Bearing Electrical Erosion Signal Detection Technique Based on Current Signal Analysis for Electric Motors

The efficiency and performance improvement of electric motors has become a critical requirement in modern industries. In particular, pulse-width modulation (PWM) inverters equipped with high-speed switching devices have been widely employed in electric motors for variable speed control and energy efficiency enhancement [1–3]. However, such inverter control systems generate complex shaft voltages and bearing currents, leading to electrical erosion issues [2]. This electrical erosion deforms the metallic surfaces of the bearings, causing early bearing failures and increasing vibration and noise in motors [4,5]. It also leads to a decrease in the durability and reliability of electric motors, as well as a reduction in overall system performance and lifespan [5–7].

Numerous studies have been conducted to prevent such bearing electrical erosion. Researchers such as Tawfiq et al. [1] have discussed various failure mechanisms in electric motors, the sources and definitions of different types of bearing currents, and shaft voltages. He et al. [2] and Xiao et al. [5] also summarized the processes of shaft voltage and bearing current generation that lead to bearing electrical erosion, and reviewed methods for mitigating such erosion.

Research on methods for detecting electrical erosion signals has also been conducted. Traditional bearing fault diagnosis and condition monitoring methods primarily relied on vibration analysis, which detects faults caused by various factors such as mechanical wear, overload, heat, and lubrication issues [8,9]. However, among these factors, to reliably assess the impact of electrical erosion on bearing durability and lifespan, several studies have specifically focused on electrical measurement methods and electrical erosion signal detection techniques [8]. Plazenet et al. [10] have analyzed various measurement methods for shaft voltages and bearing currents, discussing the pros and cons of fault diagnosis methods through shaft voltage signals, and suggested future improvements for bearing condition monitoring technologies. Ahola et al. [11,12] and Muetze et al. [13] used noninvasive radio frequency-based methods to detect bearing discharge pulses, summarizing the number of pulses that exceeded a threshold within a specific frame. However, they did not present the accuracy of the results or the performance evaluation of the proposed methods.

Researchers have extensively studied bearing electrical erosion and proposed various countermeasures to detect and mitigate it. However, studies on classifying electrical erosion signals by type and accurately detecting the occurrence frequency of each type in real-world environments have not been conducted extensively.

This study introduces a detection technique for electrical erosion signals in bearings of electric motors. The proposed technique utilizes common mode voltage (CMV) and bearing current data and identifies relationships between measured physical quantities through data analysis. Low-pass filters and band-pass filters are applied for preprocessing the measured signals, and differentiation is used to facilitate the extraction of features at the point of occurrence of electrical erosion signals. The interquartile range (IQR) method is then used for outlier detection, classifying electrical erosion signals caused by high-frequency (HF) circulating currents and electric discharge machining (EDM) currents. The frequency and occurrence timing of each type of electrical erosion signal are determined through data clustering at fixed time intervals. Performance evaluation confirms that the proposed technique has higher electrical erosion signal classification performance compared to traditional methods. This approach allows for the quantification of occurrence frequencies for each type of electrical erosion signal under different load conditions. These findings can be utilized for in-depth analysis of bearing electrical erosion across a wider range of motor operating scenarios and for conducting durability assessments.

Electrical erosion in bearings refers to damage caused by arcs generated through the conduction of leakage currents in metallic components such as bearing races and rotors [2,14,15]. The primary sources of electrical erosion in electric motor bearings are shaft voltages and the resulting bearing currents [1]. The generation of shaft voltages can be divided into three main sources: magnetic asymmetry, electrostatic discharge, and CMV [1,2]. Magnetic asymmetry is caused by design, manufacturing, or installation issues such as eccentricity and asymmetry in the rotor windings [16]. Mechanical or electrical asymmetry in the motor windings creates an unbalanced magnetic flux around the shaft, thereby forming voltages. This results in low-frequency bearing currents circulating through the motor [1].

The second source, electrostatic discharge, arises from composite and polymer materials widely used for lightweight applications [2]. Frictional charging between these material surfaces leads to significant charge separation and the accumulation of static charges [17], resulting in a voltage difference between the shaft and the ground. If the bearing voltage exceeds the breakdown strength of the insulating lubricant film, the accumulated charge discharges, producing EDM currents [18].

Bearing currents are categorized into “circulating” and “noncirculating” currents based on their source of generation [7]. Noncirculating currents include $dv/dt$ currents generated by rapid voltage differences and EDM currents [23]. These currents flow unidirectionally from the rotor to the stator through the bearings, thus termed “noncirculating” [24]. The $dv/dt$ currents flow from the stator windings through the rotor and bearings, ultimately reaching the frame. However, the magnitude of $dv/dt$ currents is relatively small, is mostly harmless, and constitutes a minor fraction of the total bearing currents [2]. EDM bearing currents occur when the bearing voltage exceeds the dielectric strength of the insulating lubricant, releasing capacitor energy through current and arc discharges [7,25,26]. These currents make up the majority of noncirculating bearing currents, and unlike dv/dt currents, they are not directly related to the fluctuation times of the CMV pulses [27,28]. The generation of circulating currents involves complex mechanisms including magnetic induction, inductive coupling, and capacitive coupling. The first type of low-frequency circulating current arises from magnetic asymmetry and is characterized by a frequency corresponding to the rotational speed of the shaft. When the induced shaft voltage is strong enough to damage the bearing's lubricant film, bearing currents circulate through a conductive loop comprising “stator—drive-end bearing—rotor shaft—nondrive-end bearing—stator” [29].

The second type of HF circulating current is derived from the CMV of the high-speed switching three-phase inverters. The parasitic capacitance between the stator windings and the frame is excited by the abrupt changes in the high-frequency switching CMV, producing high-frequency common mode currents flowing through the stator core [30,31]. The path of the current flow is the same as that of the first type of circulating current, but the frequency of this current can reach megahertz [24,31].

Figure 1 illustrates the flow paths of each type of bearing current in an electric motor, as indicated by the arrows [5]. EDM currents and the second type of HF circulating currents constitute the largest proportion of the total bearing currents and have the most significant impact on bearing electrical erosion [2]. In particular, EDM currents create short arcs that form craters on the surfaces of bearing races and rollers, leading to pitting on the outer or inner race, thus causing increased vibration, noise, and systemic issues [2,32].

Therefore, this study proposes methods to analyze, classify, and detect the characteristics of EDM and HF circulating currents to determine the frequency and timing of each type of electrical erosion signal.

For the detection of bearing electrical erosion signals, data acquisition and analysis of signal characteristics for methodological planning were conducted. First, CMV, bearing current, and shaft voltage signal data were gathered using a bearing current measurement system. Figure 2 shows a photograph of the established bearing current measurement system. This system includes a high voltage differential probe for CMV measurement, a current probe (TCP0030) for measuring bearing currents, slip rings and carbon brushes for shaft voltage measurement, and an oscilloscope (Tektronix DP07104C). Data were captured at a sampling rate of 100 MHz, recording signals over duration of 0.5 ms per dataset.

The operational conditions of the electric motor were configured in accordance with the driving speed and required torque specifications provided by the worldwide harmonized light vehicles test cycles (WLTC) class 3b, to reflect the real-world driving scenarios of contemporary electric vehicles [33]. The motor's rotational speed was set at 3000 rpm, which corresponds with the typical rated rotational speed range for electric vehicle motors and allows for the replication of WLTC class 3b driving speeds [34]. In actual driving conditions, motors endure various loads due to factors such as rolling resistance, road inclines, and acceleration forces, which in turn significantly influence the magnitude, amplitude, and waveform of shaft voltage and bearing currents. To develop and evaluate a bearing electrical erosion signal detection methodology that operates effectively under various load conditions, motor load conditions were designated as independent variables in this study. The load conditions were established based on the torque requirements of WLTC class 3b, with specific loads set at 0, 20, 30, and 40 N·m, while maintaining a constant rotational speed throughout the experimental phase. Under these operational conditions, data acquisition was conducted. A total of 20 datasets were captured under no-load conditions, and two additional datasets were gathered for each of the load conditions at 20, 30, and 40 N·m, culminating in a comprehensive collection of 26 datasets.

Figure 3 illustrates a dataset measured under no-load conditions, displaying the CMV, bearing current, and shaft voltage signals measured through slip rings. The horizontal axis represents time, while the vertical axis sequentially displays CMV, bearing current, and shaft voltage. As shown in the figure, the CMV signal appeared as a stepped waveform, and pulses of bearing current and shaft voltage were observed at points where CMV changed sharply. This corresponds to the HF circulating currents derived from the high dv/dt of CMV, as discussed in Sec. 2, exhibiting a periodic occurrence pattern. Typically, in regions where CMV was relatively stable, nonperiodic fluctuations in bearing current and shaft voltage were observed, which are attributed to EDM currents unrelated to CMV pulse timing [23,24].

Figure 4 presents the CMV, bearing current, and shaft voltage signals measured under load conditions, displaying a single dataset for each of the 20, 30, and 40 N·m load conditions in Figs. 4(a), 4(b), and 4(c), respectively. Even under these load conditions, fluctuations in bearing currents induced by HF circulating currents and EDM currents were clearly observed, and shaft voltage variations, though not consistent, were detectable.

Based on this data analysis, the present study has defined the criteria for identifying bearing electrical erosion signals as follows. As discussed in Sec. 2, bearing electrical erosion is caused by bearing currents that include EDM and HF circulating currents, arising from various causes leading to abrupt changes in shaft voltage. Therefore, this study defines the points where rapid changes in shaft voltage coincide with the pulse signals of bearing currents as the actual occurrence points of bearing electrical erosion signals. The criteria for identification are set to meet the following two conditions: first, a sharp change in shaft voltage of more than 2.0 V must occur within a time interval of 1 µs; second, at points satisfying the first criterion, the bearing current values must exceed a sufficiently large deviation of 3.5 times the standard deviation (σ) of the data within a $±2.5μs$ radius around these points. If these two conditions are met, they are considered the actual occurrence points of bearing electrical erosion signals, as illustrated in Fig. 5(a) capturing the signal occurrence regions within the dataset. In contrast, Fig. 5(b) displays instances where the first condition is met but not the second, hence not recognized as signal occurrence points. The shaded regions in the left side of the first column graph in Fig. 5(a) and the background of the first column graph in Fig. 5(b) indicate the points where rapid changes occur in the CMV signal. In contrast, the shaded regions in the right side of the first column graph and the second column graph in Fig. 5(a), as well as the second column graph in Fig. 5(b), represent intervals where the CMV signal remains constant.

However, measuring shaft voltage in actual operating environments of electric motors poses challenges, especially in finding shaft voltage sensors that can operate under varying rotational speeds and load conditions. Furthermore, while bearing currents typically exhibit consistent pulse signals at the signal occurrence points, the shaft voltage data tend to show inconsistent waveforms and variations under load conditions, making feature extraction problematic.

Therefore, this study proposes a method that utilizes bearing current data, which clearly shows the signal variations caused by HF circulating currents and EDM currents. Moreover, by utilizing the relationship between the bearing current and CMV signal data, this study outlines a method for classifying the signals induced by EDM and HF circulating currents, thereby deriving their occurrence frequency and timing.

Figure 6 presents a flowchart of the proposed method for detecting bearing electrical erosion signals. The technique proceeds sequentially through preprocessing, outlier detection, classification, and derivation of the results. Functionally, it comprises signal processing, feature extraction, and classification. Detailed descriptions of each step and function are provided in the following sections.

The preprocessing stage involves refining and numerically transforming the voltage and current data collected via sensors to clearly reveal the overall trends and characteristics. In particular, voltage and current data collected at high sampling rates contain substantial noise. Through signal processing, unnecessary components are removed, and the data are transformed into a format convenient for application in the proposed technique.

The CMV signal exhibited a periodic pulse waveform with staircase patterns that recurred at regular intervals. The points of rapid voltage increase or decrease within these signals can trigger the generation of HF circulating currents. To facilitate feature extraction at these abrupt transition points of the CMV signal, a low-pass filter was initially applied. With a cutoff frequency set at 3.5 MHz, unnecessary noise and short-term fluctuations were removed to clarify the overall staircase waveform trend. Subsequently, differentiation of the filtered signal over time was performed to calculate the instantaneous change rate. This process maximizes the values at sharp transition points of the CMV signal, while areas of lesser change retain smaller values, thus simplifying feature extraction.

For bearing current signals, which need to be detected using this technique, both EDM and circulating currents consist of high-frequency current components. To accurately identify the signals caused by these EDM and circulating currents, a band-pass filter was employed. The lower and upper cutoff frequency limits were set at 1 MHz and 10 MHz, respectively, allowing primarily the frequencies within this bandwidth to pass while attenuating or eliminating other frequencies. The rationale for setting the lower cutoff frequency at 1 MHz is that signals below this threshold predominantly exhibited repetitive patterns, which were determined not to contain pulse signals induced by axial voltage changes. These signals also introduced significant amplitude variability, thereby hindering the identification of pulse signals. The upper cutoff frequency was set based on the criteria defined in Sec. 3 for signal detection, ensuring that the absolute magnitude of the bearing pulse signals, which occur simultaneously with changes in axial voltage, is not excessively reduced while effectively removing high-frequency noise. Table 1 presents the filters used in this study along with their respective cutoff frequencies.

In Eq. (2), $xscaled$ denotes the normalized data and x represents the input data. The terms $xmin$ and $xmax$ are the minimum and maximum values of the data, respectively.

In the preprocessing phase, the refined CMV and bearing current data exhibit pulse signals at points where rapid changes occur in the CMV signal and electrical erosion signals caused by HF circulating currents and EDM currents. Therefore, outlier detection methods were employed for feature extraction at these pulse signal points.

Traditional outlier detection methods involve setting thresholds, identifying values that exceed upper limits or fall below lower limits as outliers [10,28]. However, the bearing current data in this study, which are significantly influenced by the operating conditions and load states of electric motors, vary greatly in amplitude and waveform, making it challenging to accurately identify the desired signals as outliers using this method. Consequently, this study utilized the IQR method, which is unaffected by the distribution shape of the data and capable of detecting outliers [36].

In this study, the multiplier for the IQR was selected considering the distribution of each dataset and the elements that needed to be identified as outliers. The processed CMV and bearing current data are typically concentrated around the median, exhibiting very small IQR values. Additionally, considering that the filtered bearing current signals are high frequency, it is crucial to identify the few pulse signals with significantly larger amplitudes as outliers, requiring the detection of more extreme outliers than usual. Considering these factors, the multipliers for the IQR of the processed CMV and bearing current data were set to 3 and 4.5, respectively, to compute outliers. Figure 7 illustrates the application of the IQR method on processed CMV and bearing current data to detect outliers, simplifying the process in a diagrammatic representation. This approach enabled the extraction of features at sharp transition points in the CMV signal and pulse signal occurrence points in the bearing current, determining the timing of these outliers as potential electrical erosion signal occurrences.

Outliers in the processed CMV data indicate points of rapid change in the CMV signal, which correspond to potential sites for the generation of HF circulating currents. Similarly, outliers in the processed bearing current data represent electrical erosion signals induced by HF circulating currents and EDM currents. Outlier comparison, time-series clustering, and counting were performed to classify the signals caused by HF circulating currents and EDM currents separately and to determine the occurrence and timing of each type of signal. Figure 8 illustrates a schematic of the feature extraction and signal classification process.

Outlier data in the bearing current signal that occur simultaneously with outliers in the CMV signal are classified as electrical erosion signals resulting from the HF circulating currents triggered by sharp changes in the CMV. Conversely, outlier data in the bearing current signal that do not coincide with the timing of the CMV signal outliers are classified as electrical erosion signals caused by the EDM currents. First, data points where bearing current outliers coincided with CMV outliers within a tolerance of 0.25 µs were identified, and other noncoincident bearing current outlier data were classified separately.

The classified outlier data consist of numerous data points within short time intervals. Thus, to derive the occurrence frequency and timing of the electrical erosion signals, it is necessary to cluster the time-series data according to a set criterion. In this study, data points within a 0.25 µs interval were grouped into a single cluster, and the frequency of occurrence was determined by counting these clusters. Additionally, the average time of data within each cluster was calculated to determine the timing of each electrical erosion signal event.

Utilizing the data obtained through the bearing current measurement system, the bearing electrical erosion signal detection technique was applied, facilitating the extraction of the frequency and timing of electrical erosion signals caused by HF circulating currents and EDM currents across each dataset. Figure 9 presents a graph illustrating processed CMV and bearing current signals during intervals of EDM current occurrences among a total of 26 datasets. The x-axis represents time, while the y-axis shows the derivative values of the processed CMV and bearing currents for different loading conditions: (a) no-load, (b) 20 N·m, (c) 30 N·m, and (d) 40 N·m. As indicated, pulse signals in the bearing currents that coincided with sharp transitions in the processed CMV data were classified as signals induced by HF circulating currents. Moreover, pulse signals in the processed bearing currents that did not coincide with the CMV pulse timings were identified as electrical erosion signals induced by EDM currents. In Fig. 9, signals attributed to HF circulating currents are highlighted with shaded backgrounds, whereas those associated with EDM currents are delineated using rectangular outlines.

The counts of electrical erosion signals extracted using the proposed technique were organized and divided between no-load and load conditions, as detailed in Tables 2 and 3. These results were compared with the actual frequencies of electrical erosion signals. The actual electrical erosion signals were defined based on the criteria outlined in Sec. 3, which stipulate that both a rapid change in shaft voltage and a pulse signal in bearing current must occur simultaneously. Points that satisfied these two criteria were manually counted to compile the frequency of occurrence.

Under no-load conditions, a total of 442 HF circulating current erosion signals and two EDM current signals were detected over 10 ms, aligning accurately with the actual frequencies. Under a 20 N·m load condition, within 1 ms, a total of 15 HF circulating current erosion signals and one EDM current signal were detected, resulting in more frequent counting than actual occurrences with an error rate of 7.14%. Under the 30 N·m load condition, within 1 ms, 13 HF circulating current erosion signals were detected, and no EDM current signals were detected, resulting in fewer counts than the actual frequency with an error rate of 6.67%. Under a 40 N·m load condition, within 1 ms, 16 HF circulating current erosion signals and 3 EDM current signals were detected, matching the actual occurrence rates precisely.

Upon analysis, the detection counts of HF circulating current erosion signals exhibited a low error rate of within 8.00% relative to the actual values, and the detection of EDM current erosion signals consistently matched the actual counts, demonstrating high overall accuracy. The primary cause of discrepancies identified was the increase in high-frequency noise in the bearing currents under load conditions, which were also classified as outliers. Furthermore, the clustering of multiple HF circulating current signals within very short intervals into a single count was identified as another source of error. These results facilitated the quantification and comparison of electrical erosion signal occurrences induced by HF circulating currents and EDM currents under various load conditions against actual occurrences. However, this comparison was limited to the total counts, without numerically verifying the precise classification of each signal as either an electrical erosion or a normal signal. Therefore, a performance evaluation of the proposed bearing electrical erosion signal detection technique was conducted.

To evaluate the performance of the proposed bearing electrical erosion signal detection technique and compare it with traditional methods, we classified the normal and electrical erosion signals and compared the results with the actual values. The performance metrics used included precision, recall, the harmonic mean of precision and recall known as the F1-score, and the weighted harmonic mean of precision and recall known as the F2-score. To emphasize the assessment of detection performance for electrical erosion signals, the final evaluation was based on the F2-score, which prioritizes recall [38,39].

Additionally, the performance metrics of the conventional bearing current pulse detection method, based on threshold setting, were calculated and compared with those of the proposed method to evaluate the difference in performance between the two approaches. Thresholds were determined by applying multiples of the median's absolute value. Specifically, in this study, the upper and lower thresholds were set by, respectively, adding and subtracting 3.5 times the absolute value of the median from the median itself. The performance comparison, focused on the recall to assess the accuracy of detecting actual electrical erosion signals, was quantified using the F2-score.

Table 4 presents the performance metrics under different load conditions for each method, with the proposed technique labeled BEDT (bearing electrical erosion signal detection technique) and the threshold method labeled DWT (detection with threshold). Under no-load conditions, both methods demonstrated high performance, with F2-score exceeding 96%. However, under load conditions, while the proposed method maintained an F2-score above 96.39%, the threshold method showed significantly lower performance, achieving only a 58.56% F2-score. This discrepancy is attributed to the amplitude and waveform variations of the bearing current under different load conditions, which the fixed threshold method fails to accommodate effectively. The use of a fixed threshold leads to performance variation under different load conditions, whereas adjusting the threshold dynamically for each load condition could potentially improve accuracy, but poses challenges in accurately setting these thresholds for all conditions.

Furthermore, the threshold method encountered issues in detecting multiple electrical erosion signals or misclassifying normal signals as electrical erosion owing to minor threshold deviations in data with large amplitude variations. Figure 10 illustrates instances of misclassification in which normal signals exceeded the set thresholds due to significant amplitude changes. In contrast, the proposed method, which utilizes the IQR method to detect outliers based on data frequency, demonstrated a high detection rate of electrical erosion signals across various load conditions.

Therefore, the method proposed in this study can be directly applied to experiments generating electrical erosion using a dynamometer under various real driving conditions, without any adjustments. In contrast, the threshold setting method requires the reanalysis of data and resetting of accurate thresholds whenever the range of bearing current values changes under different load conditions. This advantage highlights the applicability and efficiency of the proposed method in dynamic testing environments.

Additionally, the inability of the threshold method to clearly differentiate between circulating current and EDM current signals made detailed classification challenging, whereas the proposed technique efficiently classified EDM and circulating currents, as detailed in Sec. 5. Consequently, the proposed technique can be utilized in evaluating the durability of electric motors, offering the ability to assess the impact of bearing electrical erosion caused by EDM currents and circulating currents on the durability of electric motors individually.

This article proposes a new methodology for detecting electrical erosion signals in bearings of electric motors and evaluates its performance. The proposed methodology utilizes CMV and bearing current data to effectively classify and detect high-frequency circulating currents and EDM currents. Noise was mitigated through the application of low-pass and band-pass filters, enabling precise detection of significant signal changes during data preprocessing. Additionally, outlier detection using the IQR method identified pulse signal occurrence points, and comparative analysis of outliers in CMV and bearing current data facilitated the classification of the physical characteristics of each current type into circulating and EDM currents. Finally, time-series clustering was employed to determine the frequency and timing of electrical erosion signals caused by high-frequency circulating and EDM currents.

The methodology maintained consistent performance under load conditions, exhibiting higher recall and precision compared to traditional threshold-setting methods, thus overcoming the limitations of fixed threshold approaches. This study demonstrates that the proposed technique can be practically applied to the maintenance and prediction of modern electric motors, contributing to the enhancement of system durability and reliability. Future research will explore the feasibility of applying this method across various electric motors and real-world operational conditions, including the evaluation of its effectiveness in high-speed scenarios, which is considered essential. Furthermore, additional studies are necessary to develop real-time detection systems. This approach is expected to significantly improve the maintenance efficiency and safety of electric motor systems.

This study has been conducted with the support of the Korea Planning & Evaluation Institute of Industrial Technology as part of the Development of a 600 kW-33,000 rpm ultra-high-speed motor and dynamometer control system (KEIT 20024404). It was also supported by the Korea Institute of Industrial Technology through the Autonomous Manufacturing Technology based on DNA Platform (KITECH EH-24-0002), and by the Korea Institute of Energy Technology Evaluation and Planning (KETEP), and the Ministry of Trade, Industry and Energy (MOTIE), Republic of Korea (No. 20227310100010).

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.

$VA$ =

phase voltages of phases $A$

$VB$ =

phase voltages of phases $B$

$VC$ =

phase voltages of phases $C$

$Q1$ =

lower quartile corresponding to the 25th percentile

$Q3$ =

upper quartile corresponding to the 75th percentile

$CMV$ =

common mode voltage

$IQR$ =

difference between the third and the first quartile in interquartile range

$σ$ =

standard deviation

$x$ =

input data

$xmin$ =

minimum value of the data

$xmax$ =

maximum value of the data

$xscaled$ =

normalized data