Incremental learning method and system for cable partial discharge based on graph convolutional neural network
By using a graph convolutional neural network-based incremental learning method for cable partial discharge, the adaptability of the cable partial discharge identification model in dynamic environments and the identification of new defects are solved. This method achieves efficient cross-domain feature mapping and adaptive updating, thereby improving identification accuracy and learning ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG VOCATIONAL COLLEGE
- Filing Date
- 2026-03-19
- Publication Date
- 2026-07-31
AI Technical Summary
Existing cable partial discharge identification models are ill-suited to adapting to changes in data distribution under dynamic operating conditions, are unable to identify new partial discharge types in a timely manner, and are prone to catastrophic forgetting during continuous learning. Furthermore, existing incremental learning methods cannot simultaneously adapt to similar data from different domains and learn new defect types.
A method for incremental learning of partial discharge in cables based on graph convolutional neural networks is adopted. By freezing the encoder parameters, the incremental learning task type is determined to be either class increment or domain increment. Feature mapping and adaptive weight update are then performed in the corresponding dynamic subspace to achieve cross-domain feature alignment and balance between new and old knowledge.
In cable partial discharge identification, the model has achieved adaptation to similar cross-domain data and learning of new defect types, maintaining high identification accuracy while reducing computational overhead, reducing catastrophic forgetting, and improving the model's continuous learning ability.
Smart Images
Figure CN122490148A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment condition monitoring and intelligent fault diagnosis technology, specifically to a method and system for incremental learning of partial discharge in cables based on graph convolutional neural networks. Background Technology
[0002] With the development of power grid digitalization and intelligent operation and maintenance, cable partial discharge monitoring data continues to accumulate. Although deep learning-based partial discharge pattern recognition methods have made progress, most existing methods are based on offline training and the premise that the training and test sets are independently and identically distributed, making it difficult to adapt to dynamic scenarios in actual operation. In practical applications, on the one hand, partial discharge is greatly affected by detection conditions, operating environment, equipment aging, and noise interference, and the distribution of data features collected at different times and under different operating conditions often changes significantly. On the other hand, as the service life of cables increases and defects evolve, new types of partial discharge samples will constantly emerge, making it difficult for the original fixed-category recognition model to cover new defects. At the same time, partial discharge samples often suffer from problems such as insufficient samples and class imbalance, further limiting the model's generalization ability.
[0003] Existing incremental learning methods are typically designed for single scenarios. Domain-incremental methods, oriented towards environmental changes, primarily address the migration of similar samples across different distributions, but lack sufficient support for learning new defect types. Class-incremental methods, oriented towards new categories, focus more on knowledge preservation and fusion between old and new categories, often ignoring the impact of data distribution drift under different operating conditions. Therefore, in cable partial discharge identification tasks, existing technologies struggle to simultaneously address the dual requirements of "continuous adaptation to similar data across different domains" and "continuous learning of new defect types," easily leading to problems such as decreased identification accuracy, significant catastrophic forgetting, and excessively high retraining costs. Therefore, it is necessary to propose an intelligent partial discharge identification method and system that can uniformly handle domain-incremental and class-incremental scenarios while balancing identification accuracy, computational efficiency, and continuous learning capabilities. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for incremental learning of partial discharge in cables based on graph convolutional neural networks, in order to solve the problems of existing partial discharge identification models relying on offline static training, being difficult to adapt to changes in data distribution under different operating environments, being unable to identify new partial discharge types in a timely manner, and being prone to catastrophic forgetting during continuous learning.
[0005] To achieve the above objectives, this invention provides an incremental learning method for cable partial discharge based on a graph convolutional neural network, comprising the following steps: acquiring historical cable partial discharge data and corresponding type labels, and pre-training a graph convolutional neural network basic model based on the historical partial discharge data to obtain a pre-trained model; freezing the encoder parameters in the pre-trained model, inputting the historical partial discharge data into the encoder to extract historical features, and simultaneously inputting new input partial discharge data into the encoder to extract new input features; based on the distribution difference and feature distance between the historical features and the new input features, determining the incremental learning task type corresponding to the new input partial discharge data as a class incremental learning task, a domain incremental learning task, or a task without incremental learning. The process involves several steps: When the task is classified as a class-based incremental learning task, a dynamic subspace is constructed based on historical features, and the embedded features corresponding to the new input partial discharge data are mapped to the dynamic subspace to achieve feature integration of the new partial discharge type; when the task is classified as a domain-based incremental learning task, dynamic subspaces are constructed based on source domain features and target domain features respectively, and the embedded features corresponding to the new input partial discharge data are mapped to the target domain subspace through adversarial dynamic mapping to achieve cross-domain feature alignment; then, based on the distribution differences corresponding to the class-based incremental learning task or the domain-based incremental learning task, the model weight update step size is adaptively determined, and the model weights are iteratively updated using the new input partial discharge data; finally, the updated model is used to output the partial discharge type identification result.
[0006] Furthermore, the base model of the graph convolutional neural network is preferably a residual graph convolutional neural network (ResGCN); the pre-training process uses optimization algorithms such as Adam for training, and updates the model parameters with a cross-entropy loss function; the output of the base model is:
[0007]
[0008] in, For encoder parameters, Let be the classifier weight matrix. The corresponding cross-entropy loss function is:
[0009]
[0010] in, This represents the total number of partial discharge types. This is the indicator function. After training, the encoder weights are frozen, making them a fixed feature extractor for the incremental learning phase.
[0011] Furthermore, during the incremental learning task type determination process, features from historical partial discharge data are extracted using the frozen encoder. Features of newly input partial discharge data ,Right now:
[0012]
[0013] Maximum Mean Difference (MMD) Measured and Differences in characteristic distribution:
[0014]
[0015] Simultaneously, calculate the average Euclidean distance between the features of the new input data and the features of the historical data:
[0016]
[0017] Set distribution difference threshold and distance threshold, if Greater than the preset threshold and If the distance exceeds a preset threshold, it is determined to be an incremental learning task; if Greater than the preset threshold and If the distance is less than a preset threshold, it is determined to be a domain incremental learning task; if neither exceeds the threshold, it is determined that no incremental learning task is required.
[0018] Furthermore, when the task is classified as an incremental learning task, historical data features are utilized. Construct dynamic subspaces. Divide historical data into subspaces using clustering algorithms. There are clusters, and the mean vector of each cluster. As a representative vector. For the newly input task data. The initial embedding vector is obtained through encoder and flattening operations. :
[0019]
[0020] calculate With each cluster representative vector Cosine similarity:
[0021]
[0022] Select the one with the highest similarity Each cluster constructs a dynamic subspace associated with the new data. Further, within this dynamic subspace, the initial embedding vector... Perform a local linear transformation with regularization terms and solve for the transformation matrix. :
[0023]
[0024] in, For regularization parameters, Let be the subspace matrix formed by the selected cluster basis vectors. Its solution is:
[0025]
[0026] This leads to the mapped embedding vector:
[0027]
[0028] This is to better integrate the features of newly input partial discharge data into the existing type system and reduce the disruption to the feature structure of old categories.
[0029] Furthermore, when the task is determined to be a domain incremental learning task, the source domain features are first processed separately. and target domain features Construct a dynamic subspace. Obtain the representative vectors of the source domain clusters using a clustering algorithm. and target domain cluster representative vector The new input data is encoded and flattened to obtain the initial embedding vector. Calculate the cosine similarity between it and the representative vectors of each cluster in the source domain:
[0030]
[0031] Through a dynamic mapping module based on adversarial learning, Mapped to the target domain subspace. Let the domain discriminator be... The mapping function is Its objective function is:
[0032]
[0033] in, and Data distribution for the source and target domains, respectively. This is the regularization coefficient. By minimizing... The distance between the vectors representing the most similar clusters in the target domain and the vectors in the target domain is used to obtain the domain-adapted embedding features. This enables cross-domain feature alignment.
[0034] Furthermore, when the incremental learning task type is a quasi-incremental learning task, JS divergence is used to represent the distribution of historical task features. Distribution of new task features Differences between them:
[0035]
[0036] The weight update step size is based on the incremental learning of the JS divergence calculation class:
[0037]
[0038] in, This is the initial step size parameter. Then, the loss function is calculated based on the new task data. And its gradient with respect to the model weights, and complete the model weight update according to the weight update step size:
[0039]
[0040] Furthermore, when the incremental learning task is a domain incremental learning task, the loss function of the model in the source and target domains is calculated using feedback information from the domain discriminator. and And calculate the difference between the two:
[0041]
[0042] Combining the aforementioned loss difference and the maximum mean difference (MMD) between the source domain features and the target domain features, calculate the weight update step size for domain incremental learning:
[0043] ;
[0044] in, To balance the parameters, the gradient of the loss function with respect to the model weights is calculated based on the target domain data, and the model weights are updated according to the stated weight update step size.
[0045] .
[0046] This invention also provides a cable partial discharge incremental learning system based on graph convolutional neural networks, comprising: a basic model pre-training module, a feature extraction module, a task type discrimination module, a class incremental mapping module, a domain incremental mapping module, an adaptive weight update module, and a recognition output module; wherein, the basic model pre-training module is used to pre-train the graph convolutional neural network basic model based on historical partial discharge data and freeze the encoder parameters; the feature extraction module is used to extract features from historical partial discharge data and new input partial discharge data; the task type discrimination module is used to determine the incremental learning task type based on the distribution difference and feature distance between historical features and new input features; the class incremental mapping module is used to construct a dynamic subspace when the task is determined to be a class incremental learning task and embed the new input partial discharge data into the corresponding feature mapping of the dynamic subspace; the domain incremental mapping module is used to construct a source domain subspace and a target domain subspace when the task is determined to be a domain incremental learning task and achieve cross-domain feature alignment through adversarial dynamic mapping; the adaptive weight update module is used to adaptively determine the model weight update step size and update the model based on the difference information corresponding to the class incremental learning task or the domain incremental learning task; and the recognition output module is used to output the partial discharge type recognition result.
[0047] Compared with the prior art, the beneficial effects of the present invention are:
[0048] 1. This invention can uniformly handle two types of continuous learning problems in cable partial discharge identification: domain increment and class increment; it can automatically match learning strategies under different increment scenarios through a task type discrimination mechanism; it can reduce catastrophic forgetting in the process of learning new categories through dynamic subspace mapping, and improve the adaptability of samples from different domains of the same type through adversarial cross-domain mapping; 2. It can balance the absorption of new knowledge and the retention of old knowledge through an adaptive weight update mechanism, thereby maintaining high recognition accuracy under conditions of few samples, class imbalance and changes in the operating environment, and reducing the computational overhead caused by overall retraining. Attached Figure Description
[0049] Figure 1 This is a flowchart of the incremental learning method for partial discharge of cables based on graph convolutional neural networks according to the present invention.
[0050] Figure 2 This is a block diagram of the incremental learning system for cable partial discharge based on graph convolutional neural networks according to the present invention.
[0051] Figure 3 This is a comparison chart of the recognition effect of newly added data before and after incremental update in the domain incremental learning embodiment of the present invention, wherein... Figure 3 (a) is the confusion matrix of the new data before the incremental update. Figure 3 (b) is the confusion matrix of the new data after incremental update.
[0052] Figure 4 This is a comparison chart of feature distribution and recognition performance at different stages of incremental learning in an incremental learning embodiment of the present invention. Figure 4 (a) is a scatter plot of the features of Task1 stage. Figure 4 (b) is the confusion matrix diagram for Task 4. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0054] In the following two embodiments, Embodiment 1 focuses on how the present invention processes newly added partial discharge data from different domains of the same type, mainly corresponding to claims 1, 2, 3, 6, and 8; Embodiment 2 focuses on how the present invention processes continuously learned data of newly added defect types, mainly corresponding to claims 1, 2, 3, 4, 5, and 7. The two embodiments complement each other, highlighting the implementation steps and technical effects of domain incremental branching and class incremental branching, respectively; Table 1 below shows the sample library and task settings in the embodiments.
[0055] Table 1: Sample Library and Task Settings
[0056]
[0057] The methods and modules used are as follows: Figures 1 to 2 As shown, Figure 1 This is a flowchart of the cable partial discharge domain-class incremental learning method based on graph convolutional neural network of the present invention; the flowchart mainly shows the overall steps of historical data input, basic model pre-training, feature extraction, task type discrimination, class incremental / domain incremental branch processing, adaptive weight update and recognition output. Figure 2 This is a block diagram of the cable partial discharge domain-class incremental learning system based on graph convolutional neural network of the present invention. The diagram mainly shows the connection relationship between the basic model pre-training module, feature extraction module, task type discrimination module, class incremental mapping module, domain incremental mapping module, adaptive weight update module and recognition output module.
[0058] Example 1:
[0059] This embodiment is an example of incremental learning for partial discharge domain in cables. It uses two sets of data representing four typical partial discharge models: set a (tip, air gap, suspension, and surface discharge) serves as historical domain samples, and set b (corresponding to the four discharge types) serves as new target domain samples. This embodiment illustrates the domain incremental learning implementation process of this invention when the category semantics remain unchanged but the test conditions change.
[0060] Step 1: Pre-train the basic model on historical domain samples. Input group a of samples into ResGCN, and write the model output as...
[0061]
[0062] And with cross-entropy loss
[0063]
[0064] The parameters are optimized. The first equation maps the localized features after graph convolutional encoding to classification probabilities, while the second equation uses a set of known labels to establish the basic classification boundary. After training, the encoder parameters are frozen so that they only perform the function of stable feature extraction in the subsequent domain increment stage, thereby reducing the computational overhead caused by overall retraining.
[0065] Step 2: Perform task type discrimination on the newly added group b samples. After freezing the encoder, extract historical features and new domain features respectively:
[0066]
[0067] Then calculate the distribution difference between the two.
[0068]
[0069] and average Euclidean distance
[0070]
[0071] MMD is used to determine whether there has been a significant shift in the overall statistical distribution between the new sample and the historical sample. This is used to distinguish whether the drift is more likely to originate from a "domain change" or a "category change". Since groups a and b still correspond to the same defect semantic category, in this embodiment, this step determines the new task as a domain increment branch rather than a class increment branch.
[0072] Step 3: Perform dynamic subspace mapping of the target domain. First, map the source domain features... and target domain features Clustering was performed separately to obtain the prototype of the source domain cluster. and target domain cluster prototype Then map the new samples to the initial embedding.
[0073]
[0074] And calculate its similarity to the source domain cluster prototype.
[0075]
[0076] Subsequently, a domain discriminator was introduced. and mapping function ,according to
[0077]
[0078] Adversarial alignment is performed on the newly added samples. This formula gradually pulls the samples from group b into the target domain subspace, making the representation of the same defect more consistent in both domains, thereby reducing the recognition bias caused by environmental changes.
[0079] Step four: Perform domain incremental adaptive weight update. First, the source domain loss is obtained by the domain discriminator. and target domain loss ,remember Press again
[0080]
[0081] Calculate the incremental update step size of the domain and use
[0082]
[0083] Update the model parameters. Specifically, MMD is used to characterize the performance difference of the model in the source domain and the target domain. MMD is used to characterize the difference in feature distribution between the two domains. Together, they determine the weight update magnitude to balance the adaptability of the target domain and the stability of historical domain recognition.
[0084] Step 5: Output the recognition results and complete the verification of the embodiment. Before the incremental update, the model had an accuracy rate of 99.13% on the old data, but only 78.02% on the newly added group b data, with recognition rates of only 57% for type 7 floating discharge and 66% for type 5 tip discharge. After adopting the domain incremental strategy of this invention, the overall recognition accuracy on the newly added group b data was significantly improved, with recognition rates of all four types of discharge reaching over 97%. Meanwhile, the updated model maintained a recognition accuracy of 97% to 99% on historical data, indicating that this embodiment can rapidly adapt to target domain samples while preserving historical domain knowledge. Figure 3 This is a comparison chart of the recognition effect of newly added data before and after incremental update in the domain incremental learning embodiment of the present invention, wherein... Figure 3 (a) is the confusion matrix of the new data before the incremental update. Figure 3 (b) is the confusion matrix of the new data after incremental update.
[0085] Example 2:
[0086] This embodiment is an incremental learning implementation for cable partial discharge. It uses eight types of cable terminal defect data as the object, dividing them into four sequential tasks: Task 1 adds insulation surface transverse cuts and suspended particles on the insulation surface; Task 2 adds outer semiconducting layer residue and metal burrs; Task 3 adds stress cone misalignment and moisture absorption; and Task 4 adds insulation surface circumferential cuts, copper shielding corrosion, and grounding wire corrosion. This embodiment illustrates the implementation process of incremental learning for newly added defect types while maintaining the ability to identify historical categories.
[0087] Step 1: First, use the Task1 samples to complete the basic model training. The training method is the same as in Example 1, that is, through...
[0088]
[0089] and cross-entropy loss
[0090]
[0091] Determine the initial parameters and freeze the encoder. The purpose of this step is to allow the model to first master the stable feature representation of the known old classes, and subsequent class increments will only expand on this basis without destroying the already learned underlying graph structure representation.
[0092] Step two, when Task 2, Task 3, or Task 4 arrives, first press...
[0093]
[0094] Extract historical and new features, and continue calculating MMD and In this embodiment, the newly added samples are no longer just from different domains within the same category, but rather new defect types that did not exist in the historical tasks. Therefore, they exhibit both a significant distribution shift and a greater semantic distance. Hence, this step switches the task to the class incremental branch.
[0095] Step 3: Construct a dynamic subspace for newly added categories. First, cluster historical features to obtain cluster prototypes. Then calculate the initial embedding for the newly added task samples.
[0096]
[0097] And calculate its cosine similarity with each cluster prototype.
[0098]
[0099] The purpose of this step is not to forcibly embed the new class features directly into the existing decision boundary, but to first find several historical clusters that are closest to it, thereby constructing a locally accommodating dynamic subspace, providing a stable reference for subsequent new class mapping.
[0100] Step 4: Obtain the local linear transformation matrix within the dynamic subspace.
[0101]
[0102] And obtain the new class embedding after mapping.
[0103]
[0104] The first equation constrains the smooth embedding of new categories within a local subspace, while the second equation provides the final feature representation of the new class. The core of this approach lies in reducing disruption to the existing class structure through local mapping rather than global rewriting, thereby mitigating catastrophic forgetting.
[0105] Step 5: Perform incremental adaptive weight updates. First, use JS divergence to measure the distribution of historical tasks. With current task distribution Differences:
[0106]
[0107] Then determine the update step size based on this.
[0108]
[0109] And complete The weights are adjusted accordingly. JS divergence is used to characterize the distribution difference between the newly added task and historical tasks, and the weight update intensity is adjusted based on this to suppress excessive perturbation of the decision boundary of historical categories when introducing new categories.
[0110] Step Six: Output the recognition results and complete the verification of the embodiment. In Task 1, the model's recognition accuracy for Types 1 and 2 was close to 100%; in Task 2, after introducing Types 3 and 4, the overall recognition accuracy still reached 98.78%; in Task 3, after further introducing Types 5 and 6, the recognition accuracy reached 97.69%; and in Task 4, after further introducing Types 7 and 8, the model's average accuracy remained at 95.61%. This demonstrates that this embodiment can gradually incorporate new defect types without retraining the overall model, while maintaining a high recognition capability for historical categories. Figure 4 This is a comparison chart of feature distribution and recognition performance at different stages of incremental learning in an incremental learning embodiment of the present invention. Figure 4 (a) is a scatter plot of the features of Task1 stage. Figure 4 (b) is the confusion matrix diagram for Task 4; Table 2 below summarizes the key results of incremental learning.
[0111] Table 2: Summary of Key Results of Incremental Learning
[0112]
[0113] As can be seen from the above two embodiments, the present invention can perform domain incremental learning for partial discharge data of the same type but different domains, and can also perform class incremental learning for newly added defect types, thereby realizing the continuous updating and intelligent evolution of the cable partial discharge identification model.
[0114] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.
Claims
1. A cable partial discharge incremental learning method based on graph convolutional neural network, characterized in that, Includes the following steps: S1. Obtain historical cable partial discharge data and corresponding type labels, and pre-train the graph convolutional neural network basic model based on the historical cable partial discharge data to obtain the pre-trained model. S2. Freeze the encoder parameters in the pre-trained model, input historical cable partial discharge data into the encoder to extract historical features, and simultaneously input new input partial discharge data into the encoder to extract new input features; S3. Based on the distribution difference and feature distance between the historical features and the new input features, determine the incremental learning task type corresponding to the new input partial discharge data as a class incremental learning task, a domain incremental learning task, or a task that does not require incremental learning. S4. When it is determined to be an incremental learning task, a dynamic subspace is constructed based on historical features, and the embedded features corresponding to the new input partial discharge data are mapped to the dynamic subspace to realize the feature integration of the new partial discharge type. S5. When it is determined to be a domain incremental learning task, a dynamic subspace is constructed based on the source domain features and the target domain features respectively. The embedded features corresponding to the new input partial discharge data are mapped to the target domain subspace through adversarial dynamic mapping to achieve cross-domain feature alignment. S6. Based on the distribution differences corresponding to the class incremental learning task or the domain incremental learning task, adaptively determine the model weight update step size, and use the new input partial discharge data to iteratively update the model weights. S7. Use the updated model to output the cable partial discharge type identification results.
2. The cable partial discharge delta learning method based on graph convolutional neural network according to claim 1, wherein, In step S1, the basic model of the graph convolutional neural network is a residual graph convolutional neural network (ResGCN). The pre-training process uses an optimization algorithm for training and updates the model parameters with a cross-entropy loss function. The output of the basic model consists of an encoder, a flattening operation, and a classifier. After training, the encoder weights are frozen so that they can be used as a fixed feature extractor in subsequent incremental learning.
3. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S3, the determination of the incremental learning task type specifically involves: extracting features from historical partial discharge data using the frozen encoder. Features of newly input partial discharge data The maximum mean difference (MMD) metric was used. and The characteristic distribution differences are analyzed, and the average Euclidean distance is used. measure and Feature distance; Set distribution difference threshold and distance threshold. When MMD is greater than the preset threshold and When the distance exceeds a preset threshold, it is classified as an incremental learning task; when the MMD exceeds a preset threshold and When the distance is less than the preset distance threshold, it is determined to be a domain incremental learning task; when neither exceeds the corresponding threshold, it is determined that no incremental learning task is required.
4. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S4, constructing a dynamic subspace based on historical features includes: dividing the historical features into... Each cluster is represented by its mean vector; the new input partial discharge data is encoded and flattened to obtain the initial embedding vector. ; Calculate the initial embedding vector The cosine similarity between the vector and the representative vector of each cluster is used to select the vector with the highest similarity. Each cluster constructs a dynamic subspace associated with the new input partial discharge data.
5. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S4, the embedding features corresponding to the newly input partial discharge data are mapped to the dynamic subspace. Specifically, within the dynamic subspace, the initial embedding vector is... Perform a local linear transformation with regularization terms and solve for the transformation matrix. The transformed embedding vector is obtained so that the features of the new input partial discharge data can be better integrated into the existing type system, while reducing the damage to the feature structure of the old category.
6. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S5, dynamic subspaces are constructed based on source domain features and target domain features, respectively, and cross-domain feature alignment is achieved through adversarial dynamic mapping. This includes: clustering source domain features and target domain features respectively to obtain source domain cluster representative vectors and target domain cluster representative vectors; and obtaining initial embedding vectors by encoding and flattening new input partial discharge data. Calculate the cosine similarity between it and the representative vectors of each cluster in the source domain; set up a domain discriminator. and mapping function Using adversarial learning to transform the initial embedding vector Map to the target domain subspace; by minimizing the distance between the mapped features and the corresponding cluster representative vector in the target domain, the domain-adapted embedded features are obtained, achieving cross-domain feature alignment and domain invariance enhancement.
7. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S6, when the incremental learning task type is a class-incremental learning task, the JS divergence between the historical task data features and the new task data features is used to represent the distribution difference between the old and new tasks, and the weight update step size corresponding to the class-incremental learning is calculated based on the JS divergence; then, the loss function and its gradient with respect to the model weights are calculated based on the new task data to complete the model weight update.
8. The cable partial discharge incremental learning method based on graph convolutional neural network according to claim 1, characterized in that, In step S6, when the incremental learning task type is a domain incremental learning task, the loss function of the model in the source domain and the target domain is calculated through the feedback information of the domain discriminator, and the difference between the two is calculated; combined with the difference and the maximum mean difference (MMD) between the source domain features and the target domain features, the weight update step size corresponding to the domain incremental learning is calculated; then, based on the target domain data, the gradient of the loss function with respect to the model weights is calculated to complete the model weight update.
9. A cable partial discharge incremental learning system based on graph convolutional neural networks, characterized in that, include: The basic model pre-training module is used to pre-train the graph convolutional neural network basic model based on historical cable partial discharge data and freeze the encoder parameters. The feature extraction module is used to extract features from historical partial discharge data and features from newly input partial discharge data. The task type discrimination module is used to determine the incremental learning task type based on the distribution differences and feature distance between historical features and new input features; The incremental mapping module is used to construct a dynamic subspace and map the new input partial discharge data to the corresponding embedded features when the task is determined to be an incremental learning task. The domain incremental mapping module is used to construct source domain subspace and target domain subspace when the task is determined to be a domain incremental learning task, and to achieve cross-domain feature alignment through adversarial dynamic mapping. The adaptive weight update module is used to adaptively determine the model weight update step size and update the model based on the difference information corresponding to the class incremental learning task or the domain incremental learning task. The identification output module is used to output the identification results of the partial discharge type of the cable.
10. The system according to claim 9, characterized in that, The task type discrimination module uses a combination of maximum mean difference (MMD) and average Euclidean distance to determine the incremental type; the class incremental mapping module uses a combination of clustering, cosine similarity filtering, and local linear transformation to map the new category features; the domain incremental mapping module uses a combination of source / target domain bi-subspace construction and adversarial learning of the domain discriminator to complete cross-domain mapping; and the adaptive weight update module determines the weight update step size based on either JS divergence or a joint metric of loss difference and MMD.