are described in Table 1.

Table 1.

Query criteria of SLR preprocessing.

No	Topic	Query criteria
1	Feature selection	(“Feature selection”) AND (“PRONOSTIA” OR “FEMTO” OR “CWRU”)
2	Normalization	(“Normalization”) AND (“PRONOSTIA” OR “FEMTO” OR “CWRU”)
3	Outlier detection	(“Outlier$;) AND (“PRONOSTIA” OR “CWRU” OR “FEMTO”)
4	Feature extraction	(“Feature extraction”) AND (“PRONOSTIA” OR “FEMTO” OR “CWRU”)
5	Imbalanced data	(“Imbalanced$;) AND (“PRONOSTIA” OR “CWRU”)
6	Imputation	(“Imputation”) AND (“Bearing” OR “CWRU” OR “PRONOSTIA”)

Selecting relevant papers

Duplicates were removed, and the remaining articles were screened based on predefined inclusion and exclusion criteria. This process selected peer-reviewed journal articles in English that specifically focused on preprocessing. Conference papers were excluded from the study. The titles and abstracts of the selected articles were then reviewed to identify those for a full-text assessment. The inclusion and exclusion criteria are presented in Table 2.

Table 2.

Review exclusion and inclusion criteria adapted from Trilles et al. ²⁸

Exclusion	A	Articles that are not peer-reviewed, such as conference papers, book chapters, notes, books, or theses/dissertations
	B	Articles not written in English
	C	Articles without full-text availability
	D	Articles that are not related to PHM
Inclusion	E	Articles published in the period 2020–2024
	F	Articles must include a title, abstract, and keywords
	G	Articles on engineering topics

Full-text evaluation

The full texts of the shortlisted articles were thoroughly evaluated to ensure their quality and relevance. Additionally, the reference lists of these articles were examined using the same inclusion and exclusion criteria.

Analysis and synthesis

The final set of articles underwent detailed analysis and synthesis to extract findings, which are presented in the following sections. These steps, along with the final selection process, are summarized in Figure 6, which outlines the methodology used to identify the core papers for this review.

Figure 6.

Flowchart for literature review with exclusion and inclusion criteria.

By using these search terms, the publication information including the title, abstract, year of publication, journal source, volume, and publisher was downloaded as a tab-delimited text file. This initial search yielded a total of 3358 research articles. After applying Criteria A, B, and C, the dataset was reduced to 1506 articles. Applying the additional Criteria D and E further narrowed the selection to 361 articles. Following a full-text assessment of these 361 publications, only the most relevant and significant works were retained. The final literature set, consisting of 89 references, was organized into two major contribution areas: preprocessing techniques and PHM.

An additional perspective was gained by analyzing the relationships among the selected publications through keyword co-occurrence analysis. Using VOSviewer, these relationships were visualized in a co-occurrence map that reveals the thematic structure of the field. In total, 707 keywords were extracted, but only 56 keywords appeared at least four times and were therefore included in the mapping. VOSviewer grouped these frequently occurring keywords into six clusters, each represented by a distinct color and defined by its internal association strength, as shown in Figure 7. Based on the thematic analysis of the clusters, they can be organized into two overarching groups, each focusing on different aspects of machinery fault detection, diagnostics, and prognostics:

Group 1: ML and fault diagnosis

This group focuses on advanced ML techniques, fault diagnosis, and predictive maintenance, which are crucial in modern industrial and engineering systems.

• Cluster 1 (red): Anomaly detection, attention mechanism, condition monitoring, DL, domain adaptation, edge computing, fault diagnosis, feature engineering, PHM, RUL, rolling bearing, transfer learning.

• Cluster 2 (green): Bearing, bearing fault diagnosis, data mining, extraction, fault diagnosis, feature extraction, feature selection, ML, machinery.

• Cluster 4 (yellow): Failure analysis, fault diagnosis method, intelligent fault diagnosis, learning systems, rolling bearings, rotating machinery.

Figure 7.

Network visualization of preprocessing-related keywords, illustrating their relationships and co-occurrence patterns.

Group 1 emphasizes the use of various ML methodologies to monitor and diagnose faults in machinery, particularly in bearings. The clusters within this group cover a wide range of topics, from multiple preprocessing methods to more sophisticated techniques such as DL. These methods are applied to predict the RUL of components, monitor the condition of equipment, and detect faults before they lead to significant failures. The group also includes intelligent systems that can autonomously learn and improve their fault diagnosis capabilities over time.

Group 2: DL, CNN, and uncertainty in fault diagnosis

This group highlights the role of CNN and DL techniques in fault detection and prediction, as well as the integration of IoT and uncertainty analysis.

• Cluster 3 (blue): Convolution, CNN, fault detection, LTSM, roller bearings, support vector machine, time domain analysis, vibration signal.

• Cluster 5 (purple): CNN, data fusion, fault prediction, fuzzification, Internet of Things.

• Cluster 6 (cyan): CNN, uncertainty, unsupervised learning.

Group 2 focuses on leveraging advanced DL techniques to detect and predict faults in mechanical systems. Central to this group is the use of CNN, which are powerful tools for analyzing complex data such as vibration signals. This group also tackles the challenge of uncertainty in fault diagnosis, developing methods that can operate effectively even when the data are incomplete or ambiguous. Additionally, the integration of IoT with ML models, such as CNN, allows for real-time monitoring and prediction of equipment failures, making maintenance more proactive and less reliant on scheduled downtime.

Data acquisition

To identify bearing faults, researchers typically employ a data acquisition step, utilizing various types of sensors such as vibration, temperature, and acoustic sensors. Each of these sensors possesses distinct characteristics that contribute to the measurement process. Through the SLR, we compiled a comprehensive list of the datasets commonly used by researchers to facilitate the analysis of bearing conditions. An overview of several key datasets is illustrated in Figure 8 and all the collected data are described in Table 3. This table provides a new description of the common processes and tools used during the experiments. For better understanding, we illustrate them in Figure 9.

Figure 8.

Overview of widely used bearing fault datasets: (a) FEMTO, (b) CWRU, (c) XJTU-SY, and (d) IMS.

Table 3.

Bearing datasets and description.

No.	Dataset	Prediction	Data type	Description	Amount of data	Freq	Fault type	Status	References
1	CWRU	Classification	Tabular data	This dataset is often referenced as a benchmark for bearing fault classification and is widely used as an open-source resource for rotating machinery. It was generated under three operational conditions	8000 (training sample)	12, 48 kHz	IF, OF, and BF	OA	29
2	PRONOSTIA	Fault detection, RUL	Tabular data	This dataset was publicly released at the IEEE International Conference on PHM in 2012. It includes 17 run-to-failure datasets of rolling element bearings from the PRONOSTIA platform under three operating conditions	28 (samples)	25.6 kHz	Unknown	OA	12, 25 a , 30
3	IMS	Fault detection, RUL	Tabular data	Provided by the NSF I/UCR Center for IMS at the University of Cincinnati and included in NASA’s Prognostics Data Repository, this bearing dataset comes from three test-to-failure experiments. It includes four conditions: rolling element fault, inner ring fault, outer ring fault, and healthy status, with an imbalance ratio of ∼10:1	984 (bearing vibration signal samples)	20 kHz	Unknown	OA	31, 32
4	PU dataset	Classification	Tabular data	The PU bearing dataset provides high-resolution data with six healthy bearing groups and 12 damaged groups. Damage was introduced in three ways: (i) EDM, creating grooves (0.25 mm length, 1–2 mm depth); (ii) Drill holes (0.9 mm, 2 mm, 3 mm aperture); (iii) Manual electric engraving (1–4 mm damage length). It also includes real damage from 14 accelerated life tests	1200 (samples)	64 kHz	• IF• OF	OA	33
5	IITR dataset	RUL	Tabular data	The dataset includes three operational conditions: OC1 (B1_1–B1_3) at 2400 rpm, OC2 (B2_1–B2_3) at 3000 rpm, and OC3 (B3_1–B3_3) at 3600 rpm. The test bearing, an NBC NUC205E cylindrical roller bearing, has an OR diameter of 52 mm, an IR diameter of 25 mm, a pitch circle diameter of 38.5 mm, a roller diameter of 7.5 mm, 13 rollers, and a 0° contact angle	Unknown	51.2 kHz	Unknown	OA	30
6	IPF dataset	RUL	Tabular data	Provided by the PHM Society in 2015, this dataset contains eight-dimensional time-series data classified into multiple classes, including six fault types. Faults 1–5 are the primary focus, while type 6 represents other fault categories. Measurements were recorded every ∼15 min over 3–4 years. The goal is to predict future failure events (types 1–5) and their occurrence times based on historical data. The imbalance ratio can reach up to 1:1000	106,184 (sample)	–	Unknown	OA	34
7	NUE dataset	Classification	Image dataset	The experiment used an ER-16 K rolling bearing with 10 balls. A piezoelectric acceleration sensor captured vibration signals at a 12 kHz sampling frequency. EDM created pitting faults on the IR, OR, and rolling element. The proposed method analyzes vibration signals under four motor-end bearing conditions. Based on different imbalance ratios, four datasets were created, each condition had 2000 RGB images (32 × 32 pixels), totaling 8000 images (7200 for training, 800 for validation). Imbalance ratios (normal:fault) were: training A–D = 100:50, 100:25, 100:20, 100:10; testing A/B/C/D = 1:1	13,125 (samples)	12 kHz	OR, IR, Rolling Element (RE)	CA	35
8	Bogie dataset	Classification	Tabular data	This dataset was collected from an experimental high-speed multi-unit train bogie bearing system. An accelerometer placed on the load module captures vibration data at a sampling rate of 5 kHz. Operating at a rotational speed of 1950 rpm, the system corresponds to a train speed of 320 km/h. The dataset contains data from bearings with faults, including RF, OF, and IF	4000 (samples)	5 kHz	IF, OF, RF	CA	36
9	Centrifugal pump bearing dataset	Clustering	Tabular data	In the experiment, a centrifugal pump test rig was used, featuring a pump with a closed impeller that has three vanes, driven by a 0.37 kW motor. The impeller is attached to a rotor, which is supported by two bearings inside the motor housing. The pump’s discharge capacity is 1.6 L/s. A fault was introduced in the bearing closest to the impeller, labeled bearing number 6203ZZ	Unknown	–	IF, OF	CA	37
10	XJTU-SY dataset	RUL	Tabular data	Comprising 15 sets of bearing life-cycle vibration signals collected by an acceleration sensor, the datasets consist of six training sets and nine testing sets. The signals are sampled at a frequency of 25.6 kHz, with each sample lasting 1.28 s, amounting to a total sampling period of 1 min. Horizontal vibration signals from this dataset are also included in the experimental analysis	32,768 (data points)	25,6 kHz	Unknown	OA	32
11	Wind turbine data set	RUL	Tabular data	The dataset consists of sensor data from the high-speed bearing shaft of a 2.2 MW wind turbine. It records IR energy measurements at 10-min intervals over ∼50 days. The investigated bearing, an SKF 32222 J2 tapered roller bearing, has an IR fault. This TRB has a 200 mm outer diameter, 110 mm bore, 56 mm length, 20 rolling elements at a 16° taper angle, and weighs ∼20 pounds	Unknown	20 kHz	IF, OF, RF	CA	12
12	NEPU dataset	Classification	Tabular data	The testing platform mainly consists of a motor, gearbox, and bearings. The dataset was obtained at a sampling frequency of 10 kHz, a rotor speed of 1400 rpm, and a load of 1 HP. Operating states and instance quantities are consistent with the CWRU dataset	Unknown	10 kHz	IF, RF, BF	CA-OR	38
13	MFPT dataset	Classification	Tabular data	The laboratory setup consisted of an induction motor (2 HP, four poles, 380 V, 60 Hz), a data miner (NI PXI-1037), a servo motor (FUKUTA), an oscilloscope, and a personal computer	5170 (samples)	48 kHz	IF, OF	OA	39
14	UCI dataset	Classification	Tabular data	Vibration signals from a two-stage gearbox with replaceable gears were recorded using a DSPACE system at a 20 kHz sampling rate. The dataset includes nine gear conditions, such as healthy (label 1), missing tooth (label 2), root crack (label 3), spalling (label 4), and tooth tip fragmentation	936 (samples)	20 kHz	• Missing tooth• Root crack• Spalling• Chipping tip	OA	33
15	JNU dataset	Classification	Tabular data	An experimental system with a 3.7 kW three-phase inductor motor was used to collect vibration signals in the vertical direction of a rolling bearing. The signals were recorded with an accelerometer (PCB MA352A60). The motor operated at speeds of 600, 800, and 1000 rpm under four bearing conditions	775 (samples)	50 kHz	OF, IF, RF	OA	40
16	XIU dataset	Classification	Tabular data	An induction motor with a bearing-rotor system featured an adjustable bolt device to apply radial load on the middle bearing. Two bearings were mounted at the rotor ends: one normal and one test bearing. Accelerometers were symmetrically placed in the housings to capture vibration data. The test bearing (6206-2RS) had faults induced by EDM in the ball element, IR, and OR. Vibration signals were recorded at 1484 rpm with a 20 kHz sampling frequency. Data points were divided into 60 non-overlapping sets, each containing 2048 samples	522 (samples)	20 kHz	BF, OF, IF	CA	41
17	KU dataset	Classification	Tabular data	An industrial system at Konkuk University was utilized to evaluate a feature selection algorithm for diagnosing bearing faults. A vibration sensor was mounted on the infeeding roll, rotating at 636 rpm, to monitor a bearing (6006VVP5) with a maximum speed of 1765 rpm. Defects in the IR, OR, and ball were introduced using electro-discharge machining. Vibration signals were recorded using a piezoelectric accelerometer (356A01, PCB) with a sensitivity of 0.51 mV/(m/s2) and captured with a DAQ system (NI9234). Data for each condition, including normal, and combined defects, were collected over 5 min	–	25.6 kHz	IF, OF, BF	CA	42
18	CYCU dataset	Classification	Tabular data	The motor fault dataset was measured by a three-phase squirrel-cage induction motor. Motor specifications are as follows: rated voltage 220/380 V, rated horsepower 2 HP, 1715 rpm and the number of poles is 4. The signal was measured under four conditions: health, bearing damage, SF, and RB damage. Data miner (NI PXI-1033) was used for data current collection. The measurement conditions are as follows: load rate 0%, torque 0 N/m; load rate 25%, torque 2 N/m; load rate 50%, torque 4 N/m; load rate 75%, torque 6 N/m and load rate 100%, torque 8 N/m	70 (samples)	–	IF, OF, BF, SF, RB	CA	24
19	Northeastern University	Classification	Tabular data	The motor operated at speeds ranging from 1400 to 1600 rpm, with vibration signals recorded at five specific speeds: 1400, 1450, 1500, 1550, and 1600 rpm. Data was collected using an NI-9234 acquisition card, with a sampling frequency of 1024 Hz. Four and 10 output class	1500 (samples)	1 kHz	Umbalanced, misalignment, and rubber	CA	43
20	NCEPU	Classification	Tabular data	The experimental setup consists of an induction motor, accelerometer, testing bearings, and a loading module. Two faulty bearings were tested: one with a visible trench on the IR (IF) and another with a BF not externally visible. The experiment included six conditions at 1470 rpm (no load) and 1440 rpm (1 HP load). Each sample contains 2048 points, reshaped into a 60 × 37 matrix for analysis	120 (samples)	12 kHz	IF, OF, BF	CA	44
21	HUST	Classification	Tabular data	The experimental dataset was collected using a custom MFS that includes an electric motor, bearing, gearbox, and brake. SKF-6206 deep groove ball bearings were tested under both normal and faulty conditions, with defects introduced to the inner and ORways using EDM to a depth of 0.2 mm. The drive train operated with varying load torques from 1 to 3.9 Nm to capture different operating conditions	4380 (samples)	10 kHz	IF, OF, BF	CA-OR	45
22	Key Laboratory of Traffic Safety on Track CSU	Classification	Tabular data	The laboratory bearing fault simulation test rig includes a motor, motor controller, coupling, and test bearing (MB ER-12K). Faulty bearings were created using EDM pitting with diameters of 1 and 2 mm, simulating varying fault severities in the outer ring, ball, inner ring, and compound faults. This setup includes nine bearing types: one normal and eight faulty	200 (samples)	32 kHz	IF, BF, OF, compound fault	CA	46
23	AUT	Classification	Tabular data	The experimental setup used the BVT-5 bearing fault vibration measuring instrument. A 1210 self-aligning ball bearing was tested, with pitting faults introduced on the inner ring, outer ring, and rolling elements using an SG electric engraving machine. Vibration signals were collected at 1800 rpm. Each condition had 100 signals, with 2048 sampling points/signal	400 (vibration signals)	5 kHz	IF, OF, BF	CA	29
24	CUP	Classification	Tabular data	The dataset was collected from a laboratory test platform consisting of a motor, gearbox, and bearing. The bearing has 13 balls, each with a diameter of 0.295 inches, and the motor operates at 1400 rpm. A fault size of 0.0126 inches was introduced. Four datasets were gathered under different conditions	100 (samples)	1, 10 kHz	IF, OF, BF	CA	47
25	QPZZ-II platform	Classification	Tabular data	The testbed was powered by a driving motor and connected to the spindle rotor via a coupling. The spindle rotor was supported by left and right bearings. The left bearing was a standard bearing, while the right bearing was a replaceable fault-bearing. Vibration acceleration sensors were positioned in both the horizontal and vertical directions of the fault-bearing seat	80 (groups)	12.8 kHz	IF, OF, BF	CA	48
26	RHV	Classification	Tabular data	An experiment on the wheel bearing of a China RHV. The test stand includes two wheels, a shaft, an axle box, and two driving wheels. Three health states were tested: normal, outer ring failure, and cage failure, with cage failure replacing the rare inner race fault. The vehicle ran at 3, 50, and 100 km/h	–	10 kHz	IF, OF, cage fault	CA	49
27	OtW	Classification	Tabular data	Experiments were conducted using a SpectraQuest MFS-PK5M. Two ER16K ball bearings support the shaft: the left is healthy, and the right is replaced with bearings of varying health conditions. An ICP accelerometer (model 623C01) on the housing collects vibration data, while an incremental encoder (EPC model 775) measures the shaft’s rotational speed. The experiments were run in four conditions: increasing speed (OC1), decreasing speed (OC2), increasing then decreasing speed (OC3), and decreasing then increasing speed (OC4)	36 (datasets)	200 kHz	IF, OF	OA	50

BF: ball fault; IF: inner fault; MFS: machinery fault simulator; OF: outer fault; RB: rotor bar; RF: roller fault; SF: stator faults.

a

Edited 2025.

Figure 9.

Common experimental processes and tools used in bearing fault diagnosis research.

Preprocessing method

After collecting the bearing vibration data, preprocessing is necessary to ensure data quality, improve model performance, and reduce uncertainty. The following subsections present the preprocessing techniques in a logical sequence, first detecting outliers, then reducing noise, handling missing entries, assessing data distribution, transforming scales, addressing class imbalance, and finally extracting or reducing features for subsequent machine learning tasks.

Outlier detection

The preprocessing sequence begins with outlier detection, as anomalous or extreme data points can distort subsequent analysis and compromise feature quality. Outliers may arise from sensor faults, sudden mechanical impacts, or environmental disturbances. The LOF compares the local density of a data point to that of its neighbors, identifying points with significantly lower density as outliers. ⁵¹ When combined with LS-SVM, this approach enables effective outlier detection and allows the data to be smoothed, as illustrated in Figure 10. Finally, the Mahalanobis Distance method measures the distance of a point from the mean in multivariate datasets, effectively identifying outliers by considering correlations between variables. ⁵²

Figure 10.

Structure of the PSLOF method, combining LOF and LS-SVM: (a) With prediction data and (b) After improved outlier. ⁵¹

Outlier detection is particularly relevant for datasets such as PRONOSTIA and IMS, where sudden spikes or sensor dropouts occur during long-duration run-to-failure tests. These datasets frequently exhibit irregular points caused by mechanical shocks, making methods such as LOF and Mahalanobis distance especially effective.

Table 4 summarizes the methods of outlier detection that were used.

Table 4.

Summary of the methodologies and techniques for anomaly detection.

Method	Application	Machine learning	References
LOF	PRONOSTIA	Cross-Transformer, LightGBM, LS-SVM, LSTM	51, 53, 54
Adaptive K-Means	PRONOSTIA	LSTM	54
Quartile (Q1 and Q3)	PRONOSTIA, IMS, Wind Turbine	HMM regression	55
Mahalanobis distance	PRONOSTIA	DL	52

Removing outliers ensures a cleaner dataset before noise reduction, which is addressed in the next subsection.

Smoothing/noise elimination

Following the removal of outliers, the next step is to eliminate noise to ensure that the signals accurately reflect the true behavior of the bearing. Noise may originate from environmental vibration, electrical interference, or interactions with other machine components. Consequently, smoothing and noise elimination techniques are essential for enhancing signal clarity, preserving important fault-related characteristics, and improving the accuracy of subsequent predictive models in bearing health monitoring.

Simple exponential smoothing provides a basic approach to reduce noise by applying a weighted average to past observations. ⁵⁶ More advanced methods such as double exponential smoothing or Holt exponential smoothing account for trends in the data, ⁵⁷ while Holt–Winters exponential smoothing further incorporates seasonality. ⁵⁶ Simple MA offers a straightforward method for smoothing data by averaging a fixed number of observations, and MA by weight refines this approach by assigning different weights to observations. ⁵⁸

The exponential MA gives more importance to recent observations, making it responsive to recent changes. ⁵⁹ A smoothing spline fits a smooth curve through data points to minimize noise, ⁶⁰ while SGF is particularly useful for preserving the signal’s features while smoothing. ⁶¹ Extended KF linearizes the system using the first-order term of the Taylor series expansion for the non-linear functions. ¹² Locally estimated scatterplot smoothing and locally weighted scatterplot smoothing are non-parametric methods that fit smooth curves to data using local regression techniques. Lastly, kernel smoothing regression applies a kernel function to weight data points, creating a smooth estimate of the regression function. ⁵⁶ All the techniques described in Table 4 are crucial for preprocessing data in bearing health monitoring, ensuring that the models are built on clear, noise-free signals.

Noise elimination is essential for high-frequency datasets such as CWRU and XJTU-SY, where vibration signals contain electrical and environmental noise. Smoothing filters are widely applied to these datasets to ensure that fault-related frequency signatures remain identifiable. A summary of methodologies and techniques for noise elimination is presented in Table 5.

Table 5.

Summary of the methodologies and techniques for noise elimination.

Method	Application	ML	References
MA	C-MAPSS, PRONOSTIA	Denoising generative adversarial network 1D-PHM	59, 62–65
Exponential MA	PRONOSTIA, CWRU, SU, Pardebon, MFPT, MFAULT, XJTU-SY	CNN-LSTM, error based evolving Takagi–Sugeno fuzzy model, gradient-aligned domain generalization, self-adaptive CNN, RNN	53
EWMA	PRONOSTIA	Weibull proportional hazards model	66
Linear rectification technology—EWMA	PRONOSTIA	Read-first-LSTM	67
Improved KF	PRONOSTIA	Wiener based exponential model	68
SGF	PRONOSTIA, XJTU-SY, IITR	LSTM	30, 69, 70
Extended KF	PRONOSTIA, IMS	–	12
Moving median	PRONOSTIA, Wind Turbine	–	71
Adaptive extended KF	PRONOSTIA, XJTU-SY	Bi-GRU	72
Permutation invariant training	PRONOSTIA	DL	52

Handling missing data

After improving signal quality, it is essential to address missing or incomplete data to ensure the accuracy and reliability of predictive models. Missing data may result from sensor dropouts, transmission issues, or experimental interruptions. Various techniques can be employed to manage and fill in gaps within datasets. These methods can be broadly categorized into interpolation and imputation strategies. Interpolation methods involve estimating missing values based on known data points using various mathematical approaches. Linear interpolation, for instance, estimates missing values using a straight line between two known data points. ⁷³ Spline interpolation employs piecewise polynomials to provide a smoother estimation across data points, while the modified Akima interpolation method enhances spline interpolation by reducing oscillations between points. ⁷⁴ Cubic Hermite spline interpolation preserves monotonicity and provides a smooth curve through data points. ⁷⁵ Inverse distance weighted interpolation estimates missing values by assigning weights that are inversely proportional to the distance from known points. ⁷⁶

Imputation methods, on the other hand, fill in missing data by estimating the most likely values based on information from other parts of the dataset. Mean imputation replaces missing values with the mean of the available data, while hot deck imputation fills in missing values using a similar record from the dataset. ⁷⁷ KNN imputation uses the average value of the k-nearest data points. ⁷³ Regression imputation estimates missing data using regression models based on other variables. ⁷⁷ Mean imputation utilizes a statistical approach to predict missing values, ⁷³ while support vector regression employs support vector machines to perform regression for imputation. ⁷⁵

FCM imputation applies fuzzy logic to cluster data points, imputing the data based on cluster centroids. ⁷⁸ Singular value decomposition uses matrix factorization techniques for data imputation. eXtreme gradient boosting is a ML approach that uses boosted trees to estimate missing data. ⁷³ Predictive mean matching matches missing values to observed values with similar predicted means, ⁷⁸ while multilayer perceptrons utilize neural networks to estimate missing values. ⁷⁴ Multiple RF regression imputations use an ensemble of decision trees to predict and impute missing data. ⁷⁹ The KF is a recursive algorithm used to estimate the state of a dynamic system and fill in missing data. Exponential weighted MA applies weighted averages to past observations to estimate missing values. ⁷³ In this part, no methods used by researchers for handling missing data were identified.

Handling missing data is crucial for long-running datasets such as IMS and Wind Turbine recordings, where sensor interruptions often create gaps. Interpolation and model-based reconstruction methods are therefore commonly applied to maintain time continuity in these datasets.

Normality test

Before applying transformations and scaling procedures, it is important to determine whether the data follow a normal (Gaussian) distribution, as many machine learning algorithms assume or benefit from normally distributed features. Normality tests evaluate whether a dataset conforms to a Gaussian distribution, providing essential guidance for subsequent statistical analyses and preprocessing decisions. The Shapiro–Wilk test is widely used, especially for small samples. Skewness and kurtosis assess asymmetry and peakedness of the data. The Anderson–Darling test focuses on tail behavior, while the Shapiro–Francia test is suited for larger samples. The Jarque–Bera test combines skewness and kurtosis to evaluate normality. The Lilliefors test adapts the Kolmogorov–Smirnov test for unknown population parameters, ⁸⁰ and the Ajne test offers a non-parametric approach based on data distances. ⁸¹ These tests help ensure that the data meet the normality assumption for accurate statistical analysis. We did not identify any methods used by researchers to find normality data.

Certain datasets, such as PRONOSTIA and PU, show significant deviations from normality due to varying operating conditions. Conducting normality tests guides the selection of proper normalization or transformation strategies for these datasets.

Normalization

Based on the results of the normality assessment, normalization ensures that all features operate within a comparable numerical range. Normalization improves model stability, accelerates training, and prevents variables with larger scales from dominating others. Normalization and transformation techniques are crucial for preparing data by adjusting their scale and distribution. Common techniques include z-score standardization, min–max scaling, and log transformations. Standardization shifts data to have a mean of 0 and a standard deviation of 1. ⁸² Min–max normalization scales data to a range, typically (0, 1). ²³ The Box–Cox transformation helps approximate normality by stabilizing variance, while the Yeo–Johnson transformation extends this method to handle zero and negative values. The John–Draper transformation and Manly transformation offer alternative approaches for variance stabilization. The modified Box–Cox and Lambert W × F transformations further enhance flexibility and handle specific data distributions. ⁸³ The methods used for handling data normalization are described in Table 6.

Table 6.

Summary of the methodologies and techniques for normalization.

Method	Application	ML	References
Min–max normalization	PRONOSTIA, CWRU, IMS, MFAULT, XJTU-SY, SEU	LTSM, CNN, RNN, extended attention signal transformer	23, 30, 82, 84–87
Z-score	PRONOSTIA	Transformer architecture	88

Normalization is especially important for multi-sensor datasets such as HUST, where different measurement channels (i.e. vibration, temperature, and speed) operate at different scales. Proper scaling ensures that no single feature dominates model training.

Imbalanced methods

Once the data have been normalized, class imbalance must be addressed, especially in bearing fault diagnosis where normal samples typically far outnumber fault samples. Imbalanced data can lead to biased models that disproportionately favor the majority class. Therefore, imbalance-handling methods are essential to ensure that underrepresented classes are properly learned and that the resulting models achieve balanced and reliable predictive performance.

Oversampling methods aim to balance the class distribution by generating synthetic samples for the minority class (Figure 11).

Figure 11.

Imbalanced method for Imbalanced data (adapted from Li et al. ⁸⁹ ).

SMOTE is widely used to create synthetic instances by interpolating between existing minority class samples. ⁸⁹ Borderline SMOTE focuses on generating synthetic samples near the decision boundary to improve classifier performance, ⁹⁰ while Kmeans SMOTE clusters the data first and then applies SMOTE within each cluster to maintain the natural structure. ⁹¹ SMOTE SVM combines SMOTE with SVM to enhance the synthetic sample generation process. ⁹² Adaptive synthetic sampling adjusts the sampling strategy dynamically based on the data distribution. ⁹¹

In contrast, cost-sensitive approaches directly address the imbalanced learning problem by assigning higher costs to misclassifications of the minority class. Cost-sensitive resampling modifies the resampling process by incorporating these costs. ⁹¹ Cost-sensitive algorithms integrate the cost function into the learning process itself, ⁹³ and cost-sensitive ensembles combine multiple models, each trained with cost-sensitive considerations, to improve classification performance on imbalanced datasets. ⁹⁴ All the methods used for imbalanced datasets are described in Table 7.

Table 7.

Summary of the methodologies and techniques for imbalanced data.

Method	Application	ML	References
SMOTE	CWRU, Bogie	MLP, SVM, CNN, CNN-1D, GAN	36, 91
Adaptive synthetic sampling	CWRU, NU	Deep residual shrinkage network, MLP, SVM, CNN, CNN-1D, LS-SVM	36, 89, 91
Borderline SMOTE	Gearbox bearing Wind Turbine, CWRU, IMS, IPF	Bi-GRU, LS-SVM, GAN	34, 90, 95
Cluster-MWMOTE	CWRU and IMS	LS-SVM	31
Sample-characteristic oversampling technique	CWRU, IMS	LS-SVM	95
Cluster SMOTE	CWRU, IMS	LS-SVM	95
Adaptive semi-unsupervised weighted oversampling	CWRU, IMS	LS-SVM	95
MWMOTE borderline SMOTE	CWRU, IMS	LS-SVM	95
Mixture GAN	CWRU, IMS, WTFF, IPF	GAN	34
Capsule auto-encoder	CWRU, IMS, WTFF, IPF	GAN	34
Residual joint adaptation adversarial network	CWRU, IMS, WTFF, IPF	GAN	34
Integrated class-imbalanced learning	CWRU, IMS, WTFF, IPF	GAN	34
Noise-immunity MWMOTE	CWRU	LS-SVM	96
Extended NI-MWMOTE	CWRU, PHM 2010 milling tool data, TTWD	LS-SVM	96
Conditional variational AutoEncoder generative adversarial network with self-modulation	CWRU	SVM, CNN	92

Class imbalance is a prominent issue in datasets such as CWRU and PRONOSTIA, where normal samples vastly outnumber fault samples. Techniques such as SMOTE and ADASYN are often applied to these datasets to achieve more balanced classification performance.

Feature selection

After ensuring data quality and class balance, the next step is feature selection, which identifies the most informative variables for model training. Eliminating redundant or irrelevant features reduces overfitting, accelerates computation, and improves model interpretability. This critical preprocessing step enhances overall model performance by retaining only the most relevant features, as illustrated in Figure 12.

Figure 12.

Overview of feature selection approaches used to identify the most informative features for fault diagnosis.

In supervised models, where labeled data guide the selection process, various techniques are employed. Sparse multinomial logistic regression identifies significant features by applying regularization to prevent overfitting. ⁹⁷ Automatic relevance determination regression automatically determines the relevance of each feature by adjusting its associated weight. ⁹⁸

The relief algorithm ranks features based on how well they distinguish between instances of different classes. ⁹⁷ RFE iteratively removes the least important features to optimize model performance, ⁹⁰ with SVM-RFE being a specific implementation that uses SVM. ⁹⁹ The chi-squared score evaluates the independence of features from the target variable, while the G Fisher score assesses the discriminative power of each feature. ⁹⁷ Fast correlation-based filter and correlation-based feature selection remove highly correlated features, retaining only those with strong predictive power. ¹⁰⁰ Weighted nearest neighbors feature selection and temporal redundancy maximum relevance feature selection further refine feature sets by considering proximity to class boundaries and temporal relevance. ⁹⁹ In cases where only a limited subset of data is labeled, semi-supervised feature selection can be employed to leverage both labeled and unlabeled data. ¹⁰¹ The majority of existing semi-supervised feature selection algorithms are based on constructing a similarity matrix and selecting the features that best align with this matrix.

In unsupervised models, where no labeled data are available, feature selection relies on intrinsic properties of the data. Mean absolute deviation measures feature variability to prioritize more informative features. ¹⁰² The dispersion ratio identifies features with distinct clustering potential. ¹⁰³ The Laplacian score selects features by evaluating their locality-preserving power, ⁷⁷ often combined with distance-based entropy for enhanced performance. ¹⁰⁴ Multi-cluster feature selection optimizes features for multiple clustering tasks simultaneously. ¹⁰⁵ Sequential feature selection incrementally adds or removes features to improve clustering or classification outcomes, ¹⁰⁶ while neighborhood components analysis focuses on feature selection for nearest neighbor algorithms. ⁹⁹ Techniques such as ℓ2,1-norm regularize discriminative feature selection. ¹⁰¹ All the methods for feature selection are detailed in Table 8.

Table 8.

Summary of the methodologies and techniques for feature selection.

Method	Application	ML	References
Minimum redundancy maximum relevance	CWRU, KU	Decision tree, RF, SVM, KNN, NB, ANN, RNN	42
RRelief	CWRU, KU	Decision tree, RF, SVM, KNN, NB, ANN, RNN	42
Feature selection based on Mahalanobis and density distance matrix	CWRU, KU	Decision tree, RF, SVM, KNN, NB, ANN, RNN	42
Monotonicity	CWRU, XJTU-SY, QPZZ-II simulation, IMS	SVVD	23, 107
Correlation	CWRU, XJTU-SY	SVDD	107
Robustness	CWRU, XJTU-SY	SVDD	107
Comprehensive index	CWRU, XJTU-SY	SVDD	107
Mahanabolis distance	PRONOSTIA	RNN	86
Unsupervised discriminative feature selection	CWRU and XIU	SVM	41
Memory space computation genetic algorithm	CWRU, CYCU	SVM, KNN	24
Binary salp swarm algorithm	CWRU, MFPT, UCI, IMCM-IITR	KNN, SVM	39
Binary whale optimization algorithm	CWRU, MFPT, UCI, IMCM-IITR	KNN, SVM	39
Binary gray wolf optimization extended repository	CWRU, MFPT, UCI, IMCM-IITR	KNN, SVM	39

Feature selection is widely applied to datasets such as IMS and Wind Turbine, where large numbers of statistical and spectral features are extracted. Eliminating redundant features improves computational efficiency and enhances model accuracy for these datasets.

Feature extraction

With the key features identified, feature extraction transforms the data into more descriptive representations that capture the underlying signal behavior. Extraction methods are especially valuable for vibration data because of their time-frequency characteristics. Feature extraction is a crucial process in signal processing and data analysis, particularly when dealing with both stationary and non-stationary signals. Non-stationary signals require specialized techniques due to their changing statistical properties over time.

In the time domain, methods such as the STFT allow for time-frequency representation by segmenting the signal into short intervals. ¹⁰⁸ The process of transforming a signal into STFT is detailed in Figure 13. The Wigner–Ville distribution provides a high-resolution time-frequency analysis but can suffer from cross-term interference. ¹⁰⁵ EMD breaks down signals into intrinsic mode functions, making it useful for analyzing non-linear and non-stationary processes. ⁸⁴ Spectral kurtosis analysis is employed to detect and characterize non-Gaussian transients in signals, ¹⁰⁸ while cyclostationary analysis focuses on signals with periodic statistical properties, making it effective in detecting modulated signals. ⁴⁶

Figure 13.

Visualization of the STFT procedure, showing the windowing process, FFT calculation, and resulting time–frequency representation.

In the frequency domain, various wavelet transforms are used to analyze non-stationary signals at different scales. The CWT provides a detailed time-frequency representation, ideal for identifying features at various resolutions. ⁸² The DWT, a more computationally efficient version, is commonly used for feature extraction in practical applications. ¹⁰⁹ WPT offers a richer time-frequency representation by decomposing both the high and low-frequency components. ⁴³ The Morlet wavelet, a complex wavelet, is particularly suited for analyzing oscillatory signals such as those found in power systems. ¹⁰⁸ Lastly, the Hilbert–Huang transform combines EMD with Hilbert spectral analysis, providing an adaptive method for analyzing complex, non-linear, and non-stationary signals. ¹¹⁰

For stationary signals, where statistical properties do not change over time, different approaches are taken. In the time domain, statistical methods such as RMS, skewness, Kurtosis, crest factor, clearance factor, and peak factor are used to quantify the amplitude, shape, and distribution of the signal. ¹¹¹ Model-based approaches, such as stationary HMM and autoregressive models, are employed to capture the underlying structure and predict future values of stationary signals. ¹¹² Signal processing techniques such as fractal analysis and thermoelastic stress analysis are also applied to extract features that characterize the complexity and stress patterns within stationary signals. ¹¹³ In the frequency domain, spectral analysis evaluates the frequency content of the signal, ⁸² EA is used to extract features related to the amplitude modulation, ¹⁰⁸ Cepstrum analysis helps in identifying periodicities in the frequency spectrum, ¹¹⁴ higher order spectrum provides insights into the non-linear interactions within the signal, ¹¹⁵ and the Fourier transform remains a fundamental tool for transforming time-domain signals into their frequency components. ¹¹⁶ These techniques together form a comprehensive toolkit for extracting meaningful features from both stationary and non-stationary signals. The methods used for handling feature extraction are described in Table 9.

Table 9.

Summary of the methodologies and techniques for feature extraction.

Method	Application	ML	References
EMD	PRONOSTIA, XJTU-SY, centrifugal pump bearing	LSTM	84
Ensemble EMD	CWRU, KU	FCM, Gath-Geva clustering, decision tree, RF, SVM, KNN, NB, ANN, RNN, ELM	42, 48, 117, 118
Adaptive local iterative filtering	CWRU	ELM	119
FFT	CWRU, MFPT, UCI, PRONOSTIA, centrifugal pump bearing dataset	LSTM, active federated local subdomain adaptation, LSSVM, CNN-SVM, SVM, LDA, KNN, SVVD	39, 110, 116, 120
Multi resolution analysis	CWRU, MFPT, UCI, and experiment	SVM, LDA, LSTM, LSSVM, CNN-SVM, SVM, KNN,	39, 116
Hilbert Huang transform	CWRU, UCI, CYCU, PRONOSTIA, XJTU-SY	KNN, SVM, LSTM, SVVD	24, 84, 110
Variational mode decomposition	CWRU, UCI, CYCU, IMCM-IITR, NCEPU	KNN, SVM, BPNN, ELM, WNN, MLP, BPNN, ELM, SVM, improved genetic algorithm—SVM	24, 44, 121
EA	CWRU, UCI, CYCU	KNN, SVM, CNN	24, 39, 82
Time-domain (mean, maximum, minimum, average absolute amplitude, waveform-factor, standard deviation, RMS, variance, Kurtosis, skewness, crest factor, clearance factor, shape factor, form factor, energy, line integral, impulse factor, margin factor, entropy, Kurtosis index, skewness index, wave index, margin index, pulse index, peak index, abs mean, impulse)	CWRU, QPZZ-II simulation, XJTU-SY, PRONOSTIA and IMCM-IITR, IMS, JNU, IITR, NCEPU	LSTM, CNN, RNN, Weibull proportional hazards model, BPNN, ELM, SVM, cross-transformer, Gaussian progress regression, Fusion Graph, LightGBM, SVVD, CNN, AdaBoost-WSO-SVM, DL	23, 30, 42–44, 52, 53, 66, 85, 86, 107, 111, 117, 122–127
CWT	CWRU, SEU	CNN	82, 111
Frequency domain (peak to peak, energy, power spectral density, mean frequency, median frequency, RMS frequency, spectral entropy raw signal, spectral entropy envelope signal, dispersion and concentration of the spectrum, change of the main frequency position index, frequency static values 3–12, frequency center of gravity, standard deviation frequency)	CWRU, XJTU-SY, PRONOSTIA, and IMCM-IITR, IMS, XJTU-SY, PRONOSTIA, JNU, IITR, NCEPU	LTSM, BPNN, ELM, SVM, cross-transformer, LightGBM, SVDD, AdaBoost-BP neural network, AdaBoost-WSO-SVM, DL	30, 39, 43, 44, 47, 52, 53, 107, 111, 117, 124
Time-frequency (WPT-Haar, energy ratio)	PRONOSTIA, CWRU, UCI, JNU	Cross-transformer, LightGBM, SVM, KNN, AdaBoost-WSO-SVM	30, 53, 124
t-SNE	CWRU, XJTU-SY	1D CNN, SVM	128
WPT	CWRU	WNN, MLP, KNN, BPNN, ELM, SVM	44, 109
DWT	CWRU	WNN, MLP, KNN, BPNN, ELM, SVM,	44, 109
1D-CNN	XJTU-SY	simple recurrent unit	129
Euclidean feature values	CWRU	–	130
Spectral Kurtosis	CWRU, IMS	CNN	108
STFT	CWRU, IMS	CNN, LSTM, deep domain confusion, deep adaptation network	108, 131, 132
Wavelet decomposition	CWRU, LVCAPSM	AdaBoost-BP neural network	43
Empirical wavelet decomposition	CWRU	–	133
Modified hierarchical time-shift multi-scale amplitude-aware permutation entropy	CWRU, XIU	SVM	41

Feature extraction plays a major role in non-stationary datasets such as PRONOSTIA, XJTU-SY, and PU, where time–frequency methods are required to capture the evolving fault signatures under variable load conditions.

Feature reduction

The final preprocessing stage involves reducing the dimensionality of high-dimensional feature sets. This step minimizes redundancy, lowers computational cost, and enhances visualization while preserving essential information. Dimensionality reduction converts data from a high-dimensional space into a lower-dimensional one while retaining the most important characteristics of the original data. AutoEncoder is effective in transforming high-dimensional features into more compact representations, ⁴⁰ as shown in Figure 14. The goal is to maintain a representation that reflects the data’s true underlying structure or intrinsic dimension. All the methods for feature reduction are detailed in Table 10.

Figure 14.

Operational flow of an AutoEncoder, including encoding, bottleneck representation, and decoding for feature learning.

Table 10.

Summary of the methodologies and techniques for feature reduction.

Method	Application	ML	References
Stacked AutoEncoder	CWRU, QPZZ-II, PU, HUST	LTSM, CWRU, IMS	23, 45
Transfer learning	CWRU, RHV	CNN	49
Topology learning and out-of-sample embedding	AUT, CWRU	SVM	29
LPP	CSU, CWRU, AUT	ELM, MPA-SVM	29, 46
NPE	CSU, CWRU	ELM, WNN, MLP, KNN	46, 109
Orthogonal LPP	CSU, CWRU	ELM	46
Multi-level feature fusion	NUE, CWRU	Multi-layer pooling classifier—CNN	134
PCA	CWRU, KU, PU, MFPT, PU, NCEPU	MLP, KNN, stacked denoising AutoEncoders, wavelet scattering network, IKELM, WNN, MLP, KNN, ELM	33, 44, 46, 47, 109, 117, 135
Kernel PCA	CSU, CWRU, NCEPU	IKELM, ELM	44, 46
Weighted PCA	CWRU and OtU	GMM cluster	50
ISOmap	CWRU, NCEPU	IKELM, WNN, MLP, KNN	44, 109
t-SNE	CWRU, SEU, PU, IMS, WTFF, IPF	Back propagation—incremental broad learning system, transferable Siamese network, envelope spectrum neural network, cross-domain adaptation model based on sparse filtering, improved residual neural network, CNN, MLP, SVM, CNN-1D, improved multi-scale convolutional neural network, federated transfer learning, depthwise separable convolution-CNN, GAN, WNN, MLP, KNN	34, 82, 109, 122, 136–147
AutoEncoder	CWRU, JNU, XJTU-SY, and PRONOSTIA	Kmeans, GMM, FCM, LTSM, CNN	40, 111, 114
LLE	CWRU, NEPU, CUP	SVM, NB, RF	38, 47, 109
Supervised LLE	CWRU, NEPU	SVM, NB, RF	38
Semi-supervised LLE	CWRU, NEPU	SVM, NB, RF	38
Local mutual representation LLE	CWRU, NEPU	SVM, NB, RF	38
Robust LLE	CWRU, CUP	–	47
LLE vote	CWRU, CUP	SVM, NB, RF	38, 47
Probabilistic PCA, generalized discriminant analysis, stochastic proximity embedding	CWRU	WNN, MLP, KNN	109
LTSA, semi-supervised ISOmap	CWRU, AUT	MPA–SVM	29
Topology learning and out-of-sample embedding	CWRU, AUT	MPA–SVM	29

Dimensionality reduction methods such as PCA and stacked AutoEncoders are particularly effective for high-dimensional datasets such as CWRU, HUST, and QPZZ-II, where hundreds of time–frequency or deep-learning features are generated. Reducing dimensionality improves model training stability and interpretability.

As shown in Figure 15, certain preprocessing and modeling methods have become dominant because they effectively address the challenges inherent in vibration-based bearing diagnostics. Techniques such as FFT, PCA, t-SNE, and min–max normalization are widely used because they support the two most common preprocessing objectives: feature extraction and feature reduction. FFT enables the transformation of raw signals into the frequency domain, where fault signatures are more pronounced, while MA smoothing helps reduce noise. PCA and t-SNE are popular for their ability to compress high-dimensional vibration features into compact and separable spaces, facilitating more accurate model training. On the modeling side, CNNs and SVMs stand out. CNNs are preferred in deep learning frameworks because they excel at learning multi-level features from spectrograms, FFT maps, or raw signals without requiring extensive manual engineering. Conversely, SVMs remain strong among traditional ML techniques due to their robustness on smaller datasets and their ability to effectively classify reduced or handcrafted feature sets derived from PCA or FFT.

Figure 15.

Distribution of preprocessing objectives, techniques, and ML models used across reviewed studies.

Figure 15 also indirectly highlights the trade-offs between traditional machine learning and deep learning methods. Traditional approaches such as SVM, KNN, and MLP are often simpler, computationally efficient, and perform well when combined with strong preprocessing and feature extraction—making them suitable for scenarios with limited data or lighter hardware. In contrast, deep learning methods such as CNN and LSTM provide superior performance when sufficient data are available, as they automate feature extraction and can capture more complex nonlinear degradation patterns. However, they require greater computational resources and are less interpretable compared to traditional models. The distribution in Figure 15 reflects this balance: studies relying heavily on hand-crafted features tend to prefer traditional ML models such as SVM, while those incorporating richer signal representations or large datasets gravitate toward CNNs for their ability to learn directly from raw or minimally processed inputs.

Our analysis revealed inconsistencies in the preprocessing methods employed by researchers, despite the dominance of DL techniques in their data-driven models. In summary, we conducted a comprehensive evaluation of all models and quantified the percentage usage of different preprocessing objectives, techniques, and ML models, as illustrated in Figure 15. From this SLR, we identified that the data-driven approaches used for bearing degradation prediction can be broadly categorized into three groups: statistical methods, ML, and DL. A detailed summary of these findings is presented in Figure 16.

Figure 16.

Collection of data-driven approaches applied in bearing fault diagnosis research.

By completing the dimensionality reduction step, the preprocessing pipeline ensures that the data are clean, compact, and ready for subsequent analysis.

For better understanding, all preprocessing methods identified in the reviewed literature are presented in Figure 17.

Figure 17.

Summary of all preprocessing methods identified in the reviewed literature.

Summary and conclusion

Industrial machines increasingly rely on vibration and acoustic emission monitoring to detect bearing faults and prevent costly downtime. However, fault detection remains challenging due to uncertainties in data acquisition, sensor accuracy, instrument precision, model predictions, and varying operating conditions. To mitigate these issues and improve data quality, preprocessing has become a critical step in PHM workflows.

This paper has presented a systematic literature review and bibliometric analysis of preprocessing methods for vibration-based bearing fault diagnosis, focusing on high-quality studies from 2020 to 2024 that use the PRONOSTIA and CWRU benchmark datasets. The review highlights publication trends and research themes, and maps how preprocessing is integrated into data-driven PHM pipelines.

Across the surveyed works, commonly used preprocessing operations include noise elimination, outlier detection, handling imbalanced data, feature extraction, feature selection, and feature reduction. Feature reduction is particularly widespread, with t-SNE and related techniques frequently used for dimensionality reduction and visualization. At the modeling level, CNNs have emerged as the dominant deep learning approach, while PRONOSTIA and CWRU remain the primary reference datasets for benchmarking and comparison.

A clear pattern is that deep learning studies typically rely on relatively simple preprocessing (e.g. normalization and outlier detection), whereas traditional statistical and shallow learning methods employ richer preprocessing chains, combining several of the above techniques. This reflects a trade-off between model complexity and preprocessing effort. At the same time, several gaps persist: most studies still rely on controlled laboratory datasets rather than real-time industrial data; multi-sensor fusion at the preprocessing stage is limited; explainable AI for understanding learned representations is underexplored; and there is a need for more diverse public datasets that capture complex and early-stage faults under realistic noise and operating conditions.

For practitioners, the findings suggest that preprocessing should be chosen in conjunction with the modeling paradigm and application constraints. In data or resource-limited settings, carefully designed traditional methods remain attractive, while deep learning approaches can benefit from simple but well-chosen preprocessing when sufficient data and computation are available. Overall, this review consolidates current knowledge on preprocessing and datasets for bearing PHM and outlines promising directions toward more robust, interpretable, and industrially relevant fault diagnosis systems.