Switch cabinet incremental partial discharge mode identification method based on adaptive clustering distillation

Abstract

The invention relates to a switch cabinet incremental partial discharge mode identification method based on adaptive clustering distillation. The method comprises the following steps: acquiring a historical partial discharge sample set to train a partial discharge mode identification model; performing density peak detection on each historical sample in the historical partial discharge sample set, and adaptively selecting an initial clustering center and the number of clustering centers; carrying out clustering operation on the historical partial discharge sample set by adopting a clustering method, obtaining playback samples and calculating training weights of the playback samples; weighting the distillation loss function of the historical sample according to the training weight of the playback sample to obtain a weighted distillation loss function of the playback sample; and according to the weighted distillation loss function and the training loss function of the newly-added sample, obtaining comprehensive loss to learn the newly-added sample, updating the partial discharge mode recognition model, classifying the newly-added sample into a historical training set, and carrying out fault diagnosis by adopting the updated partial discharge mode recognition model. According to the invention, the recognition accuracy of the partial discharge mode of the switch cabinet can be improved under the condition of low computing resources.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostics and condition monitoring technology for switchgear, and in particular to a method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation. Background Technology

[0002] Due to factors such as overheating, insulation aging, and mechanical wear, some live parts of switchgear often experience exposure problems, leading to partial discharge. Partial discharge damages the switchgear insulation, and gradually progresses to complete insulation failure, causing a breakdown accident and posing a significant threat to equipment and personnel safety. Discharge modes such as point discharge, floating discharge, internal discharge, and surface discharge have different degrees of impact on equipment insulation and occur in different locations. Identifying the partial discharge mode of switchgear can provide important reference information for operation and maintenance. Performing partial discharge mode identification for switchgear is crucial for ensuring the safe and stable operation of switchgear.

[0003] Partial discharge is a process of partial insulation ionization and recombination caused by local electric field distortion. Its discharge model, discharge frequency, and single discharge magnitude are directly affected by the electromagnetic environment. The diverse factors contributing to partial discharge, along with the varied shapes and electrical parameters of insulation defects, result in diverse electromagnetic environments for partial discharge. Consequently, the characteristics of some real-world partial discharge samples differ significantly from those of experimentally collected samples. This leads to the artificial intelligence diagnostic model trained on the experimental partial discharge sample training set no longer being suitable, resulting in a substantial decrease in the accuracy of partial discharge pattern recognition.

[0004] To address this issue, the commonly used approach is to learn new partial discharge faults using incremental learning methods. For example, Zhang Yi et al. used knowledge distillation incremental learning to learn new topological features of partial discharges, while Wang Chunxin et al. proposed a partial discharge incremental fault diagnosis method based on lossless estimation and balanced training. These two methods are typical representatives of the playback method and the generation method, respectively. The playback method is characterized by low computational resource consumption but insufficient accuracy, while the generation method is the opposite. Since the cost of switchgear is relatively low, the corresponding computing resources of the monitoring host computer are also small, so the learning method for new partial discharge faults does not easily consume too many resources.

[0005] In existing technologies, the main difficulty in learning new partial discharge faults lies in the selection of historical samples for playback. Playback-based incremental learning methods, such as knowledge distillation, select historical samples that are not representative enough, leading to catastrophic forgetting of historical features and a severe drop in accuracy. While generative incremental learning methods, such as generative replay, can select more representative historical samples, they consume significant computational resources. Therefore, a new method for learning new partial discharge faults is needed to select representative historical samples with low computational cost, minimizing computational resource requirements while maintaining recognition accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an incremental partial discharge pattern recognition method for switchgear based on adaptive clustering distillation. By analyzing the distribution density of historical samples and calculating the weights of replay samples, the model can quickly learn new faults and memorize historical samples without relying on all historical samples.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A method for incremental partial discharge pattern recognition in switchgear based on adaptive clustering distillation includes:

[0009] The partial discharge pattern recognition model is trained by acquiring a historical partial discharge sample set;

[0010] Density peak detection is performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers are adaptively selected.

[0011] Based on the initial cluster centers and the number of cluster centers, a clustering method is used to perform clustering operations on the historical partial discharge sample set to obtain replay samples and calculate the training weights of the replay samples.

[0012] The distillation loss function of the historical samples is weighted according to the training weights of the replay samples to obtain the weighted distillation loss function of the replay samples.

[0013] Based on the weighted distillation loss function and the training loss function of the new samples, a comprehensive loss is obtained to learn the new samples, the partial discharge pattern recognition model is updated, the new samples are classified into the historical training set, and the updated partial discharge pattern recognition model is used for fault diagnosis.

[0014] Optionally, obtaining the historical partial discharge sample set includes: using two methods, namely, extracting historical operation datasets of switchgear and conducting partial discharge experiments, to establish the historical partial discharge sample set.

[0015] Optionally, density peak detection is performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers are adaptively selected, including:

[0016] Calculate the density peak value of each historical sample, select the local extreme points of the density peak value, and use the selected local extreme points of the density peak value and the number of them as the initial cluster centers and the number of cluster centers, respectively.

[0017] Optionally, clustering the historical partial discharge sample set using a clustering method to obtain replay samples and calculate the training weights for the replay samples includes:

[0018] Using Euclidean distance as the decision variable, clustering operations are performed on the historical partial discharge sample set to obtain cluster centers and cluster sets;

[0019] The cluster centers are used as the replay samples, and the ratio of the training gradient length of the samples in each cluster to the training gradient length of the replay sample set is used as the training weight of the replay samples. The training gradient of each sample point is calculated using the original pattern recognition model.

[0020] Optionally, the training weights for the replay samples are:

[0021] ;

[0022] in, For replay samples The training weights; For replay samples The training gradient; For sample points The training gradient; Modulo operation; Let c be the c-th cluster set.

[0023] Optionally, the weighted distillation loss function of the replay sample for:

[0024] ;

[0025] in, Let be the training loss function for the pattern recognition model. For new pattern recognition models after online learning, This is the original pattern recognition model, where c is the c-th cluster set. The number of cluster centers. It serves as the cluster center.

[0026] Optionally, the training loss function of the newly added samples for:

[0027] ;

[0028] in, For new samples The tag.

[0029] The beneficial effects of this invention are as follows: Addressing the dynamic changes in data distribution during switchgear operation and the catastrophic forgetting problem inherent in traditional models, this invention proposes an incremental partial discharge pattern recognition method for switchgear based on adaptive clustering distillation. Through historical sample distribution density analysis and replay sample weight calculation, this invention achieves rapid learning of new faults and memorization of historical samples without relying on the entire historical sample set, demonstrating significant technical and economic advantages. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of an incremental partial discharge pattern recognition method for switchgear based on adaptive clustering distillation, according to an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, this embodiment proposes a method for incremental partial discharge pattern recognition in switchgear based on adaptive clustering distillation, including:

[0035] The partial discharge pattern recognition model is trained by acquiring a historical partial discharge sample set;

[0036] Density peak detection is performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers are adaptively selected.

[0037] Based on the initial cluster centers and the number of cluster centers, a clustering method is used to perform clustering operations on the historical partial release sample set to obtain replay samples and calculate the training weights of the replay samples.

[0038] The distillation loss function of historical samples is weighted according to the training weights of the replay samples to obtain the weighted distillation loss function of the replay samples.

[0039] Based on the weighted distillation loss function and the training loss function of the new samples, the comprehensive loss is obtained to learn the new samples, update the partial discharge pattern recognition model, classify the new samples into the historical training set, and use the updated partial discharge pattern recognition model for fault diagnosis.

[0040] Furthermore, obtaining the historical partial discharge sample set includes two methods: extracting historical operation datasets of switchgear and conducting partial discharge experiments to establish the historical partial discharge sample set.

[0041] Furthermore, density peak detection is performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers are adaptively selected, including:

[0042] Calculate the density peak value of each historical sample, select the local extreme points of the density peak value, and use the selected local extreme points of the density peak value and the number of them as the initial cluster centers and the number of cluster centers, respectively.

[0043] Specifically, density peak detection is performed on each historical sample, and the sample distribution density around each sample is detected. Local extreme points of density peaks are selected as the initial cluster centers of the historical sample training set, and the number of clusters is consistent with the detected local extreme points of density peaks.

[0044] Calculate the density peak value for each historical sample, select local extreme points of the density peak value, and use these local extreme points and their number as the initial cluster centers and the number of cluster centers, respectively. The density peak value calculation method is as follows:

[0045] (1);

[0046] In the formula, This represents the density peak. The number of regions to be divided is set to 10; The i-th annular region is defined by the sample as the center and the distance to the farthest sample as the radius. The distance is calculated using Euclidean distance. For the region Number of samples within, For the region The outer diameter.

[0047] The local extreme points of the density peak are determined by the following formula:

[0048] (2);

[0049] In the formula, Local extreme points The density peak; For sample points The density peak; For local extreme points The i-th annular region is centered at a point on the radius, with the furthest sample distance being the distance divided equally by the radius.

[0050] The local extreme points of the density peak represent the densest distribution of samples in each cluster set, close to the cluster center of each cluster set. By taking each local extreme point as the cluster center point and their number as the number of clusters, the initial cluster center and the number of cluster centers can be adaptively selected.

[0051] Furthermore, clustering methods are used to perform clustering operations on the historical partial discharge sample set to obtain replay samples and calculate the training weights for the replay samples, including:

[0052] Using Euclidean distance as the decision variable, clustering operations are performed on the historical partial discharge sample set to obtain the cluster centers and cluster sets;

[0053] Cluster centers are used as replay samples, and the ratio of the training gradient length of samples in each cluster to the training gradient length of the replay sample set is used as the training weight of the replay samples. The training gradient of each sample point is calculated using the original pattern recognition model.

[0054] Specifically, Euclidean distance is used as the decision variable for clustering. That is:

[0055] (3);

[0056] In the formula, Let c be the c-th cluster set; The number of cluster centers; Calculated for Euclidean distance; Let be the cluster center of the j-th cluster set.

[0057] After clustering each sample, the cluster centers are reset using the centroids of each cluster space, i.e.:

[0058] (4);

[0059] In the formula, Let be the total number of samples in the c-th cluster set.

[0060] Repeat equations (3) and (4) until convergence. Then, use each cluster center as a replay sample. The training weights of each replay sample are calculated as follows:

[0061] (5);

[0062] In the formula, For replay samples The training weights; For replay samples The training gradient; For sample points The training gradient; Modulo operation.

[0063] The training gradient of each sample point is calculated using the original pattern recognition model, and the training weight of each playback sample point can be calculated according to Equation (5).

[0064] Furthermore, to ensure that the training gradient generated by the replay samples approximates the training gradient of the complete historical sample set as closely as possible, thereby enhancing the model's memory of historical samples when learning new fault samples, this paper, based on the knowledge distillation incremental learning method, weights the distillation loss function of historical samples and applies a weighted processing to the distillation loss function of the replay sample training. for:

[0065] (6);

[0066] in, Let be the training loss function for the pattern recognition model. For new pattern recognition models after online learning, This is the original pattern recognition model, where c is the c-th cluster set. The number of cluster centers. It serves as the cluster center.

[0067] The pattern recognition model used in this embodiment is the repVGG model. Furthermore, for newly added samples that require relearning, the training loss... for:

[0068] (7);

[0069] in, For new samples The tag.

[0070] Furthermore, once a certain number of new samples have accumulated, the model is trained on these new samples. A weighted distillation loss function constrains the model's learning process, ensuring that the training gradients generated by the replay samples closely approximate the gradient distribution of the complete historical sample set, thus preventing the model from forgetting historical knowledge when learning new knowledge. The loss function expression is:

[0071] (8);

[0072] in, The training loss function for the new samples. The weighted distillation loss function for the replay samples. The linear combination weight between the two training functions is set to 0.5, and the weight coefficient is based on equation (5).

[0073] Static relationship: The original model parameters are used as a frozen layer to generate old knowledge distillation output; the new model updates parameters under the same architecture.

[0074] Dynamic relationship: During the training process, the model receives "new samples + replay samples" in batches and updates the parameters in real time to minimize the overall loss.

[0075] Effects: Enables adaptive learning and knowledge retention of the model, ensuring that it maintains a high ability to recognize historical samples even after learning from new fault samples.

[0076] Once the online learning process converges, the updated model parameters replace the original pattern recognition model, and the newly added samples are incorporated into the historical training set. The updated pattern recognition model is then put into operation to perform partial discharge pattern recognition on the newly collected switchgear samples.

[0077] Static relationship: The model update module connects to the diagnostic server and replaces the old model files through the API interface.

[0078] Dynamic relationship: After the update operation is completed, the system automatically refreshes the threshold parameters and starts the next round of real-time monitoring.

[0079] Effects: Enables continuous self-evolution and long-term stable operation of the switchgear partial discharge pattern recognition system, improving the accuracy and response speed of fault detection.

[0080] To verify the effectiveness of the proposed model, operational sample data from a 35kV switchgear was used for validation. After five rounds of online learning, the model achieved an average diagnostic accuracy of over 99%, reduced the catastrophic forgetting rate to below 0.81%, and shortened the training time by approximately 62% compared to the full retraining method. Furthermore, the model maintained stable performance during 120 days of continuous operation in the field. This method can significantly improve the intelligence level of switchgear operation and maintenance and the reliability of power supply.

[0081] In this embodiment, comparative experiments were conducted on a sample set of partial discharge and mechanical faults in switchgear, with similar technologies including joint training, fine-tuning, EWC (elastic weight solidification), and LwF (standard knowledge distillation). The performance comparison results of each method under the same sample conditions are shown in Table 1.

[0082] Table 1

[0083]

[0084] As shown in Table 1, the diagnostic accuracy of the method in this embodiment is 98.31%, the catastrophic forgetting rate is 1.34%, the training time is reduced by 64.17%, the pattern recognition accuracy is significantly higher than other methods, and the training resources required are also higher than most methods. Its performance is significantly better than existing incremental partial discharge pattern recognition algorithms. While ensuring good model accuracy and memory rate of historical samples, it significantly reduces the consumption of computing and storage resources.

[0085] From a technical perspective, this embodiment innovatively introduces an adaptive clustering and gradient-weighted training strategy for historical partial discharge samples based on the traditional knowledge distillation framework. This achieves adaptive fitting of the gradient distribution of historical samples, enabling the model to simultaneously learn new fault knowledge and retain historical knowledge during incremental updates. Using this method, the model achieves near-full retraining performance with less than 10% of the total sample training load. Verification using measured partial discharge data from switchgear shows that the online model update cycle is shortened by approximately 67%, and the model redeployment time is reduced from 2 days to less than 16 hours, significantly improving the system's response speed and intelligence.

[0086] From an economic perspective, the method in this embodiment eliminates the need for full-sample retraining, significantly reducing system operation and maintenance costs and model upkeep. In practical applications of switchgear in typical 35 kV substations, this method reduces model maintenance frequency by approximately 70% and fault misjudgment rate by approximately 74%, effectively mitigating economic losses caused by unplanned power outages. Based on average grid outage losses, this method can save a single substation an average of annual operation and maintenance costs, demonstrating significant economic benefits.

[0087] From a social benefit perspective, the method described in this embodiment provides a new intelligent diagnostic model for power systems, possessing the capabilities of autonomous learning, rapid adaptation, and self-evolution. This technology enables cross-site and cross-regional model sharing and collaborative updates in the operational monitoring of large-scale switchgear clusters, providing key technical support for building a "smart grid—self-learning sensing system." Through widespread application, this method is expected to increase the coverage of switchgear health monitoring by more than 30%, which has profound significance for the intelligent and digital transformation of the power grid.

[0088] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for incremental partial discharge pattern recognition in switchgear based on adaptive clustering distillation, characterized in that, include: The partial discharge pattern recognition model is trained by acquiring a historical partial discharge sample set; Density peak detection is performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers are adaptively selected. Based on the initial cluster centers and the number of cluster centers, a clustering method is used to perform clustering operations on the historical partial discharge sample set to obtain replay samples and calculate the training weights of the replay samples. The distillation loss function of the historical samples is weighted according to the training weights of the replay samples to obtain the weighted distillation loss function of the replay samples. Based on the weighted distillation loss function and the training loss function of the new samples, a comprehensive loss is obtained to learn the new samples, the partial discharge pattern recognition model is updated, the new samples are classified into the historical training set, and the updated partial discharge pattern recognition model is used for fault diagnosis.

2. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 1, characterized in that, Obtaining the historical partial discharge sample set includes two methods: extracting historical operation datasets of switchgear and conducting partial discharge experiments.

3. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 1, characterized in that, Density peak detection was performed on each historical sample in the historical partial emission sample set, and the initial cluster centers and the number of cluster centers were adaptively selected, including: Calculate the density peak value of each historical sample, select the local extreme points of the density peak value, and use the selected local extreme points of the density peak value and the number of them as the initial cluster centers and the number of cluster centers, respectively.

4. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 1, characterized in that, Clustering is performed on the historical partial discharge sample set to obtain replay samples and calculate the training weights for the replay samples, including: Using Euclidean distance as the decision variable, clustering operations are performed on the historical partial discharge sample set to obtain cluster centers and cluster sets; The cluster centers are used as the replay samples, and the ratio of the training gradient length of the samples in each cluster to the training gradient length of the replay sample set is used as the training weight of the replay samples. The training gradient of each sample point is calculated using the original pattern recognition model.

5. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 4, characterized in that, The training weights for the replay samples are: ； in, For replay samples The training weights; For replay samples The training gradient; For sample points The training gradient; Modulo operation; Let c be the c-th cluster set.

6. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 5, characterized in that, The weighted distillation loss function of the replay sample for: ； in, Let be the training loss function for the pattern recognition model. For new pattern recognition models after online learning, This is the original pattern recognition model, where c is the c-th cluster set. The number of cluster centers. It serves as the cluster center.

7. The method for incremental partial discharge pattern recognition of switchgear based on adaptive clustering distillation according to claim 6, characterized in that, The training loss function of the newly added samples for: ； in, For new samples The tag.

Images