High-resistance fault detection and explainability analysis method for distribution network

Abstract

The present invention provides a method for detecting and interpreting high-resistance faults in a distribution network, and belongs to the field of distribution network fault diagnosis and artificial intelligence deep learning technology. The method decomposes and reconstructs the original transient zero-sequence current signal through an improved adaptive noise complete set empirical mode decomposition to generate a reconstructed transient zero-sequence current signal; a time convolution network model is constructed based on the reconstructed transient zero-sequence current signal to output the fault detection result; a fractionally weighted class activation mapping method is further constructed to carry out qualitative analysis and quantitative calculation on the detection basis of the model, and the interpretability of the closed-loop feedback data-driven fault detection scheme is obtained by characterizing the matching degree between the high-resistance "zero-rest" characteristics and the model decision-making focus area. The present invention adopts the above-mentioned method for detecting and interpreting high-resistance faults in a distribution network, which has high detection accuracy and interpretability under different operating conditions, and is suitable for high-resistance fault detection and analysis in complex distribution network environments.

Description

Technical Field

[0001] The present invention relates to the field of distribution network fault diagnosis and artificial intelligence deep learning technology, and in particular to a distribution network high-resistance fault detection and explainability analysis method. Background Art

[0002] Distribution systems have numerous, complex lines and are located close to the ground. These lines are prone to high-resistance ground faults caused by non-metallic conductive media such as grass, concrete, and branches. Due to the large transition resistance of high-resistance faults and their susceptibility to harmonic interference, their fault signatures are subtle and resemble disturbance signals caused by conventional switching events such as load switching, capacitor switching, and magnetizing inrush current, making them difficult to effectively detect and promptly address. If high-resistance faults persist for a long time, they can damage equipment, cause fires, and even further expand the fault range, resulting in severe losses. Therefore, developing sensitive and reliable high-resistance fault detection schemes is crucial to ensuring the safe and reliable operation of distribution systems.

[0003] Currently, high-resistance grounding fault detection schemes in distribution networks can be primarily categorized into two types based on their feature analysis: indicator threshold methods based on electrical quantity characteristics, and data-driven artificial intelligence methods. Indicator threshold methods quantitatively analyze the differences in electrical information such as voltage and current before and after the fault in the time, frequency, and time-frequency domains to set feature thresholds for detecting high-resistance grounding faults. However, these methods all rely on single or local features to construct criteria, and their thresholds are typically set based on empirical experience. This can lead to classification blind spots in complex operating scenarios, such as increasingly complex system structures and measurement noise interference, and their versatility needs further expansion. In recent years, data-driven artificial intelligence technologies have continued to develop, providing new research approaches for addressing high-resistance fault detection. AI methods, independent of complex mechanism analysis, can fit nonlinear mappings between input samples and output results from massive amounts of data, thereby achieving rapid and accurate fault detection. Researchers have applied AI to high-resistance fault detection, using methods such as support vector machines, artificial neural networks, and convolutional neural networks.

[0004] However, existing approaches generally focus on the application of AI algorithms, often modeled as "black boxes" with low confidence and weak interpretability of decisions. This poses risks when AI algorithms are used for safety-sensitive tasks such as distribution network fault diagnosis. Furthermore, they fail to quantitatively analyze and evaluate power operation and maintenance knowledge from an interpretability perspective, and fail to comprehensively consider multiple quantitative indicators to assess the effectiveness and reliability of model decisions. Overall, while research on interpretability in power AI has been initially explored in various fields, it is still in its infancy in the field of distribution network fault diagnosis. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for high-resistance fault detection and interpretability analysis in distribution networks to overcome the problems in the background technology. The effectiveness of the proposed solution and its applicability in real scenarios have strong application value.

[0006] To achieve the above objectives, the present invention provides a method for detecting and interpreting high-resistance faults in a distribution network, comprising the following steps:

[0007] Step S1, using improved adaptive noise complete set empirical mode decomposition to decompose and reconstruct the original transient zero-sequence current signal, and generate a reconstructed transient zero-sequence current signal;

[0008] Step S2: Based on the reconstructed transient zero-sequence current signal generated in step S1, a time convolution network model is constructed to output a fault detection result;

[0009] Step S3: Based on the fault detection result output in step S2, a score-weighted class activation mapping method is used to generate an attribution heat map, analyze the correlation between the reconstructed transient zero-sequence current signal and the fault detection result, and construct a quantitative evaluation index;

[0010] Step S31, the class activation mapping method generates a class activation map by linearly weighted fusion of feature fusion weights and feature maps;

[0011] Step S32, calculating the attribution value of each sampling point in the reconstructed transient zero-sequence current signal time series;

[0012] Step S33: further constructing a quantitative evaluation index based on the attribution value generated in step S32.

[0013] Preferably, step S1 specifically includes:

[0014] Step S11: adding Gaussian white noise to the original transient zero-sequence current signal, and calculating the residual value and modal component of the first decomposition;

[0015] Step S12: Repeat the step of superimposing Gaussian white noise until further decomposition is impossible, and decompose the original transient zero-sequence current signal into the sum of multiple intrinsic mode function (IMF) components and residuals;

[0016] Components above IMF5 are selected for signal superposition to form the reconstructed transient zero-sequence current signal.

[0017] Preferably, the temporal convolutional network model includes an input layer, multiple TCN modules, a 1×1 convolutional layer, a Flatten layer, a Dense layer, and a Softmax layer connected in sequence.

[0018] Preferably, each TCN module includes: a causal hole convolution layer, a weight normalization layer, a ReLU activation function, and a Dropout unit.

[0019] Preferably, the class activation map is calculated as follows:

[0020]

[0021] Where L represents the class activation map; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; c represents the category of interest; Represents the activation output of the kth channel of the lth convolutional layer in the model.

[0022] Preferably, the steps for constructing the quantitative evaluation index are as follows:

[0023] For a zero-sequence current waveform with T sampling points, define the signal zero-crossing time as t0, take the time interval [t0-ΔT, t0+ΔT] of each ΔT sampling point before and after this time as the analysis window, and define it as the waveform zero-crossing key area set Ω1:

[0024] Ω1={t|t∈[t0-ΔT,t0+ΔT]};

[0025] Calculate the attribution value S(t)∈[0,1] of the feature at the sampling point t in the transient zero-sequence current time series and define the high attribution region set Ω2:

[0026] Ω2={t|S(t)>τ};

[0027] Among them, τ represents the attribution value threshold;

[0028] Define the ZAM indicator for the matching degree between high attribution areas and key areas, and the KAR indicator for the global attribution ratio of key areas:

[0029]

[0030] Here, |·| represents the number of elements in the set.

[0031] Therefore, the present invention adopts the above-mentioned distribution network high-resistance fault detection and interpretability analysis method, and the beneficial technical effects are as follows:

[0032] This method can perform high-precision high-resistance fault detection in the complex operating environment of new power distribution systems and visually analyze the model's decision-making mechanism. This visualization intuitively presents the model's focus on the zero-sequence current sequence, providing qualitative and tuning guidance for hyperparameter selection. Furthermore, combined with quantitative evaluation metrics, it explains the model's reliance on zero-period waveform distortion in the detection results, enhancing the interpretability of the model's decision-making process. This research finding provides technical support for the application of deep learning-based high-resistance fault detection methods in practical systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a 10kV distribution network topology;

[0034] Figure 2 For the Emanuel model;

[0035] Figure 3 Flowchart of the method for high-resistance fault detection and interpretability analysis in distribution network;

[0036] Figure 4 The performance of the model on the training set and test set during training;

[0037] Figure 5 is the confusion matrix of the model on the training set and test set; Figure 5 (a) is the confusion matrix on the training set; Figure 5 (b) is the confusion matrix on the test set;

[0038] Figure 6 is the t-SNE visualization result; among them, Figure 6 (a) is the visualization distribution result of the original data set after dimensionality reduction by the t-SNE algorithm; Figure 6 (b) shows the visual distribution of the dataset processed by one TCN module and reduced in dimension using the t-SNE algorithm; Figure 6 (c) shows the visual distribution of the dataset processed by two TCN modules after dimensionality reduction using the t-SNE algorithm; Figure 6 (d) is the visualization distribution result of the dataset processed by the three TCN modules and reduced in dimension by the t-SNE algorithm;

[0039] Figure 7 In order to use Score-CAM to analyze the decision mechanism of samples under different operating conditions, Figure 7 (a) is the attribution heat map of the high resistance fault sample; Figure 7 (b) is the attribution heat map of the capacitor switching sample; Figure 7 (c) in the figure is the attribution heat map of the load shedding sample; Figure 7 (d) is the attribution heat map of the excitation inrush current sample;

[0040] Figure 8 is the decision basis for different convolution kernel sizes; among them, Figure 8 (a) is the attribution heat map when the convolution kernel size is 1; Figure 8 (b) is the attribution heat map when the convolution kernel size is 3; Figure 8 (c) in the figure is the attribution heat map when the convolution kernel size is 5;

[0041] Figure 9 It is the residual block structure in TCN. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0043] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0044] Example 1

[0045] like Figure 3 As shown in the figure, the flow chart of the distribution network high-resistance fault detection and interpretability analysis method includes four stages: data preprocessing, model training and parameter tuning, practical application and interpretability analysis.

[0046] Data preprocessing stage: The original transient zero-sequence current signal under high-resistance fault and normal disturbance conditions is collected using a measurement device. Random noise is added to simulate the complex operating environment of the distribution system. The noisy original transient zero-sequence current signal is decomposed and reconstructed using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) algorithm. Considering the uncertainty of the actual system fault occurrence time and the action delay of the recording device startup, the reconstructed transient zero-sequence current signal is processed using sliding window segmentation (the time window used in this embodiment is 40ms), and the segmented transient zero-sequence current signal x is normalized using formula (1):

[0047]

[0048] Among them, x max and x min They represent the maximum and minimum values ​​of the transient zero-sequence current signal x before normalization; x i and x i′ represents the value of the i-th element in the signals x and x' before and after normalization, respectively. This operation can reduce the magnitude differences in electrical quantities, allowing the model to focus more on data trends and improve model performance.

[0049] Unlike traditional empirical mode decomposition (EMD) methods, ICEEMDAN improves the stability and accuracy of the decomposition process by optimizing the noise addition and decomposition process. It can effectively resolve modal aliasing while reducing the residual noise of the intrinsic mode components and further highlighting the time and frequency domain characteristics of each component. The present invention uses the ICEEMDAN method to decompose and reconstruct the noisy transient zero-sequence current signal. The steps are as follows:

[0050] Step S11: Define the transient zero-sequence current signal to be decomposed as i0, E K (·) represents the K-order intrinsic mode function (IMF) after EMD decomposition. Gaussian white noise is added to i0, and the residual value R1 and the first modal component D1 of the first decomposition are calculated:

[0051]

[0052] D1=i0-R1(4);

[0053] Among them, ε represents the signal-to-noise ratio of Gaussian white noise, w (n) It represents the nth group of Gaussian white noise added, and N represents the total number of times added.

[0054] Step S12: Repeat the noise superposition step in step S11 to calculate the residual value R after the j-th decomposition. j and modal component D j Through the above steps, the transient zero-sequence current signal i0 to be decomposed is decomposed into the sum of multiple IMF components and residuals.

[0055] Spectral analysis of each decomposed IMF component reveals that some contain signal characteristics characteristic of measurement noise, which are weakly correlated with the original transient zero-sequence current signal. Therefore, the present invention selects components above IMF5 for signal superposition to form a reconstructed signal, effectively filtering out noise interference while fully preserving the fault characteristics.

[0056] Model training and parameter tuning: The preprocessed dataset is divided into training and test sets in a 4:1 ratio. A temporal convolutional network model is constructed, and the training set samples are fed into the model for model training and hyperparameter tuning. Cross entropy is used as the loss function, and minimized using the Adam optimizer. When the loss function and accuracy stabilize, the model weights are saved for subsequent testing and application analysis.

[0057] like Figure 9 As shown in Figure 2, the temporal convolutional network model includes the following modules connected in sequence:

[0058] The input layer is used to receive the reconstructed transient zero-sequence current signal;

[0059] Multiple TCN modules;

[0060] 1×1 convolutional layer, used to adjust the feature dimension;

[0061] Flatten layer, used to flatten multidimensional features;

[0062] Dense layer, used for fully connected operations;

[0063] Softmax layer, used to output classification probability

[0064] Each TCN module includes:

[0065] Causal atrous convolutional layer, used to extract time series features;

[0066] Weight normalization layer, used to stabilize the training process;

[0067] ReLU activation function, used to introduce nonlinearity;

[0068] Dropout unit is used to prevent overfitting.

[0069] Practical application stage: The transient zero-sequence current signal to be detected is collected, and after data preprocessing, it is input into the trained and saved model to output the classification result.

[0070] Explainability Analysis: The Score-CAM method is used to analyze the model's decision results. The attributed value of the features at each sampling point in the transient zero-sequence current time series is calculated to further clarify which parts of the time series dominate or influence the sample classification results. Visualization is used to explain the model's decision-making basis and the impact of hyperparameter settings on the classification results, thereby intuitively reflecting the inherent relationship between input features and output categories.

[0071] In practical applications, appropriate thresholds can be set for ZAM and KAR to quantify the model's focus on the zero-point distortion characteristics of high-resistance faults and assess the reliability of the model's classification results. By comparing the model output with the attribution features from interpretable analysis, the controllability and reliability of the fault detection process can be enhanced. Specifically, if the model outputs a high-resistance fault classification result, but both the ZAM and KAR metrics for the sample are below the set thresholds, this indicates that the model's decision is not primarily based on features in the key areas, indicating insufficient attention to the core characteristics of high-resistance faults, potentially leading to a risk of misclassification. If the model outputs a non-high-resistance fault classification result, but its metrics are above the set thresholds, this indicates that the model has accurately focused on the zero-point distortion characteristics, but the final classification result is not a high-resistance fault, potentially indicating a risk of missed detection. In these cases, a manual intervention mechanism can be triggered, with technical personnel conducting an in-depth review of the entire fault diagnosis process to further analyze the model's judgment basis and identify potential sources of problems. This approach effectively reduces the risk of both false and missed detections, improving the accuracy and reliability of high-resistance fault detection.

[0072] Score-CAM uses the linear weighted fusion of feature fusion weights and feature maps to generate class activation maps to achieve a more stable interpretation effect. For a given temporal convolutional network model, let its input and output be x and Y respectively, satisfying Y = f(x). Select the kth channel of the lth convolutional layer, and the corresponding activation is For a known input x i , Contribution to output Y It can be defined as:

[0073]

[0074] in, represents the Hadamard product; is a with x i Vectors of the same shape are defined as follows:

[0075]

[0076] Among them, the function s[.] represents the normalization operation, mapping each element to the interval [0, 1]; the function Up(.) represents the normalization operation. Upsampling.

[0077] Select the class c of interest, and the calculation formula of the class activation map of Score-CAM can be defined as:

[0078]

[0079] Where L represents the class activation map; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; Represents the activation output of the kth channel of the lth convolutional layer in the model;

[0080] The weight of each channel It can determine the specific category information contained in the class activation map; ReLU(·) is the activation function used to remove neurons that are useless for the category c of interest.

[0081] The present invention uses Score-CAM to analyze the decision results of the fault detection model based on the time convolution network, and calculates the attribution value of the characteristics at each sampling point in the transient zero-sequence current time series using equations (5)-(7). This allows the correlation between the input samples and the detection results to be analyzed.

[0082] The class activation maps generated by Score-CAM provide a stable visualization of the key areas of focus of the model during classification decisions, thereby providing guidance for model hyperparameter tuning. Building on this qualitative analysis, the present invention further constructs quantitative evaluation metrics based on the generated attribution results, providing more intuitive decision support for operations and maintenance personnel.

[0083] First, for the transient zero-sequence current waveform containing T sampling points, the signal zero-crossing moment (the time point when the current sign of the adjacent sampling points changes) is defined as t0, and the time interval of each ΔT sampling point before and after this moment [t0-ΔT, t0+ΔT] is taken as the analysis window, and it is defined as the waveform zero-crossing key area set Ω1 as shown in formula (8):

[0084] Ω1={t|t∈[t0-ΔT,t0+ΔT]}(8);

[0085] The attribution value S(t)∈[0,1] of the feature at the sampling point t in the transient zero-sequence current time series is calculated by equations (5)-(7), and the high attribution region set Ω2 is defined as equation (9):

[0086] Ω2={t|S(t)>τ} (9);

[0087] Here, τ is the attribution value threshold, which is set to 0.6 in this embodiment.

[0088] The matching index between high-attribution areas and key areas (ZAM) and the global proportion index of key area attribution (KAR) are further defined as follows:

[0089]

[0090] Here, |·| represents the number of elements in the set.

[0091] Analyzing equations (10) and (11), we can see that ZAM measures whether the model accurately focuses on the distortion features near the zero-crossing point of the waveform by evaluating whether the high-contribution attribution value points are concentrated in the key area. The larger the ZAM value, the more the model's high-contribution attribution value points are distributed in the key area, indicating that the model has extracted more discriminative features in this area. Conversely, if the ZAM value is low, it means that the model's high-contribution attribution value points are relatively scattered, and it may rely on information from other areas for classification, reducing its dependence on the key features of the zero-crossing point. KAR reflects the proportion of the key area attribution value in the global attribution value and measures the degree of dependence of the model on the zero-crossing key area in the overall decision-making process. The larger the KAR value, the more the model relies on the waveform features of the zero-crossing key area in the classification decision, rather than the noise or redundant information in other non-key areas, thus reflecting the global importance of the model to the key area features.

[0092] Overall, KAR reflects the proportion of key area features in the overall attribution value, measuring the model's global focus, while ZAM reflects whether the model accurately focuses on high-contribution points in the key area, measuring its ability to capture local features. Combining these two can effectively evaluate the model's attribution feature distribution and decision-making basis in high-resistance fault detection.

[0093] The present invention will be further described below through simulation experiments.

[0094] Build in MATLAB / Simulink Figure 1 The 10kV distribution network simulation model shown in the figure is used to simulate high-resistance faults and disturbances and conduct subsequent analysis. The model includes power source G, transformer (using DYn11 connection, D on the high voltage side represents delta connection, Y on the low voltage side represents star connection), line (L1~L 10 ), loads (grid-connected via transformers), and distributed generation (DG1-DG3). The transformer capacity is 250MVA, with a transformation ratio of 110kV / 10.5kV. The neutral point is grounded using an arc suppression coil (R and L represent the arc suppression coil's resistance and inductance, respectively, with an 8% compensation). The line lengths are marked in the diagram, and the line parameters are shown in Table 1. Transient zero-sequence current signals are acquired at the monitoring points using a μPMU with a 10kHz sampling frequency.

[0095] Table 1 Line parameters

[0096]

[0097] Based on the above power distribution system simulation, a simulation sample library can be established, and the parameters are shown in Table 2. Among them, the high resistance fault is as follows Figure 2 Emanuel model simulation shown. Figure 2 Middle,U pand U n It is a DC voltage source with a ±10% fluctuation, which is used to simulate the asymmetry and nonlinearity of arc voltage and fault current; R p and R n is a time-varying resistance used to simulate the fault arc resistance; D p and D n is an ideal diode, and U p and U n Together they constitute the positive and negative half-cycle current paths of the circuit.

[0098] Table 2 Sample parameters

[0099]

[0100] By varying the access location, initial phase angle, and sample parameters, simulation samples can be fully captured for conditions with varying transition resistances, line types, and locations. Data within a 0.2s time window before and after the switching event (including two cycles before the fault and eight cycles after the fault) is uniformly captured and segmented into 40ms time windows. After decomposition and reconstruction using the ICEEMDAN method, labels 0, 1, and 2 are assigned based on the simulation settings, corresponding to normal, disturbance (including capacitor switching, load switching, and magnetizing inrush current), and high-resistance faults, respectively. Finally, the dataset is divided into training and test sets in a 4:1 ratio. Given the imbalance between training and test samples, solely evaluating model performance based on overall accuracy is inherently biased. Therefore, precision, recall, and F1 score are supplemented for a comprehensive evaluation of model performance. The F1 score is used to assess the classification effectiveness of the fault location model.

[0101] pass Figure 4 The performance of the model on the training set and test set during training is shown in the figure. The horizontal axis is the number of iterations, which is 50 rounds in total; the vertical axis is the F1 score and loss. Figure 4 It can be observed that the loss curve drops significantly in the early stage of training, and stabilizes after 40 iterations. To ensure the stability of the model effect, this embodiment selects the model saved at the 48th iteration as the final model for subsequent testing. The confusion matrix of the model on the training set and test set is exported as follows: Figure 5 As shown in the figure, the model achieved an overall accuracy of 96% on the test set, a precision of 94.75%, a recall of 96.01%, and an F1 score of 95.32%. This demonstrates that the model maintains a high level of generalization under different operating conditions and does not misjudge or miss high-resistance faults, preliminarily verifying the effectiveness of the proposed improved resistance fault detection scheme.

[0102] To further verify the effectiveness of the proposed scheme, the t-distributed random neighbor embedding algorithm is used to reduce the dimension of the original data and the data processed by different TCN modules. The results are shown in the figure below. Figure 6 Different colors represent different categories; the coordinate axis only represents the distribution of data points on the two-dimensional plane after data dimensionality reduction and is dimensionless. Figure 6 (a) shows that the original samples show considerable disorder in both feature dimensions and distribution patterns, randomly scattered in the low-dimensional visualization space, and lack obvious aggregation trends. However, as the depth of the feature extraction layer increases (e.g. Figure 6 As shown in (b)-(d) in Figure 3, the sample points begin to cluster. The distances between sample points of the same class gradually decrease, forming a distinct cluster structure. At the same time, the distances between different classes increase significantly, gradually establishing clear classification boundaries. This transformation process clearly demonstrates the increasing clustering effect of the samples, demonstrating that the model designed in this paper can fully exploit the implicit features in transient zero-sequence current time series, demonstrating superior feature extraction and classification capabilities in fault detection tasks.

[0103] The proposed method was compared with common fault detection models to fully validate its advantages, including model classification performance and test time. Considering the randomness inherent in the model training and testing process, the model was tested repeatedly and the averaged values ​​were calculated. The results are shown in Table 3. The test time is the average computation time for the test set samples. The results demonstrate that the proposed method offers substantial advantages in both computational efficiency and accuracy compared to methods based on Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs). Furthermore, while the CNN model exhibits superior performance in terms of test time, the TCN model demonstrates significant superiority in terms of classification performance. It is worth noting that, when considering the safe and reliable operation of actual distribution networks, the consequences of missed and false detections are more serious than those of delayed detection. Therefore, sacrificing a certain amount of test time to improve fault detection accuracy is still reasonable.

[0104] Table 3 Comparison of the effects of different classification methods

[0105] plan Accuracy Accuracy Recall F1 score Testing time Support Vector Machine 90.86％ 93.09％ 85.30％ 87.41％ 1.5ms ANN 88.29％ 93.20％ 80.48％ 82.41％ 3ms CNN 91.14％ 94.59％ 85.24％ 87.58％ 0.6ms Method of the present invention 96.00％ 94.75％ 96.01％ 95.32％ 0.9ms

[0106] The Score-CAM algorithm is used to analyze the operating mechanism of the temporal convolutional network. Specifically, for the reconstructed transient zero-sequence current signal, the contribution of each time series segment to the classification result is calculated. Based on this, an attribution heatmap is constructed to clearly identify the characteristic segments that play a key role in the decision-making outcome, thereby helping operation and maintenance personnel understand the decision-making basis of the model. In the attribution heatmap, the attribution value range is in the (0, 1) interval. Larger attribution values ​​indicate that the sampling point has a more significant impact on the model's decision, and the size of the attribution value is represented by different colors. Based on this criterion, the impact of each segment of the transient zero-sequence current time series on the model's decision-making results can be qualitatively evaluated. Furthermore, a quantitative analysis is performed in conjunction with the ZAM and KAR methods to quantitatively assess the model's focus on the waveform's zero-crossing intervals, thereby analyzing the differences in attribution patterns among different sample types.

[0107] Figure 7 The operation mechanism of the TCN model for fault classification through Score-CAM is demonstrated. It can be seen that the key factor affecting the model classification result is not the peak or trough area, but the different degrees of distortion near the zero crossing point of the waveform. Figure 7 In the high-resistance fault condition shown in Figure (a), the transient zero-sequence current exhibits a horizontal trend for a period of time after crossing zero and then returns to a sinusoidal state. This zero-sequence current characteristic is effectively detected by the model and serves as a key foundation for high-resistance fault detection. This demonstrates that the TCN model can capture key waveform features of different sample types, providing a visual basis for the model to make appropriate category decisions.

[0108] From the perspective of quantitative analysis, Figure 7 For the high-resistance fault sample shown in (a), the values ​​of the indicators ZAM and KAR are 0.69 and 0.7 respectively; Figure 7 For the capacitor switching sample shown in (b), the values ​​of the indicators ZAM and KAR are 0.61 and 0.61 respectively; Figure 7 For the load shedding sample shown in (c), the values ​​of the indicators ZAM and KAR are 0.59 and 0.63 respectively; Figure 7 For the magnetizing inrush current sample shown in (d), the values ​​of the indicators ZAM and KAR are 0.58 and 0.56 respectively. It can be seen that the indicators of the high-resistance fault sample are significantly higher than those of other disturbance samples, indicating that the high contribution attribution value of the model is mainly concentrated in the zero-crossing area of ​​the waveform, which is consistent with the electrical characteristics of the high-resistance fault, indicating that the attribution results of Score-CAM have a certain physical interpretability. In contrast, the two indicators of the disturbance sample are lower than those of HIF (high-resistance fault), and the indicator of the IC sample (magnetizing inrush current sample) is the smallest. Figure 7As can be seen from (d) in the figure, the high attribution value areas of the excitation inrush current samples are relatively evenly distributed on the entire time axis, indicating that the model's classification of such disturbance samples relies more on global features and is not limited to the zero-crossing area of ​​the waveform.

[0109] Overall, Figure 7 The ZAM and KAR metrics also show a strong positive correlation. Specifically, when the proportion of attribution values ​​within the zero-crossing region of the waveform is high, the high-contribution attribution points in the classification output tend to be more concentrated in the high-contribution region. Therefore, these quantitative metrics collectively reflect the model's focus on key areas during classification decisions. They can serve as important quantitative indicators for measuring the rationality of the model's decision-making basis and provide more intuitive and reliable decision-making support for operations and maintenance personnel.

[0110] Under normal circumstances, the selection of model hyperparameters (such as convolution kernel size, expansion coefficient, etc.) is blind due to the lack of clear theoretical guidance, and researchers mostly rely on continuous combination and trial and error to determine suitable hyperparameters. This process requires a large amount of computing resources and time cost, and it is difficult to obtain the optimal parameter combination. The interpretability method used in the present invention can intuitively analyze the inherent correlation between hyperparameters and model performance, while effectively avoiding blind trial and error, and provides a clear and traceable basis for the reasonable selection of hyperparameters, further improving model optimization efficiency.

[0111] During the optimization process for the convolution kernel size, we found that models with different kernel sizes misclassified high-resistance faults as disturbances. For a specific high-resistance fault sample, models with kernel sizes of 1 and 5 misclassified it, while a model with a kernel size of 3 correctly classified it. Figure 8 The visualization analysis results of the sample diagnosis results of the models trained with the above three different convolution kernel sizes are shown. Figure 8 As can be seen, as the convolution kernel size gradually increases, the model can extract features over longer time series segments. If the convolution kernel is too small, only local features in the sample are captured, resulting in excessive fragmentation. However, when the convolution kernel is set too large, it is prone to capturing interfering features or shifting the model's focus. When the convolution kernel size is set to 3, the model accurately covers the zero-crossing distortion region that determines the sample classification as a high-resistance fault, thereby providing a correct classification result.

[0112] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0113] Therefore, the present invention adopts the above-mentioned distribution network high-resistance fault detection and explainability analysis method, which has high detection accuracy and explainability under different operating conditions and is suitable for high-resistance fault detection and analysis in complex distribution network environments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting and interpreting high-resistance faults in a distribution network, characterized in that: The following steps are involved: Step S1, using improved adaptive noise complete set empirical mode decomposition to decompose and reconstruct the original transient zero-sequence current signal, and generate a reconstructed transient zero-sequence current signal; Step S2: Based on the reconstructed transient zero-sequence current signal generated in step S1, a time convolution network model is constructed to output a fault detection result; Step S3: Based on the fault detection result output in step S2, a score-weighted class activation mapping method is used to generate an attribution heat map, analyze the correlation between the reconstructed transient zero-sequence current signal and the fault detection result, and construct a quantitative evaluation index; Step S31, the class activation mapping method generates a class activation map by linearly weighted fusion of feature fusion weights and feature maps; Step S32, calculating the attribution value of each sampling point in the reconstructed transient zero-sequence current signal time series; Step S33: further constructing a quantitative evaluation index based on the attribution value generated in step S32.

2. The method for detecting and interpreting high-resistance faults in a distribution network according to claim 1, wherein: Step S1 specifically includes: Step S11: adding Gaussian white noise to the original transient zero-sequence current signal, and calculating the residual value and modal component of the first decomposition; Step S12: Repeat the step of superimposing Gaussian white noise until further decomposition is impossible, and decompose the original transient zero-sequence current signal into the sum of multiple intrinsic mode function (IMF) components and residuals; Components above IMF5 are selected for signal superposition to form the reconstructed transient zero-sequence current signal.

3. The method for detecting and interpreting high-resistance faults in a distribution network according to claim 1, wherein: The temporal convolutional network model includes an input layer, multiple TCN modules, a 1×1 convolutional layer, a Flatten layer, a Dense layer, and a Softmax layer connected in sequence.

4. The method for detecting and interpreting high-resistance faults in a distribution network according to claim 1, wherein: Each TCN module includes: causal hole convolution layer, weight normalization layer, ReLU activation function, and Dropout unit.

5. The method for detecting and interpreting high-resistance faults in a distribution network according to claim 1, wherein: The class activation map is calculated as follows: Where L represents the class activation map; ReLU represents the activation function; represents the channel weight; k represents the index of the channel; c represents the category of interest; Represents the activation output of the kth channel of the lth convolutional layer in the model.

6. The method for detecting and interpreting high-resistance faults in a distribution network according to claim 1, wherein: The steps to construct quantitative evaluation indicators are as follows: For a zero-sequence current waveform with T sampling points, define the signal zero-crossing time as t0, take the time interval [t0-ΔT, t0+ΔT] of each ΔT sampling point before and after this time as the analysis window, and define it as the waveform zero-crossing key area set Ω1: Ω1={t|t∈[t0-ΔT,t0+ΔT]}; Calculate the attribution value S(t)∈[0,1] of the feature at the sampling point t in the transient zero-sequence current time series and define the high attribution region set Ω2: Ω2={t|S(t)>τ}; Among them, τ represents the attribution value threshold; Define the ZAM indicator for the matching degree between high attribution areas and key areas, and the KAR indicator for the global attribution ratio of key areas: Here, |·| represents the number of elements in the set.

Images