A Data-Driven-Based Grounding Fault Location Method for the Auxiliary Power Supply System in an Electric Locomotive

Abstract

Grounding faults are a common type of fault in train auxiliary power supply systems (APS). Timely identification and localization of these faults are crucial for ensuring the stable operation of electric locomotives and the safety of passengers. Therefore, this paper proposes a fault diagnosis method for grounding faults (GFs) that integrates mechanistic insights with data-driven feature extraction. Firstly, this paper analyzes the mechanisms of grounding faults and summarizes the characteristics of their time–frequency distribution. Then, a Short-Time Fourier Transform (STFT) is employed to derive a frequency signature vector enabling classification into three principal categories. Concurrently, a time series sliding window approach is applied to extract time domain indicators for further subdivision of fault types. Finally, a time–frequency hybrid-driven diagnostic model framework is constructed by integrating the frequency distribution with the spatiotemporal map, and validation is conducted using an experimental platform that replicates system fault scenarios with a hardware-in-the-loop (HIL) simulation and executes the real-time diagnostic frameworks on a DSP diagnostic board card. The results demonstrate that the proposed method can detect and accurately locate grounding faults in real time.

1. Introduction

The rapid evolution of the electric locomotive (EL) industry, characterized by increasing speeds and expanding operational distances, poses new challenges for high-speed railway technology. In addition to the safe and stable operation of trains, comfortable riding experiences are receiving increasing attention. The electric traction drive system, which is at the heart of electric locomotives and powers both the locomotion and internal loads, comprises the main transformer, traction drive apparatus, auxiliary power supply system, and electrical control system [1]. The auxiliary power supply system (APS) is a typical AC–DC–AC power electronic conversion system that supplies power to ventilation, heating, and lighting and charges the on-board backup batteries [2]. Therefore, the stable and reliable power supply of an APS is crucial for the safe construction of trains and a good riding experience for passengers.

In recent years, the extended operating hours of electric locomotives have rendered the auxiliary power supply system more susceptible to sudden electrical faults, attributable to component aging and environmental stressors [3]. Premature component failures can gradually reduce the overall performance of the system at a slow gradient trend [4], potentially resulting in a degraded riding experience and safety concerns. However, in actual engineering practice, a fault limit alarm [5] is capable of detecting faults but is inadequate for pinpointing their locations, increasingly falling short of the operational demands of locomotives. Consequently, there is an urgent need for a timely and reliable fault location method [6] to ensure the safe operation of locomotives and to enable maintenance personnel to swiftly and effectively address the fault, thereby minimizing disruptions to the train’s operational schedule.

In fact, the integration of fault diagnosis technology in electric locomotives lags significantly behind advancements in other domains, such as control [7], information [8], materials, and management [1]. The characteristics or parameters of the electric locomotive system are not allowed to deviate from their normal state, leading to system “faults” or “failures” in fault diagnosis [9]. The collection of raw data for fault diagnosis is inherently challenging, further complicated by the interdependent relationships among various system components [10]. Additionally, variations in load and speed [11], attributable to fluctuating working conditions, necessitate enhanced adaptability in fault diagnosis methodologies.

Typical electrical component failures encompass open and short circuit faults of Insulated Gate Bipolar Transistors (IGBTs) [12,13], performance degradation of DC link electrolytic capacitors, abnormal changes in passive components, etc. [14]. These issues are predominantly related to individual device levels. C. Yang et al. in [12] proposed a fault diagnosis method for a three-level converter in an electric traction system based on voltage difference residuals to identify device-level faults through a hierarchical diagnostic approach. However, during actual operation, electric locomotives are more susceptible to system-level faults, such as grid-side faults [15], main circuit faults [16], traction motor faults [17,18], etc. The fluctuation of the pantograph-contact network severely affects the data collection quality and operational safety of high-speed electric locomotives. H. Wang et al. in [6] proposed Bayesian Adaptive Reinforcement Learning (CBAMRL) based on contrastive learning to adapt to the non-stationary environment of the contact network and efficiently collect data.

System-level faults within the main circuit, particularly grounding faults, can significantly impair the operation of the traction motor and, consequently, the operational safety of the electric locomotive. For grounding faults in the direct–alternating current circuit of electric vehicles, J. M. Guerrero et al. in [19] employed a spectral analysis of the detected half-voltage to distinguish whether the fault occurs on the AC or DC side. And, the team in [20] estimated fault location based on the calculation of voltage parameters from histograms; however, this method encounters significant inaccuracies during load variations. The team in [21] detected grounding faults on the AC grid side, DC positive or negative terminals, or the AC inverter side by analyzing the waveform of the terminal voltage signal. Nonetheless, relying solely on waveform diagnosis without extracting additional waveform features reduces the accuracy of identifying different types of grounding faults. It is evident that the diagnosis of main circuit grounding faults mainly relies on the grounding detection half-voltage of the DC detection circuit. However, analyzing only the detected voltage and ignoring the impact of variables in each area makes it difficult to locate faults within the area. Additionally, the effects of load variations and fault degradation must also be considered. For a traction system main circuit grounding fault, Z. Chen et al. in [22] employed the traditional correlation method and residual vector method for fault location. Q. Ni et al. in [23] proposed a Mechanism-Data Hybrid Driven (MDHD) approach for grounding fault diagnosis, while the team in [24] introduced a Frequency Domain Characteristic (FDC) method, achieving good diagnostic results. Refs. [23,24] conducted research on the traction drive system. The auxiliary power supply system and traction drive system obtain single-phase AC power from different secondary windings of the traction transformer, respectively, and they are usually isolated from each other. The traction drive system provides driving power for the train to move forward and the auxiliary power supply system provides electrical energy for passenger train carriages. Moreover, their circuit topology and working mechanism are completely different; therefore, the signal selection, variable construction, and feature indexes are different. At the same time, both [23,24] are based on fault samples and were completed offline in a computer environment, without conducting online diagnostic research based on the on-board DSP platform. Therefore, the real-time detection and positioning method for APS grounding faults under different loads and fault degradation still needs further research.

Inspired by the aforementioned methods, this paper focuses on the grounding fault issue in the auxiliary power supply system of electric locomotives and proposes a time–frequency feature-based stepwise diagnostic method for the detection and location of grounding faults. The main contributions are as follows:

(1)

The grounding fault mechanism is analyzed, and the circuit relationship between grounding detection voltage U_h (voltage sensor SV₂) and DC- link voltage U_a (voltage sensor SV₁) is described. The diagnostic process leverages the sampling signals of an APS system (U_h, U_a, transformer secondary side AC voltage U_s, and AC current I_s) to identify fault types and investigate their corresponding characteristics.

(2)

The frequency domain features of U_h are extracted, and their correlation with fault types is determined using an extreme learning machine. Additionally, the feature variables V₁ and V₂ are reconstructed from signals U_a and I_s, and the concept of a time sliding window is introduced to extract time domain feature indicators M₁ and M₂.

(3)

A diagnostic framework that integrates signal time–frequency domain analysis and data-driven methods is proposed, which first divides faults into three categories: before rectifier bridge, before the filtering inductor, and after filtering inductor. Then, both before the rectifier bridge faults and after the inductor faults are further divided into two subtypes of faults each, achieving the identification of five types of faults. The effectiveness of the diagnostic strategy is verified online through a hardware-in-the-loop (HIL) platform and a Digital Signal Processor (DSP) diagnostic board card.

The rest of this paper is organized as follows. In Section 2, the time–frequency mechanism of GFs in an APS is analyzed. In Section 3, the time–frequency hybrid-driven-based GF diagnosis method is proposed. In Section 4, the proposed method is verified by the HIL and the experimental results are analyzed. Section 5 concludes this paper.

2. Mechanism of Ground Faults in a Power Supply System

2.1. Main Circuit of an APS

The schematic of the main circuit of the auxiliary power supply system for an electric locomotive is shown in Figure 1. The single-phase 25 kV AC power, drawn from the traction substation, is transmitted through the overhead contact line. After passing through the pantograph and high-voltage circuit breaker, it is stepped down by the traction transformer to provide a single-phase AC power source for the auxiliary power supply system’s dedicated rectifier (diodes V1 and V2, and thyristors V3 and V4). This converts the single-phase AC 860 V to a pulsating DC 600 V. Subsequently, a stable DC 600 V is output to the integrated control cabinet through the intermediate filtering stage (smoothing reactor L and supporting capacitor C), which controls and distributes the flow of energy (different voltage levels for lighting, heating, ventilation, charging, etc.). The control system collects the transformer secondary side voltage U_s (voltage transformer TV), current I_s (current transformer TA), DC link voltage U_a (voltage sensor SV₁), and DC-side ground detection voltage U_h (voltage sensor SV₂) to regulate the output of the auxiliary power supply system and detect ground faults. R represents a fixed discharge resistor with equal resistance values to achieve an even pressure effect on the DC side, and R_f is the equivalent ground resistance, ensuring effective collection of U_h. Referring to the circuit topology and the actual fault location maintenance experience of field engineering, the ground fault location types are divided into five categories across the AC and DC sides. Numbers 1 to 5 indicate five different ground fault locations, as shown in

Table 1. Types of ground faults.

Number	Fault Point	Code	Ground Faults	Code
①	DC link	$F_1$/$F_2$	Positive side of the DC link	$F_1$
②	DC link	$F_1$/$F_2$	Negative side of the DC link	$F_2$
③	Reactor link	$F_3$	Front side of the Reactor	$F_3$
④	RC Input	$F_4$/$F_5$	RC Input Positive side	$F_4$
⑤	RC Input	$F_4$/$F_5$	RC Input Negative side	$F_5$

.

2.2. Time–Frequency Mechanism of Ground Faults

In the auxiliary power supply system of electric locomotives, an online insulation detection device is equipped with a control system. The signals from the TV, TA, SV₁, and SV₂ are fed into the system for online analysis.

As shown in Figure 1, five grounding fault locations are marked, with R_f1 to R_f5 representing the equivalent grounding insulation resistances at these locations. When the electric locomotive is operating normally, the equivalent grounding resistance R_f at each location is at the MΩ level, and since the resistance values are equal for R₁ and R₂, it can be determined that no grounding fault has occurred in the electric locomotive when U_h = U_a/2.

(1) When direct-current-side positive-end grounding (F₁) occurs in the system, according to circuit theory, after star-delta transformation of the grounding detection circuit, the expressions for U_h and U_a are as follows:

$$U_h={\displaystyle \frac{R_{f1}}{R+2R_3+2R_{f1}}}{U}_a$$

(1)

R represents the fixed discharge resistor with equal resistance values for R₁ and R₂ to achieve a uniform pressure effect on the DC side, and R_f1 is the equivalent ground resistance when the positive side of the DC link is grounded. R₃ is the ground detection resistor, ensuring effective collection of DC-side ground detection voltage U_h. U_a represents the DC link voltage. From a frequency domain perspective, U_h is a direct current that is directly proportional to U_a, and its frequency composition only contains a DC component. From a time domain perspective, as the degree of grounding insulation degradation increases (R_f1 gradually decreases to 0), U_h gradually decreases from 0.5U_a to 0.

(2) When direct-current-side negative-end grounding (F₂) occurs in the system, the expressions for U_h and U_a are as follows:

$$U_h={\displaystyle \frac{R+2R_3+R_{f2}}{R+2R_3+2R_{f2}}}{U}_a$$

(2)

R represents the fixed discharge resistor with equal resistance values for R₁ and R₂ to achieve a uniform pressure effect on the DC side, and R_f2 is the equivalent ground resistance when the negative side of the DC link is grounded. R₃ is the ground detection resistor, ensuring effective collection of DC-side ground detection voltage U_h. U_a represents the DC link voltage. From a frequency domain analysis perspective, U_h is a direct current that is proportional to U_a, with its frequency composition consisting solely of a DC component. From a time domain analysis perspective, as the degree of degradation of the grounding insulation increases (with R_f2 gradually decreasing to 0), U_h increases from 0.5U_a to U_a, exhibiting a trend that is opposite in polarity to that observed in the time domain for F₁.

(3) When a grounding fault occurs at the front end of the reactor (F₃) in the system, the expressions for U_h and U_a are as follows:

$$U_h={\displaystyle \frac{R+2R_3+R_{f3}}{R+2R_3+2R_{f3}}}{U}_a-{\displaystyle \frac{R+2R_3}{R+2R_3+2R_{f3}}}{U^{\prime }}_s$$

(3)

$${U^{\prime }}_s=\left\{\begin{array}{cc}{U}_s,& I_s>0\\ -U_s,& I_s<0\\ 0,& I_s=0\end{array}\right.$$

R represents the fixed discharge resistor with equal resistance values for R₁ and R₂ to achieve a uniform pressure effect on the DC side, and R_f3 is the equivalent ground resistance when the front side of the reactor is grounded. R₃ is the ground detection resistor, ensuring effective collection of DC-side ground detection voltage U_h. U_a represents the DC link voltage. U_s^’ is a piecewise function that represents the voltage on the AC side. From the frequency domain perspective, U_h is composed of a direct current (DC) component and an alternating current (AC) component. The DC part is constituted by the steady DC current corresponding to U_h during a grounding fault at the negative end of the DC side. The AC part is made up of AC quantities that are strongly related to U_s and I_s. Additionally, the frequency composition only includes even multiples of the fundamental frequency. When I_s = 0, the AC component is 0, and U_h contains only the DC component. From the time domain perspective, as the degree of degradation of the grounding insulation increases (with R_f3 gradually decreasing to 0), the waveform amplitude range increases and the periodic temporal features become more pronounced.

(4) When a grounding fault occurs at the AC-side positive-end (F₄) in the system, the expressions for U_h and U_a are as follows:

$$U_h={\displaystyle \frac{R+2R_3+R_{f4}}{R+2R_3+2R_{f4}}}{U}_a-{\displaystyle \frac{R+2R_3}{R+2R_3+2R_{f4}}}{U^{\prime }}_s$$

(4)

$${U^{\prime }}_s=\left\{\begin{array}{cc}0,& I_s<0\\ U_s,& I_s\ge 0\end{array}\right.$$

R represents the fixed discharge resistor with equal resistance values for R₁ and R₂ to achieve a uniform pressure effect on the DC side, and R_f4 is the equivalent ground resistance when the RC input positive side is grounded. R₃ is the ground detection resistor, ensuring effective collection of DC-side ground detection voltage U_h. U_a represents the DC link voltage. U_s^’ is a piecewise function that represents the voltage on the AC side. From the frequency domain perspective, U_h consists of a direct current (DC) component and an alternating current (AC) component. Additionally, the frequency composition includes both even and odd multiples of the sub-harmonic. When I_s < 0, the AC component is 0 and U_h contains only the DC component. From the time domain perspective, as the degree of degradation of the grounding insulation increases (with R_f4 gradually decreasing to 0), the waveform amplitude span increases and the periodic negative polarity temporal characteristics become more pronounced.

(5) When a grounding fault occurs at the AC-side negative-end (F₅) in the system, the expressions for U_h and U_a are as follows:

$$U_h={\displaystyle \frac{R+2R_3+R_{f5}}{R+2R_3+2R_{f5}}}{U}_a-{\displaystyle \frac{R+2R_3}{R+2R_3+2R_{f5}}}{U^{\prime }}_s$$

(5)

$${U^{\prime }}_s=\left\{\begin{array}{cc}-U_s,& I_s<0\\ 0,& I_s\ge 0\end{array}\right.$$

R represents the fixed discharge resistor with equal resistance values for R₁ and R₂ to achieve a uniform pressure effect on the DC side, and R_f5 is the equivalent ground resistance when the RC input negative side is grounded. R₃ is the ground detection resistor, ensuring effective collection of DC-side ground detection voltage U_h. U_a represents the DC link voltage. U_s^’ is a piecewise function that represents the voltage on the AC side. From the frequency domain perspective, U_h consists of a DC component and an AC component, which are similar to the frequency composition of F₄, including both even and odd multiples of the sub-harmonic. When I_s > 0, the AC component is 0, and U_h contains only the DC component. From the time domain perspective, as the degree of degradation of the grounding insulation increases (with R_f5 gradually decreasing to 0), the waveform amplitude range increases, and the periodic positive polarity temporal characteristics become more pronounced, showing a trend opposite to the time domain polarity characteristics of F₄.

3. The Proposed Diagnosis Method

3.1. Frequency Domain Feature Analysis

From the aforementioned mechanism analysis, it is evident that under various grounding fault conditions, U_h demonstrates variability in the composition of its direct current and alternating current components. Concurrently, the operating state of the electric locomotive fluctuates with changes in road conditions. To obtain the non-stationary transient time–frequency characteristics, this study utilizes Short-Time Fourier Transform (STFT) [25] to extract distinctive features at varying times and frequencies in pursuit of identifying robustly correlated feature indicators.

The principle of the STFT is illustrated in Figure 2, where the continuous signal x(t) is combined with a window function w(t) of a fixed length. By sliding the window function, the frequency characteristics of the signal at different time segments are analyzed. The formula is defined as follows:

$$STFT\left\{x(t)\right\}\left(f,\tau \right)={\displaystyle {\int }_{-\infty }^{\infty }x(t)w(t-\tau )}{e}^{-j2\pi ft}dt$$

(6)

In the equation, t represents the current moment, and x(t) represents the continuous signal. w(t − τ) denotes the window function, which moves with time τ. f signifies the frequency variable, while τ is the time variable, representing the central position of the window.

The core characteristic of the STFT is the trade-off between time and frequency resolution. The width of the window function affects the resolution: a wider window leads to better frequency resolution, while a narrower window results in better time resolution. During the time–frequency transformation process, the sampling frequency fs is set to 25 kHz, consistent with the sampling frequency of the experimental data. A Hamming window function is chosen to reduce spectral leakage. The formula is as follows:

$$w(n)=\alpha -\beta \cos ({\displaystyle \frac{2\pi n}{N-1}})$$

(7)

where w(n) represents the nth sample value of the window function. $\alpha =0.54$ and $\beta =1-\alpha =0.46$ are the coefficients that adjust the shape of the window. N is the total number of samples in the window function, and n is the position of the current sample. $n=0,1,\cdots ,N-1$.

$$\left[S,f,t\right]=spectrogram\left(y,win,noverlap,nfft,fs\right)$$

(8)

The function of the STFT is depicted in (8), where S is the matrix generated after performing the STFT transformation on the signal y. Each column represents the result of the Fast Fourier Transform (FFT), each row represents the time series, f denotes the frequency value corresponding to each point of the FFT, and t indicates the center time of the sliding window. The type of window function is denoted by win. The step size of the sliding window is noverlap = 50, the number of sampling points for the sliding window FFT is nfft = 512, and the sampling frequency is fs = 25 kHz.

As depicted in Figure 3, a frequency domain analysis is conducted on the U_h for five types of grounding faults. There are obvious differences in the time–frequency domain performance of F₁–F₅. In terms of frequency domain characteristics, the fault signal on the DC side (F₁ and F₂) only contains the DC part, the fault signal at the front end of the reactor (F₃) contains both DC and AC parts, and both are even harmonics; additionally, the fault signal on the AC side (F₄ and F₅) contains both DC and AC parts, and there are both odd and even harmonics. By observing the distribution of frequency composition, the main frequencies of the five types of grounding faults are analyzed: A₀ (DC component), A₁ (first harmonic amplitude), and A₂ (second harmonic amplitude). Relying solely on the frequency values for fault classification presents a problem due to the variation in fault severity, which can lead to differences in frequency amplitude. Through a numerical analysis, the characteristic value of A₂/A₀ is introduced, forming a feature vector T = [A₀, A₁, A₂, A₂/A₀].

The 4-dimensional dataset T is imported into the input layer of the extreme learning machine (ELM) [26], and the optimal input weights (I_w), bias (B), and output weights (L_w) are generated through random training by setting the number of hidden layer nodes to 10 and activating the function to Sigmoid. The output layer outputs the fault label corresponding to the operation of the matrix, which can distinguish between grounding faults in three areas: the DC side (F₁/F₂), the front end of the reactor (F₃), and the input side of the rectifier (F₄/F₅). However, because the frequency domain vector of the signal alone can only be divided into three categories, it is not possible to accurately locate the specific position. Therefore, it is necessary to further analyze the characteristics of the specific fault type.

3.2. Time Domain Feature Analysis

As shown in Figure 3, a time domain analysis was conducted on the U_h for five types of grounding faults. The fault signals on the DC side (F₁ and F₂) have stepped numerical characteristics, and with aggravation of the faults, the U_h signal of F₁ tends to 0 and the U_h signal of F₂ tends to 600. The U_h signals at the front end (F₃) and AC side (F₄ and F₅) of the reactor show periodic numerical characteristics. By observing the characteristics of the time domain waveforms, it was determined that the two types of faults (F₁/F₂) exhibit a feature of opposite polarities in U_h, and the other two types of faults (F₄/F₅) show localized regular waveform features.

In order to further distinguish the polarity characteristics of faults on the DC side, a new variable V₁ was reconstructed to normalize the polarity threshold range to the standard 0 on either side. The expression for V₁ is given by the following:

$$V_1=U_h-0.5U_a$$

(9)

The V₁ feature fails to distinguish between faults F₄ and F₅. U_h represents the DC-side ground detection voltage and U_a represents the DC link voltage. I_s represents the transformer secondary side AC current. Through the mechanism analysis, when I_s < 0 (F₄) and I_s > 0 (F₅), the influence of the AC quantities of F₄ and F₅ is respectively eliminated. Therefore, the I_s variable is introduced, in conjunction with the variable V₁, to reconstruct a new variable V₂. The expression is given by the following:

$$V_2=V_1\ast I_s=(U_h-0.5U_a)\ast I_s$$

(10)

Due to the complex and variable operating conditions, instantaneous indicators cannot accurately describe fault characteristics. It is necessary to extract continuous fluctuation feature vectors over a period of time to dynamically represent the health status of the system. Variables V₁ and V₂ exhibit periodic temporal features; hence, a time-sliding window algorithm is employed for feature extraction.

As shown in Figure 4, a window(win(t)) [27] of length L (20 ms) is used to slide and segment the waveform of the data segment of length H. At the same time, the sliding step S (1 ms) is set and moved along the direction of the time series. Each slide results in a data segment of length L. When the sliding reaches a point where the remaining data segment is less than the set step, the sliding is stopped.

After applying the sliding window treatment to variables V₁ and V₂, the characteristic indicators M₁ and M₂ of the variables within the period are extracted, with the following formulas:

$$M_1(k)={\displaystyle \frac{1}{N}}{\displaystyle \sum _{t=1}^{N+k-1}{V}_1(t)}$$

(11)

$$M_2(k)={\displaystyle \frac{1}{N}}{\displaystyle \sum _{t=1}^{N+k-1}{V}_2(t)}$$

(12)

where k is each point of the feature indicators; N is the number of feature variables within the sliding window; i = 1, 2, …, N; t represents the current moment; M₁ is the mean of V₁; and M₂ is the variance of V₂.

As shown in

Table 2. Internal ground fault diagnostic rules.

Indicator	Diagnostic Rules	Ground Faults
M1	M1 < Jth1	$F_1$
M1	M1 > Jth2	$F_2$
M2	M2 < Jth3	$F_4$
M2	M2 > Jth4	$F_5$

, the judgment based on the threshold values of M₁ and M₂ can distinguish between F₁ or F₂, and F₄ or F₅. In this context, J_th1, J_th2, J_th3, and J_th4 are threshold values around 0. Specifically, J_th1 is less than 0, J_th2 is greater than 0, J_th3 is less than 0, and J_th4 is greater than 0.

After judging the fault to be a DC-side fault (F₁/F₂) through the frequency domain characteristics, the specific fault type is judged by the time domain index diagnosis rules in Table 2; when M₁ < J_th1, it is judged to be a DC-side positive ground fault (F₁), and when M₁ > J_th2, it is judged to be a DC-side negative ground fault (F₂). Similarly, after judging that the fault is an AC-side fault (F₄/F₅) based on frequency domain characteristics, when M₂ < J_th3, it is judged to be a ground fault at the positive end of the RC input side (F₄), and when M₂ > J_th4, it is judged to be a ground fault at the negative end of the RC input side (F₅).

3.3. Fault Diagnosis Scheme Based on Time–Frequency Features

Based on the aforementioned time–frequency analysis, a diagnostic approach that clarifies fault area determination in conjunction with internal localization is established. The fault diagnosis scheme based on time–frequency features is divided into two phases—offline and online—as shown in Figure 5.

In the offline phase, by integrating the topological model with the actual parameters of the locomotive and using an offline waveform acquisition system to collect signals, the original fault detection signal U_h is subjected to frequency domain analysis. The feature vector T = [A₀, A₁, A₂, A₂/A₀] is extracted as the dataset. The classification of the three fault areas (F₁/F₂, F₃, F₄/F₅) is accomplished through an extreme learning machine (ELM) classification model. The reconstructed variables V₁ and V₂ are analyzed in the time domain to extract feature indicators M₁ and M₂. With threshold judgment, precise fault location can be achieved under the condition of clearly defined fault areas. Ultimately, the differentiation of five types of grounding faults is completed.

In the online phase, the DSP diagnostic board card collects the relevant analog signal of the hardware-in-the-loop system in real time, extracts the sampling data segment, and compares it with the threshold range of the offline normal working conditions to determine whether to enter the fault diagnosis process; when the fault detection flag changes from 0 to 1, the time–frequency characteristics are each extracted to determine the type of fault.

4. Experiment and Data Analysis

To verify the effectiveness of the proposed diagnostic framework, this section conducts an experimental validation and a result analysis.

4.1. Description of the Experiment

This paper utilizes data from a hardware-in-the-loop (HIL) experimental platform, where the system parameters are consistent with the actual locomotive parameters. As shown in Figure 6, RT-LAB (OP5600) is responsible for compiling the hardware circuit model and outputting the required analog monitoring signals to the dSPACE (DS1202) control platform; dSPACE is responsible for compiling the control strategy and outputting the necessary digital control signals to RT-LAB to complete the information interaction. Ultimately, the fault detection signals are output by the RT-LAB analog I/O board.

Referring to the national standards of the DC 600 V supply power system in railway passenger trains and the experience of engineering practice, the grading of the system’s healthy status is shown in

Table 3. Health status grading explanation form.

Grade	Health Status	Ground Insulation Resistance Value
A	normal	$R_f\ge 20k$
B	degradation	$20k>R_f\ge 1k$
C	abnormal	$1k>R_f\ge 50$
D	fault	$50>R_f\ge 0$

. In the experiment, the real-time control signals of dSPACE are stabilized, and the equivalent grounding resistance R_f of the RT-LAB circuit and its connection position are changed in sequence to simulate different types of grounding faults with increasing degrees of degradation. Specifically, R_f is set to 10 kΩ, 5 kΩ, 1 kΩ, 500 Ω, and 30 Ω, corresponding to different health statuses, with a sampling time of 20 μs, and the simulation time is 4 s.

The fault diagnosis program execution unit can be used as a diagnostic board in traction control units (TCUs). The AD7606 module continuously collects system signals (U_a, U_h, and I_s) by setting a data buffer for the sampling frequency, and the DSP concurrently extracts segments of sampled data for fault detection. These data segments are compared with thresholds, which are obtained through a data statistical analysis of normal working conditions. If the set conditions are not met, the DSP continues to extract data segments for further fault detection. Once the conditions are satisfied, the fault detection is halted, and the online diagnostic model is executed. The current data segment undergoes data processing, followed by analysis through the model’s computation, ultimately outputting a fault location tag to complete the online diagnosis.

4.2. Experimental Results

As shown in Figure 7, according to the load power requirements of the actual design of the auxiliary power supply system, under the three loading conditions of 9 kW (fault emergency load), 100 kW (light load), and 400 kW (heavy load), the four frequency domain indicators of the five types of faults are extracted and spliced into a continuously varying sample curve. Among them, the number of specific fault samples for each type is 205, and the number of fault samples for a total of five types is 1025 (205*5). Through horizontal comparison, there are differences in the amplitude of the frequency domain indexes T = [A₀, A₁, A₂, A₂/A₀] on the DC side (F₁/F₂), the front end of the reactor (F₃), and the rectifier input side (F₄/F₅) under the conditions of a single load, and the change trends are different; specifically, the two indicators of A₀ and A₂ are obviously different. The sample curves are clearly divided into three major fault categories. At the same time, through longitudinal observation, the above phenomenon also exists under a load with different powers, which indicates that it is effective for dividing the frequency domain index into three types of fault categories, so the method of identifying the fault categories of the frequency domain index can be summarized. Firstly, it is evident that the faults can be categorized into the DC side (F₁/F₂), the front end of the reactor (F₃), and the rectifier input side (F₄/F₅). Secondly, although there are differences in frequency amplitude under different loads, the overall distribution trend remains consistent. The A₂/A₀ indicator proposed in this paper is unaffected by load variations. Lastly, by integrating the frequency domain indicator vector T = [A₀, A₁, A₂, A₂/A₀] and by learning the mapping relationship through a classification model, the fault areas can be reliably distinguished.

As shown in Figure 8, in order to test the accuracy of the model in identifying three types of fault categories and to compare it with the identification of five types of specific faults, the same dataset and division ratio were used. The sample dataset (1025) is divided into 70% (717) as the training set for training the model and 30% (308) as the test set for testing the model. In the test set, the model has an accuracy of 59.6% for five types of specific faults and 99.3% for three types of fault categories. The classification accuracy of the locations of the five types of grounding faults is significantly lower than the classification accuracy of the three fault areas. Therefore, it is still necessary to rely on time domain features for classifying faults within the area.

As shown in Figure 9, under the three loading conditions of 9 kW, 100 kW, and 400 kW, the time domain characteristic indicators are extracted for the DC side (F₁/F₂) and the rectifier input side (F₄/F₅) grounding faults, respectively. In Figure 9a,b, under the condition of a single load, the current I_s on the AC side is a sinusoidal wave curve, and the voltage U_a on the DC side is basically stable at about 600 V. The indicators (U_h, V₁, and M₁) for F₁/F₂ exhibit a step-like trend with the degree of fault degradation. Under different loading conditions, the current I_s on the AC side is a sine wave curve, and with the increase in load power, the current amplitude rises. The voltage U_a on the DC side is basically stable at about 600 V, and the fluctuation range is smaller. The indicators (U_h, V₁, and M₁) for F₁ show a downward trend with the increase in load, while the indicators (U_h, V₁, and M₁) for F₂ show an upward trend. The indicator M₁ for F₁ is significantly less than the threshold (less than 0), and the indicator M₁ for F₂ is significantly greater than the threshold (greater than 0), with the range of changes in the indicator and the threshold remaining unaffected by changes in load.

In Figure 9c,d, under the condition of a single load, the indicators for F₄/F₅ exhibit an increasing trend in amplitude with the degree of fault degradation. Under different loading conditions, the waveforms of the F₄/F₅ fault also show an increasing trend in amplitude with the increase in load. The indicator M₂ for F₄ is significantly less than the threshold (less than 0), and the indicator M₂ for F₅ is significantly greater than the threshold (greater than 0). As the load increases, the range of variation in the indicator increases, but the threshold is minimally affected by changes in load.

Through HIL offline experiments, we analyzed the frequency domain distribution characteristics and time domain waveform numerical characteristics of different ground faults under different loads and different ground degradation degrees. Specifically, the time–frequency feature is used to scan for regional faults and then accurately locate internal faults. Experiments show that the proposed method is basically not affected by load changes, thus achieving rapid and precise localization of grounding faults. Therefore, the real-time effectiveness of the proposed method is verified online using a DSP diagnostic board card.

4.3. Online Diagnostics and Analysis

As shown in Figure 10a, at t₀, the system begins an operation and reaches a stable condition. At t₁, the system starts to experience a grounding fault. The period from t₁ to t₂ is the degradation phase (R_f = 10 kΩ). From t₂ to t₃, the degradation intensifies (R_f = 5 kΩ). The period from t₃ to t₄ is the abnormal phase (R_f = 1 kΩ). From t₄ to t₅, the abnormality intensifies (R_f = 500 Ω), and at t₅, the fault phase occurs (R_f = 30 Ω). During normal operation of the locomotive, V₁ fluctuates within the range [J_UL, J_LL] due to signal interference. Real-time sampling detection identifies any deviation outside this range as a grounding fault in the system, and the fault detection label F changes from 0 to 1.

As depicted in Figure 10b, the frequency domain indicators T = [A₀, A₁, A₂, A₂/A₀] are extracted. Before the fault occurs, the state of each indicator is stable within the safe range, and when the fault occurs, the indicator deviates from the normal range and increases with aggravation of the fault. After model computation, the fault area label AL outputs 3 (F₄/F₅).

As shown in Figure 10c, V₂ is stable at about 0 before the fault occurs, and after the fault occurs, it oscillates violently and tends to be less than 0. However, due to the large range of oscillation, after extracting the indicator M₂, the fault information is more accurate, and the value of the threshold J_th3 can be determined to be about −30. After the fault area is identified, the fault localization label LL outputs 4 (F₄) based on the judgment of the time domain characteristic indicator M₂ and the polarity threshold J_th3.

Therefore, through an offline experimental analysis and online experimental verification, it is shown that the proposed method can be divided into two steps—the frequency domain feature division fault category and the time domain feature location fault specific sub-category—and complete the location and diagnosis of a ground fault in an auxiliary power supply system.

5. Conclusions

This paper proposes a method for the accurate diagnosis and localization of grounding faults in auxiliary power supply systems by combining fault mechanisms with a time–frequency feature analysis. Through a mechanism analysis, frequency domain features and time domain indicators are selected; the frequency domain features are used to determine the fault area, and the time domain indicators are used for fault localization within the area. By combining a data-driven model with judgments of load threshold changes, the method effectively captures variations in frequency distribution and spatiotemporal mapping, enhancing the distinguishability of different types of faults. The proposed time–frequency hybrid diagnostic model was validated using an HIL and OMAP-L138 experimental platform. The experimental results show that this diagnostic strategy can track fault locations online, improving maintenance efficiency.

The contributions of this paper are the introduction of a fault mechanism analysis on the basis of a traditional data-driven model; making data processing interpretable; and at the same time, a simplification of the operation of machine learning models by extracting frequency domain indicators, the construction of time domain variables and the extraction of feature indicators, and the clarification of the diagnostic criteria. More importantly, this paper proposes a combination mode method of offline mechanism analysis and hardware online diagnosis on the basis of limited hardware and achieves real-time fault matching DSP online diagnosis based on an HIL platform, which can be extended to other industrial application scenarios.

In the future, this method will be ready to be adapted to the on-board platform, and at the same time, it is necessary to continuously optimize the online sampling mechanism of real-time fault data and to be able to quickly scan the sampling points, so as to reduce the detection of data points and complete the hardware online diagnosis more quickly. In addition, research on the assessment technology for the severity of grounding faults will be conducted to further enhance the level of train intelligence.