Partial discharge mode recognition method based on residual neural network and attention module

Abstract

The invention discloses a partial discharge mode recognition method based on a residual neural network and an attention module, and relates to the technical field of partial discharge signal processing and recognition. The method comprises the following steps: acquiring partial discharge signal data to be detected; carrying out multi-scale denoising processing on the partial discharge signal by an improved ensemble empirical mode decomposition adaptive noise algorithm and wavelet packet decomposition combined method so as to remove noise interference and reserve main components of the signal; the time domain, the frequency domain and the phase information of the denoised partial discharge signals are converted into grey-scale maps through a Grubby angle field algorithm, a Grubby angle difference field algorithm and a Markov transfer field algorithm, and the grey-scale maps are fused to generate a feature RGB image. Inputting the RGB image into a residual neural network combined with a convolutional block attention module for feature extraction and classification; and finally, outputting a partial discharge mode identification result for fault detection and early warning of power equipment. According to the method, multi-scale denoising processing and multi-dimensional feature fusion are utilized, and depth feature extraction capabilities of the residual network and the attention module are combined, so that the recognition accuracy and robustness of the partial discharge mode are remarkably improved, and the method can be widely applied to fault diagnosis and early warning of power equipment.

Description

Technical Field

[0001] The invention relates to the technical field of partial discharge pattern recognition, in particular to a partial discharge pattern recognition method based on a residual neural network and an attention module. Background Art

[0002] In the fault diagnosis and maintenance of power equipment, partial discharge has attracted extensive attention as an early fault signal. The partial discharge phenomenon occurs in the insulation defect area of ​​the power equipment. If it is not monitored and processed in time, it may lead to accelerated insulation aging, or even cause equipment failure or major power system failure. Therefore, online monitoring and identification of partial discharge is crucial for the safe operation of power equipment. Traditional partial discharge pattern recognition methods mainly rely on signal processing and feature engineering methods to extract pattern information of partial discharge signals through time domain, frequency domain or time-frequency domain features. However, due to the influence of complex environment and noise, the characteristics of partial discharge signals are highly nonlinear and unstable, and traditional methods are difficult to obtain robust recognition effects. With the development of deep learning, the application of neural networks, especially in pattern recognition, has made significant progress, and the analysis and classification of partial discharge signals have been gradually introduced. The simple use of convolutional neural networks to deal with the diversity and complexity of partial discharge signals still has limitations. For this reason, residual neural networks are introduced as improved deep network structures for partial discharge pattern recognition. Through residual connections, ResNet effectively alleviates the gradient vanishing and explosion problems of deep networks, enabling the network to learn richer signal features. At the same time, the introduction of the attention mechanism can improve the recognition of features and the focusing ability of the network. It can adaptively allocate feature weights and highlight signal features with recognition value, thereby improving the accuracy and robustness of recognition. Summary of the invention

[0003] In view of the above problems, the purpose of the present invention is to provide a local discharge pattern recognition method based on a residual neural network and an attention module to solve the problems of low recognition rate and poor generalization ability in the prior art, and to provide an efficient and accurate local discharge pattern recognition method.

[0004] To achieve the above object, the present invention adopts the following technical solution: a partial discharge pattern recognition method based on residual neural network and attention module, characterized by comprising the following steps:

[0005] S1: Collecting the partial discharge signal data to be detected;

[0006] S2: The local discharge signal is denoised by combining the improved ensemble empirical mode decomposition adaptive noise algorithm and wavelet packet decomposition, thereby achieving multi-scale denoising processing of the local discharge signal and laying a reliable data foundation for subsequent local discharge pattern recognition.

[0007] S3: Use the Gram angle sum field, Gram angle difference field and Markov transfer field to convert the time domain, frequency domain and phase information into a two-dimensional grayscale image, and map the obtained three grayscale images to the R channel, G channel and B channel of the image to form a feature fusion RGB image to build a partial discharge feature database.

[0008] S4: The generated RGB image is input into the ResNet50-CBAM model for training, so that the network can learn the characteristic pattern of the partial discharge signal and identify the partial discharge type according to the trained model.

[0009] Further, the step S2 is specifically: firstly, the partial discharge signal is decomposed by ICEEMDAN, and then denoising is performed by an improved wavelet threshold method, which specifically includes the following steps:

[0010] Add multiple groups of Gaussian white noise to the original signal, construct multiple noise-enhanced signals and perform EMD decomposition to obtain the first-layer modal component IMF. By adjusting the noise amplitude and multiple iterations, the K-order IMF is obtained. k Serving size:

[0011] IMF k =r k-1 -r k

[0012] Among them, r k is the K-order residual.

[0013] The IMF components obtained by screening are decomposed by wavelet packets, and the signal after wavelet transformation is expressed as:

[0014]

[0015] Among them, ψ i (t) is the wavelet basis function, c i is the corresponding wavelet coefficient. By selecting an appropriate threshold, the wavelet coefficient c i Processing is performed to filter out noise signals and retain the main components of partial discharge signals.

[0016] Further, the step S3 is specifically: converting the time domain information of the partial discharge signal into a grayscale image G using the Gram angle and field T , the frequency domain information is converted into a grayscale image G using the Gram angle difference field F , the phase information is converted into a grayscale image G using the Markov transfer field P The grayscale image G obtained above T , G F , G P As the three channels of the RGB image, respectively, to form a fused image I RGB .

[0017] Further, the step S4 is specifically as follows: the residual neural network is used to extract features and recognize patterns of the graph structure of the partial discharge signal. A convolutional attention module is added to the key layer of the network. The CBAM module combines the channel attention and spatial attention mechanisms, so that the network can focus on the feature areas that contribute significantly to the partial discharge pattern recognition.

[0018] Channel attention module: The channel attention module focuses on the important information on the channel. It aggregates the spatial information of the feature map by using average pooling and maximum pooling at the same time, thereby compressing the feature map in the spatial dimension. Its attention expression mechanism can be expressed as:

[0019]

[0020] Among them, F is the input feature, M channel is the channel attention weight.

[0021] Spatial attention module: further introduces attention in the spatial dimension and emphasizes the salient areas of the feature map. Its attention expression mechanism can be expressed as:

[0022]

[0023] Among them, F is the input feature, M' channel is the spatial attention weight.

[0024] The beneficial effects of the present invention are:

[0025] (1) The noise of partial discharge signals is suppressed by wavelet packet decomposition and improved wavelet threshold function, which effectively retains the main components of the signal and enhances the signal quality.

[0026] (2) Combined with the RGB image fusion method, the time domain, frequency domain and phase information of the local discharge signal are integrated into a multidimensional feature map, which fully mines the multidimensional information of the local discharge signal and enhances the expressiveness of the signal characteristics.

[0027] (3) By introducing the residual neural network (ResNet) and convolutional block attention module (CBAM), the local discharge signal is extracted and enhanced, which greatly improves the model's ability to recognize local discharge patterns. ResNet effectively alleviates the gradient vanishing problem in deep network feature extraction, and the CBAM module dynamically focuses on important feature areas through the attention mechanism in the channel and spatial dimensions, effectively improving the generalization ability and classification accuracy of the model under complex and diverse local discharge patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 A flowchart of a partial discharge pattern recognition method based on a residual neural network and an attention module provided in an embodiment of the present application;

[0030] Figure 2 A partial discharge signal denoising flow chart provided in an embodiment of the present application;

[0031] Figure 3 A schematic diagram of a ResNet-CBAM network structure provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of this application clearer, the relevant technical solutions will be described in detail below with reference to the accompanying drawings. It should be pointed out that the embodiments described here are only part of the content of this application, not all embodiments. In this field, all other embodiments that can be obtained based on the embodiments in this application by ordinary technicians without creative work should be included in the scope of protection of this application.

[0033] Figure 1 A flowchart of a local discharge pattern recognition method based on a residual neural network and an attention module is provided in an embodiment of the present application. The local discharge pattern recognition method based on a residual neural network and a convolutional attention module specifically includes the following steps:

[0034] Step S101, inputting the partial discharge signal collected by the sensor.

[0035] As an embodiment, the partial discharge signal may include a voltage waveform, an ultrasonic wave signal, an electromagnetic wave signal, and the like.

[0036] Step S102: performing improved overall empirical mode decomposition adaptive noise decomposition on the partial discharge signal.

[0037] As an embodiment, step S102 includes: adaptively decomposing the partial discharge signal by using the ICEEMDAN method to obtain IMF components of different frequencies of the signal and a residual component.

[0038] Add Gaussian white noise E1(ω (i) ), get the signal s i:

[0039] s i =s+β0E1(ω (i) )

[0040] Through EMD decomposition, the first-layer modal component IMF1 and the first-order residual r1 are obtained. By adjusting the noise amplitude and multiple iterations, the K-order IMF is obtained. k Components and k-order residuals:

[0041] IMF k =r k-1 -r k

[0042] r k = <M(r k-1 +β k-1 E k (ω (i) ))>

[0043] Repeat the iteration until the decomposition is completed and all IMF components and residual components are obtained.

[0044] Step S103: Filter the IMF components based on energy entropy and kurtosis.

[0045] As an embodiment, step S103 includes: setting energy entropy and kurtosis thresholds based on the characteristics of the partial discharge signal; calculating the energy entropy and kurtosis threshold of each IMF component, and dividing the IMF component into a noisy IMF component and a noise-free IMF component.

[0046] Energy entropy describes the disorder of the energy distribution of IMF components. IMF components with smaller energy entropy usually contain regular signal components. The calculation formula of energy entropy is:

[0047]

[0048] Among them, p i is the normalized probability density of the energy at the i-th sampling point, and n is the length of the IMF component.

[0049] Kurtosis is a statistic that measures the peak state of a signal and is used to describe the sharpness of a signal. Partial discharge signals are often accompanied by spikes or mutations, and IMF components with higher kurtosis usually represent the instantaneous characteristics of partial discharge. The calculation formula for kurtosis is:

[0050]

[0051] Where E is the expected value, μ is the mean of the signal, and x is the sample value of the IMF component.

[0052] Step S104, performing wavelet packet decomposition and denoising on the screened noisy PD components.

[0053] As an embodiment, step S104 includes: selecting a wavelet basis; improving wavelet threshold denoising and signal reconstruction.

[0054] The db8 wavelet basis function is selected to perform wavelet packet decomposition on the IMF components obtained by decomposition, and the improved wavelet threshold function is used for denoising. The threshold function is:

[0055]

[0056] Among them, d j,k represents the kth wavelet coefficient of the jth layer in the wavelet packet decomposition; σ1 is the threshold.

[0057] The noisy IMF component and the noise-free IMF component after wavelet packet decomposition and denoising are reconstructed to obtain the denoised partial discharge signal.

[0058] Among them, the partial discharge signal denoising flow chart is as follows Figure 2 shown.

[0059] Step S105: converting the denoised one-dimensional partial discharge signal into a two-dimensional image.

[0060] As an embodiment, step S105 includes: converting the time domain information of the partial discharge signal into a grayscale image using the Gram angle sum field; converting the frequency domain information into a grayscale image using the Gram angle difference field; converting the phase information into a grayscale image using the Markov transfer field. The three images obtained are fused to generate an RGB feature image.

[0061] Define the time series X = {x1,x2,...x n}, and scaled to the interval [-1,1] by normalization, as follows:

[0062]

[0063] The normalized time series X value is used as the angle cosine, and the timestamp is used as the radius. The coordinate system conversion formula is:

[0064]

[0065] In the formula, φ i is the polar angle; r i is the polar axis; t i Is the timestamp.

[0066] The two image encoding methods GADF and GASF are defined as:

[0067]

[0068]

[0069] The one-dimensional time series vibration signal is divided into L regions, and the signal in any region can be mapped to the corresponding subdomain. Define a matrix R to represent the probability that a subdomain signal is followed by another subdomain signal in all subdomains. The probability calculation formula in R is:

[0070] r i,j =P(x i ∈l i |x i-1 ∈l j )

[0071]

[0072] Where: r i,j is the probability value of point ii being followed by point j in the probability following matrix.

[0073] By calculating the follow-up probability value of each element, we can further obtain the L×L weighted adjacent migration probability matrix R.

[0074]

[0075] Introduce time information to correlate vibration signal amplitude information and define the MTF matrix M.

[0076]

[0077] m ij =P{r ij ||x i ∈l i ,x j ∈l j}

[0078] Where: m ij for l i To l j The transition probability of a state transition occurring.

[0079] The three images obtained are mapped to the R channel, G channel and B channel respectively to form a feature fusion RGB feature image.

[0080] Step S106, using a residual neural network combined with a convolutional attention module to extract features and classify the graph constructed in step S105, and outputting a partial discharge pattern recognition result.

[0081] Among them, the ResNet-CBAM network structure diagram is as follows Figure 3 shown.

[0082] As an embodiment, step S106 includes: neural network training; classification loss function; classification result.

[0083] The neural network training described herein refers to inputting the RGB image in the feature database into a difference neural network embedded in a convolutional block attention module to learn the key features in the image.

[0084] The classification loss function uses the cross entropy loss function to calculate the classification loss, and the formula is:

[0085]

[0086] Among them, y i is the actual label, p i is the predicted probability.

[0087] The classification result output is to output the classification result of the fully connected layer, including the recognition accuracy of each partial discharge pattern and the overall recognition accuracy.

[0088] Step S107, outputting the final partial discharge pattern recognition result for use in fault detection and early warning of power equipment.

[0089] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A partial discharge pattern recognition method based on residual neural network and attention module, characterized in that: The following steps are involved: Step 1: Collecting partial discharge signal data to be detected, wherein the partial discharge signal may include a voltage waveform, an ultrasonic signal, and an electromagnetic wave signal; Step 2: Perform improved overall empirical mode decomposition, adaptive noise decomposition and wavelet packet decomposition on the partial discharge signal, including ICEEMDAN decomposition; screening of appropriate IMF components; wavelet threshold denoising and signal reconstruction. Remove the noise in the signal and retain the main components of the signal; Step 3: The time domain, frequency domain and phase information of the denoised partial discharge signal are converted into grayscale images through the Gram angle field, Gram angle difference field and Markov transfer field algorithms respectively, and the three grayscale images are used as the R channel, G channel and B channel of the RGB image respectively to generate a feature-fused RGB image; Step 4: embed the RGB image input in step S105 into the residual neural network of the convolutional block attention module, perform feature extraction and classification, and output the partial discharge pattern recognition result; Step 5: Output the final partial discharge pattern recognition result.

2. The method according to claim 1, characterized in that In step 2, by adding Gaussian white noise E1(ω (i) ) and perform multiple iterative decompositions to obtain multi-order IMF components and corresponding residual components to improve the denoising effect. i =s+β0E1(ω (i) ) The first-layer modal component IMF1 and residual r1 are obtained by empirical mode decomposition, and the K-order component is iterated layer by layer:

3. The method according to claim 1, characterized in that: In step 2, the IMF components are screened according to energy entropy and kurtosis. The energy entropy is calculated as: Among them, p i is the normalized probability density of the energy at the i-th sampling point, and n is the length of the IMF component. The calculation formula of kurtosis is:

4. The method according to claim 1, characterized in that In step 2, the db8 wavelet basis is used to perform wavelet packet decomposition on the IMF component, and the improved wavelet threshold function is used to denoise the wavelet coefficients, thereby filtering out the noise signal and retaining the main components of the partial discharge signal. The threshold function is: Among them, d j,k represents the kth wavelet coefficient of the jth layer in the wavelet packet decomposition; σ1 is the threshold.

5. The method according to claim 1, characterized in that In step 3, the time domain information is converted into a grayscale image through the Gram angle field, the frequency domain information is converted into a grayscale image through the Gram angle difference field, the phase information is converted into a grayscale image through the Markov transfer field, and the three grayscale images are used as different channels of the RGB image to construct a fused feature image. The time domain information Gram angle field is expressed as: G T (i,j)=cos(θ i +θ j ) The frequency domain information Gram angle difference field is expressed as: G F (i,j)=sin(θ i -θ j ) Among them, θ i =cos -1 (X i ), X is the normalized time series. The phase information Markov transfer field is expressed as: P i,j =p i,j (x t ∈q i |x t-1 ∈q j )。 6. The method according to claim 1, characterized in that In step 4, the residual neural network uses the ResNet structure and embeds a convolutional block attention module in the key layer of the network, wherein the CBAM includes a channel attention module and a spatial attention module to adaptively enhance the attention to important feature areas, thereby improving the recognition accuracy.

7. The method according to claim 1, characterized in that The spatial attention module generates a spatial weight map by performing a convolution operation on the feature map to enhance the important spatial regions in the feature map.

8. The method according to claim 1, characterized in that In step 4, the channel attention module obtains the channel weight distribution of the feature map through global average pooling and global maximum pooling, and adjusts the weights of different channels in the feature map according to these weights, which is expressed as: s channel =σ(W1(ReLU(W0z avg ))+W1(ReLU(W0z max ))) Among them, z avg and z max They are the feature vectors after average pooling and maximum pooling, respectively.

9. The method according to claim 1, characterized in that: In step 5, the classification loss function is a cross entropy loss function, which is used to calculate the error between the predicted result and the actual label to guide model optimization and is defined as: Among them, y i is the actual label, is the predicted probability, N is the total number of categories.

Images