A cable early fault diagnosis method and system

Abstract

The present invention discloses a method and system for diagnosing early faults of cables, and relates to the technical field of fault diagnosis of distribution network equipment. The method comprises the following steps: obtaining current data in a cable to be tested in real time and preprocessing the data; extracting fault features of the preprocessed current data by combining wavelet transform and bispectrum theory to obtain a two-dimensional fault feature map; inputting the feature map into a fault detection model based on an improved residual network, and outputting early fault detection and classification results of cables; wherein the fault detection model uses a convolutional neural network as a backbone network, and adds an improved residual module to the backbone network, and the improved residual module includes three branches, each branch performs convolution of different scales on the feature map and then fuses it to extract global features; the global features are then classified through a fully connected layer, and the early fault detection and classification results of cables are output. The present invention can extract more comprehensive signal features, and achieve higher-precision fault detection through an improved detection model.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network equipment fault diagnosis, and in particular to a cable early fault diagnosis method and system. Background Art

[0002] At present, the classification and identification methods of early cable faults are mainly divided into threshold differentiation method and neural network method. Among them, the threshold differentiation method distinguishes the fault type by comparing the threshold value. However, in the actual power system, the line structure is complex, there is a strong correlation and coupling relationship between the components, and the operating conditions are uncertain. It is difficult to determine the appropriate threshold according to the actual power system using the threshold differentiation method. Due to the uncertainty in the selection of thresholds, the neural network method is currently used. The neural network method refers to extracting features from fault data through machine learning or deep learning, and completing fault discrimination based on the extracted features. Although this method does not require a threshold, it usually requires the use of signal processing methods such as wavelet transform to analyze the time and frequency domain characteristics of the fault signal, and use this as a basis to infer the fault type. However, this method still has certain problems: on the one hand, the fault signal of the early cable fault is too weak in the entire cable transmission signal, and its fault features are difficult to extract. In addition, there is a certain similarity between the fault features of different fault types. It is difficult to extract comprehensive and effective fault features using only wavelet transform, which affects the accuracy of the final fault identification and classification; on the other hand, the traditional classifier has poor classification and recognition capabilities for the extracted features, resulting in low accuracy in the final fault identification. Summary of the invention

[0003] In order to solve the shortcomings of the above-mentioned prior art, the present invention provides a cable early fault diagnosis method and system, which extracts features of the cable current signal based on the characteristics of the current signal in the early cable fault, integrates wavelet transform and bispectrum theory, and can take into account the time and frequency domain characteristics, and can more effectively detect intermittent signals in the signal and extract more accurate fault signal features; adopts a fault detection model based on an improved residual network to deeply mine more potential features in the fault data, and avoids the loss of local features to a certain extent, so as to further improve the performance and accuracy of fault identification and classification.

[0004] In a first aspect, the present invention provides a method for early fault diagnosis of a cable.

[0005] A method for diagnosing early cable faults, comprising:

[0006] Acquire the current data in the cable to be tested in real time and pre-process the current data;

[0007] Combined with wavelet transform and bispectrum theory, the fault characteristics of pre-processed current data are extracted to obtain a two-dimensional fault characteristic map;

[0008] The two-dimensional fault feature map is input into the fault detection model based on the improved residual network, and the cable early fault detection and classification results are output; wherein, the fault detection model based on the improved residual network uses a convolutional neural network as the backbone network, and an improved residual module is added to the backbone network. The improved residual module includes three branches, each branch performs convolution on the feature map at different scales and then fuses them to extract global features; the global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

[0009] In a second aspect, the present invention provides a cable early fault diagnosis system.

[0010] A cable early fault diagnosis system, comprising:

[0011] The data acquisition and preprocessing module is used to acquire the current data in the cable to be tested in real time and preprocess the current data;

[0012] The feature extraction module is used to extract the fault features of the pre-processed current data by combining wavelet transform and bispectral theory to obtain a two-dimensional fault feature map;

[0013] The early fault diagnosis module is used to input the two-dimensional fault feature map into the fault detection model based on the improved residual network, and output the cable early fault detection and classification results; wherein, the fault detection model based on the improved residual network uses a convolutional neural network as the backbone network, and an improved residual module is added to the backbone network. The improved residual module includes three branches, each branch performs convolution on the feature map at different scales and then fuses them to extract global features; the global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

[0014] One or more of the above technical solutions have the following beneficial effects:

[0015] 1. The present invention provides a method and system for diagnosing early cable faults. According to the characteristics of the current signal in the early cable fault, the fault characteristics of the current signal of the cable to be tested are extracted by integrating wavelet transform and bispectrum theory for joint analysis. This method can not only take into account the time-frequency domain characteristics, but also more effectively detect the intermittent signals in the signal and effectively extract the fault signal characteristics; a fault detection model based on an improved residual network is adopted to more comprehensively and deeply mine the fault characteristics, and to a certain extent avoid the loss of local features, so as to further improve the performance and accuracy of fault identification and classification.

[0016] 2. The present invention realizes the identification of early-stage cable faults in the distribution network, and can effectively identify early-stage cable faults from overcurrents caused by capacitor switching, excitation inrush current, capacitor switching, etc., and has a wide range of applications in the identification of early-stage cable faults.

[0017] 3. The fault detection model based on the improved residual network proposed in the present invention designs and improves the residual network, and deeply extracts hidden features in the data through multiple branches in the residual network, which has higher accuracy than traditional pattern recognition methods. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed descriptions are exemplary only, are intended to describe specific embodiments, are intended to provide further explanation of the present invention, and are not intended to limit exemplary embodiments according to the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art to which the present invention belongs. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0019] Embodiment 1

[0020] In view of the problems of incomplete feature extraction and poor detection model recognition ability in the existing wavelet transform-based early fault diagnosis method for cables, this embodiment proposes a method for early fault diagnosis of cables, which specifically includes the following steps:

[0021] Step S1, obtaining the current data in the cable to be tested in real time, and preprocessing the current data;

[0022] Step S2, combining wavelet transform and bispectrum theory to extract fault features of the preprocessed current data to obtain a two-dimensional fault feature map;

[0023] Step S3, input the two-dimensional fault feature map into the fault detection model based on the improved residual network, and output the cable early fault detection and classification results; wherein, the fault detection model based on the improved residual network uses a convolutional neural network as the backbone network, and adds an improved residual module to the backbone network, the improved residual module includes three branches, each branch performs convolution on the feature map at different scales and then fuses them to extract global features; the global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

[0024] The following content introduces the cable early fault diagnosis method proposed in this embodiment in more detail.

[0025] In the above step S1, the current data in the cable to be tested is obtained in real time, and the current data is preprocessed. Specifically, firstly, the current data in the cable to be tested is obtained according to the set sampling frequency, and the current data is actually a current signal time series (i.e., a current waveform); then, the obtained current data is preprocessed, that is, the current signal time series is divided to obtain a plurality of current time series signals of length N, so as to facilitate the subsequent signal feature extraction and detection.

[0026] Generally, the duration of early cable faults is very short, and their fault characteristics are relatively weak in the current signal. In addition, there are certain similarities between the fault characteristics of different fault types, which makes fault identification and classification difficult. To address this problem, in this embodiment, current waveforms of various early fault types (including cable half-cycle early faults, multi-cycle early faults, permanent faults, excitation inrush currents, capacitor switching, and load surges, etc.) are obtained through simulation, wherein the sampling frequency when the waveform is obtained is 10kHz, that is, there are 200 sampling points in each power frequency cycle, and the current waveforms of various cable early faults are analyzed. According to the simulation waveform, it can be seen that for a half-cycle early fault in the cable, the current waveform will undergo a disturbance process of a certain duration (about 1 / 4 cycle duration), and the duration and amplitude of this disturbance are close to the disturbance during capacitor switching. Therefore, the overcurrent disturbance caused by the half-cycle early fault of the cable and the capacitor switching is easily confused, and it is necessary to distinguish the capacitor switching when identifying the half-cycle early fault; for multi-cycle early faults, the amplitude of the waveform disturbance and the duration of the disturbance are similar to the characteristics of the excitation inrush current generated when the transformer is no-load closed, so it is necessary to distinguish the excitation inrush current when identifying the multi-cycle early fault. In addition, this embodiment also simulates overcurrent disturbances caused by permanent faults and sudden load increases to verify the ability of the proposed method to distinguish other disturbance waveforms.

[0027] In view of the problem that the above fault features are weak and there are certain similarities between different fault features, this embodiment proposes a joint analysis method of wavelet transform and bispectrum theory, through which more comprehensive fault features are accurately extracted, and the ability to distinguish non-stationary signals is improved, solving the problem of difficulty in identification due to high similarity. That is, in the above step S2, after extracting the corresponding current time series signal, the fault features of the pre-processed current data are extracted by combining wavelet transform and bispectrum theory to obtain a two-dimensional fault feature map.

[0028] In this embodiment, a joint analysis method of wavelet transform and bispectral theory is introduced in feature extraction, and instantaneous information is enhanced by wavelet transform, and the limited representation of non-stationary signals by the second-order spectrum represented by the power spectrum is compensated by the high-order spectrum. Compared with the single wavelet transform method, this embodiment organically integrates wavelet transform and bispectral theory, which can take into account the time-frequency characteristics, more effectively detect the secondary nonlinear coupling characteristics existing in the signal, and improve the resolution of intermittent signals (i.e., early fault signals), making it more suitable for the analysis of transient short data signals, and can effectively distinguish the above-mentioned multiple early fault types.

[0029] Specifically, the fault characteristics of the pre-processed current data are extracted by combining wavelet transform and bispectrum theory, including the following steps:

[0030] First, the current time series signal is decomposed in the time-frequency domain using wavelet transform to obtain sub-signals in different frequency bands. Specifically, this embodiment uses wavelet transform for decomposition, uses wavelet coefficients as the central element of calculation, and converts the scaled mother wavelet function With current timing signal Perform convolution operation, the formula is:

[0031] ;

[0032] in, Indicates a moment, represents the translation factor, Indicates the current timing signal The total number of samples in is the scaling item, Indicates frequency, represents the scaling factor, Indicates conjugation.

[0033] Secondly, the bispectral density function of the sub-signal on each frequency band is calculated to obtain a bispectral density matrix. In this embodiment, the formula of the bispectral density matrix obtained by combining wavelet transform and bispectral theory analysis can be expressed as:

[0034] ;

[0035] in, , , Indicates that the frequency Decomposed frequency.

[0036] Afterwards, the bispectral density matrices of the sub-signals on different frequency bands are combined to obtain a global bispectral density matrix; the global bispectral density matrix is ​​decomposed to obtain a wavelet bispectral feature map, that is, a two-dimensional fault feature map.

[0037] The fault characteristic diagram of the current signal is obtained in the above manner, and then step S3 is performed, that is, the two-dimensional fault characteristic diagram is input into the fault detection model based on the improved residual network, so as to output the early fault detection and classification results of the cable.

[0038] This embodiment builds a fault detection model based on an improved residual network. The fault detection model uses a convolutional neural network as the backbone network, and adds an improved residual module to the backbone network. The improved residual module includes three branches. Each branch performs convolution on the features at different scales and then fuses them, so that the global features can be deeply extracted. Finally, the global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

[0039] Convolutional neural networks (CNN) are feedforward and constrained neural networks with convolution calculations and deep structures. CNN has three main characteristics: local receptive field, weight sharing, and pooling operation. It is usually composed of an input layer, a hidden layer, and an output layer, and the hidden layer includes a convolution layer, a pooling layer, a fully connected layer, etc. In this embodiment, the fault detection model based on the improved residual network includes an input layer, a convolution layer, a BN layer (Batch Normalization, batch normalization layer), a pooling layer, an improved residual module, a fully connected layer, and an output layer connected in sequence; wherein the improved residual module includes three branches, the first branch sequentially uses a plurality of convolution kernels of different sizes to convolve the input features; the second branch uses a single convolution kernel to convolve the input features and then stacks them with the output features of the first branch, and performs convolution operation again after stacking; the third branch directly splices the input features with the features output by the second branch, and outputs the spliced ​​features.

[0040] Specifically, the convolution layer is the most important component of CNN, and the convolution layer is composed of several convolution kernels. In this embodiment, 1 to 3 convolution layers can be set in the convolutional neural network. Generally, the more convolution layers there are, the more complex the extracted features are, and the stronger the nonlinear expression ability of the features is. Each convolution layer performs a convolution operation on the output of the previous layer through the convolution kernel, and then extracts the upper layer features. A pooling layer is connected after each convolution layer. Considering that the dimension of the features extracted by the convolution layer is still large, the dimension size of the features is reduced by pooling operation, thereby further reducing the amount of network calculation. A BN layer is set between the convolution layer and the pooling layer. Before the result of the output of the previous convolution layer is input to the next layer of the network, the input is normalized to solve the problem of changes in the distribution of data in the middle layer during training. BN can prevent the gradient from disappearing or exploding, and reduce the risk of overfitting of the network. The fully connected layer is set at the end of the improved residual module. After the input sample is extracted by multiple convolutional layers and reduced in dimension by the pooling layer, all feature information is integrated through the fully connected layer and passed to the output layer, which outputs the final detection result. The output of the fully connected layer is:

[0041] ;

[0042] In the above formula, X is the output of the previous layer of the fully connected layer, is the weight matrix from the input layer to the hidden layer, is the corresponding bias vector, is the activation function.

[0043] The output value Y of the fully connected layer is used as the input of the output layer, and the output result of the output layer is:

[0044] ;

[0045] In the above formula, is the Softmax regression function, which is used to classify the input results of the fully connected layer. is the weight matrix from the hidden layer to the output layer, is the corresponding bias vector.

[0046] Furthermore, considering that when the number of network layers exceeds a certain number, the generalization performance of the network reaches saturation, or even degrades, which will cause the training accuracy and test accuracy to decrease at the same time. This is often caused by the disappearance of gradients or the explosion of gradients, which makes it impossible to update the weight coefficients of the network. This problem limits the depth of the network model to a certain extent. To this end, on the basis of the above-mentioned convolutional neural network, this embodiment also adds a residual module to solve the problem of network performance degradation. Based on the residual module, the features of the upper-layer input can be retained through constant mapping processing during network training, avoiding the loss of features caused by excessive extraction and compression of the upper-layer input features after multi-layer convolution operations, thereby reducing the risk of overfitting. That is, by adding a residual module, the training error of the network model will not increase even if the number of network layers is increased.

[0047] Considering that the horizontal and vertical directions of the feature map input in this embodiment respectively contain the time domain information and frequency domain information of the feature, the size of the convolution kernel in each convolution layer in the conventional residual structure is symmetrical, and the feature size extracted each time is fixed, and the time domain and frequency domain feature information cannot be well distinguished. Therefore, this embodiment improves the residual module and adds the improved residual module to the CNN backbone network. Through the improved residual module, not only a large amount of computing resources can be saved, but also the depth of the network can be further deepened.

[0048] Specifically, the improved residual module includes three branches: the third branch directly outputs the input without processing through constant mapping; the second branch is convolved through a 1×1 convolution kernel to reduce the dimension of the upper layer input and reduce the number of channels; the first branch is convolved through 1×1, 1×3, and 3×1 convolution kernels in sequence, and then the output results of the latter two branches are stacked, and then a convolution with a convolution kernel size of 1×1 is performed, and finally the output result is stacked with the output of the third branch. Based on the above settings, the third branch ensures that the network does not degenerate, the first branch and the second branch are used to reduce the optimization parameters, and the first branch enables the network to have local receptive fields of different scales, thereby fully mining the time domain information and frequency domain information of the input features. By stacking the results of the first branch and the second branch, all feature information is integrated as much as possible to improve the efficiency of feature extraction.

[0049] In addition, in the above fault detection model, the convolution kernel sizes of the improved residual module are different, which is conducive to the convolution layer to extract fault features of different sizes, thereby fully extracting the time domain information and frequency domain information of the input features: the parameters required to be optimized during training of the model are greatly reduced, which improves the network training speed and reduces convolution operation resources; by improving the identity mapping of the residual module, it is ensured that the performance of the network model does not degrade, and at the same time, the fault features are deeply mined through the convolution layer to ensure the network classification effect.

[0050] Furthermore, the training process of the above-mentioned fault detection model based on the improved residual network includes:

[0051] Simulate various types of early-stage cable faults to obtain current simulation data; specifically, according to the characteristics of different types of early-stage cable faults, use simulation software to establish a circuit model, simulate and obtain current waveforms of different disturbance types, thereby forming a simulation data set;

[0052] Preprocessing the current data with early fault types marked in the simulation data set;

[0053] Combined with wavelet transform and bispectrum theory, the fault characteristics of pre-processed current data are extracted to obtain a two-dimensional fault characteristic map;

[0054] The two-dimensional fault feature map is used for model training, and a fault detection model based on an improved residual network is obtained. The fault detection model can be used to detect and identify early cable faults.

[0055] The improved residual network-based fault detection model trained in this embodiment is used to perform early cable fault detection, thereby achieving higher performance and accuracy in fault identification and classification.

[0056] Embodiment 2

[0057] This embodiment provides a cable early fault diagnosis system, including:

[0058] The data acquisition and preprocessing module is used to acquire the current data in the cable to be tested in real time and preprocess the current data;

[0059] The feature extraction module is used to extract the fault features of the pre-processed current data by combining wavelet transform and bispectrum theory to obtain a two-dimensional fault feature map;

[0060] The early fault diagnosis module is used to input the two-dimensional fault feature map into the fault detection model based on the improved residual network, and output the cable early fault detection and classification results; wherein, the fault detection model based on the improved residual network uses a convolutional neural network as the backbone network, and an improved residual module is added to the backbone network. The improved residual module includes three branches, each branch performs convolution on the feature map at different scales and then fuses them to extract global features; the global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

[0061] The steps involved in the above embodiment 2 correspond to those in the method embodiment 1. For the specific implementation method, please refer to the relevant description part of the embodiment 1.

[0062] The above description is only a preferred embodiment of the present invention. Although the specific implementation mode of the present invention is described, it is not a limitation of the protection scope of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the protection scope of the present invention.

Claims

1. A method for diagnosing early cable faults, characterized in that: include: Acquire the current data in the cable to be tested in real time and pre-process the current data; Combined with wavelet transform and bispectrum theory, the fault characteristics of pre-processed current data are extracted to obtain a two-dimensional fault characteristic map; The two-dimensional fault feature map is input into the fault detection model based on the improved residual network, and the early cable fault detection and classification results are output; wherein, the fault detection model based on the improved residual network uses a convolutional neural network as the backbone network, and an improved residual module is added to the backbone network, and the improved residual module includes three branches, the first branch sequentially uses a plurality of convolution kernels of different sizes to convolve the input features; the second branch uses a single convolution kernel to convolve the input features and then stacks them with the output features of the first branch, and then performs convolution operation again after stacking; the third branch directly splices the input features with the features output by the second branch, and outputs the spliced ​​features to extract global features; The global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

2. A cable early fault diagnosis method as claimed in claim 1, characterized in that: According to the set sampling frequency, the current data in the cable to be tested is obtained, wherein the current data is a time series of the current signal; The acquired current data is preprocessed, and the preprocessing includes dividing the current signal time series to obtain a plurality of current time series signals with a length of N.

3. A cable early fault diagnosis method as claimed in claim 1, characterized in that: Combined with wavelet transform and bispectrum theory, the fault characteristics of the pre-processed current data are extracted to obtain a two-dimensional fault feature map, including: The current time series signal is decomposed in the time-frequency domain using wavelet transform to obtain sub-signals in different frequency bands. Calculate the bispectral density function of the sub-signal on each frequency band to obtain a bispectral density matrix; Merge the bispectral density matrices of the sub-signals in different frequency bands to obtain the global bispectral density matrix; The global bispectral density matrix is ​​decomposed to obtain the wavelet bispectral feature map, that is, the two-dimensional fault feature map.

4. A cable early fault diagnosis method as claimed in claim 1, characterized in that: The fault detection model based on the improved residual network includes an input layer, a convolutional layer, a BN layer, a pooling layer, an improved residual module, a fully connected layer and an output layer which are connected in sequence.

5. A cable early fault diagnosis method as claimed in claim 1, characterized in that: The training process of the fault detection model based on the improved residual network includes: Simulate various types of early cable faults to obtain current simulation data and build simulation data sets; Preprocessing the current data with early fault types marked in the simulation data set; The fault characteristics of the pre-processed current data are extracted by joint analysis of wavelet transform and bispectrum theory, and a two-dimensional fault characteristic diagram is obtained; The two-dimensional fault feature map is used for model training, and a fault detection model based on the improved residual network is obtained through training.

6. A cable early fault diagnosis method as claimed in claim 5, characterized in that: The early fault types of cables include half-cycle early fault, multi-cycle early fault, capacitor switching, excitation inrush current, load surge and permanent fault.

7. A cable early fault diagnosis system, characterized in that: include: The data acquisition and preprocessing module is used to acquire the current data in the cable to be tested in real time and preprocess the current data; The feature extraction module is used to extract the fault features of the pre-processed current data by combining wavelet transform and bispectrum theory to obtain a two-dimensional fault feature map; An early fault diagnosis module is used to input a two-dimensional fault feature map into a fault detection model based on an improved residual network, and output cable early fault detection and classification results; wherein the fault detection model based on the improved residual network uses a convolutional neural network as a main network, and adds an improved residual module to the main network, and the improved residual module includes three branches, the first branch sequentially uses a plurality of convolution kernels of different sizes to convolve the input features; the second branch uses a single convolution kernel to convolve the input features and then stacks them with the output features of the first branch, and then performs a convolution operation again after stacking; the third branch directly splices the input features with the features output by the second branch, and outputs the spliced ​​features to extract global features; The global features are classified through the fully connected layer in the backbone network, and the cable early fault detection and classification results are output.

8. A cable early fault diagnosis system as claimed in claim 7, characterized in that: Acquire current data in the cable to be tested, wherein the current data is a time series of current signals; The acquired current data is preprocessed, and the preprocessing includes dividing the current signal time series to obtain a plurality of current time series signals with a length of N.

9. A cable early fault diagnosis system as claimed in claim 7, characterized in that: Combined with wavelet transform and bispectrum theory, the fault characteristics of the pre-processed current data are extracted to obtain a two-dimensional fault feature map, including: The current time series signal is decomposed in the time-frequency domain using wavelet transform to obtain sub-signals in different frequency bands. Calculate the bispectral density function of the sub-signal on each frequency band to obtain a bispectral density matrix; Merge the bispectral density matrices of the sub-signals in different frequency bands to obtain the global bispectral density matrix; The global bispectral density matrix is ​​decomposed to obtain the wavelet bispectral feature map, that is, the two-dimensional fault feature map.

10. A cable early fault diagnosis system as claimed in claim 7, characterized in that: The fault detection model based on the improved residual network includes an input layer, a convolutional layer, a BN layer, a pooling layer, an improved residual module, a fully connected layer and an output layer which are connected in sequence.