High-voltage cable partial discharge identification method based on incremental learning

Abstract

The invention belongs to the technical field of picture classification, discloses a high-voltage cable partial discharge identification method based on incremental learning, and provides a method combining incremental learning, knowledge distillation and a graph attention network for the problem of performance reduction caused by sample distribution change in partial discharge classification. And a dynamic temperature parameter adjustment strategy based on data driving is introduced. The dynamic temperature parameter adjustment strategy adjusts the temperature according to the feature distribution difference of new and old data, and dynamically sets the temperature by using the PRPD image distribution change condition in each batch of training. Specifically, depth feature vectors of new and old data are calculated, and the temperature is adjusted according to the feature distribution difference of the depth feature vectors, so that the temperature is increased when the distribution difference is large, and inter-class relationship learning of old classes is enhanced; when the difference is small, the temperature is reduced to promote new feature learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of image classification, and in particular relates to a method for identifying partial discharge of high-voltage cables based on incremental learning. Background Art

[0002] As the operation time of the cable increases, the distribution of partial discharge (PD) data will change significantly, mainly due to the combined influence of multiple factors such as cable aging, changes in external environmental conditions, and upgrades to monitoring equipment. The partial discharge characteristics of the cable at different stages may show significant differences in frequency, amplitude, and duration. Traditional classification models usually assume that the distribution of training data and test data is consistent, but this assumption is often not true in actual scenarios, resulting in a significant decrease in the classification performance of the classification model on new data. In particular, when the partial discharge signal shows distribution similarity or a certain degree of difference before and after the change, the existing classification model may find it difficult to accurately distinguish different types of partial discharge signals.

[0003] In order to cope with the challenges brought by changes in data distribution, incremental learning has gradually attracted attention and become an effective strategy to adapt to dynamic data distribution. Knowledge distillation is one of the key technologies of incremental learning. For example, existing research Rebuffi SA, Kolesnikov A, Sperl G, et al. Incremental classifier and representation learning[C] / / Conference on Computer Vision and Pattern Recognition (CVPR). 2024: 5533-5542. and Castro FM, Marín-Jiménez MJ, Guil N, et al. End-to-end incremental learning[C] / / Proceedingsof the European conference on computer vision (ECCV). 2018: 233-248., by replaying some representative historical data, transfer prior knowledge to the new model to help it learn new data and avoid the phenomenon of "catastrophic forgetting". However, these methods have shortcomings when facing the distribution difference between new and old data in incremental learning. They do not fully consider the small sample problem in the newly generated partial discharge data and the impact of the distribution difference between new and old data on the knowledge distillation effect. When there are fewer new samples, the incremental learning effect of the model is often not ideal, and the classification performance is difficult to guarantee. Summary of the invention

[0004] The purpose of the present invention is to provide a high-voltage cable partial discharge identification method based on incremental learning, to improve the identification performance in small sample scenarios, and to have better classification performance when processing new partial discharge data and historical data under changes in the cable operating environment.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A method for identifying partial discharge of a high-voltage cable based on incremental learning, the method comprising: Use the randomly initialized AlexNet network as the original model , using historical datasets to train the original model Until convergence, save the original model The optimal parameters of Using the original model Initialize and use randomly initialized The layer replaces the fully connected layer at the end to get a new model ; A representative sample selection strategy based on Euclidean distance is used to retain part of the data in the historical data set to construct representative old data, and the representative old data and new data are combined into a new data set; Feed the new dataset into the original model and new models , based on the original model and new models The characteristic vector output by the penultimate layer dynamically adjusts the temperature parameter, and calculates the original model according to the temperature parameter and new models The soft label of the original model is measured by the distillation loss function and new models The difference between the soft labels of The output prediction value and the true label calculate the classification loss function, aggregate the distillation loss function and the classification loss function, and correct the new model until convergence; Utilize a new model trained to convergence Identify partial discharge in high voltage cables.

[0006] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.

[0007] Preferably, the The execution process of the layer is as follows: First, construct the initial node feature matrix: The output of the previous layer of the layer is used as the initial node feature matrix ,in is the number of nodes, is the feature vector dimension, each node The eigenvector of It is expressed as: in Indicates Nodes No. Dimensional features, ; Then, an adaptive adjacency matrix is ​​generated based on the initial node feature matrix: for any two nodes and The similarity between them is calculated by the cosine similarity of the feature vectors of the nodes: in Representation Node and nodes The similarity between and Node and The corresponding eigenvector, Represents the feature vector and The splicing operation is performed by calculating the similarity between all node pairs to obtain the adaptive adjacency matrix , each element in the adaptive adjacency matrix Representation Node and nodes The weight of the edge between them; Finally, an adaptive attention mechanism is introduced to dynamically calculate the weight between each node and its neighboring nodes based on the initial node feature matrix and the adaptive adjacency matrix, and update the initial node feature matrix based on the weight to obtain the updated node feature matrix; The updated node feature matrix is ​​passed through the fully connected layer and activation function to output the local discharge recognition result of the high-voltage cable.

[0008] Preferably, the adaptive attention mechanism is introduced to dynamically calculate the weight between each node and its neighboring nodes based on the initial node feature matrix and the adaptive adjacency matrix, and the initial node feature matrix is ​​updated based on the weight to obtain an updated node feature matrix, including: Map the feature vector of each node in the initial node feature matrix to the new feature space through linear transformation to obtain the preliminary feature vector of the node: in is the weight matrix to be trained, is the feature vector Preliminary eigenvector after linear transformation; Compute Node and nodes The attention coefficient between : in is the attention weight vector to be trained, express The transpose of || represents the vector concatenation operation. is the activation function, is the feature vector Preliminary eigenvector after linear transformation; For all neighbor nodes, use The function normalizes the attention coefficient to obtain the normalized attention weight of each neighbor node: in Representation Node and nodes The normalized attention weights between Representation Node The set of all neighbor nodes of Representation Node and nodes The attention coefficient between Use the normalized attention weights to perform weighted summation on the preliminary feature vectors of neighbor nodes to obtain a single-head feature vector: in For Node The corresponding single-headed eigenvector, is a nonlinear activation function; Aggregate the outputs of multiple attention heads to obtain the updated node feature vector: in For Node The updated node feature vector, is the number of attention heads of the adaptive attention mechanism, Indicates Attention heads are used to normalize the attention weights Normalized attention coefficients of the output.

[0009] Preferably, the method of retaining part of the data in the historical data set and constructing representative old data by using a representative sample selection strategy based on Euclidean distance includes: Preprocess the PRPD images of partial discharge of high-voltage cables and use the pre-trained AlexNet network to extract the feature representation of each PRPD image , Represents each sample in the historical data set The eigenvector of For each category , take the category Each sample in , calculate the sample The eigenvector of , then for the category The feature vectors of all samples in are averaged to get the category The characteristic mean : in Yes Category The number of samples in ; Calculate each sample With sample The feature mean of the category The Euclidean distance between , and then sort the samples in ascending order by Euclidean distance: For each category , select the distance mean Recent samples as the representative old data of the category, forming the category A representative old data set : in Indicates the selected samples, Representation sample With sample The feature mean of the category The Euclidean distance between .

[0010] Preferably, the original model and new models The characteristic vector output by the penultimate layer dynamically adjusts the temperature parameter, and calculates the original model according to the temperature parameter and new models The soft label of the original model is measured by the distillation loss function and new models The differences between the soft tags include: Take the original model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector The model, is the feature vector dimension, the new model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector First, we calculate the cosine similarity of the two feature vectors. , the calculation formula is: Next, the calculated cosine similarity is substituted into the following formula to calculate the temperature parameter: : in is an adjustment factor, is the upper limit of temperature parameter; According to the adjusted temperature parameters , original model For Category The soft label output is: in Represents the original model The corresponding soft label probability distribution after temperature adjustment is The probability value corresponding to each category, The original model For Category Unnormalized output , Represents the original model For Category Unnormalized output , is the sum index, traversing all categories; New Model For Category The soft label output is: in Representing the new model The corresponding soft label probability distribution after temperature adjustment is The probability value corresponding to each category, It is a new model For Category Unnormalized output , Representing the new model For Category Unnormalized output ; Calculate the distillation loss function based on soft labels as follows: in, is the KL divergence, which measures the original model Output soft label distribution and sexual model Output soft label distribution The difference between.

[0011] As a preference, the new model based The output prediction value and the true label are used to calculate the classification loss function, including: in represents the classification loss function, represents the total number of samples in the new dataset, Representation sample Angle, where It is a sample The corresponding correct category label, represents the angle correction factor, represents the scaling factor, represents the total number of categories, Representation sample All incorrect categories of Representation sample In category angle.

[0012] Preferably, the aggregation distillation loss function and the classification loss function include: in is the total loss function of the aggregate distillation loss function and the classification loss function, is the distillation loss function, is the classification loss function, with parameters Used to balance the distillation loss and classification loss.

[0013] The present invention provides a high-voltage cable partial discharge identification method based on incremental learning. Aiming at the performance degradation problem caused by changes in sample distribution in partial discharge classification, a method combining incremental learning, knowledge distillation and graph attention network (GAT) is proposed, and a data-driven dynamic temperature parameter adjustment strategy is introduced.

[0014] The dynamic temperature parameter adjustment strategy adjusts the temperature according to the difference in feature distribution between the new and old data, and dynamically sets the temperature using the distribution changes of the PRPD images (i.e., the phase distribution (PRPD) atlas) in each batch of training. Specifically, the deep feature vectors of the new and old data are calculated, and the temperature is adjusted according to the difference in their feature distribution, so that when the distribution difference is large, the temperature is increased to enhance the inter-class relationship learning of the old categories; when the difference is small, the temperature is reduced to promote the learning of new features.

[0015] By introducing the attention mechanism, GAT can dynamically allocate weights between nodes, thereby improving the information interaction between new and old data. Specifically, GAT can adaptively allocate weights between historical data and new data nodes, allowing the model to more effectively utilize useful information in historical data and improve the ability to classify new data. In addition, GAT's dynamic weight allocation mechanism can flexibly adapt to the dynamic changes in data distribution, ensuring the stability of the model's classification performance at different cable operation stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method for identifying partial discharge of a high-voltage cable based on incremental learning according to the present invention; Figure 2 This is the wiring diagram of the signal acquisition system in the experiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0018] It should be noted that when a component is referred to as being “connected” to another component, it may be directly connected to the other component or there may be a central component; when a component is referred to as being “fixed” to another component, it may be directly fixed to the other component or there may be a central component.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0020] like Figure 1 As shown, this embodiment provides a method for identifying partial discharge of a high-voltage cable based on incremental learning, comprising the following steps: Step 1: Initialize the randomly The network (which consists of 5 convolutional layers and 3 fully connected layers) is used as the original model , using historical datasets Training the original model Until convergence, save the original model The optimal parameters of .

[0021] Training the original model When Input to In the network, training is performed; during the training process, the cross entropy loss function is used ( ) to minimize the model's classification error on historical data sets, using an optimization algorithm ( ) Continuously update the model's weight parameters; train until convergence, and save the original model with the best performance during the training process The weight parameters ensure the best classification performance.

[0022] Step 2: Use the original model Initialize and use randomly initialized The layer replaces the fully connected layer at the end to get a new model .

[0023] Generating a new model When , the optimal original model is used first Initialization of weight parameters The network is then randomly initialized Layer replacement after initialization The fully connected layer at the end of the network is used as the output layer to maintain exploration capabilities. The specific operations of the layer are as follows:

[0024] (1) Construct the initial node feature matrix: Take The output of the previous layer of the layer is used as the initial node feature matrix ,in is the number of nodes, is the feature vector dimension, each node The eigenvector of It is expressed as: in Indicates Nodes No. Dimensional features, .

[0025] (2) An adaptive adjacency matrix construction method is used to generate an adaptive adjacency matrix based on the initial node feature matrix. Specifically, for any two nodes and The similarity between them is calculated by the cosine similarity of the feature vectors of the nodes:

[0026] in Representation Node and nodes The similarity between and Node and The corresponding eigenvector, Represents the feature vector and The splicing operation is performed by calculating the similarity between all node pairs to obtain the adaptive adjacency matrix , each element in the adaptive adjacency matrix Representation Node and nodes The weight of the edge between .

[0027] (3) An adaptive attention mechanism is introduced to dynamically calculate the weight between each node and its neighboring nodes based on the initial node feature matrix and the adaptive adjacency matrix, and the initial node feature matrix is ​​updated based on the weight to obtain the updated node feature matrix.

[0028] (3.1) Map the feature vector of each node in the initial node feature matrix to the new feature space through linear transformation to obtain the preliminary feature vector of the node: in is the weight matrix to be trained, is the feature vector Preliminary eigenvector after linear transformation.

[0029] (3.2) Computing nodes and nodes The attention coefficient between , attention coefficient Reflects the node and nodes Importance: in is the attention weight vector to be trained, express The transpose of || represents the vector concatenation operation. is the activation function, is the feature vector Preliminary eigenvector after linear transformation; (3.3) For all neighbor nodes, use The function normalizes the attention coefficient to obtain the normalized attention weight of each neighbor node: in Representation Node and nodes The normalized attention weights between Representation Node The set of all neighbor nodes of Representation Node and nodes The attention coefficient between (3.4) Use the normalized attention weights to perform weighted summation on the preliminary feature vectors of neighboring nodes to obtain a single-head feature vector: in For Node The corresponding single-headed eigenvector, is a nonlinear activation function; This embodiment adopts the learning process of the multi-head attention mechanism (actually using 4 attention heads), so it is necessary to aggregate the outputs of multiple attention heads to obtain the updated node feature vector: in For Node The updated node feature vector, is the number of attention heads of the adaptive attention mechanism, Indicates Attention heads are used to normalize the attention weights Normalized attention coefficients of the output.

[0030] The introduction of the attention mechanism can adaptively assign different weights to important connections between new and old data, which has significant advantages, especially when dealing with small samples and changes in data distribution in partial discharge scenarios. The attention mechanism dynamically adjusts the weight of each node, allowing the model to more accurately capture the subtle relationship between new data and historical data. Compared with traditional methods, GAT can identify key features in new data and strengthen its connection with historical data, thereby improving the generalization ability and classification accuracy of the model. Especially when the category distribution of new and old data is similar, the attention mechanism can effectively highlight the connection between key nodes, ensuring that the model still has excellent classification performance in the face of distribution changes.

[0031] (4) The updated node feature matrix is ​​passed through a fully connected layer and an activation function to output the local discharge recognition result of the high-voltage cable.

[0032] Step 3: Use the representative sample selection strategy based on Euclidean distance to retain part of the data in the historical data set to construct representative old data, and combine the representative old data and new data to form a new data set.

[0033] Step 3.1: Data preprocessing and feature extraction: Preprocess the PRPD images of high-voltage cable partial discharge and use the pre-trained AlexNet network to extract the feature representation of each PRPD image. , Represents each sample in the historical data set The feature vector of .

[0034] Step 3.2, calculate the feature mean of each category: for each category , take the category Each sample in , calculate the sample The eigenvector of , then for the category The feature vectors of all samples in are averaged to get the category The characteristic mean : in Yes Category The number of samples in .

[0035] Step 3.3: Calculate each sample With sample The feature mean of the category The Euclidean distance between , and then sort the samples in ascending order by Euclidean distance: For each category , select the distance mean Recent samples as the representative old data of the category, forming the category A representative old data set : in Indicates the selected samples, Representation sample With sample The feature mean of the category The Euclidean distance between Thus, each category A representative old data set in Can effectively represent the main characteristics of the category.

[0036] Step 3.4, the representative old data and new data are combined into a new data set: the representative old data of all categories and the newly collected new data of each category are combined into a new data set. The old and new data are divided according to time, and the time boundary is adjusted as needed. For example, the current time of this identification can be used as the time boundary, the data within a preset time length before the current time is used as the old data, and the data collected at the current time is used as the new data. Of course, the time boundary can also be any historical moment.

[0037] Step 4: Input the new dataset into the original model and new models , based on the original model and new models The characteristic vector output by the penultimate layer dynamically adjusts the temperature parameter, and calculates the original model according to the temperature parameter and new models The soft label of the original model is measured by the distillation loss function and new models The difference between the soft labels of The output prediction value and the true label calculate the classification loss function, aggregate the distillation loss function and the classification loss function, and correct the new model until convergence.

[0038] For the soft labels output by the model, this embodiment introduces a temperature parameter that is dynamically adjusted based on the feature difference of the deep model. The temperature is dynamically adjusted by calculating the cosine similarity between the feature vector output by the penultimate layer (fully connected layer) of the original AlexNet model and the feature vector output by the corresponding penultimate layer (fully connected layer) in the new model. Prior knowledge is transferred by constraining and approximating the soft label output obtained based on the adjusted temperature for representative old data to protect previous learning and avoid catastrophic forgetting.

[0039] Among them, the distillation loss function is calculated as follows: Model output Represents the unnormalized score of the model for each category. In order to obtain soft labels, this embodiment introduces a temperature parameter T to soften the output distribution. Different from the fixed temperature set in traditional knowledge distillation, the present invention designs a data-driven dynamic temperature strategy, which adjusts the temperature according to the difference in feature distribution of new and old data, and dynamically sets the temperature based on the distribution changes of the PRPD graph in each batch of training. Specifically, take the original model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector The model, is the feature vector dimension, the new model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector First, we calculate the cosine similarity of the two feature vectors. , the calculation formula is:

[0040] Next, the calculated cosine similarity is substituted into the following formula to calculate the temperature parameter: : in is an adjustment factor used to adjust the amplitude of temperature change. is the upper limit of the temperature parameter. , The soft label output operation in the knowledge distillation process is performed according to the calculated temperature, and the feature vector is re-extracted and the cosine similarity and temperature are calculated in each training iteration to ensure that the temperature dynamically reflects the distribution changes of the new and old data. In this way, as the distribution difference between the new and old data increases, the temperature will automatically increase, allowing the model to pay more attention to the inter-class relationship of the old categories; when the distribution difference is small, the temperature will automatically decrease, which is conducive to the model to learn the new category features in a refined manner.

[0041] According to the adjusted temperature parameters , original model For Category The soft label output is: in Represents the original model The corresponding soft label probability distribution after temperature adjustment is The probability value corresponding to each category, The original model For Category Unnormalized output , Represents the original model For Category Unnormalized output , is the temperature parameter, which is used to control the smoothness of the soft label; is the sum index, traversing all categories.

[0042] New Model For Category The soft label output is: in Representing the new model The corresponding soft label probability distribution after temperature adjustment is The probability value corresponding to each category, It is a new model For Category Unnormalized output , Representing the new model For Category Unnormalized output .

[0043] Calculate the distillation loss function based on soft labels as follows: in, is the KL divergence, which measures the original model Output soft label distribution and sexual model Output soft label distribution The difference between.

[0044] The classification loss function is calculated as follows: The classification loss function of this embodiment refers to the new model The difference between the predicted value and the true value The loss function is: in represents the classification loss function, Represents the total number of samples in the new dataset, that is, how many samples are there in the batch. Representation sample Angle, where It is a sample The corresponding correct category label, Represents the angle correction coefficient, by introducing Make the decision boundary of classification more strict and enhance the distance between classes. Represents the scaling factor, usually a hyperparameter, used to enlarge To improve the classification performance of the model, represents the total number of categories, Representation sample All incorrect categories of Representation sample In category angle, used to calculate the scores of other categories, Representation sample For its correct category The cosine similarity of Representation sample For categories The cosine similarity of Indicates the sample The scores of all incorrect categories are summed.

[0045] This embodiment The loss function replaces the traditional softmax classification loss function and significantly improves the performance of the model when processing small sample partial discharge data. The loss function enhances the distinction between features of different categories by introducing angle intervals, especially when the new and old data categories are highly similar, which can effectively widen their classification boundaries. Compared with the traditional softmax classification loss function, The loss function not only improves the model's ability to recognize new data, but also effectively avoids category confusion and ensures stable memory of old data knowledge. The substitution effect of the loss function in the incremental learning framework not only improves the classification accuracy of the model, but also solves the problem of small samples and similar categories being difficult to distinguish in partial discharge scenarios.

[0046] The loss function is constructed by adding parameters Will Distance changes to ,Will Change to , and normalize the weight vector .exist On the basis of The loss function is . Indicates that the parameters Corrected sample To its correct category This correction term increases the boundaries between classes, making it more difficult for the classifier to classify samples into the wrong category, thereby improving classification accuracy.

[0047] Therefore, the total loss function is calculated as follows: in is the total loss function of the aggregate distillation loss function and the classification loss function, is the distillation loss function, which is used to keep the new model consistent with the teacher model on the old data. is the classification loss function for representative old and new data, ensuring that the new model can effectively learn the characteristics of new data. It is used to balance the distillation loss and classification loss, so that the new model can better adapt to new data while ensuring the recognition ability of old data.

[0048] Step 5: Use the new model trained to convergence Identify partial discharge in high voltage cables.

[0049] As the high-voltage cables increase in service life, the distribution of partial discharge data will change, making it difficult for traditional models trained based on historical data to effectively classify new data. To solve this problem, the present invention proposes a partial discharge classification method that combines incremental learning, knowledge distillation, and graph attention networks and introduces a data-driven dynamic temperature adjustment strategy. This method gradually updates the model through incremental learning, without the need to retrain all data, reducing computing resource requirements; combined with knowledge distillation and graph attention networks, the model's ability to learn new data is enhanced, solving the "catastrophic forgetting" problem in small sample scenarios.

[0050] The advantages of the method of the present invention are illustrated by experiments below.

[0051] The experimental environment is prepared as follows: like Figure 2 As shown, the experiment of this embodiment is implemented based on a signal acquisition system, which includes a synchronous phase transmitter, a synchronous phase receiver, an online monitoring device for partial discharge of a high-voltage cable, a coupling electrode, an interference signal module, and a signal acquisition and control terminal (including a battery module, a display screen, etc.). The specific process of signal acquisition is as follows:

[0052] (1) Signal injection: First, the partial discharge signal and the interference signal are injected into the cable simultaneously through the coupling electrode. The partial discharge signal is used to simulate the discharge behavior caused by the cable insulation defect, while the interference signal is used to simulate the external noise and signal interference that may exist in the actual working conditions.

[0053] (2) Signal synchronization and monitoring: The system ensures that the signal acquisition phase of the partial discharge detection equipment is consistent with the phase of the system by synchronizing the phase of the transmitter and receiver, thereby ensuring the accuracy of the phase during the data acquisition process.

[0054] (3) Data transmission: The injected signal is transmitted to the PD online monitoring device through a mutual inductor at the end of the cable, and the monitoring device records and analyzes the signal in real time. 6000 cycles of data are collected each time to generate a complete PRPD spectrum.

[0055] (4) Data processing: After denoising and processing, the collected signals are used to generate a partial discharge data set for model training and testing. By collecting signals under different working conditions, simulating the "new data" generated after the cable runs for a period of time or the monitoring equipment is changed, and combining it with historical data, the effectiveness of the incremental learning algorithm is verified.

[0056] The experimental process is as follows: (1) Data preparation: The experimental task is to classify the partial discharge of high-voltage cables. The collected partial discharge types are divided into five categories, namely, air gap discharge inside insulation, air gap discharge of insulation contacting high-voltage conductors, metal particle jumping discharge on the surface of high-voltage conductors, cracking discharge of the semi-conductive layer outside insulation, and electrical tree discharge inside insulation.

[0057] use Figure 2 The signal acquisition system generates 2500 (500 / class) PRPD spectra to represent the historical data set and 1000 (200 / class) PRPD spectra to represent the new data set. 500 PRPD spectra are randomly selected from the two data sets as their respective test sets, and the rest are used as the corresponding training sets ( and ).in, Used to train the original partial discharge recognition model, is the new data needed in the incremental learning phase. In addition, from the historical data set Select 20% of the historical data as a representative sample , used as auxiliary information in incremental learning to avoid catastrophic forgetting while learning new data.

[0058] (2) Data preprocessing: All images are resized to 256*256, converted to PyTorch tensors, and normalized according to the mean and standard deviation of the ImageNet dataset.

[0059] (3) Training process: Step 1: First, from the historical data set A representative sample selection strategy based on Euclidean distance is used to select part of the data to build a representative historical data set At the same time, the newly added partial discharge PRPD spectrum is preprocessed to generate a new data set .

[0060] Step 2: From Select different types of PRPD maps from the dataset and sample the samples to be tested to form a training unit. Assume that the new model The training batch size is K, where half of the K samples come from The other half of the samples came from In order to ensure that the sample types in the training batch are relatively balanced, The samples in are supplemented by random oversampling.

[0061] Step 3: Use the The trained model is used as the original model , and use this model to generate a new model During the construction process, the weights of the feature extraction part are kept unchanged, and the last fully connected layer is replaced by a randomly initialized GAT layer as the new model The output layer of , thus enhancing its adaptability to new data.

[0062] Step 4: Use the original model during incremental learning right The samples in the dataset are used for knowledge distillation, and the soft labels they output are used as prior knowledge to assist the new model. Perform incremental training.

[0063] Step 5: In incremental training, we first construct a node feature matrix and an adaptive adjacency matrix as input to the GAT network. The node feature matrix is ​​generated by the output of the last fully connected layer of the original model as the initial feature representation of each node. Next, by constructing an adaptive adjacency matrix, the connection relationship between nodes is dynamically adjusted to reflect the correlation between different partial discharge data. Subsequently, the GAT network is used to update the node features, and by aggregating the information of neighboring nodes, the commonalities and differences between new and old data are further explored to improve the recognition ability of the model. Finally, the updated node features are used for classification tasks, and the recognition results of the samples to be tested are output. Finally, the model Output the recognition result of the sample to be tested.

[0064] Experimental results: The experimental results show that the method of the present invention is verified on a PRPD atlas dataset containing 2500 historical data and 1000 newly added data. The classification accuracy of the newly added data is increased from 80.50% to 99.40%, and the accuracy of the historical data is only reduced by 1.50%. Compared with the traditional learning without forgetting (LwF) and conventional knowledge distillation methods, the recognition performance of the present invention in small sample scenarios is improved by 3.7% and 7.6%. This shows that the present method has good classification performance when processing new partial discharge data and historical data under the changing cable operating environment.

[0065] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for identifying partial discharge of high-voltage cables based on incremental learning, characterized in that: The incremental learning-based high-voltage cable partial discharge identification method comprises: Use the randomly initialized AlexNet network as the original model , using historical datasets to train the original model Until convergence, save the original model The optimal parameters of Using the original model Initialize and use randomly initialized The layer replaces the fully connected layer at the end to get a new model ; A representative sample selection strategy based on Euclidean distance is used to retain part of the data in the historical data set to construct representative old data, and the representative old data and new data are combined into a new data set; Feed the new dataset into the original model and new models , based on the original model and new models The characteristic vector output by the penultimate layer dynamically adjusts the temperature parameter, and calculates the original model according to the temperature parameter and new models The soft label of the original model is measured by the distillation loss function and new models The difference between the soft labels of The output prediction value and the true label calculate the classification loss function, aggregate the distillation loss function and the classification loss function, and correct the new model until convergence; Utilize a new model trained to convergence Identify partial discharge in high voltage cables.

2. The high-voltage cable partial discharge identification method based on incremental learning according to claim 1 is characterized in that: Said The execution process of the layer is as follows: First, construct the initial node feature matrix: The output of the previous layer of the layer is used as the initial node feature matrix ,in is the number of nodes, is the feature vector dimension, each node The eigenvector of It is expressed as: in Indicates Nodes No. Dimensional features, ; Then, an adaptive adjacency matrix is ​​generated based on the initial node feature matrix: for any two nodes and The similarity between them is calculated by the cosine similarity of the feature vectors of the nodes: in Representation Node and nodes The similarity between and Node and The corresponding eigenvector, Represents the feature vector and The splicing operation is performed by calculating the similarity between all node pairs to obtain the adaptive adjacency matrix , each element in the adaptive adjacency matrix Representation Node and nodes The weight of the edge between them; Finally, an adaptive attention mechanism is introduced to dynamically calculate the weight between each node and its neighboring nodes based on the initial node feature matrix and the adaptive adjacency matrix, and the initial node feature matrix is ​​updated based on the weight to obtain the updated node feature matrix; The updated node feature matrix is ​​passed through the fully connected layer and activation function to output the local discharge recognition result of the high-voltage cable.

3. The high-voltage cable partial discharge identification method based on incremental learning according to claim 2 is characterized in that: The adaptive attention mechanism is introduced to dynamically calculate the weight between each node and its neighboring nodes based on the initial node feature matrix and the adaptive adjacency matrix, and update the initial node feature matrix based on the weight to obtain an updated node feature matrix, including: Map the feature vector of each node in the initial node feature matrix to the new feature space through linear transformation to obtain the preliminary feature vector of the node: in is the weight matrix to be trained, is the feature vector Preliminary eigenvector after linear transformation; Compute Node and nodes The attention coefficient between : in is the attention weight vector to be trained, express The transpose of || represents the vector concatenation operation. is the activation function, is the feature vector Preliminary eigenvector after linear transformation; For all neighbor nodes, use The function normalizes the attention coefficient to obtain the normalized attention weight of each neighbor node: in Representation Node and nodes The normalized attention weights between Representation Node The set of all neighbor nodes of Representation Node and nodes The attention coefficient between Use the normalized attention weights to perform weighted summation on the preliminary feature vectors of neighbor nodes to obtain a single-head feature vector: in For Node The corresponding single-headed eigenvector, is a nonlinear activation function; Aggregate the outputs of multiple attention heads to obtain the updated node feature vector: in For Node The updated node feature vector, is the number of attention heads of the adaptive attention mechanism, Indicates Attention heads are used to normalize the attention weights Normalized attention coefficients of the output.

4. The high-voltage cable partial discharge identification method based on incremental learning according to claim 1 is characterized in that: The representative sample selection strategy based on Euclidean distance is used to retain part of the data in the historical data set to construct representative old data, including: Preprocess the PRPD images of high-voltage cable partial discharge and use the pre-trained AlexNet network to extract the feature representation of each PRPD image , Represents each sample in the historical data set The eigenvector of For each category , take the category Each sample in , calculate the sample The eigenvector of , then for the category The feature vectors of all samples in are averaged to get the category The characteristic mean : in Yes Category The number of samples in ; Calculate each sample With sample The feature mean of the category The Euclidean distance between , and then sort the samples in ascending order by Euclidean distance: For each category , select the distance mean Recent samples as the representative old data of the category, forming the category A representative old data set : in Indicates the selected samples, Representation sample With sample The feature mean of the category The Euclidean distance between .

5. The high-voltage cable partial discharge identification method based on incremental learning according to claim 1 is characterized in that: The original model and new models The characteristic vector output by the penultimate layer dynamically adjusts the temperature parameter, and calculates the original model according to the temperature parameter and new models The soft label of the original model is measured by the distillation loss function and new models The differences between the soft tags include: Take the original model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector The model, is the feature vector dimension, the new model The feature vector output by the penultimate layer is , is the feature vector The Features, is the feature vector First, we calculate the cosine similarity of the two feature vectors. , the calculation formula is: Next, the calculated cosine similarity is substituted into the following formula to calculate the temperature parameter: : in is an adjustment factor, is the upper limit of temperature parameter; According to the adjusted temperature parameters , original model For Category The soft label output is: in Represents the original model The corresponding probability distribution of the soft label after temperature adjustment is The probability value corresponding to each category, The original model For Category Unnormalized output , Represents the original model For Category Unnormalized output , is the sum index, traversing all categories; New Model For Category The soft label output is: in Representing the new model The corresponding probability distribution of the soft label after temperature adjustment is The probability value corresponding to each category, It is a new model For Category Unnormalized output , Representing the new model For Category Unnormalized output ; Calculate the distillation loss function based on soft labels as follows: in, is the KL divergence, which measures the original model Output soft label distribution and sexual model Output soft label distribution The difference between.

6. The high-voltage cable partial discharge identification method based on incremental learning according to claim 1 is characterized in that: The new model The output prediction value and the true label are used to calculate the classification loss function, including: in represents the classification loss function, represents the total number of samples in the new dataset, Representation sample Angle, where It is a sample The corresponding correct category label, represents the angle correction factor, represents the scaling factor, represents the total number of categories, Representation sample All incorrect categories of Representation sample In category angle.

7. The high-voltage cable partial discharge identification method based on incremental learning according to claim 1 is characterized in that: The aggregation distillation loss function and the classification loss function include: in is the total loss function of the aggregate distillation loss function and the classification loss function, is the distillation loss function, is the classification loss function, with parameters Used to balance the distillation loss and classification loss.

Images