Ground fault diagnosis method based on intelligent time sequence feature extraction

Abstract

The invention relates to the technical field of intelligent monitoring and protection of a power system, and discloses a ground fault diagnosis method based on intelligent time sequence feature extraction, in a feature extraction stage, a multi-scale convolutional neural network is adopted, multi-scale permutation entropy is combined, coarse graining processing is carried out on an original signal, permutation entropy values under different scales are calculated, and a ground fault diagnosis result is obtained; after spatial distribution features of signal complexity are quantized, noise interference is suppressed and features are extracted, features extracted by two channels of a multi-scale convolutional neural network and a multi-scale permutation entropy are spliced through a feature fusion layer, and the fused features are input into a BiGRU module. Aiming at the problems of insufficient multi-scale feature extraction, weak anti-noise capability, low time sequence modeling precision and the like of a traditional fault diagnosis method under a complex working condition, the method integrates the features of multi-scale feature extraction and a dynamic time sequence, realizes high-precision diagnosis of a grounding fault, and improves the fault diagnosis accuracy. And the accuracy and robustness of grounding fault diagnosis are obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring and protection of power systems, and particularly relates to a ground fault diagnosis method based on intelligent time series feature extraction. Background Art

[0002] With the continuous advancement of urbanization, electrified rail transit systems have been widely adopted in cities around the world due to their high capacity, high passenger capacity, optimized energy utilization, ease of urban traffic congestion, reduced air pollution, and rapid transportation efficiency. The traction power supply system is a key component of urban rail transit systems, ensuring normal operation by providing stable power support for trains and other equipment. Currently, urban rail transit systems primarily use a DC traction power supply mode, in which the train's traction current flows back through the rails to the traction substation, ensuring the transmission and use of power.

[0003] However, rails, as return paths for current, have longitudinal resistance and are not completely insulated from the ground, resulting in a potential relative to the ground. Some of this current leaks through rail fasteners to the ground, generating stray currents. These stray currents not only corrode the rails themselves, tunnel reinforcement, 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 damage equipment, increase repair 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 can sometimes flow into the AC grid's grounding grid. This process can create potential differences between substations, further exacerbating interference with the AC grid system and, in severe cases, 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 failures are becoming more frequent, posing challenges to power safety and stable supply.

[0005] In this context, how to effectively identify and diagnose faults in power systems has become an important issue that needs to be addressed. Although existing fault detection and diagnosis methods can detect problems to a certain extent, they still have certain limitations:

[0006] (1) The traditional threshold method relies on fixed parameter settings and has poor adaptability to complex working conditions (such as high-impedance grounding and noise interference), resulting in a high missed detection rate;

[0007] (2) A single CNN model is difficult to simultaneously capture the multi-scale characteristics and time series dependencies of fault signals. The feature extraction capability of shallow networks is limited and it is difficult to accurately judge complex power system faults in real time.

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

[0009] With the continuous changes in the topology of power systems, relying solely on traditional methods can no longer meet the increasingly complex fault diagnosis needs.

[0010] Based on this, a ground fault diagnosis method based on intelligent time series feature extraction is provided to solve the technical defects mentioned in the background technology. Summary of the Invention

[0011] In response to the above-mentioned technical problems, the present invention provides a ground fault diagnosis method based on intelligent time series feature extraction to improve the diagnostic accuracy and robustness in complex scenarios.

[0012] The present invention provides a ground fault diagnosis method based on intelligent time series feature extraction, comprising the following steps:

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

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

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

[0016] Step 4: The features extracted by the multi-scale convolutional neural network and the multi-scale permutation entropy are spliced ​​through a feature fusion layer, and the fused features are input into a 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 distribution parameter model, i r (x) represents the rail current at x, i d(x) represents the drain network current at x, u r (x) represents the rail potential to ground at x, u d (x) represents the potential of the drainage network to the ground at the position 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 formula:

[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 relationship between each parameter in the fractional parameter model and the distributed parameter model is obtained, specifically:

[0028]

[0029] Specifically, in step 2, the specific steps are:

[0030] Step 201: After processing the experimental data obtained by the train dynamic simulation platform, a data set is obtained, the data set is divided into a test set and a training set according to a preset ratio, and training analysis is performed on the data;

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

[0032]

[0033] Where: x l j is the output of the jth channel of the convolutional layer, M j is the input feature, ω l ij is the weight matrix corresponding to the convolution kernel, b l j is the bias matrix, σ(·) is the activation function, and * is the symbol of the convolution operation;

[0034] Step 203: The convolutional layer output uses a nonlinear activation function ReLU, and the calculation process is:

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

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

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

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

[0039] The pooling method uses the maximum pooling method, and the calculation process is as follows:

[0040]

[0041] Where: OUTPUT represents the feature after pooling, α j Indicates the size of the pooling area.

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

[0043] Specifically, the coarse-graining steps are:

[0044] Step 31: For the original time signal X with a length of N, N} for coarse-graining processing, namely:

[0045]

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

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

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

[0049] Where: k is the kth reconstruction component, and the total number of reconstruction components is m is the embedding dimension; τ is the delay time;

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

[0051]

[0052] Get a set of symbol sequence S r ={j1,j2,…,j m}, calculate the probability p of each symbol sequence appearing r ;

[0053] Step 34: According to 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 neural network and the multi-scale permutation entropy extracted multi-scale convolutional features and time series features, outputs joint features, and processes the features using a bidirectional gated recurrent unit to capture the dynamic dependency between previous and next time steps. The output obtained from the bidirectional gated recurrent unit is input into the fully connected layer and processed by a nonlinear activation function for final fault diagnosis and classification.

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

[0057] Step 51: Input the current moment input feature a t and the previous hidden state h t-1 To the update gate and reset gate, the σ(·) activation function is used to calculate and control the output range. The formula is:

[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 is the update gate output, r t is the reset gate output;

[0061] Step 52: Input reset door history state r t ⊙h t-1 And input feature a t , generating the intermediate hidden state Using the tanh(·) activation function, the formula is:

[0062]

[0063] in, is the intermediate hidden state, W h is the intermediate hidden state The weight matrix of

[0064] Step 53: Input update gate z t , the hidden state h at 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 recurrent unit is composed of the forward and reverse GRUs, and the formula is as follows:

[0067]

[0068] Where: U t1 is the weight matrix of forward propagation, U t2 is the weight matrix for back propagation, b t For bias.

[0069] Specifically, in step 4, when the feature maps are spliced, 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 through 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 the present invention uses the advantages of convolutional neural networks in spatial locality to capture local features in the signal, and MPE can extract the dynamic characteristics of the signal by quantifying the complexity of the time series. The combination of the two can effectively make up for the problem of insufficient expression ability of single modal features, thereby improving the overall feature extraction capability and the accuracy of fault diagnosis.

[0072] 2. The MPE in the present invention can downsample and smooth the original signal through coarse-graining processing, effectively suppress noise interference, and has good accuracy and robustness. At the same time, the BiGRU module captures the forward and reverse features at the same time through the superposition processing of forward GRU and reverse GRU, which not only simplifies the network parameters but also solves the problem of insufficient information utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0074] Figure 1 Flowchart of the present invention;

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

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

[0077] Figure 4 This is the MPE calculation flow chart of the present invention;

[0078] Figure 5 Schematic diagram of the BiGRU structure of the present invention;

[0079] Figure 6 The train operation diagram of the present invention;

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

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

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

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

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

[0085] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are only used to explain the present invention and are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0086] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0087] First embodiment

[0088] This embodiment discloses a ground fault diagnosis method based on intelligent time series 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 recurrent unit (BiGRU).

[0089] Figure 1 FIG. 1 is a flow chart of an embodiment of the present invention, and the flow chart includes the following steps:

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

[0091] Table 1 Distribution parameter model parameters

[0092]

[0093]

[0094] for Figure 3 The distribution parameter model, i r (x) represents the rail current at x, i d (x) represents the drain network current at x, u r (x) represents the rail potential to ground at x, u d (x) represents the potential of the drainage network to the ground at the position 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] Among them, C1~C4 depend on ΔL and the parameters in the distribution parameter model, and are independent of the interval location L.

[0101] Combine Figure 3 The parameters of the lumped parameter model are shown in Table 2. In the L~ΔL section, 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 represents the rail current at L, I d1 represents the drain network current at L, I r2 represents the rail current at L+ΔL, I d2 Represents the drain network current at L+ΔL, I r Indicates the rail current between L and L+ΔL, I d Indicates the drain network current between L and L+ΔL; U r1 Indicates the rail-to-ground potential at L, U d1 Indicates the potential of the drainage network to the ground at L, U r2 Indicates the rail potential to ground at the position L+ΔL, U d2 It represents the potential of the drainage network to the ground at the point L+ΔL.

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

[0111]

[0112] In summary, the construction of the sample data set is completed based on the train dynamic simulation model.

[0113] Step 2: Based on the multi-scale convolutional neural network, three parallel convolution layers are constructed, such as 3×1, 5×1, and 7×1, respectively, using convolution kernels of different sizes to extract multi-scale local features from the track potential distribution data; through the feature fusion module, the feature maps of different scales are channel-joined and normalized 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 data set is obtained, the data set is divided into a test set and a training set according to a preset ratio, and training analysis is performed on the data;

[0115] Step 202: Hyperparameters such as the size, number, and sliding step length of the convolution kernel have a significant impact on the performance and computational efficiency of the network. The convolution kernel calculation process is as follows:

[0116]

[0117] Where: x l j is the output of the jth channel of the convolutional layer l; M j is the input feature; ω l ij is the weight matrix corresponding to the convolution kernel; b l j is the bias matrix; σ(·) is the activation function; * is the symbol of the convolution operation;

[0118] Step 203: The convolutional layer output uses a nonlinear activation function ReLU, and the calculation process is:

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

[0120] We select three convolution kernel sizes of 3×1, 5×1, and 7×1 for convolution operations. We also select the pooling window size, window sliding stride, and learning rate for the pooling operation. For the pooling layer, we select a model with a stride of 2 and a pooling window of 2×1. We also use the Adam optimizer and set the learning rate of the optimizer to 0.0004.

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

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

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

[0124] The pooling method uses the maximum pooling method, and the calculation process is as follows:

[0125]

[0126] Where: OUTPUT represents the feature after pooling, α j Indicates the size of the pooling area.

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

[0128] z l =σ(ωl x l-1 +b l );

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

[0130] Each channel passes through the convolution layer, ReLU activation, batch normalization and maximum pooling operations in sequence to extract the local features of the signal at different scales. Finally, the features output by each channel are converted into one-dimensional vector output through 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 subsequence is calculated to obtain multi-scale permutation entropy data. Feature information in the multi-scale permutation entropy data is extracted and flattened to convert it into a one-dimensional vector. The multi-scale permutation entropy calculation diagram is shown in FIG. Figure 4 As shown, the calculation method is as follows:

[0132] (The coarse-graining steps are:

[0133] Step 31: For the original time signal X with a length of N, N} for coarse-graining processing, namely:

[0134]

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

[0136] Step 32: 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] Where: k is the kth reconstruction component, and the total number of reconstruction components is m is the embedding dimension; τ is the delay time;

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

[0140]

[0141] Get a set of symbol sequence S r ={j1,j2,…,j m}, calculate the probability p of each symbol sequence appearing r ;

[0142] Step 34: According to the probability p calculated in step 33 r , calculate the permutation entropy of the coarse-grained sequence:

[0143]

[0144] The value of the scale factor s affects the calculation of the multi-scale permutation entropy. s is set to 5, 10, 15, 20, and 25 to observe the change in accuracy. Finally, the scale factor s is 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 spliced ​​through a feature fusion layer. Moreover, when the feature maps are spliced, 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 through the Concatenate function instruction, and the fused features are input into the bidirectional GRU module to achieve feature fusion.

[0146] In step 5, the bidirectional GRU module fuses the multi-scale convolutional neural network and the multi-scale permutation entropy extracted multi-scale convolution features and time series features, outputs joint features, and processes the features using a bidirectional gated recurrent unit to capture the dynamic dependency between the previous and next time steps. The output obtained from the bidirectional gated recurrent unit is input to the fully connected layer and processed by a nonlinear activation function for final fault diagnosis and classification. Compared with the traditional RNN, the GRU introduces a special "gate" structure to replace the original neuron structure. Compared with the long short-term memory network (LSTM), the GRU model reduces one gate and simplifies the network parameters. The calculation process of the GRU is:

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

[0148] Step 51: Input the current moment input feature a t and the previous hidden state h t-1 To the update gate and reset gate, the σ(·) activation function is used to calculate and control the output range. The formula is:

[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 is the update gate output, r t is the reset gate output;

[0152] Step 52: Input reset door history state r t ⊙h t-1 And input feature a t , generating the intermediate hidden state Using the tanh(·) activation function, the formula is:

[0153]

[0154] in, is the intermediate hidden state, W h is the intermediate hidden state The weight matrix of

[0155] Step 53: Input update gate z t , the hidden state h at 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 recurrent unit is composed of the forward and reverse GRUs, and the formula is as follows:

[0158]

[0159] Where: U t1 is the weight matrix of forward propagation, U t2 is the weight matrix for back propagation, b t For bias.

[0160] The final input signal is processed by the bidirectional gated recurrent unit and then output. The accuracy and confusion matrix are observed. The average value is taken from multiple experiments and compared with other common methods. Figure 7 As shown, the confusion matrix is Figure 8 shown.

[0161] Second embodiment

[0162] In order to verify the reliability and accuracy of the present invention, this embodiment takes the subway as an example and the DC traction power supply system as the research object. Figure 3 The train dynamic simulation model diagram is used to obtain the dynamic change law of track potential, such as Figure 6 shown.

[0163] The time period from 1920 seconds to 2100 seconds was selected for fault analysis. The simulation step was set to 1 second. Ten different types of ground faults were set. Under each fault type, the train generated 100 samples. The total sample size of the generated data set was 1000. The sample ratio of the training set and the test set in the data set was set to 7:3.

[0164] In the case of weak insulation fault, the three methods of Laplace-BiGRU, CNN-BiGRU and MSCNN-BiGRU are compared with the method proposed in this invention. The parameters and optimizers are the same. The results are as follows: Figure 7 As shown in the figure, from the comparison results, it can be seen that the method proposed by the present invention has improved the accuracy under weak insulation faults compared with other common methods, and is better than the other three methods. In the case of weak insulation faults, the diagnostic accuracy of the present invention can reach 97.4%, which is better than 93.07% of Laplace-BiGRU, 93.4% of CNN-BiGRU and 95.07% of MSCNN-BiGRU. It can be seen that the accuracy of the present invention in the diagnosis of weak insulation faults has been significantly improved. In addition, in order to further analyze the results, Figures 8-11 The confusion matrices of several methods are shown. Figure 8 is the Laplace-BiGRU confusion matrix, Figure 9 is the CNN-BiGRU confusion matrix, Figure 10 is the MSCNN-BiGRU confusion matrix, Figure 11 The confusion matrix of MSCNN-MPE-BiGRU is shown in Figure 2. The overall classification effects of various methods are good, but the method proposed in the present invention is better, and the classification effect can reach 100% in many cases.

[0165] Therefore, the fault diagnosis method based on MSCNN-MPE-BiGRU proposed in this paper overcomes the limitations of signal feature extraction by introducing MPE and MSCNN to capture input signal features at multiple scales. Furthermore, the BiGRU is used to integrate the forward and reverse outputs, further improving the accuracy and robustness of fault diagnosis. Experimental results show that the proposed method improves fault diagnosis accuracy and can accurately identify fault conditions compared to other methods.

[0166] The present invention discloses a grounding fault diagnosis method based on intelligent time series feature extraction, which belongs to the technical field of intelligent monitoring and protection of power systems. In view of the problems of insufficient multi-scale feature extraction, weak anti-noise ability and low time series modeling accuracy in traditional fault diagnosis methods under complex working conditions, the present invention proposes a grounding fault diagnosis method based on MSCNN-MPE-BiGRU, which integrates the characteristics of multi-scale feature extraction and dynamic time series, and realizes high-precision diagnosis of grounding faults.

[0167] In the feature extraction stage, a multi-scale convolutional neural network is adopted, and three convolution kernels of different sizes (3×1, 5×1, and 7×1) are used to extract the features of the input signal. At the same time, multi-scale permutation entropy is combined to perform coarse-graining processing on the original signal, calculate the permutation entropy values ​​at different scales, quantify the spatial distribution characteristics of the signal complexity, and suppress noise interference.

[0168] After feature extraction, the features extracted by the multi-scale convolutional neural network and the multi-scale permutation entropy are spliced ​​through the feature fusion layer. The fused features are input into the BiGRU module. The input features are calculated by the update gate and the reset gate in the BiGRU. By processing in both forward and reverse directions, feature fusion is achieved, effectively capturing the dynamic changes of the timing signal. Finally, the feature vector output by the BiGRU is processed by the fully connected layer, and the label smoothing regularization method is used to identify and classify the fault diagnosis type, and the final fault diagnosis result is output.

[0169] The present invention significantly improves the accuracy and robustness of ground fault diagnosis and classification through multi-scale feature fusion and dynamic time series modeling.

[0170] The above embodiments merely represent preferred implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A ground fault diagnosis method based on intelligent time series feature extraction is characterized by: The steps include: Step 1: Establish the DC traction power supply system structure, build a train dynamic simulation platform, obtain track potential distribution data during train operation, and construct a sample data set; Step 2: construct three parallel convolutional layers based on a multi-scale convolutional neural network, using convolution kernels of different sizes to extract multi-scale features from the orbital potential distribution data. The feature maps of different scales are channel-joined and normalized through a feature fusion module to generate a fused feature vector with multi-resolution information. Step 3: Introduce multi-scale permutation entropy to extract the characteristics of the dynamic orbital potential time series signal in the sample data set, and quantify the dynamic change law of the dynamic orbital potential time series signal at different time scales; Step 4: The features extracted by the multi-scale convolutional neural network and the multi-scale permutation entropy are spliced ​​through a feature fusion layer, and the fused features are input into a bidirectional GRU module to achieve feature fusion; 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.

2. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 1 is characterized in that: 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.

3. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 2 is characterized in that: In the distribution parameter model, i r (x) represents the rail current at x, i d (x) represents the drain network current at x, u r (x) represents the rail potential to ground at x, u d (x) represents the potential of the drainage network to the ground at the position x, and the voltage and current satisfy the following formula: In the lumped parameter model, the current and voltage also satisfy the following formula: I r =I r1 +y r1 (U r1 -U d1 ) I d =I d1 +y d1 U d1 -y r1 (U r1 -U d1 ) U r2 =U r1 -Z r1 I r U d2 =U d1 -Z d1 I d I r2 =I r +y r1 (U r2 -U d2 ) The parameter relationship between each parameter in the lumped parameter model and the distributed parameter model is obtained, specifically:

4. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 1 is characterized in that: In step 2, the specific steps are: Step 201: After processing the experimental data obtained by the train dynamic simulation platform, a data set is obtained, the data set is divided into a test set and a training set according to a preset ratio, and training analysis is performed on the data; Step 202: The convolution kernel calculation process is as follows: Where: x l j is the output of the jth channel of the convolutional layer, M j is the input feature, ω l ij is the weight matrix corresponding to the convolution kernel, b l j is the bias matrix, σ(·) is the activation function, and * is the symbol of the convolution operation; Step 203: The convolutional layer output uses a nonlinear activation function ReLU, and the calculation process is: ReLU(x)=max(0,x); Step 204: The pooling layer is located after the convolutional layer, and the calculation process is: z i =down(x,y)[i] Where z i is the output of the pooling layer, 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 method uses the maximum pooling method, and the calculation process is as follows: Where: OUTPUT represents the feature after pooling, α j Indicates the size of the pooling area; Step 205: The fully connected layer is located at the end of the network and converts the feature maps extracted by the convolutional layer and the pooling layer into the final output of the fully connected layer. The calculation process is as follows: z l =σ(ω l x l-1 +b l ); Where: z l is the output of the fully connected layer, σ(·) is the activation function, ω l is the weight of the fully connected layer, x l-1 is the output feature of the previous layer, b l is the bias term of the fully connected layer.

5. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 1 is characterized in that: In step 3, the input one-dimensional time series signal is coarse-grained at multiple scales, and the permutation entropy of the coarse-grained subsequence is calculated to obtain multi-scale permutation entropy data, and feature information in the multi-scale permutation entropy data is extracted and flattened to convert it into a one-dimensional vector.

6. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 5 is characterized in that: The steps of coarse-graining are: Step 31: For the original time signal X with a length of N, N } for coarse-graining processing, namely: Where: s is the scale factor, i = (j-1)s+1; Step 32: the coarse-grained sequence y obtained in step 31 s j Refactoring: AND s j ={and s k ,and s k+τ ,…,and s k+(m-1)τ } Where: k is the kth reconstruction component, and the total number of reconstruction components is m is the embedding dimension; τ is the delay time; Step 33: Sort the sequence obtained in step 32: Get a set of symbol sequence S r ={j1,j2,…,j m }, calculate the probability p of each symbol sequence appearing r ; Step 34: According to the probability p calculated in step 33 r , calculate the permutation entropy of the coarse-grained sequence:

7. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 1 is characterized in that: In step 5, the bidirectional GRU module fuses the multi-scale convolutional neural network and the multi-scale permutation entropy extracted multi-scale convolution features and time series features, outputs joint features, processes the features using a bidirectional gated recurrent unit, captures the dynamic dependency between previous and next time steps, and inputs the output obtained from the bidirectional gated recurrent unit into a fully connected layer, which is processed by a nonlinear activation function for final fault diagnosis and classification.

8. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 7 is characterized in that: The calculation process of the bidirectional GRU module is: Step 51: Input the current moment input feature a t and the previous hidden state h t-1 To the update gate and reset gate, the σ(·) activation function is used to calculate and control the output range. The formula is: With t =σ(W z1 and t +W z2 h t-1 ) r t =σ(W r1 a t +W r2 h t-1 ) Among them, W z1 , W z2 To update the gate weight matrix, W r1 , W r2 To reset the gate weight matrix, z t is the update gate output, r t is the reset gate output; Step 52: Input reset door history state r t ⊙h t-1 And input feature a t , generating the intermediate hidden state Using the tanh(·) activation function, the formula is: in, is the intermediate hidden state, W h is the intermediate hidden state The weight matrix of Step 53: Input update gate z t , the hidden state h at the previous moment t-1 and the intermediate hidden state Output the final hidden state h t , the formula is:

9. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 8, characterized in that: The bidirectional gated recurrent unit is composed of the forward and reverse GRUs, and the formula is as follows: Where: U t1 is the weight matrix of forward propagation, U t2 is the weight matrix for back propagation, b t For bias.

10. The ground fault diagnosis method based on intelligent time series feature extraction according to claim 1, characterized in that: In step 4, when the feature maps are spliced, 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 through the Concatenate function instruction.

Images