GIS partial discharge feature matching method and system based on deep reinforcement learning

Abstract

The invention discloses a GIS local discharge feature matching method and system based on deep reinforcement learning, comprising collecting a UHF signal of a GIS local discharge model, obtaining a PRPD feature spectrum corresponding to the UHF signal; inputting the feature spectrum into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network comprises a feature extraction network and an output network, wherein the feature extraction network is formed by superposition of a plurality of feature extraction layers, and each feature extraction layer comprises a channel attention module and a space attention module connected in sequence; constructing a data set using deep features, and adopting a deep reinforcement learning algorithm to train a GIS local discharge feature matching Markov model to obtain an optimal GIS local discharge feature matching strategy; the invention improves the accuracy of GIS local discharge feature matching results by extracting deep features of UHF signals and using them for GIS local discharge feature matching optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment, and in particular to a GIS local discharge feature matching method and system based on deep reinforcement learning. Background Art

[0002] With the accumulation of partial discharge detection cases, matching suspected partial discharge data with data in historical cases to obtain similar cases is an idea for deep mining of partial discharge data under the background of big data. Intelligently matching partial discharge detection data with detection data in a large number of historical case libraries to find similar data under similar operating conditions and similar power equipment types can provide important reference for maintenance decisions of power equipment. There are two key issues in matching calculation of partial discharge data: first, extracting effective features of partial discharge data; second, obtaining matching degree based on feature vectors.

[0003] In the related art, the patent application document with publication number CN117494037A proposes to use the attention mechanism to fuse dual-channel learning to obtain features of different scales. However, this scheme actually extracts shallow features. Since GIS partial discharge does not occur frequently, the collected fault samples are relatively small. If shallow features are extracted, it is impossible to fully utilize all the information of the small number of samples, resulting in poor matching accuracy.

[0004] In the document "Partial discharge data matching method based on variational Bayesian autoencoder, Song Hui et al., Proceedings of the CSEE", a partial discharge data matching method based on variational Bayesian autoencoder is proposed, which can more effectively detect similar partial discharge data. The training of variational Bayesian autoencoder requires a large amount of data, and it is usually assumed that the data is continuous. However, the amount of GIS partial discharge data is relatively small, and the partial discharge signal is a discrete time series signal. Therefore, the variational Bayesian encoder is not suitable for matching GIS partial discharge data, and the matching results obtained using the variational Bayesian autoencoder are not accurate. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to improve the accuracy of GIS partial discharge feature matching results.

[0006] The present invention solves the above technical problems by the following technical means:

[0007] The present invention proposes a GIS partial discharge feature matching method based on deep reinforcement learning, characterized in that the method comprises:

[0008] Collect the UHF signal of the GIS partial discharge model and obtain the PRPD characteristic spectrum corresponding to the UHF signal;

[0009] Inputting the PRPD feature map into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, and the feature extraction network is formed by superimposing a plurality of feature extraction layers, and each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence;

[0010] The deep features are used to construct a data set, and a deep reinforcement learning algorithm is used to train a GIS partial discharge feature matching Markov model to obtain the optimal GIS partial discharge feature matching strategy.

[0011] Furthermore, the collecting of the ultra-high frequency signal of the GIS partial discharge model and obtaining the PRPD characteristic spectrum corresponding to the ultra-high frequency signal includes:

[0012] The ultra-high frequency signal of the GIS partial discharge model is collected, and the ultra-high frequency signal is processed by using the partial discharge phase analysis method to obtain the PRPD characteristic spectrum.

[0013] Further, the channel attention module includes a first global average pooling layer, a first global maximum pooling layer and a multi-layer perceptron, wherein the PRPD feature map is used as the input of the first global average pooling layer and the first global maximum pooling layer, the output of the first global average pooling layer and the output of the first global maximum pooling layer are both connected to the multi-layer perceptron, the multi-layer perceptron and the PRPD feature map are input to the spatial attention module after a first fusion operation, and the multi-layer perceptron is followed by an activation function;

[0014] The first global average pooling layer and the first global maximum pooling layer are used to process the PRPD feature map, respectively, to obtain a global average pooling result and a global maximum pooling result of the PRPD feature map;

[0015] The multi-layer perceptron is used to calculate the channel attention weight matrix according to the global average pooling result and the global maximum pooling result, and the formula is expressed as:

[0016]

[0017] In the formula, is the channel attention weight matrix, Represents the Sigmoid operation, , represents the fully connected structure in the multi-layer perceptron, represents the global average pooling result, Represents the global maximum pooling result;

[0018] The channel attention weight matrix output by the multi-layer perceptron is fused with the PRPD feature map to obtain a fused feature map.

[0019] Further, the spatial attention module includes a second global average pooling layer, a second global maximum pooling layer, a splicing layer and a convolution layer, the fused feature map is used as the second global average pooling layer and the second global maximum pooling layer respectively, the outputs of the second global average pooling layer and the second global maximum pooling layer are both connected to the splicing layer, the output of the splicing layer is connected to the convolution layer, the output of the convolution layer and the fused feature map are output after the second fusion operation, and the convolution layer is followed by an activation function;

[0020] The second global average pooling layer and the second global maximum pooling layer are used to process the fused feature map respectively to obtain a global average pooling result and a global maximum pooling result corresponding to the fused feature map;

[0021] The global average pooling result and the global maximum pooling result corresponding to the fused feature map are output to the splicing layer for splicing by channel, and the channel attention weight matrix is ​​obtained by operation through the convolution layer, which is expressed as follows:

[0022]

[0023] In the formula, represents the channel attention weight matrix, Represents the Sigmoid operation, Indicates The convolution operation, , Respectively represent the global average pooling result and the global maximum pooling result corresponding to the fused feature map;

[0024] The channel attention weight matrix and the fused feature map are subjected to the second fusion operation to obtain optimized features.

[0025] Furthermore, the output network includes a Flatten layer and a fully connected layer connected in sequence, and the optimized features output by the feature extraction network are converted into one-dimensional features through the Flatten layer, and the one-dimensional features are input into the fully connected layer to obtain deep features corresponding to the ultra-high frequency signal.

[0026] Furthermore, before adopting the deep reinforcement learning algorithm to train the GIS partial discharge feature matching Markov model, the method further includes:

[0027] A GIS partial discharge feature matching Markov model based on a deep reinforcement learning algorithm is pre-built, and the state space of the GIS partial discharge feature matching Markov model for: , It is The deep features of the input Corresponding to the state, the GIS partial discharge feature matches the action space of the Markov model for: ,in, It is Actions corresponding to the number of GIS partial discharge categories;

[0028] The reward function of the GIS partial discharge feature matching Markov model for:

[0029]

[0030] In the formula, is expressed as the sample size imbalance parameter, Indicates that the GIS partial discharge category is The feature subset of ,but ;like ,but ; Indicates time step actions, Indicates The labels of the time steps.

[0031] Furthermore, the deep features are used to construct a data set, and a deep reinforcement learning algorithm is used to train a GIS partial discharge feature matching Markov model to obtain an optimal GIS partial discharge feature matching strategy, including:

[0032] Using the deep features to construct a partial discharge feature training set, and using the training labels corresponding to the deep features to construct a training label set;

[0033] The initial parameters of the GIS partial discharge feature matching Markov model are set, and the initial parameters include the greedy strategy parameters , initialize the experience replay pool K, initialize the network parameters , maximum training rounds , single round training time step , current training round And set the initialization state ;

[0034] Disrupt the GIS partial discharge feature training set , and initialize the current training time step ;

[0035] Select Action , calculate the reward function , and enter the next state ,Will Stored in experience replay pool K;

[0036] Randomly select samples from the experience revisit pool to update the label ;

[0037] Using the gradient descent algorithm to solve the loss function ;

[0038] When the training termination condition is reached, the optimal GIS partial discharge feature matching strategy is output , and the corresponding Q-value network parameters .

[0039] Furthermore, before setting the initial parameters of the GIS partial discharge feature matching Markov model, the method further includes:

[0040] Build Function, optimal GIS partial discharge feature matching strategy With the best Function, the formula is:

[0041]

[0042]

[0043]

[0044] In the formula, represents the attenuation factor, ; Indicated in strategy Expected value under It means taking the maximum value in the action space; It is The deep features of the input The corresponding status, It is The actions corresponding to the number of GIS partial discharge categories are: Indicates time The state of express The action at every moment, Indicates in status Time Action The Q function under Indicates in status Time Action The best function, Indicates in status Execute an action The corresponding optimal strategy.

[0045] Furthermore, the loss function The formula is:

[0046]

[0047] In the formula, represents the expected value, represents the network parameters, is in state Time Action The Q function under represents the attenuation factor, Indicates The labels of the time steps, Indicates in status Time Action The best function.

[0048] In addition, the present invention also proposes a GIS partial discharge feature matching system based on deep reinforcement learning, the system comprising:

[0049] The acquisition module is used to collect the UHF signal of the GIS partial discharge model and obtain the PRPD characteristic spectrum corresponding to the UHF signal;

[0050] A deep feature extraction module, used for inputting the PRPD feature map into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, the feature extraction network is formed by superimposing a plurality of feature extraction layers, and each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence;

[0051] The feature matching module is used to construct a data set using the deep features, and adopt a deep reinforcement learning algorithm to train a GIS partial discharge feature matching Markov model to obtain an optimal GIS partial discharge feature matching strategy.

[0052] The advantages of the present invention are:

[0053] (1) The present invention adopts a dual attention mechanism, that is, the spatial attention and the channel attention are integrated to extract the PRPD feature map corresponding to the UHF signal. This can not only improve the discrimination of GIS partial discharge features and reduce redundancy, but also consider the features from a spatial perspective. By considering the features from different angles, it is very helpful for the extraction of GIS partial discharge features. At the same time, through multi-layer superposition, the sample information can be fully and effectively utilized. In addition, for the extraction of GIS partial discharge features, a GIS partial discharge feature matching Markov model is trained based on a deep reinforcement learning algorithm to obtain the optimal GIS partial discharge feature matching strategy. When partial discharge occurs in GIS, the feature matching of GIS partial discharge signals is realized, effectively maintaining the normal operation of GIS.

[0054] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a GIS partial discharge feature matching method based on deep reinforcement learning proposed in one embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of the structure of a deep feature extraction network in one embodiment of the present invention;

[0057] Figure 3 is a schematic diagram of the structure of a feature extraction layer in one embodiment of the present invention;

[0058] Figure 4 It is a schematic diagram of a training process of a GIS partial discharge feature matching model based on deep reinforcement learning in one embodiment of the present invention;

[0059] Figure 5 It is a structural schematic diagram of a GIS local discharge feature matching system based on deep reinforcement learning proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] like Figure 1 As shown, the first embodiment of the present invention proposes a GIS partial discharge feature matching method based on deep reinforcement learning, and the method includes the following steps:

[0062] S10, collecting the UHF signal of the GIS partial discharge model, and obtaining the PRPD characteristic spectrum corresponding to the UHF signal;

[0063] It should be noted that, in this embodiment, a UHF sensor may be used to collect the ultra-high frequency signal of the GIS partial discharge model.

[0064] S20, inputting the PRPD feature map into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, and the feature extraction network is formed by superimposing a plurality of feature extraction layers, and each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence;

[0065] S30. Utilize the deep features to construct a data set, and use a deep reinforcement learning algorithm to train a GIS partial discharge feature matching Markov model to obtain an optimal GIS partial discharge feature matching strategy.

[0066] This embodiment adopts a dual attention mechanism, that is, the spatial attention and the channel attention are integrated to extract the features of the PRPD feature map corresponding to the UHF signal, which can not only improve the discrimination of GIS partial discharge features and reduce redundancy, but also consider features from a spatial perspective and from different angles, which is very helpful for the extraction of GIS partial discharge features. At the same time, through multi-layer superposition, the sample information can be fully and effectively utilized. In addition, for the extraction of GIS partial discharge features, a GIS partial discharge feature matching Markov model is trained based on a deep reinforcement learning algorithm to obtain the optimal GIS partial discharge feature matching strategy. When partial discharge occurs in GIS, the feature matching of GIS partial discharge signals is realized to effectively maintain the normal operation of GIS.

[0067] As a further preferred technical solution, the step S10: collecting the UHF signal of the GIS partial discharge model and obtaining the PRPD characteristic spectrum corresponding to the UHF signal specifically includes:

[0068] Collect the UHF signal of the GIS partial discharge model, and use the partial discharge phase analysis method to process the UHF signal to obtain the PRPD characteristic spectrum , , , , are the height, width, and number of channels of the input features respectively.

[0069] As a further preferred technical solution, Figures 2 to 3As shown, the channel attention module includes a first global average pooling layer, a first global maximum pooling layer and a multi-layer perceptron, wherein the PRPD feature map is used as the input of the first global average pooling layer and the first global maximum pooling layer, the output of the first global average pooling layer and the output of the first global maximum pooling layer are both connected to the multi-layer perceptron, the multi-layer perceptron and the PRPD feature map are input to the spatial attention module after a first fusion operation, and the multi-layer perceptron is followed by an activation function;

[0070] The first global average pooling layer and the first global maximum pooling layer are used to process the PRPD feature map, respectively, to obtain a global average pooling result and a global maximum pooling result of the PRPD feature map, specifically:

[0071]

[0072]

[0073] in, , Respectively represent the input PRPD feature maps along the spatial dimension of the channel attention module The global average pooling result and the global maximum pooling result of , ; Is the input The height of the channel is , the width is Features, Indicates that the height is 1, the width is 1, and the number of channels is Vector.

[0074] Will , Input into a multi-layer perceptron, which is used to calculate the channel attention weight matrix according to the global average pooling result and the global maximum pooling result. The formula is expressed as:

[0075]

[0076] In the formula, is the channel attention weight matrix, Represents the Sigmoid operation, , represents the fully connected structure in the multi-layer perceptron, represents the global average pooling result, represents the global maximum pooling result, ;

[0077] The channel attention weight matrix output by the multi-layer perceptron is fused with the PRPD feature map to obtain a fused feature map , Represents matrix multiplication.

[0078] As a further preferred technical solution, the spatial attention module includes a second global average pooling layer, a second global maximum pooling layer, a splicing layer and a convolution layer, the fused feature map is respectively used as the second global average pooling layer and the second global maximum pooling layer, the outputs of the second global average pooling layer and the second global maximum pooling layer are both connected to the splicing layer, the output of the splicing layer is connected to the convolution layer, the output of the convolution layer and the fused feature map are output after the second fusion operation, and the convolution layer is followed by an activation function;

[0079] The second global average pooling layer and the second global maximum pooling layer are used to process the fused feature map respectively to obtain the global average pooling result and the global maximum pooling result corresponding to the fused feature map, specifically:

[0080]

[0081]

[0082] in, , Represents the input feature map along the channel dimension of the spatial attention module The global average pooling result and the global maximum pooling result of Is the input The height of the channel is , the width is The fusion characteristics of , , ;

[0083] The global average pooling result and the global maximum pooling result corresponding to the fused feature map are output to the splicing layer for splicing by channel, and the channel attention weight matrix is ​​obtained by operation through the convolution layer, which is expressed as follows:

[0084]

[0085] In the formula, represents the channel attention weight matrix, Represents the Sigmoid operation, Indicates The convolution operation, , They respectively represent the global average pooling result and the global maximum pooling result corresponding to the fused feature map, , ;

[0086] After the second fusion operation, the channel attention weight matrix and the fusion feature map are subjected to the second fusion operation to obtain the optimized feature .

[0087] As a further preferred technical solution, Figure 2 As shown, the output network includes a Flatten layer and a fully connected layer connected in sequence. The optimized features output by the feature extraction network are converted into one-dimensional features through the Flatten layer, and the one-dimensional features are input into the fully connected layer to obtain the deep features corresponding to the UHF signal.

[0088] It should be noted that this embodiment performs deep feature extraction on UHF signals by stacking multiple layers of feature extraction modules optimized by dual attention mechanisms, inputs the deeply extracted features into the Flatten layer, converts the multi-dimensional input into one dimension through the Flatten layer, inputs the input into the fully connected layer, and outputs the deeply extracted GIS partial discharge features through the fully connected layer.

[0089] As a further preferred technical solution, before the deep reinforcement learning algorithm is used to train the GIS partial discharge feature matching Markov model in step S30, the method further includes the following steps:

[0090] Assume that the PRPD feature map training set is , , Partial discharge category number, , , set the maximum number of training rounds , the single round training time step is , , ;

[0091] Using deep extraction of GIS partial discharge feature training set , It is The deep extracted partial discharge features corresponding to the input PRPD feature maps;

[0092] A GIS partial discharge feature matching Markov model based on a deep reinforcement learning algorithm is pre-built, and the state space of the GIS partial discharge feature matching Markov model for: , It is The deep features of the input Corresponding to the state, the GIS partial discharge feature matches the action space of the Markov model for: ,in, It is Actions corresponding to the number of GIS partial discharge categories;

[0093] The reward function of the GIS partial discharge feature matching Markov model for:

[0094]

[0095] In the formula, is expressed as the sample size imbalance parameter, Indicates that the GIS partial discharge category is The feature subset of ,but ;like ,but ; Indicates time step actions, Indicates The labels of the time steps.

[0096] It should be noted that since the GIS local discharge feature contains little feature data and there is an imbalance in the sample data, it may be difficult to identify. The reward function set in this embodiment is used to determine whether the feature matching is accurate. However, since there is an imbalance in the samples of different categories of local discharge features, this embodiment introduces a sample imbalance parameter for adjustment. The feature categories with a smaller sample size will receive greater rewards and penalties based on whether the match is accurate than the feature categories with a larger sample size, which improves the accuracy of feature matching to a certain extent.

[0097] As a further preferred technical solution, Figure 4 As shown, the deep features are used to construct a data set, and a deep reinforcement learning algorithm is used to train a GIS partial discharge feature matching Markov model to obtain an optimal GIS partial discharge feature matching strategy, including:

[0098] (1) constructing a partial discharge feature training set using the deep features, and constructing a training label set using the training labels corresponding to the deep features;

[0099] (2) Setting the initial parameters of the GIS partial discharge feature matching Markov model, the initial parameters include the greedy strategy parameters , initialize the experience replay pool K, initialize the network parameters , maximum training rounds , single round training time step , current training round And set the initialization state ;

[0100] (3) Disrupting the GIS partial discharge feature training set , and initialize the current training time step ;

[0101] (4) Select an action , calculate the reward function , and enter the next state ,Will Stored in the experience replay pool K, where: , Indicated in strategy Next, status Execute an action , Represents the greedy strategy parameters, used to select actions;

[0102] (5) Randomly select samples from the experience revisit pool to update the label ;

[0103] (6) Use the gradient descent algorithm to solve the loss function ;

[0104] (7) If not satisfied , then let , and return to (4), otherwise go to (8);

[0105] (8) If not satisfied , then let , and return to (2), otherwise go to (9);

[0106] (9) Output optimal GIS partial discharge feature matching strategy , and the corresponding Q-value network parameters .

[0107] As a further preferred technical solution, before setting the initial parameters of the GIS partial discharge feature matching Markov model, the method further includes:

[0108] Build Function, optimal GIS partial discharge feature matching strategy With the best Function, the formula is:

[0109]

[0110]

[0111]

[0112] In the formula, represents the attenuation factor, ; Indicated in strategy Expected value under It means taking the maximum value in the action space; It is The deep features of the input The corresponding status, It is The actions corresponding to the number of GIS partial discharge categories are: Indicates time The state of express The action at every moment, Indicates in status Time Action The Q function under Indicates in status Time Action The best function, Indicates in status Execute an action The corresponding optimal strategy.

[0113] As a further preferred technical solution, the loss function The formula is:

[0114]

[0115] In the formula, represents the expected value, represents the network parameters, is in state Time Action The Q function under represents the attenuation factor, Indicates The labels of the time steps, Indicates in status Time Action The best function.

[0116] In addition, if Figure 5 As shown, the present invention also proposes a GIS partial discharge feature matching system based on deep reinforcement learning, and the system includes:

[0117] The acquisition module 10 is used to acquire the UHF signal of the GIS partial discharge model and obtain the PRPD characteristic spectrum corresponding to the UHF signal;

[0118] A deep feature extraction module 20 is used to input the PRPD feature map into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, and the feature extraction network is formed by superimposing a plurality of feature extraction layers, and each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence;

[0119] The feature matching module 30 is used to construct a data set using the deep features, and train the GIS partial discharge feature matching Markov model using a deep reinforcement learning algorithm to obtain an optimal GIS partial discharge feature matching strategy.

[0120] As a further preferred technical solution, the system further includes a training module, which is specifically used for:

[0121] Using the deep features to construct a partial discharge feature training set, and using the training labels corresponding to the deep features to construct a training label set;

[0122] The initial parameters of the GIS partial discharge feature matching Markov model are set, and the initial parameters include the greedy strategy parameters , initialize the experience replay pool K, initialize the network parameters , maximum training rounds , single round training time step , current training round And set the initialization state ;

[0123] Disrupt the GIS partial discharge feature training set , and initialize the current training time step ;

[0124] Select Action , calculate the reward function , and enter the next state ,Will Stored in experience replay pool K;

[0125] Randomly select samples from the experience revisit pool to update the label ;

[0126] Using the gradient descent algorithm to solve the loss function ;

[0127] When the training termination condition is reached, the optimal GIS partial discharge feature matching strategy is output , and the corresponding Q-value network parameters .

[0128] It should be noted that other embodiments or implementation methods of the GIS local discharge feature matching system based on deep reinforcement learning described in the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0129] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0130] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0131] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A GIS partial discharge feature matching method based on deep reinforcement learning, characterized in that: The method comprises: Collect the UHF signal of the GIS partial discharge model and obtain the PRPD characteristic spectrum corresponding to the UHF signal; The PRPD feature map is input into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, and the feature extraction network is formed by superimposing a plurality of feature extraction layers, and each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence, wherein the channel attention module includes a first global average pooling layer, a first global maximum pooling layer and a multi-layer perceptron, wherein the PRPD feature map is used as the input of the first global average pooling layer and the first global maximum pooling layer, and the output of the first global average pooling layer and the output of the first global maximum pooling layer are both connected to the A multi-layer perceptron, wherein the multi-layer perceptron and the PRPD feature map are input into the spatial attention module after a first fusion operation, and the multi-layer perceptron is followed by an activation function; the spatial attention module includes a second global average pooling layer, a second global maximum pooling layer, a splicing layer and a convolution layer, and the fused feature map output by the channel attention module is respectively used as the second global average pooling layer and the second global maximum pooling layer, and the outputs of the second global average pooling layer and the second global maximum pooling layer are both connected to the splicing layer, and the output of the splicing layer is connected to the convolution layer, and the output of the convolution layer and the fused feature map are output after a second fusion operation, and the convolution layer is followed by an activation function; The deep features are used to construct a data set, and a deep reinforcement learning algorithm is used to train a GIS partial discharge feature matching Markov model to obtain the optimal GIS partial discharge feature matching strategy.

2. The GIS partial discharge feature matching method based on deep reinforcement learning according to claim 1, characterized in that: The collecting of the ultra-high frequency signal of the GIS partial discharge model and obtaining the PRPD characteristic spectrum corresponding to the ultra-high frequency signal includes: The ultra-high frequency signal of the GIS partial discharge model is collected, and the ultra-high frequency signal is processed by using the partial discharge phase analysis method to obtain the PRPD characteristic spectrum.

3. The GIS partial discharge feature matching method based on deep reinforcement learning as claimed in claim 1, characterized in that: The first global average pooling layer and the first global maximum pooling layer are used to process the PRPD feature map, respectively, to obtain a global average pooling result and a global maximum pooling result of the PRPD feature map; The multi-layer perceptron is used to calculate the channel attention weight matrix according to the global average pooling result and the global maximum pooling result, and the formula is expressed as: In the formula, is the channel attention weight matrix, Represents the Sigmoid operation, , represents the fully connected structure in the multi-layer perceptron, represents the global average pooling result, Represents the global maximum pooling result; The channel attention weight matrix output by the multi-layer perceptron is fused with the PRPD feature map to obtain a fused feature map.

4. The GIS partial discharge feature matching method based on deep reinforcement learning according to claim 1, characterized in that: The second global average pooling layer and the second global maximum pooling layer are used to process the fused feature map respectively to obtain a global average pooling result and a global maximum pooling result corresponding to the fused feature map; The global average pooling result and the global maximum pooling result corresponding to the fused feature map are output to the splicing layer for splicing by channel, and the channel attention weight matrix is ​​obtained by operation through the convolution layer, which is expressed as follows: In the formula, represents the channel attention weight matrix, Represents the Sigmoid operation, Indicates that The convolution operation, , Respectively represent the global average pooling result and the global maximum pooling result corresponding to the fused feature map; The channel attention weight matrix and the fused feature map are subjected to the second fusion operation to obtain optimized features.

5. The GIS partial discharge feature matching method based on deep reinforcement learning as claimed in claim 1, characterized in that: The output network includes a Flatten layer and a fully connected layer connected in sequence. The optimized features output by the feature extraction network are converted into one-dimensional features through the Flatten layer. The one-dimensional features are input into the fully connected layer to obtain the deep features corresponding to the ultra-high frequency signal.

6. The GIS partial discharge feature matching method based on deep reinforcement learning according to claim 1, characterized in that: Before using the deep reinforcement learning algorithm to train the GIS partial discharge feature matching Markov model, the method further includes: A GIS partial discharge feature matching Markov model based on a deep reinforcement learning algorithm is pre-built, and the state space of the GIS partial discharge feature matching Markov model for: , It is The deep features of the input Corresponding to the state, the GIS partial discharge feature matches the action space of the Markov model for: ,in, It is Actions corresponding to the number of GIS partial discharge categories; The reward function of the GIS partial discharge feature matching Markov model for: In the formula, is expressed as the sample size imbalance parameter, Indicates that the GIS partial discharge category is The feature subset of ,but ;like ,but ; Indicates time step actions, Indicates The labels of the time steps.

7. The GIS partial discharge feature matching method based on deep reinforcement learning according to claim 1, characterized in that: The method of constructing a data set using the deep features and training a GIS partial discharge feature matching Markov model using a deep reinforcement learning algorithm to obtain an optimal GIS partial discharge feature matching strategy includes: Using the deep features to construct a partial discharge feature training set, and using the training labels corresponding to the deep features to construct a training label set; The initial parameters of the GIS partial discharge feature matching Markov model are set, and the initial parameters include the greedy strategy parameters , initialize the experience replay pool K, initialize the network parameters , maximum training rounds , single round training time step , current training round And set the initialization state ; Disrupt the GIS partial discharge feature training set , and initialize the current training time step ; Select Action , calculate the reward function , and enter the next state ,Will Stored in experience replay pool K; Randomly select samples from the experience revisit pool to update the label ; Using the gradient descent algorithm to solve the loss function ; When the training termination condition is reached, the optimal GIS partial discharge feature matching strategy is output , and the corresponding Q-value network parameters .

8. The GIS partial discharge feature matching method based on deep reinforcement learning as claimed in claim 7, characterized in that: Before setting the initial parameters of the GIS partial discharge feature matching Markov model, the method further includes: Build Function, optimal GIS partial discharge feature matching strategy With the best Function, the formula is: In the formula, represents the attenuation factor, ; Indicated in strategy Expected value under It means taking the maximum value in the action space; It is The deep features of the input The corresponding status, It is The actions corresponding to the number of GIS partial discharge categories are: Indicates time The state of express The action at every moment, Indicates in status Time Action The Q function under Indicates in status Time Action The best function, Indicates in status Execute an action The corresponding optimal strategy.

9. The GIS partial discharge feature matching method based on deep reinforcement learning according to claim 7, characterized in that: The loss function The formula is: In the formula, represents the expected value, represents the network parameters, is in state Time Action The Q function under represents the attenuation factor, Indicates The labels of the time steps, Indicates in status Time Action The best function.

10. A GIS partial discharge feature matching system based on deep reinforcement learning, characterized in that: The system comprises: The acquisition module is used to collect the UHF signal of the GIS partial discharge model and obtain the PRPD characteristic spectrum corresponding to the UHF signal; A deep feature extraction module is used to input the PRPD feature map into a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, the feature extraction network is formed by superimposing a plurality of feature extraction layers, each of the feature extraction layers includes a channel attention module and a spatial attention module connected in sequence, wherein the channel attention module includes a first global average pooling layer, a first global maximum pooling layer and a multi-layer perceptron, wherein the PRPD feature map is used as the input of the first global average pooling layer and the first global maximum pooling layer, and the output of the first global average pooling layer and the output of the first global maximum pooling layer are used as the input of the first global average pooling layer. The outputs of the channel attention module are connected to the multi-layer perceptron, the multi-layer perceptron and the PRPD feature map are input to the spatial attention module after a first fusion operation, and the multi-layer perceptron is followed by an activation function; the spatial attention module includes a second global average pooling layer, a second global maximum pooling layer, a splicing layer and a convolution layer, the fused feature map output by the channel attention module is respectively used as the second global average pooling layer and the second global maximum pooling layer, the outputs of the second global average pooling layer and the second global maximum pooling layer are both connected to the splicing layer, the output of the splicing layer is connected to the convolution layer, the output of the convolution layer and the fused feature map are output after a second fusion operation, and the convolution layer is followed by an activation function; The feature matching module is used to construct a data set using the deep features, and train a GIS partial discharge feature matching Markov model using a deep reinforcement learning algorithm to obtain an optimal GIS partial discharge feature matching strategy.

Images