Figure 1A shows the system structure of the charging station, which mainly consists of the power grid, transformer, and charging station. The internal structure of the charging station includes an AC/DC rectifier and a DC/DC converter. Figure 1B presents the topology of the converter. The AC/DC rectifier of the charging station uses a simple and reliable three-phase bridge rectifier circuit (Liserre et al., 2005; Safayatullah et al., 2022). To provide electrical isolation or extend the output voltage range, the DC/DC converter adopts an isolated Dual Active Bridge (DAB) converter (Zhao et al., 2014). The use of an isolated converter also greatly reduces the damage caused by faults. When a fault occurs on the secondary side of the isolated DC-DC converter, the presence of the isolation transformer does not affect the grid. In contrast, when a fault occurs on the primary side of the isolated DC-DC converter, the energy during the fault is supplied by the cascaded capacitor, and the grid current remains unchanged. The extended fault duration provides sufficient time for the circuit breaker to protect the circuit, thus minimizing the impact on the grid. Therefore, when a fault occurs in the bridge rectifier circuit directly connected to the grid, the impact on the grid is more severe and requires extra caution, along with adequate protective measures. This paper focuses on predictive studies of faults in bridge converter circuits.

A certain data set is required to make predictions, which is collected from the charging pile simulation system shown in Figure 1. The 600 VAC grid-side voltage is converted to 380 VAC three-phase AC voltage through a three-phase transformer. This is then rectified using a three-phase bridge rectifier circuit, converting the 380 V AC into 500 V DC. The collected data includes the three-phase voltage and three-phase current on the grid side. When different faults occur in the charging pile, the voltage and current on the grid side will change correspondingly, allowing for the identification of the fault type. To obtain data on faults occurring at different operational power levels in real scenarios, the experiments start at a power of 100 kW, increasing in steps of 5 kW, and ending at the rated power of 200 kW, resulting in a total of 21 sets of fault data at different power levels. Using Ohm’s Law, the current can be calculated from the power and voltage. For example, when the power is 100 kW, the DC side current is 200 A, and the RMS value of the AC side current is approximately 263 A.

Since the three-phase power of the grid is symmetrical, the impact of any fault of the same type in any phase on the grid is the same. Therefore, it is sufficient to collect and analyze fault data for phase a. Furthermore, in real scenarios, faults in the charging pile mainly arise from the damage to switches in the rectifier converter. At the same power level, the types of faults can be categorized into three types based on the condition of the damaged switches: Type A represents a short-circuit condition between two phases, where any two phases are directly connected due to overheating and failure of the switching devices. This failure may result from overcurrent or elevated environmental temperatures, leading to a short circuit between two AC bus phases. For instance, if the upper or lower bridge switches of phases a and b fail simultaneously, a short circuit will occur. Each phase, such as phase a, includes two switches: an upper bridge switch and a lower bridge switch. Type B indicates a situation where the upper bridge switch of phase a fails, while the rest of the external circuitry remains functional. Similarly, Type C describes a failure of the lower bridge switch in phase a, with all other external circuits operating normally. While Type B and Type C faults are more likely to occur, they can often escalate into Type A faults if not addressed promptly.

The waveforms of the three types of faults are shown in Figure 2. From the waveforms of the same fault under different power levels, it can be observed that power only affects the amplitude of the current before the fault but does not influence the current waveform after the fault. This is because the post-fault current is determined by the grid voltage and the parasitic resistance of the faulted line, which remain nearly constant across different power levels. As a result, the current waveform is essentially unchanged.

When a short circuit occurs between phases a and b of the grid, the grid current rises sharply, and the grid voltage is correspondingly affected and drops, causing significant impact on the grid, as shown in Figure 2A. Similarly, when a short circuit occurs between phase a and the positive or negative terminal of the DC side, the short-circuit current increases rapidly, also leading to severe impacts on the grid, as shown in Figures 2B, C.

As illustrated in Figure 2, when fault Type A occurs, the current in phase a increases in both forward and reverse directions. However, during faults Type B and Type C, the phase a current exhibits a larger forward or reverse magnitude, resembling the addition of a positive or negative DC offset. In practice, due to the mutual influence of the three-phase currents, a surge in one phase current unbalances the grid, causing increases in the currents of the other phases. Consequently, other switches may also be affected, potentially resulting in cascading faults that lead to more severe failures.

During the operation of electric vehicle charging stations, if faults are not detected and identified in a timely manner, and appropriate safety measures such as shutdowns are not implemented, it may lead to physical damage to the equipment, and even result in safety accidents, endangering users’ personal safety and causing economic losses. Therefore, promptly and quickly identifying faults in charging stations is a critical task of this research. To this end, we collect three-phase alternating current data from the charging station within a small time window, instead of voltage data, as voltage changes are less apparent during fault occurrences. Additionally, constructing fault features with significant discriminative power is another crucial task to ensure that the machine learning model can accurately identify faults. This not only prevents the failure to detect faults but also reduces false positives during normal operation, thereby lowering the maintenance burden.

To achieve fast and accurate fault detection in charging stations, we propose a fault detection method based on spatiotemporal information fusion. The architecture of the proposed fault identification scheme is shown in Figure 3. Since the impact of a fault on voltage data is weak, we only consider the fault features contained in the three-phase AC current. We use a sliding window to obtain input data to achieve real-time fault identification. Based on the input current, we use feature engineering to extract data spatial features and time series features represented by the frequency domain to reduce the difficulty of learning the classification model. For the fault modeling part, the decision tree model is employed to adaptively obtain the relationship between the extracted features and the fault, thereby realizing the fault identification function. The detailed fault identification process is as follows. For the three-phase current data of the charging station, we first extract the operating data within a short time frame using a sliding window. Current methods primarily focus on identifying amplitude characteristics of operating currents while neglecting the extraction of temporal features. Therefore, we use Fourier transform to extract the frequency data of each phase current as the temporal features of faults. Since the amplitude in the high-frequency range is relatively small, we retain only the low-frequency temporal features. Additionally, we calculate the mean value of each phase current as the spatial amplitude information of the fault. Finally, we process the above spatiotemporal features simultaneously using simple machine learning models, such as decision trees, to output the fault type.

To demonstrate the features in the frequency domain that are significantly related to fault types, we visualized the current frequency characteristics for different types of faults and the no-fault condition, as shown in Figure 4. Based on the frequency characteristics under various conditions, the current amplitude in each phase at low frequencies exhibits significant distinction across different states. However, the behavior in the high-frequency range is consistent across these states. Therefore, using low-frequency data as the temporal information for fault detection is a reasonable approach.

Since the fault is introduced at 0.2 s, the input data window containing this time is labeled as fault data, while other data are labeled as no-fault data. The sampling rate is set to 15 kHz, and the output window length is set to 100. This study validates the effectiveness of the proposed method using the following baseline algorithms: Adaptive Boosting (AdaBoost, ABA), SVM, Decision Tree (DT), Multilayer Perceptron (MLP), Gradient Boosting Decision Tree (GBDT), and Random Forest. The number of base classifiers in the Random Forest is set to 5. The GBDT uses a learning rate of 0.1 and has 5 base classifiers. The Multilayer Perceptron uses ReLU as the activation function, with two hidden layers, each containing 64 neurons, and a maximum of 100 iterations. The ABA algorithm also has 5 base classifiers. All other parameters for the baseline methods are set to the default values in the sklearn library.

To objectively evaluate the effectiveness and superiority of the proposed method, this paper introduces the metrics of accuracy, recall, and macro-average F-score (F-macro). The definition of accuracy is given in Equation 1: a c c u r a c y = T P + T N T P + T N + F N + F P (1) where, TP, TN, FN, and FP represent the number of true positives, true negatives, false negatives, and false positives, respectively. The definition of recall is as Equation 2: r e c a l l = T P T P + F N (2)

The F-macro metric provides a comprehensive evaluation of the classification results under imbalanced sample conditions. It is defined as Equation 3: F − m a c r o = 1 C ∑ i = 1 C F 1 i (3)

C is the number of fault categories, and F1 is defined as Equation 4: F 1 = 2 · r e c a l l · p r e c i s i o n r e c a l l + p r e c i s i o n (4)

The average ± standard deviation of the fault detection results over three trials for various baseline methods and the proposed spatiotemporal fusion method are reported in Table 1. Upon comprehensive analysis, the proposed method achieves a significant improvement in accuracy for fault detection in charging stations, with the exception of the ABA method. This improvement is evident not only in the mean values of various metrics but also in the reduction of variance, demonstrating the significant robustness of the proposed method. Specifically, accuracy improvements of 5.56%, 4.15%, 4.75%, 15.58%, and 2.73% were observed for RF, GBDT, SVM, MLP, and DT, respectively. For the decision tree algorithm, its performance is nearly identical to that of the random forest algorithm, which uses multiple base classification trees. Moreover, compared to other algorithms, except for the random forest, it achieved the best performance. Therefore, considering computational cost and real-time requirements, we selected the decision tree as the classifier for fault detection.

To analyze the fault detection algorithms’ performance on different fault types in detail, we visualized the detailed fault classification results using confusion matrices, as shown in Figure 5. According to the confusion matrices for the DT and MLP methods, both with and without the proposed method, the spatiotemporal fusion features improve the identification accuracy for each type of fault. However, various methods show confusion in recognizing Fault A and Fault C. This may be due to the similar characteristics exhibited by the rectifier phase a lower bridge arm short circuit and the grid-side phase a-b short circuit fault.

We used the RF algorithm to output the fault detection results and the importance of various fault features to demonstrate the necessity of including frequency domain features. The importance of each feature is shown in Figure 6. Here, c1 to c3 represent the spatial features of the three-phase currents. c1_f1 to c1_f3, c2_f1 to c2_f3, and c3_f1 to c3_f3 represent the three lowest frequency values in the frequency domain for the three-phase currents. According to the importance results for each type of feature, there are strong correlations between the frequency domain features of each phase current and fault types. Additionally, the spatial current amplitudes also exhibit significant importance.

Due to variations in sensor and data processing equipment across different application scenarios, the input data resolution for models may vary. Therefore, fault detection methods need to be capable of handling different data input window lengths. We collected data with various input window lengths and used DT as a baseline model to validate the adaptability of the proposed method to different data window lengths. The experimental results are shown in Figure 7. Compared to the DT algorithm using only spatial features, the proposed spatiotemporal fusion features provide better adaptability to changes in data window length for the DT algorithm. However, as the window length increases, the performance of the fault detection model decreases, which is due to longer data windows containing more non-fault data within fault samples.