A GIS partial discharge defect diagnosis method based on neural supervised decision tree

Abstract

The present invention discloses a method for diagnosing local discharge defects in GIS based on a neural supervised decision tree, which relates to the technical field of insulation status diagnosis of power equipment. The method comprises the following steps: performing agglomerative hierarchical clustering based on the AGNES algorithm to construct a decision tree structure, replacing the final linear fully connected layer of a CNN neural network as an induction layer, and using the output vector of the CNN neural network as a supervision layer to perform supervised correction on the output of the induction layer to construct an improved CNN neural network; training the improved CNN neural network by collecting a local discharge phase distribution map of the GIS as a data set, and constructing a neural network model for identifying whether a target GIS has local discharge defects. The present invention solves the problems of low accuracy of interpretable defect recognition algorithms, low trust in on-site maintenance personnel, and deep identification black box problems, thereby realizing interpretable fault diagnosis of GIS local discharge with high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of insulation status diagnosis of power equipment, and in particular to a GIS partial discharge defect diagnosis method based on a neural supervisory decision tree. Background Art

[0002] Gas-insulated metal-enclosed switchgear (GIS) has been widely used in power systems due to its advantages such as small footprint and high reliability. Its insulation performance is crucial to the safe operation of the entire power grid. Partial discharge is a key indicator of insulation degradation, and identifying insulation defect types based on PD signatures has become a crucial component of GIS fault diagnosis. Deep neural networks, with their strong generalization capabilities, high recognition accuracy, and robust feature extraction capabilities, have become a major research area in this field.

[0003] However, traditional deep neural network fault diagnosis methods rely solely on data-driven diagnostic identification, ignoring the physical connections between fault defects and failing to fully leverage these connections in the decision-making process from feature vectors to class representations to improve recognition performance. Furthermore, traditional deep learning algorithms lack interpretability, and their decision-making process is a "black box," providing only the final recognition result without providing the recognition logic. This black-box nature and inexplicability not only prevents researchers and on-site maintenance personnel from understanding and analyzing the recognition logic, leading to a lack of trust in the diagnostic results, but also prevents the algorithm from fully leveraging the physical characteristics of the fault defects to guide the construction of the fault diagnosis model, making it unable to adapt to real-world needs. Ultimately, the recognition results are suboptimal. Therefore, there is an urgent need to design a diagnostic method for GIS partial discharge defects to meet the technical requirements for GIS partial discharge fault diagnosis. Summary of the Invention

[0004] In order to overcome the problems existing in the prior art, the present application provides a GIS partial discharge defect diagnosis method based on a neural supervised decision tree, comprising the following steps:

[0005] Based on the AGNES algorithm, agglomerative hierarchical clustering is performed to construct a decision tree structure, which is used as the induction layer to replace the final linear fully connected layer of the CNN neural network. The output vector of the CNN neural network is used as the supervision layer to supervise and correct the output of the induction layer to construct an improved CNN neural network.

[0006] By collecting the partial discharge phase distribution map of GIS as a data set, the improved CNN neural network is trained and a neural network model is constructed to identify whether the target GIS has partial discharge defects.

[0007] Preferably, in the process of generating the induction layer, after the partial discharge phase distribution map is input into the CNN neural network, the weight matrix of the final linear fully connected layer is divided into leaf nodes;

[0008] According to the representative vector of the leaf node, the leaf node is constructed into a decision tree structure based on the AGNES algorithm to generate the induction layer.

[0009] Preferably, in the process of constructing the decision tree structure, based on the AGNES algorithm, through the forward recognition and backward extraction methods, the internal nodes of the decision tree structure are defined according to the representative vectors to generate the decision tree structure.

[0010] Preferably, in the process of defining the internal nodes through forward identification, the process of forward identification includes:

[0011] Based on the bag-of-words model, the characteristic meaning of each dimension of the multidimensional semantic vector is obtained according to the physical characteristics, discharge characteristics and inherent differences of GIS partial discharge defect types, and a defect semantic vector set is constructed.

[0012] Based on the defect semantic vector set, the difference subset of the defect semantic vector set is determined according to the opposition relationship between the features of each dimension;

[0013] The internal nodes of the decision tree structure are defined according to the defect semantic vector set with difference subsets.

[0014] Preferably, in the process of defining the internal nodes by reverse extraction, the reverse extraction process includes: using a convergence value to characterize the similarity between the two difference subset vectors;

[0015] The semantic features corresponding to the difference subset with the smallest convergence value are selected as the maximized difference features, and the same-layer classification is performed to determine the meaning of the node.

[0016] Preferably, in the process of generating the induction layer, under the premise of considering the decision tree path loss, the CNN neural network is adjusted by the loss function to optimize and adjust the internal nodes of the decision tree and determine the final representative vector, wherein the loss function is expressed as:

[0017]

[0018] Among them, l is the total target loss, loss(y i ) is the leaf node loss, loss(y k ) is the internal node loss, i.e., the path loss; N represents the number of defects, M represents the number of nodes except the root node; β t With ω t is the variable weight adjustment coefficient, during the retraining process β tAs the number of iterations increases, ω decreases. t Increases with the number of iterations.

[0019] The final representative vector is expressed as:

[0020]

[0021] Where i is a leaf node, k is an internal node, j is a child node, wi represents the final linear layer weight matrix, n i Indicates that the leaf node represents a vector, n k Indicates that the internal node represents the vector, n j represents the child node representative vector, and K(i) represents the child node set of node i.

[0022] Preferably, before the output of the induction layer is supervised and corrected, the induction layer is probabilized by an improved softmax function, the leaf node probability is obtained by cumulative multiplication, and the maximum value is selected as the class label identification value as the output of the induction layer, wherein the improved softmax function is expressed as:

[0023]

[0024] Where λ is the relative improvement coefficient and e is the natural base.

[0025] The output of the induction layer is expressed as:

[0026] p j =p(j|k)=softmax(y j ), j∈K

[0027] y i = <n i ,x>,i∈[1,M]

[0028] Where x represents the feature vector, and node j is the child node of node k.

[0029] Preferably, in the process of supervised correction of the output of the induction layer, the output of the induction layer is supervised based on the output vector of the CNN neural network, and the output of the induction layer with an accuracy lower than the output vector of the CNN neural network is obtained as the correction target, wherein the supervision process is expressed as:

[0030] v×G(d(i),l CNN (i))>G(d(i),l DT (i))

[0031] Where v is the threshold coefficient, G is the output matrix of the induction layer node, d is the sample index vector whose output result of the neural network is different from that of the induction layer, l is the threshold coefficient, G is the output matrix of the induction layer node, d is the sample index vector whose output result of the neural network is different from that of the induction layer, CNN With l DTIndicates the label recognition result of the sample corresponding to the index vector

[0032] Based on the correction target, correction is performed by adopting the form of threshold review, where the correction process is expressed as:

[0033] G(d(i),l DT (i))=v×G(d(t),l CNN (t)).

[0034] Preferably, a GIS partial discharge defect diagnosis system for implementing the GIS partial discharge defect diagnosis method includes:

[0035] Data acquisition module, used to collect image data of target GIS;

[0036] The identification module is used to identify whether the target GIS has partial discharge defects based on the neural network model. Specifically, agglomerative hierarchical clustering is performed based on the AGNES algorithm to construct a decision tree structure, which is used as the induction layer to replace the final linear fully connected layer of the CNN neural network. The output vector of the CNN neural network is used as the supervision layer to supervise and correct the output of the induction layer to construct an improved CNN neural network. In addition, the partial discharge phase distribution map of the GIS is collected as a data set to train the improved CNN neural network and construct a neural network model.

[0037] Preferably, the recognition module is also used to input the partial discharge phase distribution map into the CNN neural network, and then divide the weight matrix of the final linear fully connected layer into leaf nodes; according to the representative vectors of the leaf nodes and the AGNES algorithm, the leaf nodes are constructed into a decision tree structure to generate an induction layer.

[0038] The present invention discloses the following technical effects:

[0039] The present invention fully combines the physical characteristics of GIS partial discharge defects to guide deep neural network identification decisions. Through the induction layer, supervision layer and adaptive adjustment, it combines the advantages of deep neural network's strong feature self-extraction ability, strong feature fusion ability, high recognition reliability with strong hierarchical structure interpretability and high decision transparency. It solves the problems of low accuracy of explainable defect recognition algorithm, low trust of on-site maintenance personnel, and deep recognition black box, and realizes explainable fault diagnosis of GIS partial discharge with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in 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.

[0041] Figure 1 This is a schematic diagram of the 110kV multi-sensing GIS partial discharge monitoring platform of the present invention;

[0042] Figure 2 This is a schematic diagram of the neural supervised decision tree process of the present invention;

[0043] Figure 3 This is a schematic diagram of hierarchical clustering according to the present invention;

[0044] Figure 4 This is a schematic diagram of the induction layer construction process of the present invention;

[0045] Figure 5 This is a schematic diagram of the hierarchical structure of the present invention;

[0046] Figure 6 This is a schematic diagram of the supervisory layer data processing flow of the present invention;

[0047] Figure 7 This is a schematic diagram of the high-voltage end surface defect sensitivity enhancement according to the present invention;

[0048] Figure 8 The present invention is a flow chart of the method described in the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0050] like Figure 1-8As shown, in order to fully demonstrate and utilize the connection between the physical characteristics of GIS defects to guide fault diagnosis and improve the reliability and interpretability of the GIS partial discharge fault diagnosis method, the present invention proposes a neural supervision decision tree algorithm (NSDT) to achieve explainable fault diagnosis of GIS partial discharge with high accuracy. The final linear fully connected layer of the convolutional neural network is replaced with a hierarchical structure to construct an induction layer. The tree supervision loss function is used to optimize the node representative vector, and the hierarchical loss is suppressed by adjusting the softmax function. The supervision layer is constructed, and the output of the induction layer is supervised and corrected by the original neural network output vector. On the basis of retaining the interpretability, the recognition accuracy is improved. By innovatively constructing the supervision layer, adjusting the softmax function, and proposing the "forward recognition, reverse extraction" node meaning determination method, it is possible to accurately diagnose the type of GIS insulation defects under multi-source defects, while providing decision-making basis and identification path, and can improve the recognition performance of specific defects according to actual maintenance needs.

[0051] In order to achieve the above technical objectives, the present invention provides a GIS partial discharge defect diagnosis method based on a neural supervised decision tree algorithm, comprising the following steps:

[0052] S1: Pre-training the neural network. The Phase Resolved Partial Discharge (PRPD) map is input, and the recognition model is trained using the original CNN neural network. The final linear fully connected layer weight matrix is ​​divided into leaf nodes, and the output matrix is ​​obtained.

[0053] S2: The induction layer builds a decision tree structure. The leaf node representative vectors obtained in the first step are clustered using the AGNES algorithm to form a decision tree structure. The "forward recognition, backward extraction" method is used to determine the meaning of the internal nodes.

[0054] S3: The tree supervision function adjusts the representative vector. Considering the decision tree path loss, the pre-trained neural network is fine-tuned to optimize and adjust the internal nodes to determine the final representative vector.

[0055] S4: Induction layer type identification. The data to be identified passes through the pre-neural network and is probabilized in the induction layer using a modified softmax function. The leaf node probabilities are multiplied and the output matrix is ​​obtained. The maximum value is selected as the class label identification value, and the decision process is output for verification.

[0056] S5: Supervised correction of recognition results. The recognition values ​​are corrected through supervised values ​​to improve the reliability and accuracy of the results while maintaining strong interpretability.

[0057] S6: Output the decision path, hierarchy and recognition results.

[0058] S7: Adjust the hierarchical structure according to actual needs and return to S3 for retraining.

[0059] In the above step S2, the specific steps are as follows:

[0060] S21: If the final linear fully connected layer weight vector matrix is ​​W, then the leaf node represents the vector n i =w i , through the leaf node represents the vector n i , using AGENS to condense the hierarchical structure.

[0061] S22: Use the “forward recognition, backward extraction” method to determine the meaning of internal nodes.

[0062] Furthermore, in step S22, the bag-of-words model is first used to determine the feature meanings of each dimension of the q-dimensional semantic vector according to the inherent differences in the physical characteristics, discharge characteristics, and partial discharge principles of the GIS partial discharge defect types, thereby constructing the semantic vector z of each defect. i Then, according to the actual opposition relationship of each dimension feature, such as high-pressure end and low-pressure end, single source and dual source, etc., the difference subset r is determined. a i , where i is the node number and a is the difference subset number. In the forward recognition process, for an internal node k, if its child node set is K, then:

[0063] z k =z K(1) ∧z K(2) ∧...Z K(n)

[0064] In the reverse extraction process, the convergence value is used Characterizes the similarity between subset vectors. If the set of nodes in the pth layer belonging to the same parent node is D, then the convergence value of the ath difference subset in this layer is:

[0065]

[0066] The semantic features corresponding to the difference subset with the smallest convergence value are selected as the maximized difference features for same-layer classification to determine the meaning of the node.

[0067] In step S3, the specific steps of step S3 are as follows:

[0068] S31: Fine-tune or retrain using the tree-supervised loss function. In soft node traversal mode, the loss function is:

[0069]

[0070] S32: Use the final linear fully connected layer of the new network to determine the representative vector of the induction layer node. The representative vector of the leaf node is still determined using the S21 method. If the final linear fully connected layer weight matrix is M is the number of nodes excluding the root node, N is the number of defects, and for node i∈[1,M], its child node j∈K(i), then for any node i, its representative vector n i for:

[0071]

[0072] In step S4, the specific steps of step S4 are as follows:

[0073] S41: In the induction layer, the probabilistic softmax function is improved, and the leaf node probabilities are obtained by cumulative multiplication to obtain the output matrix. The feature vector x extracted by the fine-tuned model before the induction layer is input from the root node into the induction layer, and the inner product with the representative vector of each node is performed layer by layer, that is:

[0074] y i = <n i ,x>,i∈[1,M]

[0075] If for node k, its child node j∈K, then the probability value of node j under parent node k and the improved normalized exponential function are:

[0076] p j =p(j|k)=softmax(y j ), j∈K

[0077]

[0078] λ is the relative boost coefficient; higher λ values ​​increase the effect of amplifying relative disparities. The improved normalized exponential function leverages the exponential function's gradient boosting properties to expand the relative size of probabilities at the same level, while preserving the absolute size of nodes at the same level. This suppresses the probabilities of leaf nodes at lower levels that are at a competitive disadvantage, and reduces the cumulative layer loss of higher-level leaf nodes.

[0079] S42: Select the maximum value as the class label recognition value and output the decision process for testing.

[0080] t=argmax p(i), i∈[1,N]

[0081] In step S5, the specific steps of step S5 are as follows:

[0082] S51: retain the pre-trained output matrix F U×N , after the induction layer, we get the output matrix G U×N

[0083] S52: Construct a supervision layer. When the recognition results of F and G are different, the higher-accuracy F is used. U×N The output matrix G of the induction layer node is censored in the form of threshold U×N Perform supervision correction, and its discriminant and correction form are:

[0084]

[0085] Example 1: Construction Figure 1 The 110kV multi-sensor GIS partial discharge monitoring platform shown in the figure collects ultrasonic, UHF and single photon three-dimensional partial discharge signals of six single-source and dual-source defects: the tip of the guide rod, the tip of the shell, the surface along the high-voltage end, the tip of the guide rod-tip of the shell, the tip of the guide rod-high-voltage end, and the tip of the shell-high-voltage end. Figure 2 The neural supervised decision tree process shown above builds a defect type recognition model. The specific implementation process is as follows:

[0086] S1: The original partial discharge PRPD spectrum is input into the pre-convolutional neural network to obtain the output matrix F U×N .

[0087] S2: Final linear fully connected layer Figure 3 The induction layer construction process shown is similar to Figure 4 The hierarchical clustering method shown in the figure initially forms a hierarchical structure, and the "forward recognition, reverse extraction" method is used to determine the meaning of the nodes, and finally a hierarchical structure is formed as shown in the figure. Figure 5 shown.

[0088] S3: Combined with the hierarchical structure, the node representative vector is optimized using the tree supervision loss function. In the soft node traversal mode, the loss function is:

[0089]

[0090] Retraining is performed based on the tree supervision loss function, and fine-tuning is used when the retraining recognition level cannot reach the initial level. t As the number of iterations increases, ω decreases. t The tree-supervised loss function increases with the number of iterations. This is because in the early stages of the iteration, leaf nodes are not yet determined, and internal nodes are of little reference value, so adding path loss will affect the learning rate. In the later stages of the iteration, leaf nodes are fully differentiated, and internal nodes need to be optimized to determine the path structure. Using the tree-supervised loss function, fine-tuning or retraining can be performed on the pre-trained CNN model to optimize the representative vectors of each node in the induction layer.

[0091] S4: If the final linear fully connected layer weight matrix is If M is the number of nodes excluding the root node, N is the number of defects, and for node i∈[1,M], its child node j∈K(i), then for any node i, its representative vector n i for:

[0092]

[0093] The leaf node probability is obtained by cumulative multiplication, and the output matrix is ​​obtained. The feature vector x extracted by the fine-tuned model before the induction layer is input into the induction layer from the root node, and the inner product with the representative vector of each node is performed layer by layer, that is:

[0094] y i = <n i ,x>,i∈[1,M]

[0095] If for node k, its child node j∈K, then the probability value of node j under parent node k and the improved normalized exponential function are:

[0096] p j =p(j|k)=softmax(y j ), j∈K

[0097]

[0098] Finally, the output matrix G is obtained U×N ;

[0099] S5: Reference Figure 6 The supervisory layer data processing flow shown is supervised and corrected

[0100] S6: Output the recognition results, decision path and hierarchy.

[0101] S7: In the actual field, high-voltage end surface defects pose a greater threat to the insulation performance of GIS among the test defects. When such defects occur, it is often necessary to open the cover for maintenance. However, the tiny spikes of the guide rod will gradually wear away during several discharge burning processes, and the degree of damage to the insulation is relatively smaller. Based on the above considerations, it is even more necessary to improve the recognition performance of defects along the high-voltage end surface. While improving the recall rate of this type of recognition to improve the reliability of the equipment, it is also necessary to reduce the false alarm rate of this type of recognition to reduce the maintenance workload. Therefore, on the basis of the unsupervised hierarchical structure, by adding pre-division nodes for high-voltage end surface defects, the sensitivity of the hierarchical structure to defects along the high-voltage end surface can be enhanced. The final hierarchical structure correction is as follows: Figure 7 As shown. Return to S3 for retraining.

[0102] After the identification model is constructed, during actual on-site maintenance, when a partial discharge signal occurs, the partial discharge PRPD spectrum is input into the identification model to obtain the defect type and identification path.

[0103] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0104] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0105] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A GIS partial discharge defect diagnosis method based on neural supervised decision tree, characterized in that: The following steps are involved: Based on the AGNES algorithm, agglomerative hierarchical clustering is performed to construct a decision tree structure, which is used as an induction layer to replace the final linear fully connected layer of the CNN neural network. The output vector of the CNN neural network is used as a supervision layer to supervise and correct the output of the induction layer to construct an improved CNN neural network. By collecting the partial discharge phase distribution map of the GIS as a data set, the improved CNN neural network is trained to construct a neural network model for identifying whether the target GIS has partial discharge defects; In the process of generating the induction layer, after the partial discharge phase distribution map is input into the CNN neural network, the weight matrix of the final linear fully connected layer is divided into leaf nodes; According to the representative vectors of the leaf nodes, the leaf nodes are constructed into the decision tree structure according to the AGNES algorithm to generate the induction layer; In the process of generating the induction layer, the CNN neural network is adjusted by the loss function under the premise of considering the decision tree path loss, so as to optimize and adjust the internal nodes of the decision tree and determine the final representative vector, wherein the loss function is expressed as: Among them, l is the total target loss, loss(y i ) is the leaf node loss, loss(y k ) is the internal node loss, i.e., the path loss; N represents the number of defects, M represents the number of nodes except the root node; β t With ω t is the variable weight adjustment coefficient, during the retraining process β t As the number of iterations increases, ω decreases. t Increases with the number of iterations; The final representative vector is expressed as: Where i is a leaf node, k is an internal node, j is a child node, wi represents the final linear layer weight matrix, n i Indicates that the leaf node represents a vector, n k Indicates that the internal node represents the vector, n j represents the child node representative vector, K(i) represents the child node set of node i; Before the output of the induction layer is supervised and corrected, the induction layer is probabilized by an improved softmax function, the leaf node probability is obtained by cumulative multiplication, and the maximum value is selected as the class label identification value as the output of the induction layer, wherein the improved softmax function is expressed as: In the formula, λ is the relative improvement coefficient, and e is the natural base; The output of the induction layer is expressed as: p j <p(j|k)=softmax(y j ),j∈K y i =<n i ,x>,i∈[1,M] In the formula, x represents the feature vector, and node j is the child node of node k; In the process of supervised correction of the output of the induction layer, the output of the induction layer is supervised based on the output vector of the CNN neural network, and the output of the induction layer with an accuracy lower than the output vector of the CNN neural network is obtained as a correction target, wherein the supervision process is expressed as: v×G(d(i),l CNN (i))>G(d(i),l DT (i)) Where v is the threshold coefficient, G is the output matrix of the induction layer node, d is the sample index vector whose output result of the neural network is different from that of the induction layer, l is the threshold coefficient, G is the output matrix of the induction layer node, d is the sample index vector whose output result of the neural network is different from that of the induction layer, CNN With l DT Indicates the label recognition result of the sample corresponding to the index vector Based on the correction target, correction is performed by adopting a threshold review method, wherein the correction process is expressed as follows: G(d(i),l DT (i))=v×G(d(i),l CNN (i))。 2. The GIS partial discharge defect diagnosis method based on a neural supervised decision tree according to claim 1 is characterized by: In the process of constructing the decision tree structure, based on the AGNES algorithm, through the forward recognition and backward extraction methods, the internal nodes of the decision tree structure are defined according to the representative vector to generate the decision tree structure.

3. The GIS partial discharge defect diagnosis method based on a neural supervised decision tree according to claim 2 is characterized by: In the process of defining internal nodes through forward identification, the forward identification process includes: Based on the bag-of-words model, the characteristic meaning of each dimension of the multidimensional semantic vector is obtained according to the physical characteristics, discharge characteristics and inherent differences of GIS partial discharge defect types, and a defect semantic vector set is constructed. Based on the defect semantic vector set, determining a difference subset of the defect semantic vector set according to the opposition relationship between the features of each dimension; The internal nodes of the decision tree structure are defined according to the defect semantic vector set having the difference subset.

4. The GIS partial discharge defect diagnosis method based on a neural supervised decision tree according to claim 3 is characterized by: In the process of defining the internal nodes through reverse extraction, the reverse extraction process includes: using a convergence value to represent the similarity between two difference subset vectors; The semantic feature corresponding to the difference subset with the smallest convergence value is selected as the maximized difference feature, and the same-layer classification is performed to determine the meaning of the node.

5. The GIS partial discharge defect diagnosis method based on a neural supervised decision tree according to any one of claims 1 to 4, characterized in that: A GIS partial discharge defect diagnosis system for implementing a GIS partial discharge defect diagnosis method includes: Data acquisition module, used to collect image data of target GIS; An identification module is configured to identify, based on the neural network model, whether the target GIS has a partial discharge defect, wherein agglomerative hierarchical clustering is performed based on the AGNES algorithm to construct a decision tree structure, which is used as an induction layer to replace the final linear fully connected layer of the CNN neural network, and the output vector of the CNN neural network is used as a supervision layer to perform supervised correction on the output of the induction layer to construct an improved CNN neural network; and the improved CNN neural network is trained by collecting a partial discharge phase distribution map of the GIS as a data set to construct the neural network model.

6. The GIS partial discharge defect diagnosis method based on neural supervised decision tree according to claim 5 is characterized by: The recognition module is also used to input the partial discharge phase distribution map into the CNN neural network, divide the weight matrix of the final linear fully connected layer into leaf nodes; according to the representative vectors of the leaf nodes and the AGNES algorithm, construct the leaf nodes into the decision tree structure to generate the induction layer.

Images