Arc fault detection method based on wavelet packet transform and residual convolutional neural network

Abstract

The application discloses an arc fault detection method based on wavelet packet transform and residual convolutional neural network, which comprises the following steps: 1) collecting current signals in a circuit; 2) decomposing the current signals by two layers by using wavelet packet transform; 3) converting each one-dimensional wavelet packet component into a two-dimensional matrix by using a weight-truncated matrix construction method; 4) performing layer-by-layer nonlinear transformation on the two-dimensional matrix by using a residual convolutional neural network to obtain high-dimensional abstract features; and 5) inputting the high-dimensional features into a support vector machine to realize feature fusion and obtain a detection result. The application decouples fault information of different frequency band components by using wavelet packet transform. The weight-truncated matrix enables the subsequent feature extraction process to focus on key fault information and weaken the interference of redundant information. The residual convolutional neural network is used to realize adaptive extraction of features and reduce the overfitting degree of the network training process. Finally, the support vector machine is used to replace the softmax classifier, so that the high-dimensional features can be more effectively fused and the final detection accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The application relates to an arc fault detection method based on a wavelet packet transform and a residual convolutional neural network and belongs to the arc fault detection field. TECHNICAL BACKGROUND

[0002] Long-term operation of a civil low-voltage power distribution system can cause an arc fault, and high temperature generated by the arc fault can easily cause a fire, thereby causing serious personnel casualties and property losses. A parallel arc fault can generate overcurrent in a line, and protection can be achieved through a conventional circuit breaker. A series arc fault is equivalent to the introduction of impedance in the line, which can cause a slight reduction in current. The series arc fault has strong randomness and concealment, and a traditional circuit breaker is difficult to achieve protection for the series arc fault. Therefore, accurate and rapid detection of the series arc fault is of great significance to maintain the safe and stable operation of the civil low-voltage power distribution system.

[0003] Through analysis of detection methods of electromagnetic radiation signals, sound signals and light signals, the detector needs to be close to the fault position. The detection method based on the current signal is not limited by the position of the detector and is easy to implement, and is the focus of current research. Some scholars set a fixed threshold for the features extracted from the arc current signal to achieve arc fault detection. However, the threshold-based series arc fault detection method is easily affected by working conditions, and it is difficult to set a suitable threshold in practical applications.

[0004] Nowadays, more and more scholars apply machine learning methods to fuse multiple types of features to adapt to different working conditions. Joshua et al. extract different types of features from the current signal based on the discrete Fourier transform, the mel-frequency cepstral coefficient and the discrete wavelet transform, and input these features into a neural network to achieve series arc fault detection. Some researchers use random forests to analyze the importance of features and input features with high importance values into a neural network classifier. Some researchers use a stochastic resonance method to achieve feature enhancement, and extract power amplitudes and values of 7.8-15.6 kHz and 39.1-46.9 kHz, respectively, to construct a feature vector, and use a support vector machine as a classifier. Some researchers extract singular values from a time-frequency spectrum obtained from a fractional Fourier transform as inputs of a support vector machine classifier.

[0005] However, the above-mentioned machine learning-based methods have two shortcomings: 1) the manually extracted features depend on professional background experience and are difficult to comprehensively include fault information; and 2) the structure of the classifier is shallow and cannot accurately represent the complex nonlinear relationship between the features and the fault types.

[0006] At present, deep learning methods with deeper network structure are used to directly process arc current to automatically learn potential abstract fault features, thereby ensuring high detection accuracy and excellent generalization ability. Convolutional neural network is the most representative deep learning framework, in which convolution kernel simulates the bionic visual mechanism and can learn various local information in input space. Meanwhile, the weighted sharing mechanism is beneficial to reduce the complexity of convolutional neural network. Convolutional neural network is used to directly process arc current, avoiding the subjectivity of the above feature extraction method. The method can effectively extract deep abstract fault features and ensure the rapidity of online series arc fault detection, but the training process of traditional convolutional neural network is prone to overfitting problem. Some researchers introduce residual learning into convolutional neural network to weaken the overfitting phenomenon in the training process and improve the detection accuracy. Some researchers use time domain sequence and frequency spectrum energy sequence of arc current as inputs of two separate channels of residual convolutional neural network. However, only using original current time domain signal or frequency domain signal as the input of residual convolutional neural network, the fault information distributed in different frequency bands will interfere with each other, thereby negatively affecting the detection performance.

[0007] Deep learning method shows strong feature extraction ability and can detect series arc fault with high precision. However, the deep learning method for series arc fault detection currently has three defects: 1) only using original current signal or frequency spectrum energy sequence as the input of deep learning model, the overlap of fault information contained in different frequency bands increases the difficulty of feature extraction of deep learning model; 2) the above deep learning model has insufficient generalization ability and overfitting phenomenon in the training process; 3) the detection result based on the above convolutional neural network is obtained based on the softmax classifier, but the classification ability of the softmax classifier is limited. SUMMARY

[0008] To solve the above problems, the present application provides an arc fault detection method based on wavelet packet transform and residual convolutional neural network, which can realize adaptive extraction of arc fault features and improve the accuracy of arc fault detection, and avoid false alarm of normal line.

[0009] To achieve the above application purposes, the present application specifically adopts the following technical solutions:

[0010] Step 1): collect current signal data in the circuit under normal and arc fault conditions;

[0011] Step 2): divide the current signal data into sample data, then standardize the sample data, and label each sample with corresponding label;

[0012] Step 3): divide the collected sample data into training set and test set;

[0013] Step 4): decompose each sample data in the training set and the test set based on wavelet packet transform to obtain a plurality of one-dimensional wavelet packet components, and then convert the one-dimensional wavelet packet components into two-dimensional matrices;

[0014] Step 5): take the two-dimensional matrix corresponding to the training set sample data as the input of the residual convolutional neural network, and train the residual convolutional neural network;

[0015] Step 6): after the two-dimensional matrix corresponding to the training set sample data is input into the residual convolutional neural network, high-dimensional fault features can be obtained, and the support vector machine is trained using the high-dimensional fault features corresponding to the training set sample data;

[0016] Step 7): after the residual convolutional neural network and the support vector machine are trained, the two-dimensional matrix corresponding to the test set sample data is input into the residual convolutional neural network to obtain high-dimensional fault features, and then the high-dimensional fault features are input into the support vector machine to obtain the detection result, i.e., the detection label of the test set sample data;

[0017] Step 8): the detection result can be evaluated by comparing the detection label of the test set sample data with the actual label.

[0018] As a preferred, in step 1), based on a sampling frequency of 100 kHz, 2025 continuous data points are collected each time.

[0019] As a preferred, in step 2), each sample data is standardized according to formula (1):

[0020]

[0021] wherein, x i is the i-th data point in the sample data, x′ i is the normalized data point, min(x i ) is the minimum value in the sample data, and max(x i ) is the maximum value in the sample data.

[0022] As a preferred, in step 4), the current signal is decomposed by N layers (N can be adjusted according to actual application requirements) using wavelet packet transform to obtain 2 N one-dimensional wavelet packet components, which can decouple the fault components contained in different frequency bands.

[0023] As a preferred method, the 1D wavelet packet components are converted into 2D matrices using a weighted truncated matrix construction method. This method offers higher computational efficiency compared to the Hankel matrix construction method. Furthermore, compared to the traditional truncated matrix construction method, the weighted truncated matrix construction method allows the constructed matrix to focus on key fault information, reducing the interference of redundant information on the detection results. The weighted truncated matrix construction method is shown in formula (2):

[0024]

[0025] In the formula, u n×m This represents the m×n-th data point in the wavelet component. n = INT(Len / m), where INT(·) is the floor function, and Len is the length of the one-dimensional wavelet component, {λ1,λ2,…,λm}. n} represents the weight coefficients of the row vectors of different matrices.

[0026] To obtain the weight coefficients, we first need to calculate the j-th row vector M of matrix row M. j =[u (j-1)m+1 ,u (j-1)m+2 ,…,u jm The time segment signal N corresponding to the normalized current signal j ={x′ (j-1)m+1 ,x′ (j-1)m+2 ,…,x′ jm The normalized Pearson correlation coefficient between}, M j With N j The normalized Pearson correlation coefficient Pearson(j) is calculated as shown in formula (3).

[0027]

[0028] in, and N j and M j The average value, and N j and M j The standard deviation.

[0029] Then for M j The corresponding Pearson(j) can be normalized according to formula (4) to obtain λ. j

[0030]

[0031] As preferred, the two-dimensional matrix obtained in step 4) is directly input into the residual convolutional neural network through different channels to obtain deep-level abstract fault features, avoiding the subjectivity of manual feature extraction. The structure of the residual convolutional neural network comprises three convolutional modules and a fully connected layer, each convolutional module comprising a convolutional layer, a batch normalization layer, a pooling layer and a ReLU activation layer; the outputs of the first convolutional layer and the second convolutional layer are connected to the output of the third convolutional layer through the pooling layer, and two residual learning structures can simultaneously excavate arc fault features from different spatial scales, as shown in formula (5).

[0032] As preferred, the structure of the residual convolutional neural network comprises three convolutional modules and a fully connected layer, each convolutional module comprising a convolutional layer, a batch normalization layer, a pooling layer and a ReLU activation layer, the size of the convolutional kernel corresponding to the first convolutional module is (6x6), and the size of the convolutional kernel corresponding to the second convolutional module and the third convolutional module is (3x3). The number of input layer and output layer neurons of the fully connected layer is 2800 and 200, respectively. The structure of the present application adopts two residual connections, i.e. the outputs of the first convolutional layer and the second convolutional layer are connected to the output of the third convolutional layer through the pooling layer, so as to excavate features at different convolutional scales.

[0033] As preferred, the two residual learning structures of the residual convolutional neural network can simultaneously excavate arc fault features from different spatial scales, as shown in formula (5):

[0034] NN in = out conv3 + pool(out conv2 ) + pool(out conv1 ) (5)

[0035] In the formula, NN in is the input of the fully connected layer, out conv3 is the output of the third convolutional layer, out conv2 is the output of the second convolutional layer, out conv1 is the output of the first convolutional layer, and pool(·) is a pooling operator.

[0036] As preferred, a softmax classifier is used as the classifier in the training process of the residual convolutional neural network in step 5), and the specific training steps are as follows: the two-dimensional matrix corresponding to the training set sample data is input into the residual convolutional neural network, the residual convolutional neural network outputs the fault features into the softmax classifier for multi-classification; the training is stopped until the maximum number of iterations in the back propagation process is reached; after the training of the residual convolutional neural network is completed, the parameter freezing operation is performed, and the trained residual convolutional neural network is involved in the training process of the support vector machine.

[0037] As preferred, in step 5), the training process is performed by using a softmax classifier in combination with a residual convolutional neural network. After the residual convolutional neural network is trained, the softmax classifier is not involved in the subsequent steps. The subsequent steps use a support vector machine to replace the softmax classifier to obtain more accurate diagnostic results.

[0038] As preferred, the kernel function of the support vector machine adopts a Gaussian kernel function form, as shown in formula (6)

[0039]

[0040] In the formula, δ 2 is a kernel function parameter.e c , c e {1, 2, …, C} and e * are the corresponding features of the training set and the test set respectively, and C is the number of training set samples.

[0041] The present application has the beneficial effects of:

[0042] 1. The present application uses wavelet packet transform to decompose the original current into a plurality of wavelet components corresponding to different frequency bands, decouples the fault information of different frequency bands, and thus weakens the mutual interference between different component fault information. Then the weight-truncated matrix construction method is used to quickly convert these wavelet packet components into a two-dimensional matrix, which is input into the residual convolutional neural network through different channels. Compared with the traditional arc fault detection method, the method of the present application can make the residual convolutional neural network more easily learn the multi-level key fault information hidden in the time-frequency domain dynamic characteristics, and avoid the subjectivity of manual feature extraction.

[0043] 2. The weight-truncated matrix construction method proposed in the present application not only realizes fast matrix transformation, but also sets corresponding weight values for each row vector in the matrix, so that the subsequent residual convolutional neural network can focus on important fault information, which is beneficial to reduce the training difficulty of the residual convolutional neural network and improve the detection accuracy.

[0044] 3. Two residual structure units are introduced into the convolutional neural network, and the outputs of different convolutional blocks are linked to the fully connected layer, which is conducive to the learning of rich fault information under different spatial scales. In addition, this mechanism can reduce the risk of overfitting in the training process and improve the accuracy of series arc fault detection.

[0045] 4. The present application uses a support vector machine to replace the softmax classifier of the residual convolutional neural network, and compared with the softmax classifier, the support vector machine has stronger stability and generalization ability. The combination of the residual convolutional neural network and the support vector machine can further improve the arc fault detection performance. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a method flowchart of one embodiment of the present application;

[0047] Figure 2 is an AC power supply system block diagram of one embodiment of the present application;

[0048] Figure 3 is a structure diagram of a residual convolutional neural network of one embodiment of the present application;

[0049] Figure 4 is a schematic diagram of training and testing of a residual convolutional neural network and a support vector machine of one embodiment of the present application;

[0050] Figure 5 is an offline diagnosis result confusion matrix of one embodiment of the present application;

[0051] Figures 6-7 is an online diagnosis result waveform diagram of one embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] The present application utilizes wavelet packet transform to obtain 8 wavelet packet component sequences after three-layer decomposition and reconstruction of the original current signal. Based on the truncated matrix construction method, the one-dimensional wavelet packet component sequence is then constructed into a two-dimensional matrix. Finally, the 8 two-dimensional matrices are directly input into the residual convolutional neural network, the deep fault feature extraction capability of the residual convolutional neural network is utilized to obtain abstract fault features, and the fault features are input into the support vector machine to obtain the diagnosis result.

[0054] Embodiment 1

[0055] The present embodiment provides an arc fault detection method based on wavelet packet transform and residual convolutional neural network, which is suitable for residential power supply systems, such as Figure 1 As shown, the method comprises the following steps:

[0056] Step 1): First, build a Figure 2The AC power supply system shown, the system contains 1 main circuit, the main circuit includes AC power supply, load and current sensor, the arc generator is connected to the main circuit, which can simulate series arc fault. The current signal of the main circuit is collected by the current sensor, and the signal is transmitted to the microprocessor through the data acquisition board card. The sampling frequency is 100 kHz, and each sample contains 2025 data points. In this embodiment, the system uses 7 types of loads, respectively, induction cooker, resistance box, hair dryer, electric oven, vacuum cleaner, motor and AC-DC. Table 1 shows the labels corresponding to different states of each load. 2000 groups of sample data are collected under each load type, of which 1000 groups of sample data correspond to the normal state, and the other 1000 groups of data correspond to the arc state. Therefore, 14000 groups of sample data are collected for 7 types of loads. Each sample is labeled with the corresponding label, as shown in Table 1 below:

[0057] Table 1 Label assignment under different load conditions

[0058]

[0059]

[0060] Step 2): Randomly select 60% of the 14000 sample data as training data set, and the remaining 40% sample data as test set. Each sample data is standardized according to formula (1):

[0061]

[0062] Step 3): Use wavelet packet transform to decompose the sample data in the training set and test set into 2 layers, get 4 1-dimensional wavelet packet components, and use the weight truncated matrix construction method to convert the 1-dimensional wavelet packet components to 2-dimensional matrix, so each sample data can get 4 two-dimensional matrices. The weight truncated matrix construction method is shown in formula (2).

[0063]

[0064] In the formula, u n×m represents the m*n data point in the wavelet component, n = INT(Len / m), INT(·) is the down rounding operation, and Len is the length of the one-dimensional wavelet component. Since each wavelet component has the same number of data points as the original sample data, Len = 2025, m = n = 45. {λ1,λ2,…,λ n} are the weight coefficients of different matrix row vectors in the matrix, so each matrix contains n = 45 weight coefficients.

[0065] To obtain the weight coefficients, first calculate the jth row vector M j = [u(j-1)×45+1 u (j-1)×45+2 ,…,u j×45 ] and the normalized current signal corresponding time segment signal N j ={x ( ′ j-1)m+1 ,x ( ′ j-1)m+2 ,…,x j ′ m} between the normalized Pearson correlation coefficient, M j and N j The calculation method of Pearson(j) is shown in formula (3)

[0066]

[0067] Wherein, and are the average values of N j and M j , and are the standard deviations of N j and M j .

[0068] Then the Pearson(j) corresponding to M j is normalized according to formula (4) to obtain λ j

[0069]

[0070] Step 4): determine the structure of the residual convolutional neural network and the kernel function type of the support vector machine.

[0071] The residual convolutional neural network adopted by the present application can perform adaptive feature extraction, avoiding the subjectivity of manual feature extraction.

[0072] The structure of the residual convolutional neural network is as follows Figure 3As shown in Table 2, the residual convolutional neural network comprises three convolutional modules and one fully connected layer, each convolutional module comprises a convolutional layer, a batch normalization layer, a pooling layer and a ReLU activation layer, the first convolutional module corresponds to a convolutional kernel with a size of (6x6), the second convolutional module and the third convolutional module correspond to a convolutional kernel with a size of (3x3). The input and output sizes of the first convolutional layer Conv1 are 4 and 16 respectively, the input and output sizes of the second convolutional layer Conv2 are 16 and 32 respectively, and the input and output sizes of the third convolutional layer Conv3 are 32 and 64 respectively. The convolutional kernel sliding step of the three convolutional layers is 1. The number of neurons of the input layer and the output layer of the fully connected layer is 2800 and 200 respectively. The sizes of the pooling layer 1 and the pooling layer 2 are 2x2 and 4x4 respectively, and the convolutional kernel sliding steps of the pooling layer 1 and the pooling layer 2 are 2 and 4 respectively.

[0073] The application adopts a structure of two residual connections, that is, the outputs of the first convolutional layer Conv1 and the second convolutional layer Conv2 are connected to the output of the third convolutional layer Conv3 after being processed by the pooling layer 1 and the pooling layer 2 respectively, so as to mine the features under different convolutional scales. Based on the two residual learning structures, the arc fault features under different spatial scales can be simultaneously mined, as shown in formula (5).

[0074] NN in = out conv3 + pool (out conv2 ) + pool (out conv1 ) (5)

[0075] In the formula, NN in is the input of the fully connected layer, out conv3 is the output of the third convolutional layer, out conv2 is the output of the second convolutional layer, out conv1 is the output of the first convolutional layer. pool(·) is a pooling operator.

[0076] Table 2 Detailed parameters of residual convolutional neural network

[0077]

[0078] The kernel function of the support vector machine adopts a Gaussian kernel function form, as shown in formula (6)

[0079]

[0080] In the formula, δ 2 is a kernel function parameter, e c , c∈{1,2,…,C} and e * are features corresponding to the training set and the test set respectively, C is the number of training set samples, C=8400.

[0081] Step 5): Figure 4 The specific process of training and testing residual convolutional neural networks (RNNs) and support vector machines (SVMs) is presented. The training process of the RNN and SVM is shown in the figure. The two-dimensional matrix corresponding to the training set sample data is used as the input to the RNN for training. A softmax classifier is used as the classifier during the RNN training process. Training stops when the backpropagation process reaches the maximum number of iterations. After training, the parameters of the RNN are frozen, and then the trained RNN participates in the training process of the SVM.

[0082] The two-dimensional matrix corresponding to the training set sample data is input into the trained residual convolutional neural network to obtain high-dimensional fault features. These high-dimensional fault features are then used to train the support vector machine. After training, the parameters are frozen. The two-dimensional matrix corresponding to the test set sample data is then input into the trained residual convolutional neural network to obtain high-dimensional fault features. These high-dimensional fault features are then input into the trained support vector machine to obtain the detection results (i.e., the labels corresponding to the test set sample data).

[0083] This embodiment combines wavelet packet transform with a residual convolutional neural network. Wavelet packet transform decomposes the current signal into a finer structure, decoupling fault components across different frequency bands and preventing interactions between fault information from different frequency bands, thus solving the problem of mutual interference between fault information from different frequency bands. Simultaneously, this invention proposes a weighted truncated matrix construction method, which can quickly convert one-dimensional wavelet packet components into two-dimensional matrices, ensuring the speed of arc fault detection. Furthermore, by setting weight coefficients for different row vectors in the matrix, the subsequent feature extraction process can focus on key fault information, reducing the interference of redundant information on the detection results. The combination of residual convolutional neural networks and support vector machines makes it easier to extract detailed differences between different categories from the data. Based on the idea of ​​residual learning, this embodiment uses two residual module units to achieve skip connections, reducing the risk of overfitting during training. This allows the network to simultaneously mine fault information from different spatial scales, more effectively extracting abstract fault features from the current signal, improving detection performance and enhancing the diversity of fault information. This embodiment uses a support vector machine to replace the softmax classifier in the original residual convolutional neural network, which more accurately fits the complex nonlinear relationship between fault features and fault types, and further enhances the generalization ability of the proposed method.

[0084] Figure 5The confusion matrix of the off-line diagnosis result of the method is given, and the overall detection accuracy is 99.92% (4996 / 5000), and the detection accuracy under normal operation is 100% (2504 / 2504). The results show that the method can accurately identify the load type according to the current signal, and effectively avoid misjudgment. Under the condition of arc fault, the detection accuracy reaches 99.83% (2492 / 2496). The detection accuracy under the linear load (resistor box and electric oven) is 100%, and there is no misjudgment. The detection accuracy of the nonlinear load (induction cooker, electric hair dryer, vacuum cleaner, electric motor, AC-DC) is 100%, 100%, 99.72%, 99.41% and 99.71% respectively. It shows that the application can accurately distinguish the arc fault under different load conditions.

[0085] When the application does not use wavelet packet transform, the overall detection accuracy is 96.36%. The experimental results show that by separating the fault information contained in different frequency bands through wavelet packet transform, the mutual interference of the fault information contained in different frequency bands in the feature extraction process can be avoided, and the fault detection accuracy can be effectively improved.

[0086] When the application does not use the traditional truncated matrix construction method, the overall detection accuracy is 99.32%. The experimental results show that the weight truncated matrix construction method can effectively improve the detection accuracy by focusing on key fault information, and at the same time will not increase the calculation amount of online detection.

[0087] The application embeds the algorithm into the microprocessor shown in Figure 2 to realize online arc fault detection. Figure 6 and Figure 7 is the online detection result. Under normal operation, the method can avoid misjudgment. The method can detect arc fault at a speed of 0.051s and 0.63s. In this paper, 20 online experiments are carried out under each load type. The online detection accuracy is 100% (140 / 140), and the maximum detection time is 0.076s. The results show that the method can accurately detect SAF, and the detection speed meets the requirements of 1699 standard.

[0088] The basic principles and main features of the application and the advantages of the application are shown and described. Those skilled in the art should understand that the application is not limited to the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection of the application is defined by the appended claims and their equivalents.

Claims

1. An arc fault detection method based on wavelet packet transform and residual convolutional neural network, characterized in that, Includes the following steps: Step 1): Collect sample data of the current signal in the circuit under normal conditions and under arc fault conditions respectively; Step 2): Standardize the sample data and label each sample accordingly; Step 3): Divide the collected sample data into training set and test set; Step 4): Based on wavelet packet transform, decompose each sample data in the training and test sets to obtain several one-dimensional wavelet components. Then, use the weighted truncation matrix construction method to convert the one-dimensional wavelet components into two-dimensional matrices. The weighted truncation matrix construction method is shown in the following formula: ， In the formula, In the wavelet component, the first Data points, , , For floor operations, The length of the one-dimensional wavelet component, The weight coefficients for different matrix row vectors are calculated as follows: First, calculate the weight coefficient of the j-th row vector in matrix M. Time segment signal corresponding to normalized current signal The normalized Pearson correlation coefficient between them and Normalized Pearson correlation coefficient The calculation method is as follows: ， in, and They are respectively and The average value, They are respectively and Standard deviation; Then to corresponding Normalization using the following formula yields the result. : ； Step 5): Use the two-dimensional matrix corresponding to the training set sample data as the input to the residual convolutional neural network to train the residual convolutional neural network; Step 6): The residual convolutional neural network outputs high-dimensional fault features, and the support vector machine is trained using the high-dimensional fault features corresponding to the training set sample data. Step 7): Input the two-dimensional matrix corresponding to the test set sample data into the trained residual convolutional neural network to obtain high-dimensional fault features. Then input the high-dimensional fault features into the trained support vector machine to obtain the detection result, which is the detection label of the test set sample data.

2. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 1, characterized in that, In step 2), each sample data is standardized according to the following formula: ， in, For the i-th data point in the sample data, For the normalized data points, The minimum value in the sample data. This is the maximum value in the sample data.

3. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 1, characterized in that, In step 4), the current signal is decomposed into N levels using wavelet packet transform to obtain... Each 1D wavelet packet component can decouple fault components contained in different frequency bands.

4. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in any one of claims 1 to 3, characterized in that, The two-dimensional matrix obtained in step 4) is directly input into the residual convolutional neural network through different channels to obtain deep-level abstract fault features. The residual convolutional neural network contains three convolutional modules and one fully connected layer. Each convolutional module includes a convolutional layer, a batch normalization layer, a pooling layer, and a ReLU activation layer. The outputs of the first and second convolutional layers are connected to the output of the third convolutional layer through the pooling layer.

5. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 4, characterized in that, The two residual learning structures based on the residual convolutional neural network can simultaneously mine arc fault features from different spatial scales, as shown in the following equation: ， In the formula, For the input of the fully connected layer, This is the output of the third convolutional layer. This is the output of the second convolutional layer. This is the output of the first convolutional layer. This is the pooling operator.

6. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 1, characterized in that, Step 5) The training process of the residual convolutional neural network uses a softmax classifier as the classifier. The specific training steps are as follows: the two-dimensional matrix corresponding to the training set sample data is used as the input of the residual convolutional neural network. The residual convolutional neural network inputs the output fault features into the softmax classifier for multi-classification. Training stops when the backpropagation process reaches the maximum number of iterations. After the training is completed, the residual convolutional neural network performs a parameter freezing operation and participates in the training process of the support vector machine.

7. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 6, characterized in that, In step 5), the softmax classifier is used in conjunction with the residual convolutional neural network to perform the training process. After the residual convolutional neural network is trained, the softmax classifier does not participate in subsequent steps.

8. The arc fault detection method based on wavelet packet transform and residual convolutional neural network as described in claim 1, characterized in that, The kernel function of the support vector machine adopts the form of a Gaussian kernel function, as shown in the following formula: ， In the formula, These are the parameters for the kernel function. and e * These are the features corresponding to the training set and the test set, respectively, and C is the number of samples in the training set.

Images