Stall detection in hydraulic excavator operations using heuristics and machine learning: A case study

Abstract

This work aims to develop a reliable algorithm for stall detection during excavator digging by analyzing key operational variables such as velocity and angular displacements from machine monitoring data. The work develops and validates a heuristic algorithm to detect stalling events and trains a support vector machine classification algorithm to distinguish between “normal” digging cycles and cycles with stalling. This work is a novel attempt at using a classification algorithm to categorize digging cycles into normal and those with stalling events based on machine monitoring data alone. The developed classification algorithm achieved a sensitivity of 100%, indicating it correctly identified all stalling cases, and a specificity of 99.0%, demonstrating its ability to accurately classify normal cases with minimal false positives. This study provides a foundation for stall detection and facilitates further study of the effect of stalling on excavator performance, improving operational reliability, and facilitating better excavator design.

Keywords

Hydraulic excavators machine learning heuristics support vector machine classification data analysis

Introduction

Hydraulic excavators are one of the most critical equipment for construction and mining operations. It is the primary equipment for material extraction and loading (Fujita et al., 2011). Their efficiency is crucial for any operation they engage in, impacting overall productivity. Since they work in high-demand environments, even minor disruptions can cause significant operational delays and financial losses (Manyele, 2017). Furthermore, maintaining and repairing such critical and costly equipment is essential, requiring proactive measures to ensure continuous and efficient operation.

Figure 1 shows the main components of hydraulic excavators. Modern hydraulic excavators have advanced hydraulic systems and can handle heavy loads. The bucket is attached to the stick (or dipper arm), which provides reach and depth and is controlled by hydraulic cylinders. The boom, which is also powered by hydraulic cylinders, is connected to the stick and offers vertical reach and lifting capability. The cabin serves as the operator's control center.

Figure 1.

Nomenclature of hydraulic excavator (Turbosquid, 2020).

Stalling during operations is one of the most critical challenges a hydraulic excavator can experience. Stalling occurs when the bucket stops moving after encountering hard ground, causing the hydraulic system to reach its pressure limit and temporarily lose power (Savage, 1938). Stalling reduces productivity and accelerates wear on mechanical components, increasing maintenance costs and reducing equipment life. Addressing this issue is essential to maintaining optimal operational efficiency and ensuring the economic viability of numerous mining and construction projects. Developing advanced detection systems that help us better understand stalling can improve decision-making and optimize resource allocation, contributing to more sustainable operations. This approach leverages data-driven methodologies for improving excavator performance in mining and construction.

Stalling detection is vital for optimizing hydraulic excavator operations and understanding operator behavior during challenging digging conditions, including digging in hard ground. Stalling is an extreme case of difficulty in digging, making its detection a crucial indicator for identifying operational inefficiencies and potential equipment stress. Original equipment manufacturers (OEMs), and mining and construction operators rely on monitoring data to study and improve performance. However, accurately identifying difficult digging events from this data requires effective stall detection methods (Haertel et al., 2024). By identifying these events, mine managers and engineers can gain insights into how geologic conditions, machine performance, and operator techniques interact, ultimately reducing wear on equipment, improving productivity, and minimizing operational costs. Moreover, understanding the effect of stalling can guide operator training programs and facilitate the development of more efficient and durable machines.

Stall detection in hydraulic excavators is complex because of operational and data-related challenges (Dunbabin and Corke, 2006). The dynamic nature of excavator operations occurs in diverse geological conditions that result in significant variability in soil or muck pile composition, material density, and resistance to digging. These variations cause fluctuations in operating parameters, complicating the establishment of reliable thresholds for detecting stalling events. Additionally, as excavators age, sensor data may increasingly indicate performance degradation instead of actual stalling, further complicating detection. Some commercial monitoring systems lack critical parameters, such as hydraulic pressure readings, ideal for identifying stalling events. This limitation constrains the accuracy and scope of detection models. Moreover, the reliability of detection systems is affected by incomplete or noisy sensor data, which can reduce the effectiveness of both heuristic-based and machine-learning approaches. To tackle these challenges, advanced detection methods that combine domain knowledge with data-driven approaches are crucial for achieving accurate and reliable stall detection.

Sensor technology and data transmission advancements have made telemetry essential for monitoring mining equipment by offering continuous insights into excavator performance and conditions (Chetty, 1983). By integrating these technologies, mining and construction companies can optimize equipment utilization, reduce downtime, and predict potential failures before they occur, ultimately reducing operating costs. Despite these advancements, the massive volume of data generated by these systems presents data storage and analysis challenges. To address the challenges of analyzing a large volume of data, some have used artificial intelligence (AI) and machine learning to process the data and provide meaningful insights (Mahamedi et al., 2022).

While previous research addresses digging challenges (Chacko et al., 2014; Fairhurst, 2017; Li et al., 2024), detecting stalling events remains a significant challenge. There is an opportunity to apply machine learning techniques to evaluate telemetry data for stall detection. Thus, the main goal of this study is to create an algorithm for accurate stalling detection and classification. By utilizing a blend of heuristic methods and machine learning algorithms, this research seeks to overcome the limitations of traditional detection techniques, providing a solution that is both accurate and adaptable to real-world situations. This paper focuses on the following specific objectives:

Develop a reliable algorithm for stall detection during excavator operations by analyzing key operational variables such as velocity and angles, using data from a specific monitoring system as a case study.

Implement statistical and machine learning techniques using a support vector machine classification algorithm to distinguish between normal digging cycles and those with stalling.

This research is novel because it is a pioneering attempt to propose a novel methodology for using hydraulic excavator monitoring data to detect stalling events. Additionally, by making it easy to identify stalling, this work will facilitate further studies on the effects of stalling and provide practical insights and tools to enhance operational efficiency and minimize equipment downtime.

This paper is organized into six sections. Section 2 provides a detailed discussion of the relevant literature and explains related concepts. Section 3 introduces the data from the monitoring system used in this work and the framework for defining stalling events in this work. It also presents the overall approach to this work. Section 4 presents the proposed heuristic algorithm and its validation. Section 5 discusses the support vector machine (SVM) model to classify normal and stalling cases. Finally, Section 6 summarizes the key insights and suggests future directions.

Literature review

Hydraulic excavators are vital equipment in both mining and construction. Understanding how this machinery operates is essential, as it greatly assists in excavating and loading earth materials (Kirmanli and Ercelebi, 2009) (Vylomov and Kovalev, 1982). Their significance comes from their ability to perform loading tasks efficiently, contributing to optimized cycle times and ensuring smooth operations in the diverse environments where mines and construction projects use them (Victor et al., 2020). Therefore, it is crucial for hydraulic excavators to be precise, quick, and have a strong lifting capacity (Yu and Choi, 2007).

Even though hydraulic excavators are capable of functioning in challenging environments, they may experience stalling when the hydraulic system fails to overcome resistance forces, resulting in a temporary stop in operation, during digging (Jud et al., 2019). Stalling can happen for several reasons, including mechanical problems and challenging working conditions (Lever, 2001). Understanding the causes and effects of stalling is essential for maintaining productivity and minimizing unexpected downtime from failures induced by the excessive loads that lead to stalling. Thus, it is important to understand material diggability (the ease or difficulty in digging a particular material) and ensure excavators are not exposed to unduly challenging digging conditions.

Digging is the most critical and time-intensive part of hydraulic excavators’ excavation and loading process (Jeong and Phillips, 2001). It involves excavating material from the ground and transferring it into the bucket (Aluko and Awuah-Offei, 2024). This process is influenced by material density, operator skill, and machine capabilities (Wu et al., 2021). Given its significance, research has extensively examined diggability, as variations in geological conditions can impact the difficulty of excavation in construction and mining (Bell et al., 1995; Khorzoughi and Hall, 2016; Santi, 2005).

Studies have found that differences in geology, fragmentation, and bench conditions can significantly influence digging performance (Cheng and Teizer, 2013). Compact and harder rock formations result in more challenging digging conditions that typically require more energy for excavation (Chen et al., 2018). The degree of fragmentation influences diggability because materials with smaller-sized blocks and higher void ratios are easier to excavate, whereas large, unbroken rock masses with low void ratios increase resistance (Nikkhah et al., 2022; Shehu and Hashim, 2020) (Malcolm et al., 2018). Proper fragmentation improves efficiency by reducing the energy needed for excavation (Xiaohe et al., 2024). In contrast, poorly fragmented rock can result in increased wear on equipment and longer excavation times (Johnson, 2018). Particularly in mining, bench conditions also play a role in diggability (Lashgari and Kecojevic, 2013). Factors such as slope, terrain irregularities, and material consistency all affect the amount of effort required for digging and the probability of encountering stalls (Bettens et al., 2022). Consequently, these factors can increase the likelihood of operational stalls, as noted in the literature (Bell et al., 1995). Understanding these factors is crucial for optimizing digging strategies and reducing disruptions caused by challenging conditions.

Challenging digging conditions place significant loads on the hydraulic systems of excavators. These loads can lead to stalling events due to excessive resistance forces (Karpuz et al., 1992). In such conditions, the hydraulic system must operate at maximum capacity, which puts a strain on components such as pumps, cylinders, and motors (Dyorina et al., 2020). Over time, this increased stress can cause accelerated wear, leading to more frequent maintenance needs and a higher risk of mechanical failures (Frimpong and Li, 2007). One of the most common consequences hydraulic excavators face in difficult digging conditions is increased hydraulic pressure and temperature, which can lead to the system overheating (Kumar and Srivastava, 2012). Excessive wear caused by excessive loads can reduce operational lifespan and increase maintenance frequency (Redmann, 2018). Additionally, extreme loads can delay the excavation process, which reduces overall productivity (Mohammadi et al., 2015). These extended excavation times can slow project completion and result in higher fuel consumption and operational costs, ultimately affecting overall efficiency (Saeed Abolhasani H. Christopher Frey and Pang, 2008). Extended exposure to challenging digging conditions raises the likelihood of equipment stalling and inefficiencies, underscoring the necessity for strategies to prevent stalls (Dong et al., 2017).

Although numerous studies have examined diggability and its effects on hydraulic excavators, research on preventing stalling in hydraulic systems remains relatively limited. Some studies have explored methods to manage resistance forces and improve digging efficiency (Coetzee and Els, 2009; Hirano et al., 2021; Yoshida et al., 2016). A practical method for preventing stalling is to avoid using cylinder force beyond certain fixed thresholds (Hirano et al., 2021). Research indicates that adaptive control algorithms, such as the automated digging algorithm designed for hydraulic excavators, can dynamically adjust cylinder force (Yoshida et al., 2016). Simulations indicate that modifying digging trajectories and force application can decrease energy consumption and enhance performance, providing valuable insights for excavation operations (Coetzee et al., 2007; Coetzee and Els, 2009; Liu et al., 2024). However, the industry has not implemented these methods yet, leaving mines and construction projects with no means of minimizing stalling events.

Managing excessive resistance forces is another way to minimize the chances of stalling. One effective method involves real-time estimation of the resistance forces to excavation on the bucket, utilizing a two-stage approach based on kinematic and dynamic models (Palomba et al., 2019). By continuously monitoring the forces between the bucket and the material, the system can optimize digging parameters (Bonchis et al., 2011; Luengo et al., 1998) to prevent stalling and enhance overall efficiency. Research shows that adjusting digging trajectories and force application when encountering high resistance can improve efficiency and reduce strain on the hydraulic system (Yoshida et al., 2013). For example, adjusting bucket tip traces and optimizing digging speeds can significantly improve excavator performance, resulting in more efficient digging (Sakaida et al., 2006; Sano and Ichiryu, 2002). Before OEMs implement these solutions in hydraulic excavators, there is potential to use existing monitoring systems and the empirical data they generate to study the effect of stalling and operator behavior during stalling events. However, a key barrier to doing this is the availability of algorithms to detect stalling events from available monitoring data. These authors could not find any algorithm that detects stalling from commercially available monitoring system data. If such algorithms were available to engineers in mining and construction, they would use them to isolate stalling events in their monitoring data and use those incidents to understand the causes of stalling, training operators on how to mitigate its effects, and understand the effects of stalling on equipment performance, health, and longevity. This is the focus of this paper.

Case study data

This study's aim is to develop an algorithm for detecting stalling in hydraulic excavators and to train an SVM classification algorithm. The work uses the case study of data from a commercial monitoring system (not named here because the provider wants to remain anonymous) acquired from a Hitachi EX5600 hydraulic excavator. This excavator has an operating weight of 541,000–545,000 kg (depending on the front attachment), a bucket capacity of 27–31 m³, and two engines, each rated at 1119 kW (Americas, 2025). The authors examined key parameters, including hydraulic pressure, engine load, bucket position, and cycle times, to gain insights into the excavator's behavior and performance. The data is recorded at a frequency of 30 Hz, capturing operational data through a combination of sensors and telemetry systems (Kirianaki et al., 2002).

The monitoring system collected data in real-time. Employes of the provider of the monitoring system retrieved the data, processed it using a proprietary MATLAB algorithm to estimate the payload and other excavator parameters, and sent it to this research team as csv files. The raw data (Table 1) consists of 26 variables. Figures 2 and 3 show plots of some of the data from a sample cycle to illustrate the nature of the data.

Figure 2.

Sample plot of angles in one cycle.

Figure 3.

Sample plot of other variables in one cycle.

Table 1.

Raw data from monitoring system provider.

No	Variable	Description	Type
1	$ϕ_{0}$	Cab roll	Double
2	$θ_{0}$	Cab pitch	Double
3	$P_{h e a d}$	Boom head pressure	Double
4	$P_{r o d}$	Boom rod pressure	Double
5	$F$	Force for both cylinders in the boom	Double
6	$θ_{b o o m}$	Boom angle (rad)	Double
7	$θ_{s t i c k}$	Stick angle (rad)	Double
8	$θ_{b u c k e t}$	Bucket angle (rad)	Double
9	$θ_{s w i n g}$	Swing angle (rad)	Double
10	${\dot{θ}}_{b o o m}$	1^st derivative of Boom angle	Double
7	${\dot{θ}}_{s t i c k}$	1^st derivative of Stick angle	Double
8	${\dot{θ}}_{b u c k e t}$	1^st derivative of Bucket angle	Double
9	${\dot{θ}}_{s w i n g}$	1^st derivative of Swing angle	Double
10	${\dot{ϕ}}_{0}$	1^st derivative of Cab pitch	Double
11	${\dot{θ}}_{0}$	1^st derivative of Cab roll	Double
12	${\ddot{θ}}_{b o o m}$	2^nd derivative of Boom angle	Double
13	${\ddot{θ}}_{s t i c k}$	2^nd derivative of Stick angle	Double
14	${\ddot{θ}}_{b u c k e t}$	2^nd derivative of Bucket angle	Double
15	${\ddot{θ}}_{s w i n g}$	2^nd derivative of Swing angle	Double
16	$E n u m$	State of the excavator during each cycle	Double
17	$P$	Instant payload	Double
18	$M_{b}$	Moment present in the boom	Double
19	$τ$	Boom torque	Double
20	$t$	Time vector	Datetime
21	$O p$	Name of the Operator “String”	Categorical
22	$θ_{b o o m}$	Boom Angle (radians)	Double
23	$θ_{s t i c k}$	Stick Angle (radians)	Double
24	$θ_{b u c k e t}$	Bucket Angle (radians)	Double
25	$θ_{s w i n g}$	Swing Angle (radians)	Double

Stall detection in hydraulic excavators is challenging because commercial monitoring systems do not always have all the required data for optimal detection. One significant problem with the data in this work is that it did not include hydraulic pressure readings from the stick and bucket cylinders (Table 1), which are essential for identifying stalling events. These specific pressure signals were not available in the monitoring system employed in this case study, despite their direct relationship to resistance forces and stalling behavior. Monitoring hydraulic pressure in the cylinders that control bucket motion during digging could provide valuable insights into the onset and severity of stalling events. The lack of bucket and stick cylinder pressure data limited the ability to directly validate the proposed stall detection algorithms using force-based indicators. However, this limitation reflects a common constraint in industry applications, where commercially deployed monitoring systems often provide kinematic and payload information but omit detailed measurements of some important parameters due to sensor availability, cost, or system design considerations. In such contexts, developing innovative detection methods that rely on widely available data is essential. Accordingly, this work leverages domain knowledge and data-driven techniques to infer stalling events from kinematic behavior and payload stability alone, demonstrating that meaningful stall detection is achievable even when ideal sensor data are unavailable. A key step in this analysis is identifying and separating individual loading cycles, as accurate segmentation is essential for obtaining meaningful insights and precise interpretations. A standard hydraulic excavator loading cycle has four main stages: swinging into position, digging to fill the bucket, swinging away from the material to the truck, and dumping the load. The dataset includes a discrete variable called ‘Enum’, shown in Figure 4, which identifies the operational states with three distinct values: 0 for swinging (for both swing in and out), 3 for digging, and 5 for dumping. This variable, which is estimated by the provider using proprietary algorithms, is critical for distinguishing between the stages of the cycle. This work assumed that the states indicated by the ‘Enum’ variable accurately reflect the actual operational states.

Figure 4.

Enum variable.

This work uses this case study data to show how one can detect stalling using a heuristic algorithm and for training the classification algorithm.

Heuristics-based stall detection

Heuristics-based detection methods are valuable due to their simplicity, low computational requirements, and quick response times, making them ideal for mining and construction applications (Ball, 2011). These methods rely on predefined rules or thresholds based on operational parameters and provide efficient and quick detection advantages (Calle-Escobar et al., 2016). Heuristic algorithms are effective because they can deliver acceptable solutions regarding time and space complexity, even in complex scenarios where exact algorithms might require excessive computational resources or time (Kokash, 2005) (Libkin, 2014).

Because the algorithm in this work relies on the position of the bucket tip, the research team estimated the coordinates of the bucket tip and added that data to the original data file. Equations (1) and (2) show the equations for the Y and X coordinates of the bucket tip, respectively. Figure 5 illustrates the geometric relationships underlying these equations. For the Hitachi EX5600 excavator used in this work, the necessary dimensions for estimating the coordinates (Figure 5) are the boom length ( $a_{1}$ ) of 7.69 m, the stick length ( $a_{2}$ ) of 5.33 m, and the bucket length ( $a_{3}$ ) of 4.16 m. The boom joint is positioned with an offset of 5.48 m in the Y-direction ( $y_{0}$ ) and 1.10 m in the X-direction ( $x_{0}$ ). Figure 6 shows a sample bucket trajectory during a cycle using the estimated bucket tip positions. These bucket tip positions, and the estimated velocities are important in the heuristic algorithm that is described in the next section.

B u c k e t_{H t} = a_{1} * \sin (q_{1} - θ_{o}) + a_{2} * \sin (q_{1} + q_{2} - θ_{o}) + a_{3} * \sin (q_{1} + q_{2} + q_{3} - θ_{o}) + y_{0} * \cos (- θ_{o}) + x_{0} * \sin (- θ_{o})

(1)

B u c k e t_{x} = a_{1} * \cos (q_{1} - θ_{o}) + a_{2} * \cos (q_{1} + q_{2} - θ_{o}) + a_{3} * \cos (q_{1} + q_{2} + q_{3} - θ_{o}) + x_{0} * \cos (- θ_{o}) - y_{0} * \sin (- θ_{o})

(2)

Figure 5.

Bucket tip coordinates.

Figure 6.

Sample bucket tip trajectory for one cycle.

Algorithm

The algorithm for detecting stalling events from our data, which is implemented in MATLAB, is based on several constraints that must be satisfied simultaneously to classify the identified periods as stalling. Specifically, these constraints involve setting operational thresholds for key parameters, such as position and payload. This algorithm mainly relies on the fact that stalling events will lead to no motion at the bucket tip (linear or angular). Thus, the algorithm attempts to find periods during the digging phase of the cycle where there is no motion.

Table 2 defines the symbols used in the equations underlying the algorithm. Equations (3) – (9) present the conditions the algorithm tests for in each “discrete period” to determine whether there is stalling or not.

Table 2.

Heuristics algorithm symbols.

Symbol	Meaning
$v_{x}$	Bucket tip velocity (X)
$v_{y}$	Bucket tip velocity (Y)
$ω$	Bucket tip velocity (Angular)
$ϵ$	Threshold
$P$	Payload
$\frac{1}{T} \sum_{t ϵ T_{s t a l l}} x$	Average over the time
$y (t)$	Value of the variable at a specific time
$y_{m a x}$	Maximum value of y(t)
$t_{s t a r t}$	Starting time of a specific event
$t_{e n d}$	End time of a specific event
$w$	Additional time duration (delay)
$t_{s t a l l s t a r t}$	Starting time of stalling
$t_{s t a l l e n d}$	End time of stalling

The Equations are presented as follows:

| v_{x} | < ϵ

(3)

| v_{y} | < ϵ

(4)

| ω | < ϵ

(5)

V a r (P) < ϵ

(6)

y (t) < 0.9 y_{m a x}

(7)

t_{s t a l l s t a r t} = t_{s t a r t} + w

(8)

t_{s t a l l e n d} = t_{e n d} - w

(9)

The first three equations (Equations (3) – (5)) establish a threshold condition for the bucket's linear velocities in the X- and Y-directions ( $v_{x}$ and $v_{y}$ ), and angular velocity ( $ω$ ). These conditions require that $| v_{x} |, | v_{y} |,$ and $| ω |$ remain below a certain threshold ( $ϵ$ ) for a defined duration. This ensures that the bucket is not just momentarily slowing down but maintaining a near-zero velocity over time. Based on multiple experiments and our analysis, we concluded that a threshold of 0.1 m/s and 0.1 rad/s work very well. These conditions verify that the velocity is below a small threshold. These values are significantly smaller than typical bucket and stick velocities observed during normal digging operations, which are generally an order of magnitude higher, thereby representing a physically meaningful near-stationary condition rather than an arbitrary cutoff.

By applying these conditions over consecutive time steps, the algorithm reduces the likelihood of false motion detections while confirming that the bucket is genuinely stationary over a sustained period. Exploratory analysis indicated that moderate variations around these threshold values did not alter the identification of pronounced stalling events, suggesting that the algorithm is robust for extreme stalling conditions, which are the primary focus of this study.

To meet this criterion, the velocities must stay below the threshold for at least 20% of the average duration of all identified digging cycle times. This value was determined with an empirical analysis that ensures a good balance between computational time (the lower the percentage of digging time used, the more computational time is required as the number of discrete intervals increase) and the number of cases identified. This approach ensures that the threshold condition is applied consistently across the dataset, considering both the operational speed and the length of the digging intervals.

While the first three conditions are necessary, they are not sufficient to define stalling events in this case. Equation (6) ensures that the payload remains stable during the analysis. Stalling occurs when the bucket is stationary, so the payload should not fluctuate significantly during these periods. This condition helps to exclude intervals where there are changes in the payload, indicating that the excavator is not stalled. The threshold value ( $ϵ$ ) accommodates minor variations while focusing on significant payload changes.

To define the end of the interval, the algorithm has two additional conditions. Equation (7) ensures that the bucket's position remains below 90% of its maximum Y-coordinate (elevation), which ensures the period under investigation is the period where the bucket is actively engaged in the digging process. In contrast, equation (8) imposes a time constraint that redefines the end period of the digging phase ( $t_{s t a l l e n d}$ ), adjusting the actual endpoint ( $t_{e n d}$ ) by a specified window ( $w$ ). This adjustment ensures that the interval we are examining for stalling accurately reflects the digging process rather than a transition to other phases.

These two conditions are necessary because, in the later stages of digging, only bucket rotation occurs to complete the digging phase. Stalling is unlikely in this phase of digging. Furthermore, these restrictions help to prevent false detections that may arise during the repositioning of the bucket. It is worth noting that the window parameter w is intentionally defined as a configurable parameter rather than a fixed constant, allowing future users to determine it for different operating conditions. This flexibility accounts for variations in material properties, operator behavior, and machine configuration, which influence the time required for the bucket to fully engage with the material.

Equation (9) is designed to exclude the transitional phase at the start of the digging process. This condition ensures that the period we are examining for stalling events occurs when the digging is fully underway. Specifically, it guarantees that the bucket is not just rotating but is actively positioning itself in the material to initiate the digging. As a result, any early translational movements are excluded from consideration. By adapting the window size w to the specific digging conditions, the algorithm maintains consistent physical interpretation across different sites and operating regimes, allowing parameter recalibration.

After developing the code in MATLAB, we ran the algorithm using data from four data batches provided by the commercial monitoring system provider. Table 3 summarizes the results for each data batch. We used the algorithm described in Aluko and Awuah-Offei (Aluko and Awuah-Offei, 2024) to detect “normal” cycles in these same batches (Table 3). This provides a comparison of the number of “normal” cycles (this concept is further discussed in Section 5) to the number of cycles with stalling.

Table 3.

Normal and stalling cases result.

Batch	Normal Cases	Stall Cases	Total
1	493	12	505
2	490	5	495
3	637	11	648
4	79	4	83
Total	1699	32	1731

The results indicate that the algorithm identifies stalling cases in the four datasets. The stalling proportions are 2.4%, 1.0%, 1.7%, and 5.1% in Batches 1–4, respectively. The relatively small number of stalling events is consistent with what we would expect, given that stalling during hydraulic excavator operations is rare. Although the occurrence of stalling varies among the batches, the algorithm detects these instances, with Batch 4 showing the highest rate. This variation suggests possible differences in operating conditions for the data batches.

Figure 7 illustrates the difference between a sample “normal” and stalling cycle. The comparison in Figure 7 represents the digging trajectories using kernel density estimation (KDE) (Chen, 2017). KDE allows us to estimate the probability of points along the trajectory with slow parts of the trajectory indicated by high probabilities because of the higher point density.

Figure 7.

Comparison between normal and stalling cycle. (a) Normal cycle; (b) Stalling cycle.

Figure 7 shows that the normal cycle is smooth and continuous, with no interruptions in its trajectory, and the data is evenly distributed. In contrast, the stalling cycle exhibits a region with a high density of data points, which indicates that the bucket spent a significant amount of time in that region because it was unable to move or rotate. The remainder of the stall cycle flows smoothly, showing no significant differences from the normal cycle.

The algorithm described in this section effectively identifies stalling intervals during the digging process by applying the constraints described in equations (3) – (9). The next step is to validate the accuracy of the algorithm to ensure the stalling events it identifies are indeed stalling events.

Validation

Validating heuristic-based detection algorithms is crucial to ensuring their reliability and accuracy in real-world applications (Reeves, 1995). Although these methods are simple, require low computational power, and provide quick responses, they must reliably detect stalls regardless of circumstances and operational conditions. Therefore, a good validation process is necessary to give confidence to users of this algorithm.

This work uses a consistency check for validation. The technique examines the expected behavior of the variables that the heuristic algorithm does not consider seeing if those variables perform as expected. We can validate the algorithm by evaluating the behavior of the variables the algorithm does not use to see if they behave as we should expect if the predicted stalling events are indeed stalling events.

In this work, the authors examined the boom, stick, and bucket cylinder accelerations and the boom moment arm and torque. The authors examined these signals for the digging cycles of the 32 stalling cycles identified by the heuristic algorithm. All the signals are consistent with what one would expect if these digging cycles contained a stalling event. Figure 8 shows the plots of these variables for a typical stalling cycle whereas Figure 9 shows the same plots for a typical ‘normal’ cycle.

Figure 8.

Sample consistency check results: (a) boom acceleration; (b) stick acceleration; (c) bucket acceleration; (d) boom moment arm; and (e) boom torque. Labeled regions: (1) the moments before stalling; (2) the stalling region; and (3) the period after the stalling event.

Figure 9.

Normal cycle results: (a) boom acceleration; (b) stick acceleration; (c) bucket acceleration; (d) boom moment arm; and (e) boom torque.

Figure 8(a) illustrates boom acceleration for a digging cycle with stalling, which differs from that of a normal cycle (Figure 9(a)). As shown in Figure 8(a), just before stalling occurs, there are sharp variations in acceleration, leading to zero acceleration as the system attempts to and fails to overcome increasing resistance. The sharp changes in acceleration are symptomatic of the acceleration/deceleration one observes when an excavator struggles through a muck pile. Since the system cannot ultimately overcome the resistance, stalling begins to occur, which is exemplified by the zero acceleration. After the stalling phase, there is a minor change in acceleration, although it is less pronounced than the initial shift before stalling, as the system begins to recover and accelerate once again. Figure 9(a), on the other hand, shows the boom acceleration for a normal digging cycle. During digging, the boom acceleration is close to zero for most of the cycle. Unlike the stalling case, there are no sharp fluctuations in acceleration, indicating a smooth and continuous excavation process with minimal resistance fluctuations. This behavior is expected for a normal digging cycle.

Figure 8(b) and (c) illustrate stick and bucket accelerations, respectively, for a digging cycle with stalling. Both stick and bucket have similar behavior as the boom. They have a sharp variation in acceleration before stalling occurs. During stalling, there is zero acceleration. Moreover, since the excavator is digging again after this phase, there is a minor fluctuation in its value. The boom, stick, and bucket acceleration behavior are the same because when the operator is digging, it engages all three cylinders similarly to attempt to overcome the resistance. Each cylinder moves to keep the digging process smooth and balanced. On the other hand, Figure 9(b) and (c) also show almost similar behavior to the boom in a normal digging cycle. While the boom does not present significant acceleration fluctuations in this period, the stick and bucket have more variation. This makes sense because, in regular digging, the stick and bucket interact more with the material, causing slightly more variation in acceleration. However, these variations are not abrupt, as in the stalling case.

Figure 8(d) shows the boom moment arm for a digging cycle with stalling differs from a normal digging cycle, presented in Figure 9(d). Figure 8(d) shows that just before stalling occurs, there are minor variations in the moment due to slight changes in the boom's angle and position as it attempts to continue digging. These variations reflect the system's effort to adjust to increasing resistance. Once the hydraulic excavator cannot overcome the resistance, stalling begins. During this phase, the moment arm remains steady as the boom stabilizes. After stalling, there is a temporary decrease in the moment followed by an increase in the moment because the system is digging again. Instead, Figure 9(d) shows the expected behavior of the boom moment arm during digging. The moment arm remains relatively steady for most of the cycle, indicating that the boom maintains a consistent position while penetrating the material. During digging, the moment arm gradually increases until it stabilizes at a higher value. This behavior suggests that the boom is lifting the payload efficiently while digging. The system is smoothly engaging with the material, applying a steady force.

Finally, Figure 8(e) represents boom torque during stalling. As shown in Figure 8(e), just before stalling occurs, the torque briefly decreases as the system adapts to increasing resistance. The decrease in torque is a response to the hydraulic excavator's struggle through the muck pile. Since the system cannot overcome resistance, stalling begins, as represented by the constant torque value during this phase. This behavior is because the system cannot generate enough force to move the boom, causing the torque to stabilize while the excavator remains stationary. After stalling, there is a noticeable increase in the torque as the system resumes its motion. The sharp change is due to the system applying more force to overcome the accumulated resistance, enabling the boom to regain motion and continue the digging process. On the other hand, Figure 9(e) shows the boom torque for a normal digging cycle. This behavior is characterized by a decrease in the torque at the beginning of digging as the system starts digging. As the boom begins to penetrate material, the torque required to maintain motion temporarily drops due to the initial ease of movement. After that period, the torque increases as the excavator digs deeper. The increase in torque corresponds to the system applying more force to continue the excavation, especially as the resistance from the material grows with depth.

The authors present the analysis of the boom, stick, and bucket cylinder accelerations and the boom moment arm and torque, which the heuristic algorithm did not use as validation of the algorithm. The signals of these variables for the 32 digging cycles are very different from the typical “normal” cycles. The behaviors of these signals are also consistent with what one would expect for stalling events. Also, we did not use this as inputs to the heuristic detection algorithm. This independent physical verification increases confidence in the validity of the heuristic labels used in subsequent analysis.

Classification algorithm for stall detection

This section aims to implement a classification algorithm to automate stall detection, offering a scalable and efficient method to support analysis around stall detection. This addresses the second objective of this work.

This work selected a support vector machine (SVM) algorithm for this classification task because of its capability to identify the optimal hyperplane that separates different classes of data by maximizing the margin between them. Using kernel functions, SVM can effectively classify non-linearly separable data into a higher-dimensional space. This makes it a robust classifier that generalizes well, even in high-dimensional settings. It has been successfully applied in various fields, including face recognition, disease diagnostics, and text recognition (Zhang, 2012).

Classification model

This work uses the results of the heuristic stall detection algorithm (Table 3) to create labels of ‘normal’ and ‘stalling’ digging cycles for classification. As discussed earlier, we used Aluko and Awuah-Offei's (Aluko and Awuah-Offei, 2024) definition of “ideal” digging cycles (i.e., cycles with distinct stages, no bench cleanup, and a swing direction unobstructed by the boom) as our normal cycles in this work.

The classification model was developed using MATLAB. The labeling scheme assigned normal cycles a label of 1 and stalling cycles a label of 0. The research team selected 12 features for classification because we deemed these the most relevant for distinguishing between normal and stalling cycles. These features are digging time stamps, bucket X coordinates, bucket height, boom angle, stick angle, bucket angle, swing angle, bucket X velocity, bucket Y velocity, angular velocity, and payload. Some of these features are based on data from Table 1 while others are estimated based on data from Table 1.

Because the cycles have different digging cycle times, using the original data with 30 Hz resolution will result in different vector lengths for each cycle. To ensure consistency across the dataset, we used data interpolation to standardize the time-series features to a uniform length of 300 points per cycle. This approach uses the number of points for the average cycle (the mean digging time across all the cycles is 10 s and the hydraulic excavator's sensors have a frequency of 30 Hz). This method preserved all relevant features, maximizing the model's performance. Given an original time-series x of length L, equation (10) gives $x_{n e w}$ , the new interpolated time series.

x_{n e w} (k) = x (\frac{(k - 1) (L - 1)}{300 - 1} + 1), k = 1, 2, \dots, 300

(10)

The final dataset was constructed by concatenating the interpolated features into a matrix, with labels assigned to differentiate between normal and stalling cycles. In total, the data for each cycle consists of 12 features, each containing 300 points due to the interpolation process. This standardization process standardized the data and ensured consistency across all cycles. This standardization process allowed the model to reduce its variability and improve the robustness of the classification process. Due to the high dimensionality of the feature space, visualizing the hyperplane that separates the two classes directly is not feasible. Nonetheless, the model effectively utilizes these parameters to classify the data accurately.

This work trained an SVM with a linear kernel for classification using this data. To address class imbalance, we used weights based on the ratio of the two classes (Equation (11)), ensuring that the algorithm treated the minority class (stalling) with equal importance to the majority class (normal).

W e i g h t_{c l a s s} = \frac{N_{t o t a l}}{N_{c l a s s}}

(11)

Where:

$N_{t o t a l}$ is the total number of samples in the dataset.

$N_{c l a s s}$ is the number of samples in the specific class.

The MATLAB model utilizes the sequential minimal optimization (SMO) solver to identify the optimal hyperplane for classification, as shown in equation (12).

D e c i s i o n F u n c t i o n : \sum_{i} α_{i} K (x_{i}, x) + b

(12)

Where:

$α_{i}$ are the Lagrange multipliers associated with the support vectors.

$K (x_{i}, x)$ is the linear kernel function, which measures the similarity between the input x and the support vectors $x_{i}$ .

$b$ is the bias term that shifts the decision boundary.

The bias term b was determined during the training process of the SVM using the sequential minimal optimization solver. It was calculated to be 0.9542 by optimizing the decision boundary to separate normal cycles from stalling cycles effectively. This bias term helps adjust the classification threshold, ensuring accurate predictions.

The hyperparameters $C$ and $γ$ , which significantly influence the performance of SVM are the other important configuration parameters. The parameter $C$ is crucial for balancing the trade-off between maximizing the margin and minimizing classification errors. Specifically, a smaller value of $C$ permits a wider margin with some misclassifications, promoting generalization. Conversely, a larger value of $C$ emphasizes minimizing training errors, which can lead to overfitting. On the other hand, the linear kernel parameter $γ$ regulates the impact of individual training instances on the model. A smaller $γ$ value produces a smoother decision boundary, while a larger $γ$ value enables the model to fit the training data more closely.

This implementation chose the hyperparameters $C = 0.1$ and $γ = 0.01$ through a grid search method. The selected values balance the algorithm's complexity and its ability to generalize well. Essentially, they help the model avoid overfitting, ensuring it adapts effectively to the data without becoming overly tailored to specific details that might not apply to new data.

Training and validation

Proper training and validation are essential to developing a reliable classification algorithm for stall detection because it will help ensure the model's ability to generalize effectively (Chen et al., 2009). Training involves using a portion of the dataset to optimize the model's parameters, enabling it to learn the patterns and relationships between input features and target labels. In contrast, validation assesses the model's performance on a separate subset of data, which provides an unbiased evaluation of its accuracy and generalization capabilities. Typically, a part of the dataset is reserved as a test set, which serves as a final evaluation to ensure the algorithm can accurately classify new instances rather than just memorizing patterns from the training data (Meyer et al., 2003).

The data set, as outlined in Table 2, contains 1731 samples, comprising 1699 normal cases and 32 stalling cases. The data is split proportionally, with 80.0% allocated for training, 9.9% for validation, and 10.1% for testing: the training set, which includes 1385 samples (26 stalling cases and 1359 normal cases); the validation set, which consists of 172 samples (three stalling cases and 169 normal cases); and the test set, which contains 175 samples (three stalling cases and 172 normal cases). The authors trained the algorithm using the training set and then evaluated the model using the validation set to monitor its ability to generalize. We then made predictions on the test set to assess the final performance.

The model's performance is evaluated using confusion matrices for the training, validation, and test sets, as well as by combining results from all three (Figure 10). This analysis offers a comprehensive understanding of the model's accuracy and helps to identify any potential overfitting or underfitting.

Figure 10.

Confusion matrix results. (a) Training set; (b) Validation set; (c) Test set.

Figure 10 (a) demonstrates that the model accurately classified all 1361 normal cycles and 24 stalling cycles in the training set, with no misclassifications. This performance indicates a strong learning capability; however, if the accuracy drops significantly on the validation and test sets, it may also suggest overfitting. Figure 10(b) shows that the model successfully identified 165 normal cycles and five stalling cycles in the validation data set. However, it inaccurately classified one normal cycle as a stalling cycle and incorrectly labeled one stalling event as normal. Despite these two misclassifications, we can conclude that the model demonstrates strong performance overall. It effectively detects normal events while also maintaining a reasonable detection rate for stalling events.

Finally, Figure 10(c) demonstrates perfect classification, accurately identifying all 172 normal events and 3 stalling events without any misclassifications. These outcomes corroborate the model's strong performance across all sets and indicate its effective handling of class imbalance. Although class weighting and high-dimensional temporal features mitigate bias toward the majority class, the limited number of stalling samples constrains the statistical certainty of performance metrics such as sensitivity. Consequently, the interpretation of the reported 100% sensitivity should be as evidence of consistent separability of observed stalling events rather than as a definitive estimate of generalization performance.

We used stall labels generated by the heuristic detection algorithm to train the SVM classifier. Accordingly, the classifier does not aim to independently identify stalling events; instead, it generalizes and stabilizes expert-defined stall criteria by learning a continuous, data-driven decision boundary. As the heuristic approach identifies additional stalling cases, the labeled dataset expands, allowing the classifier to better represent stall-related behavior and enabling scalable deployment in practical monitoring systems.

Then, the entire dataset was analyzed simultaneously, in the same manner. The SVM model was configured to use a linear kernel. After training the model, we employed 5-fold cross-validation to evaluate its performance. This approach involves dividing the dataset into five equal parts. The model is trained on four of these parts (or folds) and tested on the remaining fold. This process is repeated five times, with each fold being used as the test set once. The final performance metric is the average of the results from these five runs, providing a reliable estimate of the model's ability to generalize new, unseen data.

This method allowed us to better estimate the model's generalization capabilities without explicitly separating a validation set. Although we did not split the dataset into distinct training, validation, and test sets, the 5-fold cross-validation provided a strong estimate of model accuracy. Figure 11 displays the results. The confusion matrix reflects the overall performance of the model across the entire dataset, rather than being specific to the training, validation, or test sets.

Figure 11.

Total confusion matrix.

To evaluate the model more effectively, we calculate sensitivity and specificity. Sensitivity (also known as recall) measures the proportion of true positive stalling cases among all actual stalling cases. Specificity measures the proportion of true negative normal cases among all normal cases. Equations (13) and (14) show the sensitivity and specificity, respectively.

\frac{T P}{T P + F N} = 100 %

(13)

\frac{T N}{T N + F P} = 99.0 %

(14)

The model achieves a sensitivity of 100%, meaning it correctly identifies all instances of stalling, while maintaining a specificity of 99.9%, which indicates it classifies normal cases accurately with very few false positives. These results demonstrate that the algorithm is well-balanced, effectively classifying both the minority class (stalling) and the majority class (normal). Additionally, this shows that the algorithm is capable of handling imbalanced datasets effectively. The high sensitivity and specificity suggest that the model is highly effective and suitable for deployment or further testing in real-world applications.

Conclusion

This research sought to develop a reliable algorithm for stall detection during excavator operations and an SVM classification algorithm to distinguish between normal digging cycles and those with stalling two significant contributions. The work successfully develops a novel heuristic-based algorithm to detect stall events in hydraulic excavators during the digging phase of a loading cycle. The algorithm utilizes seven conditions designed to detect conditions where the bucket is neither moving nor rotating. The work analyzed 1731 cases and identified 32 cases as stalling events. The results confirm that stalling is a rare event. The work validates the heuristic algorithm by examining five variables that the heuristic algorithm does not use. Furthermore, the work successfully developed a support vector machine (SVM) classification model to classify cycles into those with and without stalling. The classification model achieved nearly perfect sensitivity (100%) and specificity (99%), confirming its high reliability in distinguishing stalling events from regular loading cycles. These contributions help researchers and operators identify and address stalling events, optimizing excavator performance and minimizing downtime.

Although the algorithms presented in this work constitute significant contributions to literature, the reader should note a few limitations. First, the analysis relied on a specific case study dataset, which means further work will be necessary to adapt it to other data sets. Additionally, the model currently utilizes only positions, payloads, and angles of excavator components as features, which may limit its ability to capture the full complexity of stalling events. Adding other sensor measurements, such as the cylinder pressures of the bucket or stick, could offer deeper insights into the model's accuracy and improve its robustness across various scenarios. Nevertheless, the proposed approach provides a practical and scalable solution for monitoring systems that lack direct hydraulic pressure data, which remains common in industry. Also, a further limitation of this study is the small number of stalling events observed, which reflects real operating conditions but limits statistical robustness; expanding the dataset remains an important direction for future work. Lastly, the heuristic algorithm uses only machine kinematics to determine stalling events. While users can calibrate the window parameter, w, for different material properties, the model does not explicitly account for material properties. Future work should calibrate w for different muckpile properties or attempt to include material properties in the model.

Future research should focus on expanding the dataset to encompass other monitoring data from hydraulic excavators to overcome these limitations and enhance the system's robustness. Additionally, utilizing advanced machine learning techniques could further improve the efficiency of the detection algorithm. With these advancements, the system can enhance the reliability and effectiveness of excavation operations.

The developed algorithms are a significant step towards advancing our understanding of stalling events within the mining and construction sectors because they provide a basis for isolating stalling events for further study.

Footnotes

Acknowledgments

The authors would like to thank Dr Atta Ur Rehman, Dr Amir Hossien, Mr Abraham Hernandez and the rest of the ESCO team for their invaluable support.

ORCID iDs

Mateo Fernando Montenegro Defaz

Kwame Awuah-Offei

Ethical approval

This research did not involve human participants, human data, or human tissue. Therefore, ethical approval was not required for this study.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by ESCO funded the project.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

Data will be made available on request so long as the data provider agrees the requested data is not proprietary and will not cause competitive harm.

Statement

During the preparation of this work the author(s) used Grammarly, Inc. to enhance clarity and ensure consistency in writing. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

References

Aluko

Awuah-Offei

(2024) Using telemetry to assess operator effects on hydraulic shovel energy efficiency. Part I: Algorithms for extracting explanatory variables from sensor data. Mining, Metallurgy & Exploration 41(6): 2943–2958.

Americas

HCM

(2025) EX5600-7B, T4F. Available at: https://www.hitachicm.com/us/en/products/excavators/mining/product.ex5600-7b-t4f/.

Ball

(2011) Heuristics based on mathematical programming. Surveys in Operations Research and Management Science 16(1): 21–38.

Bell

Cripps

Culshaw

(1995) The significance of engineering geology to construction. Geological Society, London, Engineering Geology Special Publications 10(1): 3–29.

Bettens

Siegrist

McAree

(2022) How do operators and environment conditions influence the productivity of a large mining excavator? International Journal of Mining and Mineral Engineering 13(1): Inderscience: 18–36.

Bonchis

Hillier

Ryde

, et al. (2011) Experiments in autonomous earth moving. IFAC Proceedings Volumes 44(1): 11588–11593.

Calle-Escobar

Mejía-Gutiérrez

Nadeau

J-P

, et al. (2016) Heuristics-based design process. International Journal on Interactive Design and Manufacturing (IJIDeM) 10(4): 369–386.

Chacko

Cang

, et al. (2014) State of the art in excavators. In: Proceedings of the 2014 International Conference on Advanced Mechatronic Systems, 2014, pp. 481–488.

Chen

Y-C

(2017) A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology 1(1): Taylor & Francis: 161–187.

10.

Cheng

Teizer

(2013) Real-time resource location data collection and visualization technology for construction safety and activity monitoring applications. Automation in Construction 34: 3–15.

11.

Chen

Tino

Yao

(2009) Probabilistic classification vector machines. IEEE Transactions on Neural Networks 20(6): 901–914.

12.

Chen

Zou

Pang

(2018) Digging performance characterization for hydraulic excavator considering uncertainty during digging operation. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science 232(5): 857–871.

13.

Chetty

(1983) Telemetry in the mining industry. IETE Journal of Research 29(8): Taylor & Francis: 389–391.

14.

Coetzee

Basson

Vermeer

(2007) Discrete and continuum modelling of excavator bucket filling. Journal of Terramechanics 44(2): 177–186.

15.

Coetzee

Els

DNJ

(2009) The numerical modelling of excavator bucket filling using DEM. Journal of Terramechanics 46(5): 217–227.

16.

Dong

Mingyue

Guoying

(2017) Application of Internet of Things Technology on Predictive Maintenance System of Coal Equipment. In: Procedia Engineering, 2017, pp. 885–889. Elsevier Ltd.

17.

Dunbabin

Corke

(2006) Autonomous excavation using a rope shovel. Journal of Field Robotics 23(6–7): 379–394.

18.

Dyorina

Velikanov

Dyorina

(2020) Identification of failure patterns of excavator equipment failures considering the control factor. IOP Conference Series: Materials Science and Engineering 971(4): IOP Publishing: 42062.

19.

Fairhurst

(2017) Some challenges of deep mining. Engineering 3(4): 527–537.

20.

Frimpong

(2007) Stress loading of the cable shovel boom under in-situ digging conditions. Engineering Failure Analysis 14(4): 702–715.

21.

Fujita

Yasuda

Imaie

(2011) Ultra large hydraulic excavators and dump trucks for large open-pit mines. Hitachi Review 60(5): 267.

22.

Haertel

Donat

Staegemann

, et al. (2024) The application of data science at original equipment manufacturers: A literature review. IEEE Access 12: 114584–114600.

23.

Hirano

Yoshida

Tsujiuchi

, et al. (2021) Applicability of automated digging algorithm on sloping ground for hydraulic excavators. In: B and JJ and BT Oberst Sebastian and Halkon (ed.). Vibration Engineering for a Sustainable Future. Cham: Springer International Publishing, 291–297.

24.

Jeong

Phillips

(2001) Operational efficiency and effectiveness measurement. International Journal of Operations & Production Management 21(11): MCB UP Ltd: 1404–1416.

25.

Johnson

(2018) Effect of wave collision on fragmentation, throw, and energy efficiency of mining and comminution. In: Awuah-Offei

(eds) Energy Efficiency in the Minerals Industry: Best Practices and Research Directions. Cham: Springer International Publishing, 55–70. Available at: https://doi.org/10.1007/978-3-319-54199-0_4.

26.

Jud

Leemann

Kerscher

, et al. (2019) Autonomous free-form trenching using a walking excavator. IEEE Robotics and Automation Letters 4(4): 3208–3215.

27.

Karpuz

Ceylanoğlu

Paşamehmetoğlu

(1992) An investigation on the influence of depth of cut and blasting on shovel digging performance. International Journal of Surface Mining, Reclamation and Environment 6(4): Taylor & Francis: 161–167.

28.

Khorzoughi

Hall

(2016) Diggability assessment in open pit mines: A review. International Journal of Mining and Mineral Engineering 7(3): Inderscience: 181–209.

29.

Kirianaki

Yurish

Shpak

, et al. (2002) Data Acquisition and Signal Processing for Smart Sensors. England: Wiley Chichester.

30.

Kirmanli

Ercelebi

(2009) An expert system for hydraulic excavator and truck selection in surface mining. Journal of the Southern African Institute of Mining and Metallurgy 109(12): 727–738.

31.

Kokash

(2005) An introduction to heuristic algorithms. Department of Informatics and Telecommunications: Citeseer: 1–8.

32.

Kumar

Srivastava

(2012) An expert system for predictive maintenance of mining excavators and its various forms in open cast mining. In: 2012 1st International Conference on Recent Advances in Information Technology (RAIT), 2012, pp. 658–661.

33.

Lashgari

Kecojevic

(2013) Environmental impact of digging and loading equipment in surface coal mining. In: Society for Mining, Metallurgy and Exploration Annual Meeting, Denver, CO, USA, 2013, p. 5.

34.

Lever

PJA

(2001) An automated digging control for a wheel loader. Robotica 19(5): 2001/09/04. Cambridge University Press: 497–511.

35.

Wang

Chen

, et al. (2024) Identification of material excavation difficulty and uncertainty analysis based on Bayesian deep learning. Journal of Industrial Information Integration 42: 100728.

36.

Libkin

(2014) Incomplete data: What went wrong, and how to fix it. In: Proceedings of the ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems, 2014, pp. 51–62. Association for Computing Machinery.

37.

Liu

Zhou

Gan

, et al. (2024) Performance analysis of hydraulic excavators in complex excavation environments and different bucket trajectories. Powder Technology 437: 119539.

38.

Luengo

Singh

Cannon

(1998) Modeling and identification of soil-tool interaction in automated excavation. In: Proceedings. 1998 IEEE/RSJ International Conference on Intelligent Robots and Systems. Innovations in Theory, Practice and Applications (Cat. No.98CH36190), 1998, pp. 1900–1906 vol.3.

39.

Mahamedi

Rogage

Doukari

, et al. (2022) Automating excavator productivity measurement using deep learning. Proceedings of the Institution of Civil Engineers: Smart Infrastructure and Construction 174(4): ICE Publishing: 121–133.

40.

Malcolm

Lizotte

Paventi

, et al. (2018) Structural control over fragmentation. Measurement of Blast Fragmentation: 181–191.

41.

Manyele

(2017) Investigation of excavator performance factors in an open-pit mine using loading cycle time. Engineering 9(7): Scientific Research Publishing: 599–624.

42.

Meyer

Leisch

Hornik

(2003) The support vector machine under test. Neurocomputing 55(1): 169–186.

43.

Mohammadi

Rai

Oraee

(2015) A critical investigation of digging time segment of draglines in a large surface mine. Geotechnical and Geological Engineering 33(4): Kluwer Academic Publishers: 763–771.

44.

Nikkhah

Vakylabad

Hassanzadeh

, et al. (2022) An evaluation on the impact of ore fragmented by blasting on mining performance. Minerals 12(2): 258.

45.

Palomba

Richiedei

Trevisani

, et al. (2019) Estimation of the digging and payload forces in excavators by means of state observers. Mechanical Systems and Signal Processing 134: 106356.

46.

Redmann

(2018) Environmental Effects of Mining. New York, NY: Routledge.

47.

Reeves

(1995) Modern Heuristics Techniques for Combinatorial Optimization Problems. New York, NY: McGraw-Hill.

48.

Saeed Abolhasani

Christopher Frey

KKWRPL

Pang

S-H

(2008) Real-World in-use activity, fuel use, and emissions for nonroad construction vehicles: A case study for excavators. Journal of the Air & Waste Management Association 58(8): Taylor & Francis: 1033–1046.

49.

Sakaida

Chugo

Kawabata

, et al. (2006) The Analysis of Skillful Hydraulic Excavator Operation. In: 2006. Available at: https://api.semanticscholar.org/CorpusID:115023431.

50.

Sano

Ichiryu

(2002) Development of excavator operation robot. In: 2002. Available at: https://api.semanticscholar.org/CorpusID:116597882.

51.

Santi

(2005) Engineering geology and construction (fred G. Bell). Environmental & Engineering Geoscience 11(2): 177.

52.

Savage

(1938) The development of single-bucket excavators. Proceedings of the Institution of Mechanical Engineers 138(1): 177–226.

53.

Shehu

Hashim

MHM

(2020) Evaluation of blast fragmentation and its benefits: A review of the paradigm and ambiguity. International Journal of Mining and Mineral Engineering 11(4): 338–358.

54.

Turbosquid (2020) Hitachi EX5600. Available at: https://www.turbosquid.com/3d-models/3d-hitachi-ex5600-1835020 (accessed 15 January 2025).

55.

Victor

Pavel

Tatyana

(2020) Determination of the working area of a hydraulic excavator. E3S Web Conf 177: 3017.

56.

Vylomov

Kovalev

(1982) Using hydraulic excavators for selective mining of magnesite. Refractories 23(7): 414–417.

57.

Wang

, et al. (2021) Digging force and power consumption during robotic excavation of cable shovel: Experimental study and DEM simulation. International Journal of Mining, Reclamation and Environment 35(1): Taylor and Francis Ltd.: 12–33.

58.

Xiaohe

Kun

Kuangdi

(2024) Rock fragmentation. In: The ECPH Encyclopedia of Mining and Metallurgy. Singapore: Springer Nature, 1818–1819.

59.

Yoshida

Koizumi

Tsujiuchi

, et al. (2013) Dynamic analysis of an excavator during digging operation. SAE International Journal of Commercial Vehicles 6(2): 419–428.

60.

Yoshida

Tsujiuchi

Ito

, et al. (2016) Simulation-based evaluation of influence on digging performance by control parameters. JOURNAL OF JAPAN SOCIETY FOR DESIGN ENGINEERING 51(11): 821–836.

61.

Choi

(2007) Time series analysis based overload detection algorithm for excavator. In: SICE Annual Conference, 2007, 2007, pp. 1203–1208.

62.

Zhang

(2012) Support Vector Machine Classification Algorithm and Its Application. In: Information Computing and Applications (ed. L and YA Liu Chunfeng and Wang), Berlin, Heidelberg, 2012, pp. 179–186. Springer Berlin Heidelberg.