A fault detection method for partial discharge of ring main unit cables based on deep learning

Abstract

The present invention discloses a method for detecting partial discharge faults in ring main unit cables based on deep learning. The method comprises obtaining samples of the voltage signal of the cable, each sample comprising m sampling points, and sequentially performing wavelet denoising and normalization processing on the sample signals. The method for detecting partial discharge faults in ring main unit cables based on deep learning sequentially performs wavelet denoising, normalization, and grouping on the voltage signal of the cable, and then extracts features in the time domain and frequency domain respectively. The extracted features are concatenated into feature vectors, which are used as inputs of feature types of a fault detection network. The method is combined with the fault detection network to determine whether the cable has a partial discharge fault. The method is more efficient, accurate, and reliable, and ensures the reliability and safety of the power system while improving the availability of the cable system and reducing maintenance costs.

Description

Technical Field

[0001] The present invention belongs to the field of cable fault detection, and specifically relates to a method for detecting partial discharge of ring network cabinet cables based on deep learning. Background Art

[0002] With the rapid development of the national economy and the deepening reform of the power industry, the scale of power systems has continued to expand, and voltage levels have continued to rise, placing stricter demands on the safety and reliability of power systems. This makes online monitoring and fault diagnosis of power equipment particularly critical. Transmission cable systems play a vital role in energy transmission and distribution and are widely used in fields such as power transmission, power supply, and communications. However, the normal operation of cable systems is susceptible to interference from various factors, including cable aging, environmental changes, and external damage. These factors can cause partial discharge (PD) failures within the cable system, a serious electrical problem that can cause power system failures, accidents, and power outages. PD not only compromises the reliability of the cable system but can also harm the environment and even cause fires. Therefore, early detection and diagnosis of PD failures in ring main unit cables is crucial to ensuring the reliability and safety of the power system.

[0003] Traditional methods for detecting partial discharge faults in ring main unit cables mainly rely on technologies such as sensor measurement, signal processing, and threshold determination. However, these methods have some limitations: 1. Reliance on rules and thresholds: Traditional methods usually determine whether partial discharge exists based on empirical rules and pre-set thresholds. These rules and thresholds may not be applicable to different types of cable systems or working conditions. 2. Difficulty in processing complex data: Cable system data usually includes multi-channel time series information from multiple sensors, and traditional methods have difficulty in effectively processing and analyzing this complex data. 3. Difficulty in coping with changes: The performance and working conditions of the cable may change over time. Traditional methods require regular adjustment of parameters and thresholds, which has high maintenance costs. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems raised in the background technology and propose a fault detection method for partial discharge of ring main unit cables based on deep learning.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] The present invention proposes a method for detecting partial discharge faults in ring main unit cables based on deep learning, which includes obtaining samples of the cable voltage signal, each sample containing m sampling points, and sequentially performing wavelet denoising and normalization processing on the sample signals;

[0007] The normalized sample signals are grouped into n groups according to p consecutive sampling points;

[0008] Statistical features are extracted from each set of data in the frequency domain and time domain respectively, and the statistical features are concatenated into a q-dimensional feature vector;

[0009] A fault detection network is established and trained by establishing an input tensor (k, n, q) with k batches, n feature types, and q-dimensional feature vectors. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network, and a DenseNet network. The input tensor is input into the convolutional neural network and the LSTM network respectively to obtain the first and second outputs in sequence. The first and second outputs are concatenated and input into the GRU network, the attention mechanism network, and the DenseNet network in sequence.

[0010] Finally, the tensor corresponding to the voltage signal of the cable to be tested is input into the trained fault detection network, which outputs two probabilities of whether the cable to be tested has a partial discharge fault, and the one with the higher probability is selected as the final detection result.

[0011] Preferably, statistical features are extracted from each set of data in the frequency domain and the time domain, and the statistical features are concatenated into a q-dimensional feature vector, including:

[0012] Extract features in the time domain: calculate the mean, standard deviation, maximum and minimum values ​​of the signal amplitude of each group of data, as well as the percentage quantiles and relative percentage quantiles corresponding to percentages of 0, 1, 10, 25, 50, 75, 90, 99, and 100, as well as the kurtosis, skewness, number of peaks, maximum height of peaks, minimum height of peaks, average height of peaks, maximum width of peaks, minimum width of peaks, and average width of peaks for each group of data;

[0013] Perform FFT transformation on each set of data and convert it into frequency domain signal;

[0014] Extract features in the frequency domain: calculate the energy sum, maximum energy, minimum energy, standard deviation, variance, skewness, root mean square and rectified mean of the signal amplitude of each group of data converted into frequency domain signals;

[0015] Finally, the feature quantities extracted in the time domain and the feature quantities extracted in the frequency domain are concatenated in sequence to obtain a 39-dimensional feature vector, and the value of q is 39.

[0016] Preferably, the convolutional neural network includes four convolution modules, namely the first convolution module, the second convolution module, the third convolution module and the fourth convolution module, and each convolution module includes a one-dimensional convolution layer, a BN layer, a first RELU activation function and a Dropout layer arranged in sequence.

[0017] Preferably, a fault detection network is established, and an input tensor (k, n, q) is established with k batches, n feature types, and q-dimensional feature vectors to train the fault detection network. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network, and a DenseNet network. The input tensors are respectively input into the convolutional neural network and the LSTM network to obtain the first output and the second output in sequence. The first output and the second output are spliced ​​and input into the GRU network, the attention mechanism network, and the DenseNet network in sequence, including:

[0018] First, the input tensor (k, n, q) is transposed to obtain (k, q, n), and then (k, q, n) is sequentially input into the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module to obtain the first output, and passes through the one-dimensional convolution layer, the BN layer, the first RELU activation function, and the Dropout layer of each convolution module;

[0019] Input the input tensor (k, n, q) to the LSTM network to obtain the second output, and the number of layers of the LSTM network is 5;

[0020] Then, the first output is subjected to a second transposition process, the tensor after the second transposition process is concatenated with the second output to obtain a concatenated tensor, and the concatenated tensor is processed by a second RELU activation function to obtain a third output;

[0021] Then the third output is sequentially input into the GRU network and the attention mechanism network to obtain the fourth output;

[0022] Finally, the fourth output is transposed and input into the DensNet network to output two probabilities of whether there is a partial discharge fault.

[0023] Preferably, the fault detection network uses cross entropy as the loss function Loss, and the calculation formula is as follows:

[0024] Loss=-(y·log(p)+(1-y)·log(1-p))

[0025] Where y represents the actual category label, including the label of the sample with partial discharge signal and the label of the sample without partial discharge signal, and p represents the probability that the fault detection network is labeled as the sample with partial discharge signal.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] This deep learning-based method for detecting partial discharge in ring main unit cables performs wavelet denoising, normalization, and grouping on the cable voltage signal. Feature extraction is then performed in the time and frequency domains. The extracted features are concatenated into feature vectors, which are used as inputs to the feature types of the fault detection network. Combined with the fault detection network, it is determined whether the cable has a partial discharge fault. This method is more efficient, accurate, and reliable, ensuring the reliability and safety of the power system while improving the availability of the cable system and reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of a method for detecting partial discharge in a ring main unit cable based on deep learning according to the present invention;

[0029] Figure 2 This is a block diagram of the feature extraction module of the present invention;

[0030] Figure 3 Schematic diagram of the structure of the fault detection network of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0032] It should be noted that when a component is referred to as being "connected" to another component, it may be directly connected to the other component or there may be an intermediate component. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0033] like Figure 1-3 As shown, a method for detecting partial discharge of a ring main unit cable fault based on deep learning includes:

[0034] S1. Obtain samples of the voltage signal of the cable, each sample containing m sampling points, and perform wavelet denoising and normalization processing on the sample signals in sequence.

[0035] It should be noted that the cable voltage signal can be obtained using a sensor. In this embodiment, the Daubechies wavelet function is selected for wavelet denoising, with a vanishing moment of 4 and a decomposition level of 1. The sample signal after wavelet denoising is normalized and mapped to the range of [-1, 1]. In this embodiment, each sample contains 800,000 sampling points, i.e., m = 800,000. However, the specific number of sampling points is not limited and can be set according to actual needs.

[0036] S2. Group the normalized sample signals into n groups according to p consecutive sampling points.

[0037] It should be noted that, in this embodiment, the normalized sample signal is grouped into 2000 consecutive sampling points as a group, and is divided into 400 groups (800000 / 2000), that is, p = 2000, n = 800000 / 2000 = 400, but the specific number of consecutive sampling points for grouping is not limited and can be set according to actual conditions.

[0038] S3. Statistical features are extracted from each set of data in the frequency domain and time domain respectively, and the statistical features are concatenated into a q-dimensional feature vector.

[0039] Specifically as follows: Extract features in the time domain: calculate the mean, standard deviation, maximum and minimum values ​​of the signal amplitude of each group of data, as well as the percentage quantiles and relative percentage quantiles corresponding to percentages of 0, 1, 10, 25, 50, 75, 90, 99, and 100, as well as the kurtosis, skewness, number of peaks, maximum height of peaks, minimum height of peaks, average height of peaks, maximum width of peaks, minimum width of peaks, and average width of peaks of each group of data;

[0040] Perform FFT transformation on each set of data and convert it into frequency domain signal;

[0041] Extract features in the frequency domain: calculate the energy sum, maximum energy, minimum energy, standard deviation, variance, skewness, root mean square and rectified mean of the signal amplitude of each group of data converted into frequency domain signals;

[0042] Finally, the feature quantities extracted in the time domain and the feature quantities extracted in the frequency domain are concatenated in sequence to obtain a 39-dimensional feature vector, and the value of q is 39.

[0043] S4. Establish a fault detection network and establish an input tensor (k, n, q) with k batches, n feature types and q-dimensional feature vectors to train the fault detection network. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network and a DenseNet network. The input tensors are respectively input into the convolutional neural network and the LSTM network to obtain the first output and the second output in turn. The first output and the second output are spliced ​​and input into the GRU network, the attention mechanism network and the DenseNet network in turn.

[0044] Specifically as follows: The convolutional neural network includes four convolution modules, namely the first convolution module, the second convolution module, the third convolution module and the fourth convolution module. Each convolution module includes a one-dimensional convolution layer, a BN layer, a first RELU activation function and a Dropout layer arranged in sequence.

[0045] Establish a fault detection network and train it by inputting a tensor (k, n, q), where k represents the batch size, n represents the number of input features, and q represents the dimension of each feature. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network, and a DenseNet network, as follows:

[0046] First, the input tensor (k, n, q) is transposed to obtain (k, q, n), and then (k, q, n) is sequentially input into the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module to obtain the first output, and passes through the one-dimensional convolution layer, the BN layer, the first RELU activation function, and the Dropout layer of each convolution module;

[0047] Input the input tensor (k, n, q) to the LSTM network to obtain the second output, and the number of layers of the LSTM network is 5;

[0048] Then, the first output is subjected to a second transposition process, the tensor after the second transposition process is concatenated with the second output to obtain a concatenated tensor, and the concatenated tensor is processed by a second RELU activation function to obtain a third output;

[0049] Then the third output is sequentially input into the GRU network and the attention mechanism network to obtain the fourth output;

[0050] Finally, the fourth output is transposed and input into the DensNet network to output the prediction result.

[0051] It should be noted that, in this embodiment, k=1, n=400, q=39 is used as an example for explanation. First, the input tensor (1, 400, 39) is subjected to the first transposition process to obtain (1, 39, 400), and then (1, 39, 400) is sequentially input into the first convolution module, the second convolution module, the third convolution module and the fourth convolution module to obtain the first output, wherein the input tensor of the one-dimensional convolution layer of the first convolution module is (1, 39, 400), the output is (1, 128, 400), the convolution kernel is 3, the padding is 1, and then it passes through a BN layer, and the dimension is the same as the output dimension of the one-dimensional convolution layer. The BN layer is used to normalize the feature quantity obtained by convolution, and then passes through the first RELU activation function to make the training have a certain degree of randomness, and the result has stronger robustness. Finally, it passes through the Dropout layer to reduce overfitting.

[0052] Similarly, the second, third, and fourth convolutional modules all have the same structure as the first convolutional module, and the detailed process is not described here. The only difference between the second, third, and fourth convolutional modules and the first convolutional module is the input and output of each one-dimensional convolutional layer. The input tensor of the one-dimensional convolutional layer of the second convolutional module is (1, 128, 400), and the output is (1, 256, 400); the input tensor of the one-dimensional convolutional layer of the third convolutional module is (1, 256, 400), and the output is (1, 512, 400); the input tensor of the one-dimensional convolutional layer of the fourth convolutional module is (1, 512, 400), and the output is (1, 256, 400). At the same time, the convolution kernel of each one-dimensional convolutional layer is 3, and the padding is 1. The output of the convolutional neural network is a tensor of (1, 256, 400), that is, the first output is a tensor of (1, 256, 400).

[0053] The input tensor (1, 400, 39) is fed into the LSTM network, resulting in a second output of (1, 400, 256). The LSTM network input feature dimension is 39, the number of hidden layers is 256, the number of LSTM layers is 5, and the dropout ratio is 0.2 to prevent overfitting. Through the cell state and forget gate mechanisms, the LSTM network can better retain and transmit long-term information in the sequence. The forget gate, input gate, and output gate help alleviate the vanishing gradient problem.

[0054] The first output (1, 256, 400) is then transposed for the second time to obtain a tensor (1, 400, 256), and then the tensor (1, 400, 256) is concatenated with the second output (1, 400, 256) to obtain a concatenated tensor (1, 400, 512), and the concatenated tensor (1, 400, 512) is processed by a second RELU activation function to obtain a third output (1, 400, 512), so that the training has a certain degree of randomness and the result has stronger robustness.

[0055] Then the third output (1, 400, 512) is input to the GRU network, and the output is (1, 400, 128). In this process, the input feature dimension of the GRU network is 512, the output feature dimension is 128, the number of layers of the GRU network is 1, and the dropout is 0. Since the structure of the GRU network is relatively simple and the computational efficiency is high, by adjusting the parameters of LSTM and GRU, the complexity of the network can be flexibly adjusted to adapt to the requirements of different data sets.

[0056] The output of the GRU network is passed through the attention mechanism network. That is, the input of the attention mechanism network is (1, 400, 128), and the fourth output is (1, 400, 128). The dimension of the attention mechanism network is 128, the dimension of the query vector is 128, the dimension of the numeric vector is 128, the number of attention heads is 8, and the attention mechanism network uses the self-attention mechanism ScaledDotProductAttention, which calculates the dot product between the query and the key and then adjusts the attention weight by scaling. In this way, the network can learn the degree of attention to different positions or features and automatically adjust the weights to adapt to different parts of the input sequence, thereby improving the accuracy of training.

[0057] Finally, the fourth output (1, 400, 128) is transposed to obtain a tensor (1, 128, 400), and the transposed tensor (1, 128, 400) is input into the DensNet network. The number of channels of the DensNet network is 128, and the growth rate parameter of the DensNet network is 32. The DensNet network outputs (1, 2), where 1 in the output represents the batch size, and 2 in the output represents the two categories of probabilities of fault and no fault. The two probabilities are compared, and the one with a larger probability is selected as the fault detection result. 0 is used to represent the absence of partial discharge fault, and 1 is used to represent the presence of partial discharge fault.

[0058] During the training of the fault detection network, the dataset is divided into training set, validation set and test set with a ratio of 8:1:1.

[0059] The fault detection network uses cross entropy as the loss function, and the calculation formula is as follows:

[0060] Loss=-(y·log(p)+(1-y)·log(1-p))

[0061] Where y represents the actual category label, including the label of the sample with partial discharge signal and the label of the sample without partial discharge signal, and p represents the probability that the fault detection network is labeled as the sample with partial discharge signal.

[0062] It should be noted that samples with partial discharge signals are labeled 1, and samples without partial discharge signals are labeled 0. After obtaining the loss function value, it is optimized using gradient descent to minimize the loss function value. The Adam optimizer is used here to optimize the loss function. Adam (Adaptive Moment Estimation) combines momentum and adaptive learning rates to accelerate model convergence and effectively handle sparse gradients.

[0063] S5. Finally, the tensor corresponding to the voltage signal of the cable to be detected is input into the trained fault detection network, and two probabilities of whether the cable to be detected has a partial discharge fault are output, and the one with the higher probability is selected as the final detection result.

[0064] It should be noted that the voltage signal of the cable to be detected is processed by S1-S3 to obtain a feature vector of the corresponding dimension, and then the corresponding tensor is input into the trained fault detection network.

[0065] There is no restriction on the order of S1-S5; the steps can be interchanged or performed simultaneously.

[0066] This deep learning-based method for detecting partial discharge in ring main unit cables performs wavelet denoising, normalization, and grouping on the cable voltage signal. Feature extraction is then performed in the time and frequency domains. The extracted features are concatenated into feature vectors, which are used as inputs to the feature types of the fault detection network. Combined with the fault detection network, it is determined whether the cable has a partial discharge fault. This method is more efficient, accurate, and reliable, ensuring the reliability and safety of the power system while improving the availability of the cable system and reducing maintenance costs.

[0067] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The above-described embodiments merely represent specific and detailed examples of the present application and should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for detecting partial discharge in ring main unit cables based on deep learning, characterized by: The deep learning-based method for detecting partial discharge faults in ring main unit cables includes: Obtain samples of the cable voltage signal, each sample contains m sampling points, and perform wavelet denoising and normalization processing on the sample signals in sequence; The normalized sample signals are grouped into n groups according to p consecutive sampling points; Statistical features are extracted from each set of data in the frequency domain and time domain respectively, and the statistical features are concatenated into a q-dimensional feature vector; Establish a fault detection network and establish an input tensor (k, n, q) with k batches, n feature types, and q-dimensional feature vectors to train the fault detection network. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network, and a DenseNet network. The input tensors are respectively input into the convolutional neural network and the LSTM network to obtain a first output and a second output in sequence. The first output and the second output are concatenated and input into the GRU network, the attention mechanism network, and the DenseNet network in sequence. Finally, the tensor corresponding to the voltage signal of the cable to be tested is input into the trained fault detection network, which outputs two probabilities of whether the cable to be tested has a partial discharge fault, and the one with the higher probability is selected as the final detection result.

2. The method for detecting partial discharge of a ring main unit cable based on deep learning according to claim 1, characterized in that: The statistical features of each set of data are extracted in the frequency domain and the time domain respectively, and the statistical features are spliced ​​into a q-dimensional feature vector, including: Extract features in the time domain: calculate the mean, standard deviation, maximum and minimum values ​​of the signal amplitude of each group of data, as well as the percentage quantiles and relative percentage quantiles corresponding to percentages of 0, 1, 10, 25, 50, 75, 90, 99, and 100, as well as the kurtosis, skewness, number of peaks, maximum height of peaks, minimum height of peaks, average height of peaks, maximum width of peaks, minimum width of peaks, and average width of peaks for each group of data; Perform FFT transformation on each set of data and convert it into frequency domain signal; Extract features in the frequency domain: calculate the energy sum, maximum energy, minimum energy, standard deviation, variance, skewness, root mean square and rectified mean of the signal amplitude of each group of data converted into frequency domain signals; Finally, the feature quantities extracted in the time domain and the feature quantities extracted in the frequency domain are concatenated in sequence to obtain a 39-dimensional feature vector, and the value of q is 39.

3. The method for detecting partial discharge of a ring main unit cable based on deep learning according to claim 1, characterized in that: The convolutional neural network includes four convolution modules, namely the first convolution module, the second convolution module, the third convolution module and the fourth convolution module. Each of the convolution modules includes a one-dimensional convolution layer, a BN layer, a first RELU activation function and a Dropout layer arranged in sequence.

4. The method for detecting partial discharge in a ring main unit cable based on deep learning according to claim 3, characterized in that: The fault detection network is established, and an input tensor (k, n, q) is established with k batches, n feature types, and q-dimensional feature vectors to train the fault detection network. The fault detection network includes a convolutional neural network, an LSTM network, a GRU network, an attention mechanism network, and a DenseNet network. The input tensors are respectively input into the convolutional neural network and the LSTM network to obtain a first output and a second output in sequence. The first output and the second output are spliced ​​and then input into the GRU network, the attention mechanism network, and the DenseNet network in sequence, including: First, the input tensor (k, n, q) is transposed to obtain (k, q, n), and then (k, q, n) is sequentially input into the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module to obtain the first output, and passes through the one-dimensional convolution layer, the BN layer, the first RELU activation function, and the Dropout layer of each convolution module; Input the input tensor (k, n, q) to the LSTM network to obtain the second output, and the number of layers of the LSTM network is 5; Then, the first output is subjected to a second transposition process, the tensor after the second transposition process is concatenated with the second output to obtain a concatenated tensor, and the concatenated tensor is processed by a second RELU activation function to obtain a third output; Then the third output is sequentially input into the GRU network and the attention mechanism network to obtain the fourth output; Finally, the fourth output is transposed and input into the DensNet network to output two probabilities of whether there is a partial discharge fault.

5. The method for detecting partial discharge of a ring main unit cable based on deep learning according to claim 1, characterized in that: The fault detection network uses cross entropy as the loss function Loss, and the calculation formula is as follows: Loss=-(y·log(p)+(1-y)·log(1-p)) Where y represents the actual category label, including the label of the sample with partial discharge signal and the label of the sample without partial discharge signal, and p represents the probability that the fault detection network is labeled as the sample with partial discharge signal.

Images