Partial discharge mode identification method, system and device based on convolutional neural network, and medium

Abstract

The invention discloses a partial discharge mode identification method, system and device based on a convolutional neural network, and a medium, and the method comprises the steps: employing a multi-layer convolutional structure, and generating a feature map; filling the boundary of the feature map to retain information features at the boundary of the input sample; dynamically distributing the weight of the channel of the feature map and the weight of the spatial region through an attention mechanism to obtain a weight feature map; performing maximum pooling operation on the weight feature map to obtain key features of the partial discharge mode; establishing a random inactivation layer after the pooling layer, and performing dynamic sparsification on the key features; and integrating the weight feature map and the key features, and outputting a classification result of the partial discharge mode. According to the method, the extraction and classification performance of partial discharge features is effectively improved, the robustness and generalization ability of the model under complex working conditions are enhanced, and reliable technical support is provided for power cable insulation state evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge pattern recognition, and in particular to a partial discharge pattern recognition method, system, device and medium based on a convolutional neural network. Background Art

[0002] Statistical analysis of power system failures in recent years has revealed that cable insulation degradation is one of the main causes of power grid accidents, and partial discharge is the most typical early sign of insulation defects. Existing cable condition monitoring relies primarily on overall electrical parameter detection, but these methods struggle to accurately identify microscopic discharge activity within the insulation. Partial discharge patterns, as a direct reflection of the physical discharge process, contain key information such as defect type and severity, and can more fundamentally reveal the mechanism of insulation degradation. Furthermore, the discharge behavior caused by different defects exhibits distinct pattern characteristics. Accurate identification of discharge patterns can provide early classification and warning of insulation defects, providing a scientific basis for cable condition inspection and maintenance. Current methods for identifying cable partial discharge patterns primarily include traditional statistical feature methods, time-frequency analysis, and deep learning.

[0003] At present, the following problems and difficulties still exist in the cable partial discharge pattern recognition method:

[0004] (1) Traditional statistical feature methods rely on manually extracted features, which makes it difficult to fully characterize complex partial discharge patterns. They are also less robust to weak discharge signals and noise interference, and have limited generalization capabilities.

[0005] (2) Although the time-frequency analysis method can provide richer time-frequency domain features, it has high computational complexity and is sensitive to parameter selection (such as wavelet basis function, time-frequency resolution, etc.), making it difficult to achieve adaptive optimization in practical applications;

[0006] (3) Although deep learning methods can automatically learn effective discharge features, they still face some challenges in practical applications, such as the model is prone to overfitting, is sensitive to noise interference, and has poor generalization performance when the sample size is insufficient. Summary of the Invention

[0007] In view of the above existing problems, the present invention is proposed.

[0008] Therefore, the present invention provides a local discharge pattern recognition method and system based on convolutional neural network to solve the problems of limited feature extraction capability, weak anti-interference capability and insufficient model generalization performance in existing cable local discharge pattern recognition methods.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In a first aspect, the present invention provides a method for partial discharge pattern recognition based on a convolutional neural network, comprising:

[0011] A multi-layer convolution structure is used to enhance the input features and generate feature maps;

[0012] Fill the feature map boundary to retain the information features at the input sample boundary;

[0013] Dynamically assign weights to channels and spatial regions of feature maps through the attention mechanism to obtain weighted feature maps.

[0014] Establish a pooling layer, perform a maximum pooling operation on the weighted feature map, and filter out the key features of the partial discharge pattern from the spatial dimension;

[0015] A random dropout layer is established after the pooling layer to dynamically sparsify the key features;

[0016] The weighted feature map and key features are integrated to output the classification results of the partial discharge pattern.

[0017] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, wherein: the multi-layer convolution structure includes:

[0018] The first convolutional layer is used to extract the global spatial features of the input features and output the extracted global spatial features as a preliminary feature map;

[0019] The second convolutional layer is used to extract local detail features of the input features and output the extracted local detail features as a deep feature map.

[0020] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, wherein: the feature enhancement processing of the input features includes:

[0021] The feature map includes the first layer feature map and the second layer feature map;

[0022] Perform feature enhancement processing on the preliminary feature map output by the first convolutional layer to obtain the first layer feature map;

[0023] The first layer feature map is input into the second convolution layer, and the output deep feature map is enhanced again to obtain the second layer feature map.

[0024] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, filling the boundary of the feature map includes:

[0025] Add a corresponding number of zero-valued pixels at the boundary of the input sample;

[0026] The convolution operation is performed on the input features after adding zero-value pixels, so that the size of the output feature map remains consistent before and after the convolution operation.

[0027] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, the attention mechanism includes:

[0028] Assign weights to the channel dimensions of the feature map;

[0029] Assign weights to the spatial dimensions of the feature map;

[0030] Strengthen the features of the key channel spatial area of ​​the feature map and suppress the features of the irrelevant areas of the feature map.

[0031] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, wherein: the maximum pooling method is used to reduce the resolution of the feature map of each layer of convolutional neural network, and the key feature information is retained, including:

[0032] Perform maximum pooling with a pooling window of 2×2 on the first layer feature map, and reduce the resolution of the feature map by taking the maximum value in the local area, retaining the key feature information, and performing deeper feature extraction on the first layer feature map;

[0033] The maximum pooling process with a pooling window of 2×2 is applied again to the second layer feature map to further reduce the resolution of the feature map, extract deeper features, and obtain key features.

[0034] As a preferred solution of the partial discharge pattern recognition method based on convolutional neural network of the present invention, wherein: integrating the weight feature map and the key features, outputting the classification result of the partial discharge pattern includes:

[0035] Establish two fully connected layers, namely the first fully connected layer and the second fully connected layer;

[0036] The first fully connected layer flattens the feature map extracted by the convolutional layer into a one-dimensional feature vector and maps the flattened one-dimensional feature vector to the intermediate dimension;

[0037] The second fully connected layer maps the one-dimensional feature vector mapped to the intermediate dimension to the final number of categories to obtain the final classification result.

[0038] In a second aspect, the present invention provides a local discharge pattern recognition system based on a convolutional neural network, comprising: a feature extraction module for extracting local features of input data;

[0039] Feature integration module, which is used to flatten the multi-dimensional feature map into a one-dimensional vector and perform global integration on the extracted features to extract higher-level features;

[0040] The classification output module is used to convert the output of the fully connected layer into a probability distribution, representing the predicted probability of each category, and complete the final classification decision and output the classification result;

[0041] The inference module uses the trained model to infer the preprocessed data and output the corresponding classification results.

[0042] Compared with the prior art, the present invention has the following beneficial effects: the present invention realizes the nonlinear transformation and sparsification of features by constructing a two-layer convolutional neural network basic model framework and combining the ReLU activation function, thereby enhancing the model's learning ability for complex features; by adopting the same filling method, the boundary information of the input sample is retained, ensuring the stability of the feature map size; by introducing the channel space attention mechanism module to capture the subtle differences in the partial discharge image, and combining the deactivation layer mechanism to optimize the network regularization effect, effectively improving the automatic extraction and classification performance of partial discharge features; by establishing the maximum pooling layer, the feature map resolution is reduced and the key feature information is retained, thereby improving the generalization ability of the model; by establishing the random deactivation layer, the overfitting of the fully connected layer is effectively suppressed, thereby enhancing the robustness of the model; by establishing the fully connected layer, the integration and classification output of the extracted features are realized, the accurate identification of the partial discharge pattern is completed, and reliable technical support is provided for the evaluation of the insulation status of power cables. Through the above steps, the present invention effectively improves the extraction and classification performance of partial discharge features, enhances the robustness and generalization ability of the model under complex working conditions, and provides reliable technical support for the evaluation of the insulation status of power cables. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a schematic diagram of the convolution and pooling principles of a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the same filling convolution of the local discharge pattern recognition method based on convolutional neural network according to an embodiment of the present invention.

[0046] Figure 3 A schematic diagram of Valid Padding convolution of a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention.

[0047] Figure 4A schematic diagram of maximum pooling and average pooling in a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention.

[0048] Figure 5 Schematic diagram of the pooling training process of the partial discharge pattern recognition method based on convolutional neural network according to an embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram of the feature matching process of a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention.

[0050] Figure 7 Schematic diagram of the channel attention mechanism of the local discharge pattern recognition method based on convolutional neural network according to an embodiment of the present invention.

[0051] Figure 8 Schematic diagram of the spatial attention mechanism of the local discharge pattern recognition method based on convolutional neural network according to an embodiment of the present invention.

[0052] Figure 9 Schematic diagram of a standard CNN training model and a CNN training model with a random deactivation layer added for a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention.

[0053] Figure 10 Schematic diagram of an improved convolutional neural network model of a partial discharge pattern recognition method based on a convolutional neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0055] Example 1

[0056] Reference Figures 1 to 10 , as one embodiment of the present invention, provides a method for partial discharge pattern recognition based on a convolutional neural network, comprising:

[0057] S100: Use a multi-layer convolutional structure to enhance the input features and generate feature maps;

[0058] S200: Filling the feature map boundary to retain the information features at the input sample boundary;

[0059] S300: Dynamically assign weights to channels and spatial regions of the feature map through the attention mechanism to obtain a weighted feature map;

[0060] S400: establishing a pooling layer, performing a maximum pooling operation on the weight feature map, and filtering out key features of the partial discharge pattern from a spatial dimension;

[0061] S500: A random dropout layer is established after the pooling layer to dynamically sparse the key features;

[0062] S600: Integrate the weighted feature map and the key features, and output the classification result of the partial discharge pattern.

[0063] It should be noted that the recognition capability of partial discharge patterns of power cables is relatively low, especially for weak discharge signals, which are difficult to accurately identify. In practical applications, there are complexities and technical difficulties in feature extraction and feature calculation, which are mainly reflected in the following aspects:

[0064] First, partial discharge (PD) phenomena are complex and diverse, with numerous discharge types and patterns. PD signature patterns contain a variety of frequency, amplitude, and phase information, and different types of PD exhibit distinct patterns on the characteristic patterns. Furthermore, PD signature patterns may contain noise and interference signals, which can easily lead to misidentification. For example, certain high-frequency interference signals may resemble the characteristic PD signal, making identification more difficult.

[0065] Secondly, the complex operating environment of cables leads to inconsistent signal quality. In actual operation, due to the varying environmental conditions in which cables operate, partial discharge signals may be affected by factors such as electromagnetic interference, temperature fluctuations, and humidity changes. This can lead to inconsistent signal quality and strength between cables, seriously affecting the accuracy of pattern recognition. For example, in environments with strong electromagnetic interference, partial discharge signals may be drowned out, while in humid environments, the signal may be attenuated.

[0066] Finally, the pre-established partial discharge pattern datasets have limited sample diversity and low coverage. The samples in the datasets primarily come from common operating environments or equipment states, lacking coverage for diverse operating conditions such as different voltage levels, different operating ages, and varying ambient humidity. Furthermore, the datasets contain relatively few samples of rare partial discharge patterns or those under special operating conditions, making them difficult to identify when they occur.

[0067] Therefore, the present invention constructs a complete technical solution for local discharge pattern recognition based on convolutional neural networks through steps S100-S600, and realizes nonlinear transformation and sparsification of features by constructing a two-layer convolutional neural network basic model framework and combining it with the ReLU activation function, thereby enhancing the model's learning ability for complex features; by adopting the same filling method, the boundary information of the input samples is retained, ensuring the stability of the feature map size; by introducing the channel space attention mechanism module to capture subtle differences in the local discharge image, and combining the deactivation layer mechanism to optimize the network regularization effect, effectively improving the automatic extraction and classification performance of local discharge features; by establishing a maximum pooling layer, the feature map resolution is reduced and key feature information is retained, thereby improving the generalization ability of the model; by establishing a random deactivation layer, the overfitting of the fully connected layer is effectively suppressed, thereby enhancing the robustness of the model; by establishing a fully connected layer, the integration and classification output of the extracted features are realized, and the accurate identification of the local discharge pattern is completed, providing reliable technical support for the insulation status assessment of power cables. Through the above steps, the present invention effectively improves the extraction and classification performance of partial discharge features, enhances the robustness and generalization ability of the model under complex working conditions, and provides reliable technical support for the insulation status assessment of power cables.

[0068] Example 2

[0069] Reference Figures 1 to 10 , which is an embodiment of the present invention, provides a local discharge pattern recognition method based on convolutional neural network based on the above embodiment.

[0070] In the embodiment of the present application, a multi-layer convolution structure is used in step S100 to perform feature enhancement processing on the input features to generate a feature map, which includes the following steps A1 to A2:

[0071] A1: Multi-layer convolutional structure includes,

[0072] The first convolutional layer is used to extract the global spatial features of the input features and output the extracted global spatial features as a preliminary feature map;

[0073] The second convolutional layer is used to extract local detail features of the input features and output the extracted local detail features as a deep feature map.

[0074] It should be noted that the present invention constructs a two-layer convolutional neural network as a basic model framework for extracting and outputting feature maps, wherein the specific structure includes:

[0075] The first convolutional layer uses a 5×5 convolution kernel to extract the global spatial features of the input features and sets the number of output channels to 32;

[0076] The second convolutional layer uses a 3×3 convolution kernel to further capture local detail features and increases the number of output channels to 64.

[0077] It should be further explained that the present invention uses a high-definition camera with a resolution of 1920×1080 to capture data on partial discharge conditions, and crops a sub-image containing the partial discharge area from the 1920×1080 image with a cropping size of 512×512. The cropped image is then scaled to 224×224, and the image is input as an input feature to the first convolutional layer, so that the first convolutional layer performs a convolution operation on the input feature to obtain a preliminary feature map.

[0078] After the preliminary feature map is subjected to nonlinear transformation, key feature highlighting, maximum pooling, and random discarding of some neurons, the final first-layer feature map is output. At this time, the size of the first-layer feature map is 112×112 and the number of channels is 32;

[0079] The first layer feature map is input into the second convolution layer and a second convolution operation is performed to finally obtain a second layer feature map with a size of 56×56 and 64 channels.

[0080] Among them, the calculation process formula of the convolution operation is as follows:

[0081]

[0082] in, is the jth element in the output feature map of the lth layer; M is the number of output channels of the l-1th layer (that is, the number of input channels of this layer); k j is the jth convolution kernel of the convolution operation; f is the activation function used to introduce nonlinearity; Represents the i-th element of the feature map of the previous level (the l-1 layer); represents the jth incident bias of the lth layer; i is the index of the input channel (corresponding to the i-th feature map output by the previous layer); j is the index of the output channel (corresponding to the j-th feature map generated by the current layer); l is the number of the current convolutional layer.

[0083] Specifically, the process of convolution operation is as follows:

[0084] The convolution kernel k j Placed in the input feature map The upper left corner of the convolution kernel k is calculated. j With the input feature map The dot product of the covered area;

[0085] The offset Add to the dot product result;

[0086] Apply the activation function f to the biased dot product result to get the output feature map An element of .

[0087] Slide the convolution kernel to the next position of the input feature map according to the step size, and repeat the above steps until the entire input feature map is covered;

[0088] Repeat the above steps, using all convolution kernels to get the complete output feature map

[0089] It should be noted that the convolution operation process will generate a feature map for each convolution kernel, and all feature maps are stacked together to form the final output feature map, that is, the final output feature map generated by the convolution operation of the first convolution layer is the first layer feature map, and the final output feature map generated by the convolution operation of the second convolution layer is the second layer feature map. Both the first layer feature map and the second layer feature map contain the extracted features of the input features.

[0090] A2: Feature enhancement processing of input features includes:

[0091] The feature map includes the first layer feature map and the second layer feature map;

[0092] Perform feature enhancement processing on the preliminary feature map output by the first convolutional layer to obtain the first layer feature map;

[0093] The first layer feature map is input into the second convolution layer, and the output deep feature map is enhanced again to obtain the second layer feature map.

[0094] Among them, feature enhancement processing includes nonlinear transformation and sparse processing of input features.

[0095] Specifically, the present invention uses the ReLU activation function to perform feature enhancement processing on the input features, that is, the input features after each layer of convolution operation are input into the ReLU activation function (Rectified Linear Unit, linear rectification function) to perform nonlinear transformation and sparsification of the features. The specific process includes:

[0096] The input features are convolved through the first convolutional layer to obtain a preliminary feature map, which is then input into the ReLU activation function for nonlinear transformation and feature sparsification to obtain the first-layer feature map.

[0097] The first layer feature map is used as the input of the second convolution layer, and the second layer convolution operation is performed to obtain the deep feature map. The deep feature map is input into the ReLU activation function again for nonlinear transformation and feature sparsification to obtain the second layer feature map;

[0098] The feature maps include the first layer feature maps and the second layer feature maps.

[0099] Among them, the ReLU activation function is a non-saturated activation function, and its expression is as follows:

[0100]

[0101] That is, for each input feature x, if x is greater than 0, output x; if x is less than or equal to 0, output 0.

[0102] like Figure 4 As shown in the figure, the ReLU activation function completely avoids the dead zone on the positive half axis by setting the negative value to 0, retains the positive part of the input, and gives the feature nonlinear characteristics. The ReLU activation function also has linear convergence and fast calculation capabilities.

[0103] It should be noted that in the present invention, after the convolution operation, the ReLU activation function is applied to each element of the feature map, and the formula is as follows:

[0104]

[0105] Among them, y' j is the output after ReLU activation; M is the number of output channels of the l-1th layer (that is, the number of input channels of this layer); k j is the convolution kernel; It is the feature map of the previous layer; is the bias; i is the index of the input channel (corresponding to the i-th feature map output by the previous layer); j is the index of the output channel (corresponding to the j-th feature map generated by the current layer); l is the number of the current convolutional layer.

[0106] In another optional embodiment, step S100 may also use Leaky ReLU (Leaky Rectified Linear Unit, also known as LReLU, leaky linear rectifier unit) as the activation function. Leaky ReLU is an improved ReLU activation function that allows negative values ​​to have a non-zero output. The formula of Leaky ReLU is as follows:

[0107]

[0108] Where α is a small positive constant (usually between 0.01 and 0.02) that controls the slope of negative values;

[0109] x is the input feature.

[0110] Leaky ReLU can avoid neuron death by providing a non-zero output for negative input, allowing neurons to continue updating their weights during training. By adjusting the value of α, Leaky ReLU can provide more flexibility for the model.

[0111] In the embodiment of the present application, step S200 fills the feature map boundary to retain the information features at the input sample boundary, including the following steps B1-B2:

[0112] B1: Add a corresponding number of zero-value pixels at the boundary of the input sample;

[0113] Specifically, according to the size and step size of the convolution kernel, a corresponding number of zero-valued pixels are added at the boundary of the input sample.

[0114] B2: Perform convolution operation on the input features after adding zero-value pixels, so that the size of the output feature map remains consistent before and after the convolution operation.

[0115] It should be noted that the present invention uses the same filling (zero filling) method to fill the boundary of the feature map, and this method needs to be performed before the convolution operation.

[0116] Specifically, the same filling method effectively preserves the boundary information by adding zero-value pixels at the boundary of the input sample. Figure 2 As shown in Figure 1, by using the same padding operation, even if the convolution kernel slides at the boundary, it can ensure that the output feature map will not lose information due to boundary effects. Among them, the input feature map size is H×W (height H, width W); the convolution kernel size is K h ×K w (Height K h , width K w ); the step size is S, and the calculation formula for the filling amount is as follows:

[0117]

[0118] Among them, P h is the filling amount in the height direction; P w is the filling amount in the width direction; and Both indicate rounding down.

[0119] Using the same padded feature map for convolution operation, the size of the output feature map will be the same as the input feature map, that is, H output ×W output Among them, the size formula of the output feature map is as follows:

[0120]

[0121] Among them, H output and W output are the height and width of the output feature map, respectively.

[0122] In another optional implementation, step S200 may also use a ValidPadding method to extract boundary information, where ValidPadding performs a convolution operation by directly using valid pixels of the input sample.

[0123] When the feature map after Valid Padding is used for convolution operation, the size of the output feature map will be smaller than or equal to the input feature map, that is, H output ×W output Among them, the size formula of the output feature map is as follows:

[0124]

[0125] Among them, H output and W output are the height and width of the output feature map, respectively.

[0126] like Figure 3 As shown in the figure, since the convolution kernel will not exceed the boundary of the input feature map, Valid Padding will reduce the size of the output feature map. At the same time, because it will not perform calculations outside the boundary, the boundary information of the input feature map can be maintained while reducing the size. At the same time, by selecting the appropriate step size and convolution kernel size, the ValidPadding method can accurately control the size of the output feature map.

[0127] In the embodiment of the present application, step S300 dynamically allocates weights of channels and spatial regions of the feature map through the attention mechanism to obtain a weighted feature map, including the following steps C1-C3:

[0128] C1: weight distribution of the channel dimension of the feature map;

[0129] Specifically, by establishing a channel attention mechanism, global average pooling is performed on the input feature map to obtain the global information of each channel of the feature map, and the result of global average pooling is mapped to the channel space through 1×1 convolution to generate the importance weight of each channel;

[0130] The ReLU activation function is used to introduce nonlinearity into each channel. Based on the importance weight of each channel, the channel attention weight coefficient is generated through the Sigmoid function. The channel attention weight coefficient is multiplied by the input feature map to amplify important channel features and suppress unimportant channel features to obtain the channel weight feature map.

[0131] Among them, important channel features refer to feature channels that are strongly correlated with the discrimination of partial discharge patterns, and their responses significantly reflect the essential characteristics of specific discharge types. For example:

[0132] Corona discharge: high-frequency signal channel (characterizing short-duration spike pulses) and channel in the area of ​​concentrated electric field at the electrode tip;

[0133] Surface discharge: a channel with significant edge texture or gradient changes (indicating continuous discharge along the insulating surface);

[0134] Internal discharge: medium and low frequency energy distribution channel (characterizing the energy accumulation and release inside the medium).

[0135] Unimportant channel features refer to channels that have low correlation with discharge pattern discrimination or introduce noise. For example:

[0136] Channels dominated by sensor background noise (such as power frequency interference or electromagnetic noise);

[0137] Environmental signal channels unrelated to discharge (such as low-frequency fluctuations caused by equipment vibration);

[0138] Redundant feature channels (such as highly similar information between multiple channels).

[0139] Specifically, such as Figure 7 The channel attention mechanism shown includes,

[0140] ① Perform global average pooling on the input feature map to obtain the global information contained in each channel;

[0141] Among them, the global average pooling formula is as follows:

[0142]

[0143] in, Represents the eigenvalue of the c-th channel at position (i', j'); H and W represent the height and width of the feature map respectively; c represents the channel index, ranging from 1 to C, where C is the total number of channels; GAP(c) represents the global average pooling result of the c-th channel, which is a scalar representing the average value of all eigenvalues ​​of the channel; i' is the spatial position index of the feature map in the height direction (row); j' is the spatial position index of the feature map in the width direction (column).

[0144] ② The result of global average pooling is mapped to the channel space through a 1×1 convolution kernel, and nonlinearity is introduced to the feature map through the ReLU activation function;

[0145] ③ Use the Sigmoid function to generate attention weight coefficients, which determine the importance of each channel. Finally, a weight vector of size C×1 is obtained, and each weight value is between 0 and 1, where C is the number of channels.

[0146] Among them, the specific formula for generating the attention weight coefficient of the Sigmoid function is as follows:

[0147]

[0148] Among them, x is the input feature, the size is 1×H×W, and it is the result after convolution mapping;

[0149] e is Euler's constant;

[0150] σ(x) is the output of the Sigmoid function, with a size of 1×H×W. The value of each element is between 0 and 1, indicating the attention weight of each spatial position.

[0151] ④ Use the attention weight coefficient to multiply the input feature map to amplify the important channels. The specific formula is as follows:

[0152] F'=F⊙W

[0153] Among them, F' is the weighted feature map;

[0154] F is the input feature map;

[0155] W is the attention weight coefficient;

[0156] ⊙ represents element-wise multiplication.

[0157] It should be noted that through the above steps, the channel attention mechanism enables the convolutional neural network to adaptively focus on the more important feature channels in the input features, thereby improving the recognition accuracy and robustness of the model.

[0158] C2: Weight distribution of spatial dimensions of feature maps.

[0159] Specifically, by establishing a spatial attention mechanism, we perform maximum pooling and average pooling on the feature map to obtain two different feature maps;

[0160] The two feature maps are concatenated to form a feature map containing spatial information, and a convolution operation is performed on the concatenated feature map to generate a spatial attention map;

[0161] The spatial attention map is converted into a weight coefficient through the Sigmoid function, and the spatial attention weight is multiplied by the input feature map to strengthen the features of the key area and obtain the spatial weight feature map.

[0162] Specifically, the spatial attention mechanism includes:

[0163] ① Perform maximum pooling and average pooling on the feature map to obtain two different feature maps. The formulas for maximum pooling and average pooling are as follows:

[0164] Max Pooling:

[0165]

[0166] Average Pooling:

[0167]

[0168] Among them, MP(i',j') is the maximum pooling; AP(i',j') is the average pooling; is the eigenvalue of the cth channel at position (i', j'), and C is the number of channels.

[0169] ② The spliced ​​feature map is a feature map containing spatial information, and a feature map of size 2×H×W is obtained, which contains the maximum eigenvalue and average eigenvalue of each spatial position;

[0170] ③ By convolving the concatenated feature map, a spatial attention map is obtained;

[0171] ④ Use the Sigmoid function to convert the spatial attention map into a weight coefficient that represents the importance of each spatial position;

[0172] ⑤ Multiply the spatial attention weight by the input feature map to enhance the features of the key areas and suppress irrelevant areas, obtaining a weighted feature map of size C×H×W, which contains the enhanced feature information.

[0173] It should be noted that the introduction of CSAM (Channel and Spatial Attention Module) enables the convolutional neural network to adaptively assign different weights to channels and spatial regions, helping the model better capture subtle differences in partial discharge images and improve classification performance.

[0174] C3: Strengthen the features of the key channel spatial area of ​​the feature map and suppress the features of the irrelevant areas of the feature map.

[0175] The key channel spatial region refers to the local area in the feature map where discharge signals appear in a concentrated manner or have obvious morphological characteristics, for example, the bright area formed by corona discharge at the tip of the electrode.

[0176] Irrelevant regions refer to locations in the feature map that are occupied by background, noise, or non-discharge signals. Examples include: invalid pixel regions at the edge of an image; background regions of insulating materials where no discharge occurs; and random noise points generated by electromagnetic interference.

[0177] In the embodiment of the present application, a pooling layer is established in step S400, a maximum pooling operation is performed on the weight feature map, and key features of the partial discharge pattern are screened from the spatial dimension, including the following steps D1-D2:

[0178] D1: Perform maximum pooling with a pooling window of 2×2 on the first layer feature map, reduce the resolution of the feature map by taking the maximum value in the local area, retain key feature information, and perform feature extraction on the first layer feature map;

[0179] D2: Apply the maximum pooling process with a pooling window of 2×2 to the second layer feature map again to further reduce the resolution of the feature map and obtain the key features.

[0180] Among them, key features refer to the significant features of spatial dimensions that are selected from the weighted feature map and have a decisive role in the discrimination of partial discharge patterns. Its core content includes:

[0181] (1) Spatial distribution characteristics of discharge pulses:

[0182] High-response areas: These areas have significantly higher attention weights than surrounding spatial locations, typically corresponding to areas where discharge activity is concentrated. Examples include corona discharges: bright pulse clusters at the electrode tip or at locations with sudden changes in curvature; surface discharges: continuous discharge traces distributed along the surface of an insulating material; and internal discharges: discrete strong pulse points or bubble regions within a dielectric.

[0183] Spatial correlation: The specific geometric patterns (such as lines, rings, or clusters) formed by adjacent high-weight areas reflect the discharge propagation path or defect structure characteristics.

[0184] (2) Discharge intensity and amplitude characteristics

[0185] Pulse Amplitude Extreme: The local maximum value extracted by max pooling represents the instantaneous peak value of the discharge energy. For example, it can be a short-duration high-frequency spike in corona discharge or a long-duration high-energy pulse in internal discharge.

[0186] Amplitude distribution statistics: The distribution characteristics of the amplitude within the pooling window (such as variance and skewness) are used to distinguish stable discharges from random noise.

[0187] (3) Morphological features and edge information

[0188] Edge sharpness: The transition characteristics between the discharge area and the background, for example: the sharp edge of corona discharge (dominated by high-frequency components); the diffuse edge of surface discharge (propagating slowly along the surface).

[0189] Shape integrity: The contour features of the discharge area retained after pooling (such as circular, linear, or irregular shapes) reflect the discharge type and defect morphology.

[0190] (4) Statistical and frequency domain correlation characteristics

[0191] Pulse density: the number of high-weight points in a unit area, distinguishing dense discharges (such as multiple defects) from isolated discharges;

[0192] Frequency domain response pattern: indirectly reflects frequency domain characteristics through spatial distribution, such as high-frequency discharge corresponds to a fine dot matrix, and low-frequency discharge corresponds to a wide area.

[0193] Specifically, the specific steps of performing pooling processing on the first layer feature map include:

[0194] First, input feature map, size is H×W×C, where H is height, W is width, and C is the number of channels;

[0195] Secondly, a pooling window of size 2×2 is selected to slide on the input feature map;

[0196] Again, the pooling layer performs a maximum pooling operation at each pooling window position, that is, the maximum value of all values ​​in the window is selected as the output;

[0197] Finally, after the pooling operation is completed, the pooling layer outputs a new feature map whose size depends on the window size, stride, and the size of the input feature map.

[0198] like Figure 6 As shown in Figure 2, the accuracy of the maximum pooling method will continue to improve as the number of convolutional neural network iterations increases.

[0199] It should be noted that pooling layers typically follow convolutional layers to further extract and compress features. Convolutional layers extract local features from the input data, while pooling layers reduce the spatial dimensionality of these features. In this paper, pooling layers are used after each convolutional layer. This structure facilitates the progressive extraction and compression of features while maintaining the depth and expressiveness of the model.

[0200] The pooling layer is used before the channel-space attention mechanism module to provide a more compact feature representation for the attention mechanism. In this way, the model can adaptively assign different weights to channels and spatial regions, thereby better capturing the subtle differences in partial discharge images.

[0201] In the embodiment of the present application, in step S500, a random dropout layer is established after the pooling layer to dynamically sparse the key features, including:

[0202] A random dropout layer is established to prevent overfitting of the fully connected layer by randomly blocking some neuron connections.

[0203] Specifically, the Dropout method (random inactivation method) is introduced to implement dynamic random inactivation of neurons through the fully connected layer in the forward propagation stage. Its core mechanism is to temporarily remove specific neuron nodes with a preset probability P. This strategy not only enhances the generalization performance of the model, but also effectively suppresses the excessive memory effect and oversensitivity to key features during the training process. Figure 9 As shown in the figure, (a) is the standard CNN training model without the introduction of the Dropout method, and (b) is the CNN training model with the Dropout method. The specific implementation steps of Dropout are as follows:

[0204] (1) Preset probability P and temporarily discard some neurons with probability P;

[0205] (2) During the training process, only part of the training set is input through the Dropout network for forward propagation, and then the error is back-propagated. This process must ensure that only neurons that are not discarded participate in gradient calculation and weight update, and the discarded neurons will not affect the gradient propagation;

[0206] (3) Restore the temporarily discarded neurons and repeat steps (1) and (2).

[0207] In the embodiment of the present application, step S600 integrates the weighted feature map and the key features to output the classification result of the partial discharge mode, including the following steps E1-E2:

[0208] It should be noted that the classification results of partial discharge modes include corona discharge, creeping discharge, and internal discharge.

[0209] E1: Establish two fully connected layers, namely the first fully connected layer and the second fully connected layer;

[0210] The first fully connected layer flattens the feature map extracted by the convolutional layer into a one-dimensional feature vector and maps the flattened one-dimensional feature vector to the intermediate dimension;

[0211] This layer contains a large number of neurons to capture and integrate the complex features extracted from the convolutional layer and the pooling layer. The number of neurons is equal to the number of categories in the classification task, and each neuron corresponds to one category.

[0212] The flattening operation arranges the neurons in all feature maps sequentially into a long vector.

[0213] The second fully connected layer maps the one-dimensional feature vector mapped to the intermediate dimension to the final number of categories to obtain the final classification result;

[0214] E2: Determine a specific category in the final classification result of the partial discharge pattern, and output the specific category in the form of probability distribution.

[0215] The number of neurons in this layer is equal to the number of categories in the classification task. Each neuron corresponds to a category, and its output value represents the probability or score of the input belonging to that category.

[0216] The Softmax function is used to ensure that the sum of the output probabilities of all categories is 1, so that it can be directly interpreted as category probability;

[0217] Specifically, the input of the Softmax function is the output of the second fully connected layer, that is, a vector z = [z1, z2, ..., z n ..., z K ], where K is the total number of categories.

[0218] For each category n, the Softmax function calculates the probability p n The formula is as follows:

[0219]

[0220] Where n is the number of categories; is the index of the score for category n; It is the sum of all category score indices and is used for normalization. k is a summation variable, which means that when calculating the denominator, all categories (from category 1 to category K) are traversed, and the value range is k = 1, 2, ..., K.

[0221] Finally, the Softmax function outputs a probability distribution p = [p1, p2, ..., p n ..., p K ], where each p i represents the probability that the input image belongs to the nth class.

[0222] Through the back-propagation algorithm, the weights of the model are updated according to the gradient of the loss function to minimize the loss.

[0223] Specifically, the cross-entropy loss function is used during the training process to measure the difference between the probability distribution of the model output and the probability distribution of the true label, and the gradient of the loss function is used to update the weights of the model through the back-propagation algorithm to reduce the loss and improve the prediction accuracy of the model.

[0224] It should be noted that by establishing a fully connected layer and integrating features, the features extracted from the convolutional and pooling layers can be effectively mapped to the final classification results. This structure not only helps capture local features of the image, but also captures the relationship between global features and categories through the integration of fully connected layers, thereby improving classification accuracy and robustness.

[0225] Table 1 further describes the model parameters of the convolutional neural network (CNN). The input is a single-channel grayscale image of size 224×224. The architecture employs an alternating stacking of convolutional layers and channel-wise spatial attention modules. A 5×5 convolution kernel is used to extract global spatial features from the grayscale image. The CSAM1 module dynamically weights key channels and spatial regions, and 2×2 pooling reduces the image resolution to 112×112. The second convolution layer uses a smaller 3×3 convolution kernel to further capture local details. The number of channels is increased to 64, and feature importance is recalibrated using the CSAM2 module. Pooling further reduces the resolution to 56×56. Finally, the model passes through a dropout layer and then a fully connected layer, feeding into a softmax classifier for classification.

[0226]

[0227]

[0228] Table 1 CNN network model parameters

[0229] In summary, the present invention uses a two-layer convolutional neural network basic model framework to extract global and local detail features of input features using convolution kernels of different sizes, and introduces nonlinear transformation and feature sparsification through the ReLU activation function, thereby enhancing the model's ability to capture key features. Secondly, the introduction of the same filling method effectively retains the information features at the input sample boundary and avoids information loss caused by boundary effects. Then, through the channel and spatial attention mechanism, the model can adaptively assign different weights to the channels and spatial regions of the feature map, further strengthen key features, suppress irrelevant areas, and improve the model's ability to recognize subtle differences in partial discharge images. In addition, the introduction of the maximum pooling layer not only reduces the resolution of the feature map and extracts deeper features, but also effectively prevents overfitting through the random inactivation layer, thereby enhancing the generalization performance of the model. Finally, the extracted features are integrated through the fully connected layer, and the Softmax function is used to output the final classification result, ensuring that the sum of the output probabilities of all categories is 1, so that the output can be directly interpreted as the probability of each category. Overall, the method of the present invention significantly improves the accuracy of pattern classification while maintaining real-time processing capabilities, providing a reliable technical means for cable insulation status assessment.

[0230] Example 3

[0231] The above is a schematic diagram of a method for partial discharge pattern recognition based on a convolutional neural network. It should be noted that the technical solution of this system for partial discharge pattern recognition based on a convolutional neural network and the technical solution of the method for partial discharge pattern recognition based on a convolutional neural network share the same concept. For details not described in detail in the technical solution of the partial discharge pattern recognition system based on a convolutional neural network in this embodiment, please refer to the description of the technical solution of the method for partial discharge pattern recognition based on a convolutional neural network.

[0232] This embodiment further provides a partial discharge pattern recognition system based on a convolutional neural network, comprising:

[0233] Feature extraction module, used to extract local features of input data;

[0234] Feature integration module, which is used to flatten the multi-dimensional feature map into a one-dimensional vector and perform global integration of the extracted features;

[0235] The classification output module is used to convert the output of the fully connected layer into a probability distribution, representing the predicted probability of each category, and complete the final classification decision and output the classification result;

[0236] The inference module uses the trained model to infer the preprocessed data and output the corresponding classification results.

[0237] This embodiment also provides an electronic device suitable for partial discharge pattern recognition based on a convolutional neural network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the partial discharge pattern recognition method based on a convolutional neural network proposed in the above embodiment.

[0238] This embodiment further provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the method for partial discharge pattern recognition based on a convolutional neural network as proposed in the above embodiment is implemented.

[0239] The storage medium proposed in this embodiment and the method for implementing partial discharge pattern recognition based on a convolutional neural network proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0240] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general hardware, and of course can also be implemented by hardware. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer's floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0241] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for partial discharge pattern recognition based on convolutional neural network, characterized in that: include: A multi-layer convolution structure is used to enhance the input features and generate feature maps; Fill the feature map boundary to retain the information features at the input sample boundary; Dynamically assign weights to channels and spatial regions of feature maps through the attention mechanism to obtain weighted feature maps. Establishing a pooling layer, performing a maximum pooling operation on the weighted feature map, and filtering out key features of the partial discharge pattern from a spatial dimension; A random dropout layer is established after the pooling layer to dynamically sparsify the key features; The weighted feature map and key features are integrated to output the classification results of the partial discharge pattern.

2. The method for partial discharge pattern recognition based on a convolutional neural network according to claim 1, wherein: The multi-layer convolution structure includes: The first convolutional layer is used to extract the global spatial features of the input features and output the extracted global spatial features as a preliminary feature map; The second convolutional layer is used to extract local detail features of the input features and output the extracted local detail features as a deep feature map.

3. The method for partial discharge pattern recognition based on convolutional neural network according to claim 2, characterized in that: The feature enhancement processing of the input features includes: The feature map includes a first layer feature map and a second layer feature map; Perform feature enhancement processing on the preliminary feature map output by the first convolutional layer to obtain the first layer feature map; The first layer feature map is input into the second convolution layer, and the output deep feature map is enhanced again to obtain the second layer feature map.

4. The method for partial discharge pattern recognition based on convolutional neural network according to claim 3, characterized in that: Filling the boundary of the feature map includes: Add a corresponding number of zero-valued pixels at the boundary of the input sample; The convolution operation is performed on the input features after adding zero-value pixels, so that the size of the output feature map remains consistent before and after the convolution operation.

5. The method for partial discharge pattern recognition based on convolutional neural network according to claim 4, characterized in that: The attention mechanism includes, Assign weights to the channel dimensions of the feature map; Assign weights to the spatial dimensions of the feature map; Strengthen the features of the key channel spatial area of ​​the feature map and suppress the features of the irrelevant areas of the feature map.

6. The method for partial discharge pattern recognition based on convolutional neural network according to claim 5, characterized in that: The key features of screening out partial discharge patterns from the spatial dimension include: Perform maximum pooling with a pooling window of 2×2 on the first layer feature map, reduce the resolution of the feature map by taking the maximum value in the local area, retain key feature information, and perform feature extraction on the first layer feature map; The second layer feature map is again subjected to maximum pooling with a pooling window of 2×2 to further reduce the resolution of the feature map and obtain key features.

7. The method for partial discharge pattern recognition based on convolutional neural network according to claim 6, characterized in that: The integration of the weighted feature map and the key features to output the classification results of the partial discharge mode includes: Establish two fully connected layers, namely the first fully connected layer and the second fully connected layer; The first fully connected layer flattens the feature map extracted by the convolutional layer into a one-dimensional feature vector and maps the flattened one-dimensional feature vector to the intermediate dimension; The second fully connected layer maps the one-dimensional feature vector mapped to the intermediate dimension to the final number of categories to obtain the final classification result; A specific category in the final classification result of the partial discharge pattern is determined, and the specific category is output in the form of a probability distribution.

8. A partial discharge pattern recognition system based on a convolutional neural network, applying the method according to any one of claims 1 to 7, characterized in that: include: Feature extraction module, used to extract local features of input data; Feature integration module, which is used to flatten the multi-dimensional feature map into a one-dimensional vector and perform global integration of the extracted features; The classification output module is used to convert the output of the fully connected layer into a probability distribution, representing the predicted probability of each category, and complete the final classification decision and output the classification result; The inference module uses the trained model to infer the preprocessed data and output the corresponding classification results.

9. An electronic device comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the local discharge pattern recognition method based on convolutional neural network according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the partial discharge pattern recognition method based on convolutional neural network according to any one of claims 1 to 7.

Images