Ground fault diagnosis method based on intelligent timing feature extraction

Abstract

The application relates to the technical field of intelligent monitoring and protection of power systems, and discloses a grounding fault diagnosis method based on intelligent time sequence feature extraction. In the feature extraction stage, a multi-scale convolutional neural network is adopted, combined with multi-scale permutation entropy, to perform coarse-grained processing on an original signal, calculate permutation entropy values under different scales, quantify the spatial distribution characteristics of signal complexity, and inhibit noise interference. After feature extraction, the features extracted by the multi-scale convolutional neural network and the multi-scale permutation entropy two channels are spliced through a feature fusion layer, and the fused features are input into a BiGRU module. In view of the problems that the traditional fault diagnosis method has insufficient multi-scale feature extraction, weak anti-noise ability and low time sequence modeling precision under complex working conditions, the application fuses multi-scale feature extraction and dynamic time sequence features, realizes high-precision diagnosis of grounding faults, and significantly improves the precision and robustness of grounding fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and protection technology of power systems, specifically relating to a ground fault diagnosis method based on intelligent time-series feature extraction. Background Technology

[0002] With the continuous advancement of urbanization, electrified rail transit systems have been widely used in cities around the world due to their high capacity, high passenger capacity, optimized energy utilization, ability to alleviate urban traffic congestion, reduction of air pollution, and high transportation efficiency. In urban rail transit systems, the traction power supply system is a crucial component, ensuring normal train operation by providing stable power support to trains and other equipment. Currently, urban rail transit systems primarily employ DC traction power supply, where the train's traction current flows back to the traction substation through the rails, guaranteeing power transmission and usage.

[0003] However, as the return path for electric current, the rails possess longitudinal resistance and are not completely insulated from the ground, resulting in a potential relative to ground. Some rail current leaks into the ground through the rail fasteners, forming stray currents. These stray currents not only corrode the rails themselves, the reinforcing steel in tunnels, and buried pipelines, but can also interfere with the normal operation of rail transit equipment and even threaten the stability of the power supply system. The flow of stray currents between the rails and the ground can cause equipment damage, increase maintenance costs, and even affect the safe operation of trains.

[0004] Furthermore, due to the spatial overlap between urban rail transit systems and the power grid, stray currents sometimes flow into the grounding grid of the AC power grid. This process can create potential differences between substations, further exacerbating interference with the AC power grid system and potentially leading to power system instability. With the expansion of urban rail transit systems and the increasing complexity of power system structures and operating modes, power system faults occur frequently, posing challenges to power security and supply stability.

[0005] Against this backdrop, effectively identifying and diagnosing faults in power systems has become a crucial and urgent problem to be solved. While existing fault detection and diagnosis methods can identify problems to some extent, they still have certain limitations:

[0006] (1) The traditional threshold method relies on fixed parameter settings, which is not adaptable to complex working conditions (such as high impedance grounding and noise interference), resulting in a high false negative rate.

[0007] (2) A single CNN model is difficult to capture the multi-scale features and time series dependencies of fault signals at the same time. Shallow networks have limited feature extraction capabilities and are difficult to judge complex power system faults in real time and accurately.

[0008] (3) Traditional feature engineering methods (such as wavelet transform) require manual design of features, have insufficient generalization ability, and are difficult to deal with nonlinear fault signals.

[0009] With the continuous changes in the topology of power systems, traditional methods alone are no longer sufficient to meet the increasingly complex fault diagnosis needs.

[0010] Based on this, a grounding fault diagnosis method based on intelligent timing feature extraction is provided to address the technical deficiencies mentioned in the background art. Summary of the Invention

[0011] To address the aforementioned technical problems, this invention provides a grounding fault diagnosis method based on intelligent time-series feature extraction, thereby improving the diagnostic accuracy and robustness in complex scenarios.

[0012] This invention provides a ground fault diagnosis method based on intelligent timing feature extraction, comprising the following steps:

[0013] Step 1: Establish the structure of the DC traction power supply system, build a train dynamic simulation platform, obtain track potential distribution data during train operation, and construct a sample dataset;

[0014] Step 2: Construct three parallel convolutional layers based on a multi-scale convolutional neural network, using convolutional kernels of different sizes to extract multi-scale features from the orbital potential distribution data. Then, through a feature fusion module, channel-by-channel splicing and normalization of feature maps at different scales are performed to generate a fused feature vector with multi-resolution information.

[0015] Step 3: Introduce multi-scale permutation entropy to extract the features of the dynamic orbit potential time series signal in the sample dataset, and quantify the dynamic change law of the dynamic orbit potential time series signal at different time scales;

[0016] Step 4: The features extracted from the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are concatenated through a feature fusion layer, and the fused features are input into the bidirectional GRU module to achieve feature fusion.

[0017] Step 5: Process the feature vector output by the bidirectional GRU module through a fully connected layer to identify and classify the fault diagnosis type, and output the final fault diagnosis result.

[0018] Specifically, in step 1, a distributed parameter model is constructed based on the “running rail-drainage network-ground” structure of the rail transit traction power supply system.

[0019] Specifically, in the distributed parameter model, i r (x) represents the rail current at point x, i d(x) represents the drain network current at point x, u r (x) represents the rail-to-ground potential at point x, u d (x) represents the ground potential of the drainage network at point x, and the voltage and current satisfy the following formula:

[0020]

[0021] In the distributed parameter model, the current and voltage also satisfy the following formulas:

[0022] I r =I r1 +y r1 (U r1 -U d1 )

[0023] I d =I d1 +y d1 U d1 -y r1 (U r1 -U d1 )

[0024] U r2 =U r1 -Z r1 I r

[0025] U d2 =U d1 -Z d1 I d

[0026] I r2 =I r +y r1 (U r2 -U d2 )

[0027] The parameter relationships between the parameters in the fractional parameter model and the distribution parameter model are obtained as follows:

[0028]

[0029] Specifically, step 2 includes the following steps:

[0030] Step 201: After processing the experimental data obtained by the train dynamic simulation platform, a dataset is obtained. The dataset is divided into a test set and a training set according to a preset ratio, and the data is trained and analyzed.

[0031] Step 202, the convolution kernel calculation process is as follows:

[0032]

[0033] In the formula: x l j M is the output of the j-th channel of the l-th convolutional layer. j For the input features, ω l ij It is the weight matrix corresponding to the convolution kernel, b l j σ is the bias matrix, σ(·) is the activation function, and * is the symbol for convolution operation;

[0034] Step 203: The output of the convolutional layer uses the non-linear activation function ReLU, and the calculation process is as follows:

[0035] ReLU(x) = max(0,x);

[0036] Step 204: The pooling layer is located after the convolutional layer, and the calculation process is as follows:

[0037] z i =down(x,y)[i]

[0038] In the formula, z i The output of the pooling layer is y, x is the input of the pooling operation, down(x,y) is the downsampling function, and i represents the i-th element in the pooling layer.

[0039] The pooling method used is max pooling, and the calculation process is as follows:

[0040]

[0041] In the formula: OUTPUT represents the feature after pooling, α j This indicates the size of the pooling region.

[0042] Specifically, in step 3, the input one-dimensional time-series signal is coarsened at multiple scales, and the permutation entropy of the coarsened subsequence is calculated to obtain multi-scale permutation entropy data. Feature information in the multi-scale permutation entropy data is extracted and flattened into a one-dimensional vector.

[0043] Specifically, the coarsening step is as follows:

[0044] Step 31: For the original time signal X = {x1, x2, ..., xn} of length N... N} Perform coarse-graining treatment, that is:

[0045]

[0046] In the formula: s is the scale factor, i = (j-1)s+1;

[0047] Step 32: Process the coarse-grained sequence y obtained in step 31.s j Refactoring:

[0048] Y s j ={y s k ,y s k+τ ,…,y s k+(m-1)τ}

[0049] In the formula: k is the k-th reconstructed component, and the total number of reconstructed components is... m is the embedding dimension; τ is the delay time;

[0050] Step 33: Sort the sequence obtained in step 32:

[0051]

[0052] A set of symbol sequences S is obtained r ={j1,j2,…,j m} Calculate the probability p of each symbol sequence appearing. r ;

[0053] Step 34: Based on the probability p calculated in step 33... r Calculate the permutation entropy of the coarse-grained sequence:

[0054]

[0055] Specifically, in step 5, the bidirectional GRU module fuses the multi-scale convolutional features and temporal features extracted from the multi-scale convolutional neural network and the multi-scale permutation entropy, outputs joint features, processes the features using a bidirectional gated recurrent unit, captures the dynamic dependencies between consecutive time steps, and inputs the output obtained from the bidirectional gated recurrent unit into a fully connected layer. After processing by a nonlinear activation function, it is used for final fault diagnosis and classification.

[0056] Specifically, the calculation process of the bidirectional GRU module is as follows:

[0057] Step 51: Input the current time feature a t and the hidden state h from the previous moment t-1 The update and reset gates are calculated using the σ(·) activation function to control the output range. The formula is as follows:

[0058] z t =σ(W z1 a t +W z2 h t-1 )

[0059] r t =σ(W r1 a t +W r2 h t-1 )

[0060] Among them, W z1 W z2 To update the gate weight matrix, W r1 W r2 To reset the gate weight matrix, z t To update the gate output, r t To reset the gate output;

[0061] Step 52: Input the reset door history status r t ⊙h t-1 and input features a t Generate intermediate hidden states Using the tanh(·) activation function, the formula is:

[0062]

[0063] in, In the intermediate hidden state, W h In the middle hidden state The weight matrix;

[0064] Step 53: Input update gate z t The hidden state h from the previous moment t-1 and the intermediate hidden state Output the final hidden state h t The formula is:

[0065]

[0066] Specifically, the bidirectional gated loop unit is composed of a superposition of forward and reverse GRUs, as shown in the following formula:

[0067]

[0068] In the formula: U t1 U is the weight matrix for forward propagation. t2 Let b be the weight matrix for backpropagation. t For bias.

[0069] Specifically, in step 4, during feature map stitching, the feature maps of the multi-sensor signals in the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are fused using the Concatenate function instruction.

[0070] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows:

[0071] 1. The MSCNN in this invention utilizes the advantages of convolutional neural networks in spatial locality to capture local features in signals, and MPE can extract dynamic features of signals by quantizing the complexity of time series. The combination of the two can effectively make up for the problem of insufficient feature expression ability of single modes, thereby improving the overall feature extraction ability and the accuracy of fault diagnosis.

[0072] 2. The MPE in this invention can perform downsampling and smoothing of the original signal through coarse-grained processing, effectively suppressing noise interference and exhibiting good accuracy and robustness. At the same time, the BiGRU module captures both forward and backward features through the superposition of forward GRU and backward GRU, which not only simplifies the network parameters but also solves the problem of insufficient information utilization. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0074] Figure 1 This is a flowchart of the present invention;

[0075] Figure 2 This is a diagram of the overall architecture of the present invention;

[0076] Figure 3 This is a diagram of the train dynamic simulation model of the present invention;

[0077] Figure 4 This is a flowchart of the MPE calculation process of the present invention;

[0078] Figure 5 This is a schematic diagram of the BiGRU structure of the present invention;

[0079] Figure 6 This is the train operation diagram for this invention;

[0080] Figure 7 This is a comparison chart of the experimental results of this invention;

[0081] Figure 8 This is a schematic diagram of the first type of confusion matrix Laplace-BiGRU of the present invention;

[0082] Figure 9 This is a schematic diagram of the second type of confusion matrix CNN-BiGRU of the present invention;

[0083] Figure 10This is a schematic diagram of the third type of confusion matrix MSCNN-BiGRU of the present invention;

[0084] Figure 11 This is a schematic diagram of the fourth type of confusion matrix MSCNN-MPE-BiGRU of the present invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are merely illustrative of the invention and represent only a portion, not all, of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0086] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0087] First Embodiment

[0088] This embodiment discloses a ground fault diagnosis method based on intelligent temporal feature extraction, specifically a ground fault diagnosis method based on the fusion of multi-scale convolutional neural network (MSCNN), multi-scale permutation entropy (MPE), and bidirectional gated cyclic unit (BiGRU).

[0089] Figure 1 The diagram shown is a flowchart of an embodiment of the present invention, which includes the following steps:

[0090] Step 1: Establish the DC traction power supply system structure, build a train dynamic simulation platform, acquire track potential distribution data during train operation, construct a sample dataset, and build a distributed parameter model based on the "running rail-drainage network-ground" structure of the rail transit traction power supply system, such as... Figure 3 As shown, the parameters in the distributed parameter model are shown in Table 1:

[0091] Table 1 Distribution Parameters Model Parameters

[0092]

[0093]

[0094] for Figure 3 The distributed parameter model, i r (x) represents the rail current at point x, i d (x) represents the drain network current at point x, u r (x) represents the rail-to-ground potential at point x, u d (x) represents the ground potential of the drainage network at point x, and the voltage and current satisfy the following formula:

[0095]

[0096] When x = L + ΔL:

[0097] I r2 =i r (L+ΔL)

[0098] but:

[0099] I r2 =C1U r1 +C2U d1 +C3I r1 +C4I d1

[0100] C1 to C4 depend on ΔL and the parameters in the distribution parameter model, and are independent of the location L of the interval.

[0101] Combination Figure 3 The lumped parameter model is shown in Table 2. Within the L~ΔL region, the current and voltage satisfy:

[0102] Table 2 Parameters of the lumped parameter model

[0103]

[0104]

[0105] I r =I r1 +y r1 (U r1 -U d1 )

[0106] I d =I d1 +y d1 U d1 -y r1 (U r1 -U d1 )

[0107] U r2 =Ur1 -Z r1 I r

[0108] U d2 =U d1 -Z d1 I d

[0109] I r2 =I r +y r1 (U r2 -U d2 )I r1 I represents the rail current at point L. d1 I represents the drain network current at point L. r2 I represents the rail current at L+ΔL. d2 I represents the drain network current at L+ΔL. r I represents the rail current between L and L+ΔL. d U represents the drain network current between L and L+ΔL; r1 U represents the rail's potential relative to ground at point L. d1 U represents the ground potential of the drainage network at point L. r2 U represents the rail-to-ground potential at point L+ΔL. d2 This represents the ground potential of the drainage network at point L+ΔL.

[0110] The parameter relationships between the parameters in the lumped parameter model and the distributed parameter model can be obtained, specifically:

[0111]

[0112] In summary, the sample dataset was constructed based on the train dynamic simulation model.

[0113] Step 2: Construct three parallel convolutional layers based on a multi-scale convolutional neural network, such as 3×1, 5×1, and 7×1, using convolutional kernels of different sizes to extract multi-scale local features from the orbital potential distribution data; through a feature fusion module, channel concatenation and normalization of feature maps at different scales are performed to generate a fused feature vector with multi-resolution information, specifically:

[0114] Step 201: After processing the experimental data obtained by the train dynamic simulation platform, a dataset is obtained. The dataset is divided into a test set and a training set according to a preset ratio, and the data is trained and analyzed.

[0115] Step 202: The size, number, and stride of the convolutional kernels, as well as other hyperparameters, have a significant impact on the network's performance and computational efficiency. The convolutional kernel calculation process is as follows:

[0116]

[0117] In the formula: x l j It is the output of the j-th channel of the l-th convolutional layer; M j For input features; ω l ij This is the weight matrix corresponding to the convolution kernel; b l j σ is the bias matrix; σ(·) is the activation function; * is the symbol for convolution operation;

[0118] Step 203: The output of the convolutional layer uses the non-linear activation function ReLU, and the calculation process is as follows:

[0119] ReLU(x) = max(0,x);

[0120] Convolutional operations were performed using three kernel sizes: 3×1, 5×1, and 7×1. The pooling window size, window sliding stride, and learning rate were selected for the pooling operation. A pooling layer with a stride of 2 and a pooling window of 2×1 was selected. The Adam optimizer was used, and the learner rate of the optimizer was set to 0.0004.

[0121] Step 204: The pooling layer is located after the convolutional layer, and the calculation process is as follows:

[0122] z i =down(x,y)[i]

[0123] In the formula, z i The output of the pooling layer is y, x is the input of the pooling operation, down(x,y) is the downsampling function, and i represents the i-th element in the pooling layer.

[0124] The pooling method used is max pooling, and the calculation process is as follows:

[0125]

[0126] In the formula: OUTPUT represents the feature after pooling, α j This indicates the size of the pooling region.

[0127] Step 205: The fully connected layer is located after the network. The feature maps extracted by the convolutional layer and the pooling layer are transformed into the final output of the fully connected layer. The calculation process is as follows:

[0128] z l =σ(ω)l x l-1 +b l );

[0129] In the formula: z l The output of the fully connected layer is σ(·), where σ is the activation function and ω is the output of the fully connected layer. l x is the weight of the fully connected layer. l-1 It is the output feature of the previous layer, b l It is the bias term of the fully connected layer.

[0130] Each channel sequentially undergoes convolutional layers, ReLU activation, batch normalization, and max pooling operations to extract local features of the signal at different scales. Finally, the features output by each channel are converted into a one-dimensional vector output through a flattening operation.

[0131] In step 3, the input one-dimensional time-series signal is coarsened at multiple scales, and the permutation entropy of the coarsened subsequences is calculated to obtain multi-scale permutation entropy data. Feature information is extracted from the multi-scale permutation entropy data, and it is flattened and converted into a one-dimensional vector. The multi-scale permutation entropy calculation graph is shown below. Figure 4 As shown, the calculation method is as follows:

[0132] (The coarsening step is as follows:)

[0133] Step 31: For the original time signal X = {x1, x2, ..., xn} of length N... N} Perform coarse-graining treatment, that is:

[0134]

[0135] In the formula: s is the scale factor, i = (j-1)s+1;

[0136] Step 32: Process the coarse-grained sequence y obtained in step 31. s j Refactoring:

[0137] Y s j ={y s k ,y s k+τ ,…,y s k+(m-1)τ}

[0138] In the formula: k is the k-th reconstructed component, and the total number of reconstructed components is... m is the embedding dimension; τ is the delay time;

[0139] Step 33: Sort the sequence obtained in step 32:

[0140]

[0141] A set of symbol sequences S is obtained r ={j1,j2,…,j m} Calculate the probability p of each symbol sequence appearing. r ;

[0142] Step 34: Based on the probability p calculated in step 33... r Calculate the permutation entropy of the coarse-grained sequence:

[0143]

[0144] The value of the scaling factor s affects the calculation of the multi-scale permutation entropy. The scaling factor s was set to 5, 10, 15, 20 and 25 to observe the changes in accuracy. Finally, the scaling factor s was selected as 25.

[0145] In step 4, the features extracted from the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are concatenated through a feature fusion layer. Furthermore, during feature map concatenation, the feature maps of the multi-sensor signals in the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are fused using the Concatenate function instruction. The fused features are then input into the bidirectional GRU module to achieve feature fusion.

[0146] In step 5, the bidirectional GRU module fuses the multi-scale convolutional features and temporal features extracted from the multi-scale convolutional neural network and multi-scale permutation entropy, outputting joint features. It then uses bidirectional gated recurrent units to process these features, capturing the dynamic dependencies between consecutive time steps. The output from the bidirectional gated recurrent units is input into a fully connected layer, processed by a nonlinear activation function, and used for final fault diagnosis and classification. Compared to traditional RNNs, GRU introduces a special "gate" structure to replace the original neuron structure. Compared to Long Short-Term Memory (LSTM) networks, the GRU model reduces one gate, simplifying the network parameters. The GRU calculation process is as follows:

[0147] The calculation process of the bidirectional GRU module is as follows:

[0148] Step 51: Input the current time feature a t and the hidden state h from the previous moment t-1 The update and reset gates are calculated using the σ(·) activation function to control the output range. The formula is as follows:

[0149] z t =σ(W z1 a t +W z2 h t-1 )

[0150] r t =σ(W r1 a t +W r2 h t-1 )

[0151] Among them, W z1 W z2 To update the gate weight matrix, W r1 W r2 To reset the gate weight matrix, z t To update the gate output, r t To reset the gate output;

[0152] Step 52: Input the reset door history status r t ⊙h t-1 and input features a t Generate intermediate hidden states Using the tanh(·) activation function, the formula is:

[0153]

[0154] in, In the intermediate hidden state, W h In the middle hidden state The weight matrix;

[0155] Step 53: Input update gate z t The hidden state h from the previous moment t-1 and the intermediate hidden state Output the final hidden state h t The formula is:

[0156]

[0157] The bidirectional gated loop unit is composed of a superposition of forward and reverse GRUs, as shown in the following formula:

[0158]

[0159] In the formula: U t1 U is the weight matrix for forward propagation. t2 Let b be the weight matrix for backpropagation. t For bias.

[0160] The final input signal is processed by a bidirectional gated loop unit and then output. The accuracy and confusion matrix are observed. Multiple experiments are conducted, and the average value is taken for comparison with other common methods. Figure 7 As shown, the confusion matrix is ​​as follows Figure 8 As shown.

[0161] Second Embodiment

[0162] To verify the reliability and accuracy of the present invention, this embodiment takes a subway as an example, focusing on a DC traction power supply system, and combines... Figure 3 The dynamic simulation model diagram of the train was used to obtain the dynamic change law of the track potential, such as... Figure 6 As shown.

[0163] The fault analysis was conducted within a time period from 1920 seconds to 2100 seconds, with a simulation step size of 1 second. Ten different types of grounding faults were set, and 100 samples were generated for each type of fault. The total number of samples in the generated dataset was 1000, and the ratio of training set to test set in the dataset was set to 7:3.

[0164] In the case of weak insulation faults, the proposed method is compared with three other methods: Laplace-BiGRU, CNN-BiGRU, and MSCNN-BiGRU, with all parameters and optimizer selections being the same. The results are as follows: Figure 7 As shown in the comparison results, the method proposed in this invention achieves improved accuracy in weak insulation fault diagnosis compared to other common methods. It outperforms the other three methods, achieving a diagnostic accuracy of 97.4% in weak insulation fault cases, which is superior to Laplace-BiGRU's 93.07%, CNN-BiGRU's 93.4%, and MSCNN-BiGRU's 95.07%. This demonstrates a significant improvement in accuracy for weak insulation fault diagnosis. Furthermore, to further analyze the results... Figures 8-11 The confusion matrices for several methods are shown respectively. Figure 8 The Laplace-BiGRU confusion matrix. Figure 9 The CNN-BiGRU confusion matrix is... Figure 10 The confusion matrix for MSCNN-BiGRU is... Figure 11 The confusion matrix is ​​MSCNN-MPE-BiGRU. Various methods have good overall classification performance, but the method proposed in this invention performs better, achieving 100% classification performance in many cases.

[0165] Therefore, the fault diagnosis method based on MSCNN-MPE-BiGRU proposed in this invention overcomes the limitations of signal feature extraction by introducing MPE and MSCNN to capture the features of the input signal under multi-scale conditions. Simultaneously, BiGRU is employed to integrate the forward and backward outputs, further improving the accuracy and robustness of fault diagnosis. Experimental results show that the proposed method improves the accuracy of fault diagnosis and can accurately identify fault states compared to other methods.

[0166] This invention discloses a ground fault diagnosis method based on intelligent time-series feature extraction, belonging to the field of intelligent monitoring and protection technology of power systems. It addresses the problems of insufficient multi-scale feature extraction, weak noise resistance, and low time-series modeling accuracy of traditional fault diagnosis methods under complex operating conditions. The ground fault diagnosis method based on MSCNN-MPE-BiGRU proposed in this invention integrates multi-scale feature extraction and dynamic time series features, achieving high-precision diagnosis of ground faults.

[0167] In the feature extraction stage, a multi-scale convolutional neural network is used, employing three different sizes of convolutional kernels (3×1, 5×1, and 7×1) to extract features from the input signal. Simultaneously, multi-scale permutation entropy is incorporated to coarsely process the original signal, calculate permutation entropy values ​​at different scales, quantify the spatial distribution characteristics of signal complexity, and suppress noise interference.

[0168] After feature extraction, the features extracted from the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are concatenated through a feature fusion layer. The fused features are then input into the BiGRU module. The input features are processed by update and reset gates in BiGRU, and feature fusion is achieved through forward and backward processing to effectively capture the dynamic changes of time-series signals. Finally, the feature vector output by BiGRU is processed through a fully connected layer, and a label smoothing regularization method is used to identify and classify the fault diagnosis type, outputting the final fault diagnosis result.

[0169] This invention significantly improves the accuracy and robustness of grounding fault diagnosis and enhances the accuracy of classification by using multi-scale feature fusion and dynamic temporal modeling.

[0170] The above-described embodiments merely illustrate preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A ground fault diagnosis method based on intelligent timing feature extraction, characterized in that, include: S1. Establish the structure of the DC traction power supply system, build a train dynamic simulation platform, obtain track potential distribution data during train operation, and construct a sample dataset; S2. Construct three parallel convolutional layers based on a multi-scale convolutional neural network, using convolutional kernels of different sizes to extract multi-scale features from the orbital potential distribution data. Through the feature fusion module, the feature maps of different scales are channel-stitched and normalized to generate a fused feature vector with multi-resolution information. S3. Introducing multi-scale permutation entropy to extract features of dynamic orbit potential time-series signals in the sample dataset, quantifying the dynamic change patterns of dynamic orbit potential time-series signals at different time scales; S4. Concatenating the features extracted by the multi-scale convolutional neural network and multi-scale permutation entropy channels through a feature fusion layer, and inputting the fused features into a bidirectional GRU module to achieve feature fusion; S5. Processing the feature vector output by the bidirectional GRU module through a fully connected layer to identify and classify fault diagnosis types, and outputting the final fault diagnosis result; Utilizing the advantage of spatial locality of convolutional neural networks to capture local features in signals, and extracting dynamic features of signals by quantifying the complexity of time series, the two work together to improve the overall feature extraction capability and the accuracy of fault diagnosis; In S1, a distributed parameter model is constructed based on the "running rail-drainage network-ground" structure of the rail transit traction power supply system; in the distributed parameter model, This represents the rail current at point x. This represents the drain network current at point x. This represents the potential of the rail relative to ground at point x. This represents the potential of the drainage network relative to ground at point x. The longitudinal resistance per unit length of the running track. The longitudinal resistance per unit length of the drainage network. For the transition conductivity per unit length between the rail and the drainage network, For the transition conductance per unit length of the drainage network to ground, the voltage and current satisfy the following formula: ， In the lumped parameter model, current and voltage also satisfy the following formulas: ， In the formula, For the rails in Equivalent impedance within the interval, For drainage net in Equivalent impedance within the interval, For the steel rail to drain network in Equivalent conductance within the interval, For drainage network to the ground Equivalent conductance within the interval, I r1 I represents the rail current at point L. d1 I represents the drain network current at point L. r2 Indicates L+ Rail current at location I d2 Indicates L+ The drain network current at the location, I r Indicates the distance from L to L+ Rail current, I d Indicates the distance from L to L+ The drain network current between; U r1 U represents the rail's potential relative to ground at point L. d1 U represents the ground potential of the drainage network at point L. r2 Indicates L+ The rail's potential to ground, U d2 Indicates L+ The ground potential of the drainage network; The parameter relationships between the parameters in the lumped parameter model and the distributed parameter model are obtained as follows: ； S2 includes: S201, processing the experimental data obtained from the train dynamic simulation platform to obtain a dataset, dividing the dataset into a test set and a training set according to a preset ratio, and performing training analysis on the data; S202, the convolution kernel calculation process is as follows: ， In the formula: It is the output of the j-th channel of the l-th convolutional layer. As input features, It is the weight matrix corresponding to the convolution kernel. It is the bias matrix. '*' is the activation function, and '*' is the symbol for convolution operation. S203. The output of the convolutional layer uses the non-linear activation function ReLU, and the calculation process is as follows: ； S204, the pooling layer is located after the convolutional layer, and the calculation process is as follows: ， In the formula, For pooling layer output, It is a downsampling function, where i represents the i-th element in the sample; The pooling method used is max pooling, and the calculation process is as follows: ， In the formula: OUTPUT represents the pooled feature. Indicates the size of the pooling region; S205. The fully connected layer, located after the network, transforms the feature maps extracted by the convolutional and pooling layers into the final output of the fully connected layer. The calculation process is as follows: ； In the formula: For the output of the fully connected layer, It is an activation function. These are the weights of the fully connected layer. It is the output feature of the previous layer. It is the bias term of the fully connected layer.

2. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 1, characterized in that, In S3, the input one-dimensional time-series signal is coarsened at multiple scales, and the permutation entropy of the coarsened subsequence is calculated to obtain multi-scale permutation entropy data. Feature information in the multi-scale permutation entropy data is extracted and flattened into a one-dimensional vector.

3. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 2, characterized in that, The steps of coarsening are as follows: S31. For the original time signal of length N... Perform coarse-graining treatment, that is: ， In the formula: s is the scale factor, ; S32. The coarse-grained sequence obtained in step 31 Refactoring: ， In the formula: k is the k-th reconstructed component, and the total number of reconstructed components is... m is the embedding dimension; τ is the delay time. S33. Sort the sequence obtained in step 32: ， A set of symbol sequences is obtained Calculate the probability of each symbol sequence appearing. ; S34. Based on the probability calculated in step 33 Calculate the permutation entropy of the coarse-grained sequence: 。 4. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 1, characterized in that, In S5, the bidirectional GRU module fuses the multi-scale convolutional features and temporal features extracted from the multi-scale convolutional neural network and the multi-scale permutation entropy, outputting joint features. The features are then processed by the bidirectional gated recurrent unit to capture the dynamic dependencies between time steps. The output obtained from the bidirectional gated recurrent unit is input into the fully connected layer and processed by the nonlinear activation function for the final fault diagnosis and classification.

5. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 4, characterized in that, The calculation process of the bidirectional GRU module is as follows: S51, Input the current time and input features Hidden state from the previous moment To update the door and reset the door, use The activation function is calculated to control the output range; the formula is as follows: ， ， in, , To update the gate weight matrix, , To reset the gate weight matrix, To update the gate output, To reset the gate output; S52, Input to reset door history status and input features Generate intermediate hidden states ,use The activation function, with the formula: ， in, It is in a hidden state in the middle. In the middle hidden state The weight matrix; S53, Input Update Gate The state that was hidden a moment ago and the hidden state in the middle Output the final hidden state. The formula is: 。 6. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 5, characterized in that, A bidirectional gated recurrent unit is composed of a superposition of forward and reverse GRUs, as shown in the following formula: ， In the formula: This is the weight matrix for forward propagation. This is the weight matrix for backpropagation. For bias.

7. The grounding fault diagnosis method based on intelligent timing feature extraction according to claim 1, characterized in that, In S4, during feature map stitching, the feature maps of the multi-sensor signals in the two channels of the multi-scale convolutional neural network and the multi-scale permutation entropy are fused using the Concatenate function instruction.

Images