A partial discharge identification method and device based on time-frequency features and improved CNN

Abstract

The present invention relates to the technical field of electrical equipment fault identification, and in particular to a method and device for identifying partial discharge based on time-frequency features and improved CNN, including Fourier threshold filtering of collected partial discharge signals; intercepting pulses of partial discharge time domain signals, characterizing the time domain features of pulses with discrete data sampling points, and obtaining the time domain feature sequence of partial discharge pulse signals; and performing time-frequency transformation on pulse signals through improved Morlet wavelets to obtain two-dimensional time-frequency image data of corresponding signals; normalizing the time domain feature sequence and wavelet time-frequency graph of partial discharge signals; training an improved convolutional neural network model using a data set, and saving the trained network model. The present invention makes full use of the time-frequency features of partial discharge pulse signals to achieve partial discharge classification, improve the effect of partial discharge signal time-frequency feature extraction; and improves the structure of convolutional neural networks for partial discharge identification, thereby improving the accuracy of partial discharge identification.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment fault identification, and in particular to a partial discharge identification method and device based on time-frequency features and improved CNN. Background Art

[0002] The safe operation of electrical equipment is of great significance to industrial production. The deterioration and defects of equipment insulation can lead to the occurrence of partial discharge. Identifying the type of partial discharge generated by the equipment and timely evaluating the insulation condition of the equipment can prevent accidents. Common types of partial discharge include tip discharge, suspension discharge, surface discharge, air gap discharge, etc.

[0003] Feature extraction and pattern recognition of partial discharge are the key to partial discharge detection. In terms of feature extraction, the commonly used method of statistical phase and other characteristic parameters will make the network structure of the recognition algorithm more complicated, which may cause misjudgment when the statistical information is incomplete, and does not fully utilize effective information such as time domain and frequency domain waveforms. The time-frequency analysis method can provide assistance to traditional methods and is of great significance for partial discharge classification. When processing transient and unstable signals such as partial discharge, the time-frequency localization ability of wavelet transform has a good effect. However, when using wavelet transform to extract partial discharge features, it is necessary to select a suitable wavelet. In terms of pattern recognition, with the development of deep learning, artificial neural networks, support vector machines, etc. have been used in the recognition of partial discharge. Convolutional neural networks perform well in the field of image recognition and can effectively classify partial discharges.

[0004] Based on this, it is expected that the pulse waveform analysis method combined with convolutional neural network can be used to identify the type of partial discharge of electrical equipment, realize fault warning and ensure the safe operation of equipment. On this basis, it is hoped that the time domain and frequency domain characteristics of partial discharge pulse signals can be fully utilized, and the classification accuracy of convolutional neural network for partial discharge can be improved. Summary of the invention

[0005] The technical problem to be solved by the present invention is: to make full use of the time-frequency characteristics of partial discharge pulse signals to realize the classification of partial discharges and improve the effect of extracting the time-frequency characteristics of partial discharge signals; at the same time, to improve the structure of the convolutional neural network for the recognition of partial discharges and improve the accuracy of partial discharge recognition.

[0006] The technical solution adopted by the present invention is: a partial discharge recognition method based on time-frequency features and improved CNN, comprising the following steps:

[0007] S1. Perform Fourier threshold filtering on the collected partial discharge signal to remove noise interference;

[0008] Furthermore, the PD signal is directly subjected to a fast Fourier transform to obtain the corresponding frequency domain waveform, the maximum point in the Fourier spectrum is found, the spectrum amplitude of the noise interference is directly set to zero, and the interference signal is eliminated;

[0009] In order to avoid incomplete noise removal, the processing frequency band is widened, that is, the frequency band of 5KHz on both sides of the threshold is selected for amplitude zeroing operation; in order to avoid signal distortion caused by threshold zeroing, the least squares method is used to perform polynomial fitting on the zeroing area, and the optimized frequency domain signal is inversely transformed by Fourier to obtain the denoised partial discharge time domain signal;

[0010] Furthermore, the threshold filter function is as follows:

[0011]

[0012] Where f(x) is the partial discharge signal, F(f(x)) is the Fourier transform of the signal, and σ is the selected threshold.

[0013] S2, intercepting the pulse of the partial discharge time domain signal, characterizing the time domain characteristics of the pulse with discrete data sampling points, and obtaining the time domain characteristic sequence of the partial discharge pulse signal; and performing time-frequency transformation on the pulse signal by improving Morlet wavelet to obtain the two-dimensional time-frequency image data of the corresponding signal;

[0014] Further, the pulse of the partial discharge time domain signal is intercepted, the intercepted data is divided into N intervals, the average value of the data in each interval is calculated, and the time domain characteristic sequence of the partial discharge pulse signal with a length of N is obtained;

[0015] Furthermore, the PD time domain signal is subjected to continuous wavelet transform, and the wavelet time-frequency diagram corresponding to the signal data is plotted. The wavelet transform formula is:

[0016]

[0017] Among them, ψ(x) is the wavelet basis function, a is the scale factor, b is the translation factor, and * is the conjugate;

[0018] Furthermore, the improved wavelet basis function is:

[0019]

[0020] S3, normalizing the time domain characteristic sequence and wavelet time-frequency diagram of the partial discharge signal, and dividing the data set;

[0021] S4. Using the data set to train the improved convolutional neural network model, and saving the trained network model to achieve classification of partial discharges;

[0022] Furthermore, the improved convolutional neural network model includes the following:

[0023] Two parallel channels are constructed. The first channel inputs a one-dimensional time domain feature sequence, and the second channel inputs a two-dimensional wavelet time-frequency map. Both channels use convolutional layers and pooling layers to extract features. At the same time, the shallow features of the wavelet time-frequency map input by the second channel are extracted. The features of the deep and shallow layers of the time domain and time-frequency domain are stretched into feature vectors using a fully connected layer, which are fused in the fusion layer and then input into the fully connected layer. Finally, the classification of the partial discharge signal is realized through a softmax classifier.

[0024] The improved convolutional neural network is a multi-input model. The first channel is a one-dimensional network model, which specifically includes 4 convolution layers and 2 pooling layers. A 1×1 convolution layer is added after each ordinary convolution layer to increase the nonlinearity of the one-dimensional network model. The second channel is a two-dimensional network model, which specifically includes 3 convolution layers and 3 pooling layers. At the same time, an additional 1×1×1 convolution layer is used to reduce the dimension of the features output by the first pooling layer and retain the significant features to achieve the fusion of deep and shallow features. The extracted features are stretched to the same dimension using a fully connected layer and fused in the fusion layer.

[0025] Furthermore, a device for partial discharge identification method based on time-frequency characteristics and improved CNN is provided, and partial discharge signals are collected through the device, including: a high-frequency current sensor, a signal conditioning module, a high-speed AD acquisition module, an SDRAM storage module and an FPGA control module. The HFCT sensor is placed at a detection position to sense the generated partial discharge pulse signal; the signal conditioning module amplifies and filters the partial discharge pulse signal and sends the signal to the high-speed AD acquisition module; the high-speed AD acquisition module performs signal sampling; the SDRAM storage module realizes large-capacity caching of signal data; and the FPGA control module realizes data processing and read and write operations on the SDRAM.

[0026] Furthermore, the FPGA control module includes: a PLL module, a FIFO control module, an AD data processing module, a serial port control module and an SDRAM controller. The SDRAM controller completes the initialization, refresh, reading and writing of the SDRAM; the PLL module provides a clock signal; the AD data processing module controls the acquisition of the AD signal, the FIFO control module realizes the interaction with the SDRAM controller and the serial port control module; and the serial port control module realizes data transmission.

[0027] Beneficial effects of the present invention:

[0028] 1. When using wavelet transform to extract time-frequency features of partial discharge time-domain pulse data, the optimized Morlet wavelet basis function is used to improve the resolution in the frequency domain and the effect of feature extraction;

[0029] 2. The time domain and frequency domain features of the signal waveform are fully utilized in partial discharge classification. The time domain feature sequence of the partial discharge signal and the deep and shallow features of the corresponding wavelet time-frequency graph are integrated using a convolutional neural network. The improved convolutional neural network has a simple structure and fast training speed. Comparative experiments show that the improved algorithm has a higher recognition accuracy rate;

[0030] 3. The FPGA high-speed data acquisition system used can realize high-speed acquisition and large-capacity cache of partial discharge pulse signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the partial discharge experimental platform of the present invention;

[0032] Figure 2 It is a high-speed partial discharge signal acquisition system based on FPGA of the present invention;

[0033] Figure 3 It is the FPGA programming block diagram of the present invention;

[0034] Figure 4 This is a comparison diagram of the wavelet basis function effects before and after the improvement of the present invention;

[0035] Figure 5 It is a structural diagram of an improved convolutional neural network of the present invention;

[0036] Figure 6 are the convolutional neural network training accuracy curve and loss curve of the present invention;

[0037] Figure 7 This is the classification visualization effect of the improved convolutional neural network of the present invention on the test set. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0039] Using the pulse current method, the FPGA high-speed data acquisition system is used to collect and store the partial discharge signals of the simulated partial discharge models of various typical defects.

[0040] like Figure 1 A partial discharge experimental platform was built. The high-voltage experimental transformer provided the rated voltage, a 10kΩ protection resistor was used, and the voltage divider capacitance was 1000pF. The partial discharge model included electrode models of four typical partial discharge defects: tip discharge, suspended discharge, surface discharge, and air gap discharge. The HFCT high-frequency current sensor was used to sense the pulse signal generated by the partial discharge model, and the generated signal was collected through the FPGA high-speed data acquisition system.

[0041] like Figure 2 As shown, a partial discharge identification device based on time-frequency characteristics and improved CNN is used to collect partial discharge signals, including: a high-frequency current sensor, a signal conditioning module, a high-speed AD acquisition module, an FPGA control module, an SDRAM storage module and a host computer; the HFCT sensor is placed at a detection position to sense the generated partial discharge pulse signal; the signal conditioning module realizes amplification and filtering of the partial discharge pulse signal to obtain a 1-3V signal and sends it to the high-speed AD acquisition module, and the high-speed AD realizes 50MHz high-speed sampling; the SDRAM storage module realizes large-capacity caching of signal data; the FPGA control module realizes data processing and read and write operations on the SDRAM, and the cached data can be transmitted to the host computer through the serial port.

[0042] Among them, the high-speed AD chip model is AD9226, the digital interface is 12bit, and it can achieve 50MHz high-speed sampling; SDRAM is used to cache data, the SDRAM model is MT48LC32M16A2, which can store 64M bytes of data; the FPGA chip model is EP4CE10E22C8 to realize the control of each module.

[0043] like Figure 3 As shown in the FPGA program design block diagram, the SDRAM controller completes the initialization, refresh, reading and writing of SDRAM; the PLL module provides the clock signal; the AD data processing module controls the acquisition of AD signals, the FIFO control module realizes the interaction with the SDRAM controller and the serial port control module; the serial port control module realizes data transmission.

[0044] The FPGA high-speed data acquisition system collects PD data of one power frequency cycle of 0.02 seconds at a time, collects 200 pieces of PD pulse data for each PD model, and transmits them to the host computer.

[0045] A partial discharge recognition method based on time-frequency features and improved CNN, comprising:

[0046] S1. Perform Fourier threshold filtering on the collected partial discharge signal to remove noise interference;

[0047] Furthermore, the PD signal is directly subjected to a fast Fourier transform to obtain the corresponding frequency domain waveform, the maximum point in the Fourier spectrum is found, the spectrum amplitude of the noise interference is directly set to zero, and the interference signal is eliminated. The threshold function is as follows:

[0048]

[0049] Where f(x) is the partial discharge signal, F(f(x)) is the Fourier transform of the signal, and σ is the selected threshold.

[0050] Furthermore, in order to avoid incomplete noise removal, the processing frequency band is widened, that is, the frequency band of 5KHz on both sides of the threshold is selected for amplitude zeroing operation; in order to avoid signal distortion caused by threshold zeroing, the least squares method is used to perform polynomial fitting on the zeroing area, and the optimized frequency domain signal is inversely transformed by Fourier to obtain the denoised partial discharge time domain signal;

[0051] S2, intercepting the pulse of the partial discharge time domain signal, characterizing the time domain characteristics of the pulse with discrete data sampling points, and obtaining the time domain characteristic sequence of the partial discharge pulse signal; and performing time-frequency transformation on the pulse signal by improving Morlet wavelet to obtain the two-dimensional time-frequency image data of the corresponding signal;

[0052] Furthermore, the collected partial discharge time domain signal is a one-dimensional time series composed of sampling points, and the data volume is large, which is not conducive to the operation of data processing and recognition algorithms. The pulse of the partial discharge time domain signal is intercepted, and the intercepted data is divided into N intervals, and the average value of the data in each interval is calculated to obtain the time domain characteristic sequence of the partial discharge pulse signal with a length of N.

[0053] Furthermore, the pulse signal is transformed into a time-frequency image using the optimized Morlet wavelet to obtain the two-dimensional time-frequency image data of the corresponding signal.

[0054] Perform continuous wavelet transform on the PD time domain signal and draw the wavelet time-frequency diagram corresponding to the signal data. The wavelet transform formula is:

[0055]

[0056] Among them, ψ(x) is the wavelet basis function, a is the scale factor, b is the translation factor, and * is the conjugate.

[0057] The higher the matching degree between the wavelet basis function and the measured signal, the more conducive it is to the extraction of signal features. The partial discharge signal collected by the pulse current method often satisfies the exponential decay oscillation pulse mode. Several types of wavelets are studied. Morlet wavelet is more suitable for the feature extraction of partial discharge pulse signals.

[0058] The traditional Morlet wavelet basis function is:

[0059]

[0060] The traditional Morlet wavelet is composed of Gaussian function multiplied by complex trigonometric function. The improved wavelet basis function replaces the Gaussian function to slow down the exponential decay of the function and expand the support range in the time domain, thereby improving the resolution of feature extraction in the frequency domain and selecting the appropriate wavelet center frequency.

[0061] The improved wavelet basis function is:

[0062]

[0063] Select the same center frequency and use the improved wavelet basis functions to draw the wavelet time-frequency diagram of the partial discharge signal. The results are as follows: Figure 4 As shown, Figure 4 (a) is the traditional Morlet wavelet time-frequency diagram, Figure 4 (b) is the optimized Morlet wavelet time-frequency diagram. By comparison, it is found that the optimized Morlet wavelet has a better time-frequency feature extraction effect and a higher resolution in the frequency domain.

[0064] S3, normalizing the time domain characteristic sequence and wavelet time-frequency diagram of the partial discharge signal, and dividing the data set;

[0065] The amplitude of the time domain feature sequence of the partial discharge signal is normalized to obtain a one-dimensional time domain feature sequence data set; the wavelet time-frequency graph is normalized and grayed to obtain a two-dimensional time-frequency spectrum graph data set. The data set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1 as the input of the convolutional neural network.

[0066] S4. Build a convolutional neural network model. The convolutional neural network model is as follows: Figure 5 As shown in the figure, the model consists of two parallel channels; channel a inputs a one-dimensional time domain feature sequence, and channel b inputs a two-dimensional wavelet time-frequency map; both channels use convolutional layers and pooling layers to extract features; at the same time, the features of the shallow wavelet time-frequency map of channel b are extracted; these features are stretched into feature vectors using a fully connected layer, and then input into a fully connected layer after fusion in a fusion layer, and finally the classification of partial discharge signals is realized through a softmax classifier;

[0067] Furthermore, a 1×1 convolution layer is added to each channel a after the ordinary convolution layer, so as to access multiple activation functions and improve the nonlinear fitting ability of the one-dimensional network.

[0068] The first convolution layer of channel b uses a 5x5 convolution kernel to increase the receptive field and extract more features. In the feature extraction process, the shallow layer passes through fewer convolution layers, has a high feature resolution, and contains more feature information. The deep layer features have better semantics. In order to avoid feature loss, the deep and shallow features are fused to improve the recognition rate.

[0069] Traditional feature fusion directly represents the features of the two layers as feature vectors and sends them to the fully connected layer, which will result in too many parameters in the fully connected layer and make the model bloated. By adding a 1×1×1 convolutional layer, the shallow features are reduced in dimension and the significant features are retained.

[0070] Furthermore, the features output by the first pooling layer of channel b are passed through a 1×1×1 convolutional layer to achieve data dimensionality reduction, and then converged to the fusion layer through a fully connected layer for fusion.

[0071] The convolution layer uses the convolution kernel to weight the input and uses a nonlinear activation function to concatenate it into the input of the next layer. The mathematical formula is:

[0072]

[0073] Where: X k is the k-th layer output, is the k-th layer input, W i k is the weight matrix of the convolution kernel, is the bias term, f(·) is the activation function;

[0074] ReLU is selected as the activation function in the model, and the function expression is as follows:

[0075] ReLU(x)=max(0,x) (6)

[0076] In order to avoid gradient disappearance and gradient explosion, the BN layer is used after the convolution layer to normalize the data to enhance the generalization ability of the model; the pooling layer selects maximum pooling to obtain the maximum value of the local area of ​​the data to achieve dimensionality reduction; the fully connected layer is used to fuse the feature vectors of the data; in order to prevent overfitting, the Dropout operation is used with a size of 0.5; finally, the softmax classifier is used on the data output by the fully connected layer, and the classifier calculates the relative probabilities of different types of partial discharge to achieve the classification of partial discharge.

[0077]

[0078] The specific parameters of the model are shown in the table below:

[0079] Table 1 Improved convolutional neural network model parameters

[0080]

[0081]

[0082] Input the training set and validation set into the improved convolutional neural network training and save the trained network model.

[0083] The input data set is used to conduct supervised learning of the convolutional neural network model. The weights and bias values ​​of each layer of the network are adjusted using the stochastic gradient descent method. The loss function uses the cross entropy loss function:

[0084]

[0085] Among them, y i is the label value, y′ i is the predicted value.

[0086] The learning rate is set to 0.005, the number of iterations is set to 100, and the convolutional neural network training process and loss curve are shown in Figure 6 As shown, the improved convolutional neural network has a fast training speed and can obtain a higher accuracy, and the trained convolutional neural network model is saved.

[0087] Input partial discharge data with unknown fault types into the trained convolutional neural network to verify the recognition effect of the model;

[0088] The test set is used to verify the accuracy of partial discharge pattern recognition. The test results of the improved convolutional neural network are compared with those of the single-channel network with ordinary wavelet time-frequency graph as input and the single-channel network with improved wavelet time-frequency graph as input. The test accuracy is shown in the following table:

[0089] Table 1 Test results

[0090] Experimental model Time-frequency diagram + CNN Optimizing time-frequency graph + CNN The present invention Accuracy 95％ 97.5％ 99.4％

[0091] By comparison, it is found that the improved convolutional neural network of the present invention has higher recognition accuracy. Under the same number of iterations, the improved convolutional neural network converges faster during training. The classification visualization effect of the improved convolutional neural network on the test set is as follows: Figure 7 shown.

[0092] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A partial discharge recognition method based on time-frequency features and improved CNN, characterized in that: The following steps are involved: S1. Perform Fourier threshold filtering on the collected partial discharge signal to remove noise interference; S2, intercepting the pulse of the partial discharge time domain signal, characterizing the time domain characteristics of the pulse with discrete data sampling points, and obtaining the time domain characteristic sequence of the partial discharge pulse signal; and performing time-frequency transformation on the pulse signal by improving Morlet wavelet transform, and obtaining the two-dimensional time-frequency image data of the corresponding signal; The formula for Morlet wavelet transform is: Among them, ψ(x) is the improved wavelet basis function, a is the scale factor, b is the translation factor, and * is the conjugate; The formula of improved wavelet basis function is: S3, normalizing the time domain feature sequence and time-frequency image data of the partial discharge signal and dividing the data set; S4. Using the data set to train the improved convolutional neural network model, and saving the trained network model to achieve classification of partial discharges; The improved convolutional neural network model includes: Two parallel channels are constructed. The first channel inputs a one-dimensional time domain feature sequence, and the second channel inputs a two-dimensional wavelet time-frequency map. Both channels use convolutional layers and pooling layers to extract features. At the same time, the shallow features of the second channel input wavelet time-frequency map are extracted. The deep and shallow features of the time domain and time-frequency domain are stretched into feature vectors using a fully connected layer, which are fused in the fusion layer and then input into the fully connected layer. Finally, the classification of the partial discharge signal is realized through a softmax classifier. The improved convolutional neural network model also includes: The first channel is a one-dimensional network model, which specifically includes 4 convolutional layers and 2 pooling layers. A 1×1 convolutional layer is added after each ordinary convolutional layer to increase the nonlinearity of the one-dimensional network model. The second channel is a two-dimensional network model, which specifically includes 3 convolutional layers and 3 pooling layers. At the same time, an additional 1×1×1 convolutional layer is used to reduce the dimension of the features output by the first pooling layer and retain the significant features to achieve the fusion of deep and shallow features. The extracted features are stretched to the same dimension using a fully connected layer and fused in the fusion layer.

2. The partial discharge identification method based on time-frequency features and improved CNN according to claim 1 is characterized in that: The step S1 comprises: Perform fast Fourier transform on the partial discharge signal to obtain the corresponding frequency domain waveform, find the maximum point in the Fourier spectrum, and directly set the spectrum amplitude of the noise interference to zero; The processing frequency band is widened, the least square method is used to perform polynomial fitting on the zero-set area, and the frequency domain signal after optimization is subjected to inverse Fourier transform to obtain the denoised PD time domain signal.

3. The partial discharge identification method based on time-frequency features and improved CNN according to claim 2 is characterized in that: The function of the threshold filtering is: Where f(x) is the partial discharge signal, F(f(x)) is the Fourier transform of the signal, and σ is the selected threshold.

4. The partial discharge identification method based on time-frequency features and improved CNN according to claim 2 is characterized in that: The intercepting of the pulse of the partial discharge time domain signal includes: dividing the intercepted data into N intervals, calculating the average value of the data in each interval, and obtaining the time domain characteristic sequence of the partial discharge pulse signal with a length of N.

5. A device using the partial discharge identification method based on time-frequency features and improved CNN as claimed in claim 1, characterized in that: include: High-frequency current sensor, signal conditioning module, high-speed AD acquisition module, SDRAM storage module and FPGA control module. The HFCT sensor is placed at the detection position to sense the generated partial discharge pulse signal; the signal conditioning module realizes amplification and filtering of the partial discharge pulse signal and sends the signal to the high-speed AD acquisition module; the high-speed AD acquisition module performs signal sampling; the SDRAM storage module realizes large-capacity caching of signal data; the FPGA control module realizes data processing and read and write operations on the SDRAM.

6. The device for partial discharge identification method based on time-frequency features and improved CNN according to claim 5, characterized in that: The FPGA control module includes: a PLL module, a FIFO control module, an AD data processing module, a serial port control module and an SDRAM controller. The SDRAM controller completes the initialization, refresh, reading and writing of the SDRAM; the PLL module provides a clock signal; the AD data processing module controls the collection of the AD signal, the FIFO control module realizes the interaction with the SDRAM controller and the serial port control module; and the serial port control module realizes data transmission.

Images