Partial discharge identification optimization method of gas insulated enclosed switchgear for small sample set

Abstract

The invention provides a method for identifying and optimizing partial discharge of a gas insulated switchgear for a small sample set, which comprises the following steps of: S1, acquiring original data when the partial discharge phenomenon of GIS (Gas Insulated Switchgear) occurs, and converting the original data into a PRPD atlas; s2, screening out high-quality PRPD maps, carrying out classification labeling, and constructing an A-class training set and an A-class test set according to a certain proportion after preprocessing; s3, applying a K-nearest neighbor interpolation method in the A-class training set to generate new PRPD atlas data, constructing a B-class sample set, and constructing a C-class training set by using the B-class sample set and the A-class training set; s4, training the C-class training set by using a ViT model to obtain an optimized recognition model; s5, testing the trained ViT model by using the A-class test set to carry out identification and classification validity evaluation; and S6, verifying the superiority of the method by comparing the training effects of the A-class training set and the B-class training set. The new sample is generated by using the local feature information of the known sample, and the training set scale is effectively expanded, so that the learning ability and accuracy of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge detection of gas insulated switchgear (hereinafter referred to as GIS), and in particular to a GIS partial discharge identification optimization method for a small sample set. Background Art

[0002] In the operation of power equipment, partial discharge (PD) of GIS is a key factor affecting its insulation performance and safety. Effectively identifying and monitoring partial discharge phenomena is essential to ensure the reliability of equipment.

[0003] With the development of machine learning and deep learning, the current mainstream partial discharge recognition technology mainly builds sample sets and uses different models for training according to actual requirements. However, the current partial discharge recognition technology still faces many challenges under the condition of small sample data. During training, it usually needs to rely on a large number of labeled samples to ensure the model training effect, resulting in insufficient model training and reduced recognition accuracy under small sample conditions.

[0004] In this case, the generalization ability of the model is also limited, and it is difficult to effectively adapt to the new data distribution, thereby reducing the recognition rate and classification effect of partial discharge. In addition, the imbalance in the number of various partial discharge samples exacerbates this phenomenon, resulting in low recognition accuracy of certain partial discharge types, and the overall performance of the model does not meet the requirements of actual engineering applications. Summary of the invention

[0005] In order to solve the current technical problems, the main purpose of the present invention is to provide a gas-insulated enclosed switchgear partial discharge identification optimization method for a small sample set. This optimization method can well solve the problems of small sample size and unbalanced number of various samples. In order to achieve the above technical features, the purpose of the present invention is achieved as follows: a method for optimizing partial discharge identification of gas-insulated enclosed switchgear for a small sample set comprises the following steps: S1. Collect the original data when partial discharge occurs in GIS equipment and convert it into PRPD spectrum; S2, screen out high-quality PRPD maps and classify and annotate them, and after preprocessing, construct a class A training set and a class A test set according to a certain ratio; S3, applying K-nearest neighbor interpolation method in the A-type training set to generate new PRPD atlas data, construct the B-type sample set, and combine it with the A-type training set to construct the C-type training set; S4, use the ViT model to train the C-type training set to obtain an optimized recognition model; S5. Use the A-type test set to test the trained ViT model to evaluate the effectiveness of recognition and classification; S6. By comparing the training results of the A and B training sets, the superiority of this method is verified.

[0006] Preferably, in step S1, a UHF sensor is used to collect partial discharge phenomena of GIS equipment.

[0007] Preferably, the original data collected in step S1 when the GIS equipment has a partial discharge phenomenon includes: suspended discharge, tip corona discharge, insulating discharge, and free particle discharge.

[0008] Preferably, the PRPD spectrum in step S1 is formed into a two-dimensional spectrum by combining the local discharge pulse with the power frequency phase, wherein one coordinate axis represents the power frequency phase φ and the other coordinate axis represents the discharge amount q; This kind of map can clearly show the discharge intensity and occurrence number n in different phases, and then identify and classify the type of partial discharge and evaluate the insulation status of electrical equipment.

[0009] Preferably, the step S2 specifically includes: S201, manually remove unclear or erroneous maps and retain clear and high-quality PRPD maps; S202, preprocessing the image and normalizing it into a grayscale image of uniform size; S203. Construct a class A training set and a class A test set according to a certain ratio.

[0010] Preferably, in step S203, after preprocessing, a class A training set and a class A test set are constructed in a certain ratio, wherein the ratio of class A training set to class A test set is 8:2.

[0011] Preferably, applying the K nearest neighbor interpolation method in the class A training set in step S3 includes: S301, select a PRPD spectrum as a sample, and record the sample as A(x,y) ,in x and y Represent the rows and columns of the image (i.e. the spatial coordinates of the image). This sample is the reference atlas, which is used to generate new samples including; S302, calculating the K nearest neighbor samples of the selected sample, using the Euclidean distance in the K nearest neighbor algorithm for calculation; For each sample A i (x,y) , calculate its difference with the sample A(x,y) By calculating the distance of each sample, we can find the K neighbor samples with the smallest distance, which are recorded as A n1 (x,y) , A n2 (x,y) ,……, A nK (x,y) , the distance formula is as follows: ; in, d(i,s) Representation sample A i (x,y) With the benchmark sample A(x,y) The distance between A(x,y) is the pixel value of the selected benchmark sample; A i (x,y) is the pixel value of the current sample; S303, generate new PRPD map data by interpolating the selected sample and its neighbors; generate new samples by weighted average method B(x,y) , and its calculation formula is as follows: ; in: B(x,y) is the pixel value of the generated new PRPD map; A nj (x,y) It is j The pixel value of neighbor samples; K is the number of neighbors selected; Weight w j According to the definition of the distance between the neighbor and the benchmark sample, it is usually set to the inverse of the distance. The specific calculation formula is as follows: ; in: w j It is j The weight of the neighbors; d(nj,s) Indicates j Neighborhood samples A nj (x,y) With the benchmark sample A (x,y) The distance between is a small constant that prevents the denominator from being zero; S304: Combine the A-type training set and the B-type sample set to construct a C-type training set.

[0012] Preferably, when generating new PRPD atlas data in step S3, the number of samples actually generated is dynamically adjusted according to the number of sample sets of the four partial discharge types to ensure that the final number of samples of each type is balanced; for types with a small number of samples, more samples are generated; and for types with a large number of samples, the number of samples generated is reduced.

[0013] Preferably, in step S4, the ViT model, which stands for Vision Transformer, is an emerging deep learning model that applies the Transformer architecture, which has been very successful in the field of natural language processing, to computer vision tasks.

[0014] Preferably, in step S4, the ViT model is used to train the C-type training set to obtain an optimized recognition model. The training process includes selecting appropriate hyperparameters to improve the diversity of training data, and performing multiple rounds of iterative training to ensure that the final model has sufficient adaptability and recognition effect.

[0015] Preferably, in step S5, the trained ViT model is tested with a Class A test set to evaluate the effectiveness of recognition and classification, specifically including generating a loss curve and an accuracy curve by training the Class A test set to visualize performance changes during the training process and analyze the recognition effect of the model on different partial discharge types.

[0016] Preferably, the step S6 specifically includes: S601, putting the A and B training sets into the ViT model for training; S602, using the class A test set to test the two ViT models trained in step S601 respectively; S603, taking the best accuracy of each test as the best model and comparing them to verify the superiority of this method.

[0017] The present invention has the following beneficial effects: The present invention provides a GIS partial discharge identification optimization method for a small sample set. The method generates new samples by utilizing the local feature information of known samples, effectively expanding the scale of the training set, thereby improving the learning ability and accuracy of the model. At the same time, the method can also balance the number of various samples, avoid model bias, accuracy misleading and overfitting caused by sample imbalance, ensure the model's recognition ability on various samples is more balanced, and ensure the reliability and effectiveness of the model in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0019] Figure 1 It is a flow chart of a GIS partial discharge identification optimization method for a small sample set disclosed in an embodiment of the present invention.

[0020] Figure 2 These are four typical PRPD grayscale maps after preprocessing disclosed in the embodiments of the present invention.

[0021] Figure 3It is a new PRPD spectrum generated by using the K nearest neighbor interpolation method disclosed in an embodiment of the present invention.

[0022] Figure 4 These are the loss curve and accuracy curve disclosed in the embodiments of the present invention.

[0023] Figure 5 It is a comparison chart of model effects under different data set training disclosed in the embodiments of the present invention. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with embodiments. It is necessary to point out that the following embodiments are only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Some non-essential improvements and adjustments made by technical personnel in the field based on the above invention content still fall within the scope of protection of the present invention.

[0025] Embodiment 1: See also Figure 1-5 , a partial discharge identification optimization method for gas-insulated enclosed switchgear for a small sample set, comprising the following steps: S1. Collect the original data when partial discharge occurs in GIS equipment and convert it into PRPD spectrum; Wherein, in step S1, a UHF sensor is used to collect partial discharge phenomena of GIS equipment.

[0026] The original data collected in step S1 when the GIS equipment has a partial discharge phenomenon includes: suspended discharge, tip corona discharge, insulating discharge, and free particle discharge.

[0027] The PRPD spectrum in step S1 combines the local discharge pulse with the power frequency phase to form a two-dimensional spectrum, in which one coordinate axis represents the power frequency phase φ and the other coordinate axis represents the discharge amount q; this spectrum can clearly show the discharge intensity and the number of occurrences n under different phases, thereby identifying and classifying the local discharge type and evaluating the insulation status of the electrical equipment.

[0028] S2. Screen out high-quality PRPD maps and classify and annotate them. After preprocessing, construct a Class A training set and a Class A test set according to a certain ratio; specifically: S201, manually remove unclear or erroneous maps and retain clear and high-quality PRPD maps; S202, preprocessing the image and normalizing it into a grayscale image of uniform size; S203. Construct a class A training set and a class A test set according to a certain ratio.

[0029] Preferably, in step S203, after preprocessing, a class A training set and a class A test set are constructed in a certain ratio, wherein the ratio of class A training set to class A test set is 8:2.

[0030] S3. Apply K-nearest neighbor interpolation method to the A-type training set to generate new PRPD atlas data, construct the B-type sample set, and combine it with the A-type training set to construct the C-type training set; specifically: S301, select a PRPD spectrum as a sample, and record the sample as A(x,y) ,in x and y Represent the rows and columns of the image (i.e. the spatial coordinates of the image). This sample is the reference atlas, which is used to generate new samples including; S302, calculating the K nearest neighbor samples of the selected sample, using the Euclidean distance in the K nearest neighbor algorithm for calculation; For each sample A i (x,y) , calculate its difference with the sample A(x,y) By calculating the distance of each sample, we can find the K neighbor samples with the smallest distance, which are recorded as A n1 (x,y) , A n2 (x,y) ,……, A nK (x,y) , the distance formula is as follows: ; in, d(i,s) Representation sample A i (x,y) With the benchmark sample A(x,y) The distance between A(x,y) is the pixel value of the selected benchmark sample; A i (x,y) is the pixel value of the current sample; S303, generate new PRPD map data by interpolating the selected sample and its neighbors; generate new samples by weighted average method B(x,y) , and its calculation formula is as follows: ; in: B(x,y) is the pixel value of the generated new PRPD map; A nj (x,y) It is j The pixel value of neighbor samples; K is the number of neighbors selected; Weight w j According to the definition of the distance between the neighbor and the benchmark sample, it is usually set to the inverse of the distance. The specific calculation formula is as follows: ; in: w j It is j The weight of the neighbors; d(nj,s) Indicates j Neighborhood samples A nj (x,y) With the benchmark sample A (x,y) The distance between is a small constant that prevents the denominator from being zero; S304: Combine the A-type training set and the B-type sample set to construct a C-type training set.

[0031] When generating new PRPD atlas data in step S3, the number of samples actually generated is dynamically adjusted according to the number of sample sets of the four partial discharge types to ensure that the final number of samples of each type is balanced; for types with a small number of samples, more samples are generated; and for types with a large number of samples, the number of samples generated is reduced.

[0032] S4, use the ViT model to train the C-type training set to obtain an optimized recognition model; Among them, in the step S4, the ViT model, which stands for Vision Transformer, is an emerging deep learning model that applies the Transformer architecture, which has been very successful in the field of natural language processing, to computer vision tasks.

[0033] In step S4, the ViT model is used to train the C-type training set to obtain an optimized recognition model. The training process includes selecting appropriate hyperparameters to improve the diversity of training data and performing multiple rounds of iterative training to ensure that the final model has sufficient adaptability and recognition effect.

[0034] S5. Use the A-type test set to test the trained ViT model to evaluate the effectiveness of recognition and classification; Among them, in the step S5, the trained ViT model is tested with the Class A test set to evaluate the effectiveness of recognition and classification, specifically including the loss curve and accuracy curve generated by training the Class A test set to visualize the performance changes during the training process and analyze the recognition effect of the model on different partial discharge types.

[0035] S6. Verify the superiority of this method by comparing the training results of the A and B training sets; specifically including: S601, putting the A and B training sets into the ViT model for training; S602, using the class A test set to test the two ViT models trained in step S601 respectively; S603, taking the best accuracy of each test as the best model and comparing them to verify the superiority of this method.

[0036] Embodiment 2: like Figure 1 FIG. 1 is a flow chart of a GIS partial discharge identification optimization method for a small sample set. In this example, a GIS partial discharge identification optimization method for a small sample set includes the following steps: S1. Collect the original data when partial discharge occurs in GIS equipment and convert it into PRPD spectrum; S2, screen out high-quality PRPD maps and classify and annotate them, and after preprocessing, construct a class A training set and a class A test set according to a certain ratio; S3, applying K-nearest neighbor interpolation method in the A-type training set to generate new PRPD atlas data, construct the B-type sample set, and combine it with the A-type training set to construct the C-type training set; S4, use the ViT model to train the C-type training set to obtain an optimized recognition model; S5. Use the A-type test set to test the trained ViT model to evaluate the effectiveness of recognition and classification; S6. By comparing the training results of the A and B training sets, the superiority of this method is verified.

[0037] In step S1 of this embodiment, raw data when partial discharge occurs in GIS equipment is collected and converted into a PRPD spectrum.

[0038] In step S2 of this embodiment, high-quality PRPD maps are screened and classified and labeled, where the labels of suspended discharge, corona discharge, insulating discharge, and free discharge are set to 0, 1, 2, and 3, respectively; the images are preprocessed and normalized using the Python-based torchvision module, with the mean and standard deviation set to [0.5, 0.5, 0.5], and the size is adjusted to a grayscale image of 112×112×3; the typical PRPD grayscale map after preprocessing is shown in FIG. Figure 2 As shown, after that, the various types of real PRPD grayscale maps were used to construct Class A training sets and Class A test sets in a ratio of 8:2. The specific number of various discharge type data sets is shown in Table 1.

[0039] Table 1 Specific number of data sets for various discharge types

[0040] In step S3 of this embodiment, applying the K nearest neighbor interpolation method in the class A training set includes: S301, select a PRPD spectrum as a sample, and record the sample as A(x,y) ,in x and y Represent the rows and columns of the image (i.e. the spatial coordinates of the image). This sample is the reference atlas, which is used to generate new samples including; S302: Calculate the K nearest neighbor samples of the selected sample using the Euclidean distance in the K nearest neighbor algorithm. A i (x,y) , calculate its difference with the sample A(x,y) By calculating the distance of each sample, we can find the K neighbor samples with the smallest distance, which are recorded as A n1 (x,y) , A n2 (x,y) ,……, A nK (x,y) , the distance formula is as follows: ; in, d(i,s) Representation sample A i (x,y) With the benchmark sample A(x,y) The distance between A(x,y) is the pixel value of the selected benchmark sample; A i (x,y) is the pixel value of the current sample.

[0041] S303: Generate new PRPD spectrum data by interpolating the selected sample and its neighbors. Generate new samples using the weighted average method B(x,y) , and its calculation formula is as follows: ; in: B(x,y) is the pixel value of the generated new PRPD map; A nj (x,y) It is j The pixel value of neighbor samples; K is the number of neighbors chosen.

[0042] Weight w j According to the definition of the distance between the neighbor and the benchmark sample, it is usually set to the inverse of the distance. The specific calculation formula is as follows: ; in: w j It is j The weight of the neighbors; d(nj,s) Indicates j Neighborhood samples A nj (x,y) With the benchmark sample A (x,y) The distance between is a small constant that prevents the denominator from being zero.

[0043] S304: Combine the A-type training set and the B-type sample set to construct a C-type training set.

[0044] In step S3 of this embodiment, the new PRPD spectrum generated by the K nearest neighbor interpolation method is as follows: Figure 3 shown.

[0045] In step S3 of this embodiment, when generating new PRPD spectrum data, the actual number of samples generated is dynamically adjusted according to the number of sample sets of the four types of local discharges to ensure that the final number of samples of each type is balanced: for types with a smaller number of samples, more samples are generated; and for types with a larger number of samples, the number of generated samples is reduced. In this example, a total of 1,200 data are generated, including 290 suspended discharge data, 180 tip corona discharge data, 460 insulating discharge data, and 270 free particle discharge data.

[0046] In step S4 of this embodiment, the ViT model is used to train the C-type training set, wherein the main model structure hyperparameter settings are shown in Table 2.

[0047] Table 2 ViT model structure hyperparameter settings meaning Numeric image_size Image size 224 patch_size Image block size 16 num_classes Number of categories for classification tasks 4 dim Feature Dimensions in Transformer 128 depth Number of coding layers 3 heads Number of heads in multi-head self-attention 16 mlp_dim Dimensions of Multilayer Perceptron 256 In step S4 of this embodiment, the training hyperparameters are configured as follows: the batch_size of the training set and the test set is 32, the learning rate is 0.001, and the number of training cycles is 100. The classification loss function uses the cross entropy loss function to measure the difference between the model output and the true label. The optimization algorithm is the Adam algorithm, with the goal of minimizing the loss function. After every 10 training cycles, the learning rate scheduler is used to multiply the learning rate by 0.5 for adjustment, so as to dynamically update the learning rate and further improve the training effect and recognition performance of the model.

[0048] In step S5 of this embodiment, the trained ViT model is tested with the class A test set to evaluate the effectiveness of recognition and classification, which specifically includes training the class A test set to generate a loss curve and an accuracy curve to visualize the performance changes during the training process. The generated curves are as follows: Figure 4 shown.

[0049] In step S6 of this embodiment, by comparing the training effects of the A and B training sets, the superiority of the method is verified, including: S601, putting the A and B training sets into the ViT model for training; S602, using the class A test set to test the two ViT models trained in step S601 respectively; S603, taking the best accuracy of each test as the best model and comparing them to verify the superiority of this method. Figure 5 shown.

[0050] The loss curve and accuracy curve results show that when the model is trained using the C dataset, the performance increases steadily with the iteration cycle, and the accuracy does not fluctuate significantly, indicating that the model performance is suitable for actual engineering applications.

[0051] At the same time, the comparison chart of the model effects under different data set training shows that the model accuracy of data set A is 92.35%, the accuracy of data set B is lower, at 80.47%, and on data set C, the accuracy of the model is significantly improved to 98.92%. By applying the K nearest neighbor interpolation method to the A-class training set to generate class B samples and merging them with class A samples to construct the C-class training set, the diversity and sample size of the data set are significantly improved, and the generalization ability of the model is enhanced. In addition, the sample balance strategy in the C-class training set further improves the classification ability of the model on different partial discharge types, verifying the effectiveness of this method in solving the problem of insufficient data in small sample sets and poor recognition effect caused by unbalanced sample numbers.

[0052] The above only expresses the preferred embodiments of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications, improvements and substitutions can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for optimizing partial discharge identification of gas-insulated switchgear for a small sample set, characterized in that: The steps include: S1. Collect the original data when partial discharge occurs in GIS equipment and convert it into PRPD spectrum; S2, screen out high-quality PRPD maps and classify and annotate them, and after preprocessing, construct a class A training set and a class A test set according to a certain ratio; S3, applying K-nearest neighbor interpolation method in the A-type training set to generate new PRPD atlas data, construct the B-type sample set, and combine it with the A-type training set to construct the C-type training set; S4, use the ViT model to train the C-type training set to obtain an optimized recognition model; S5. Use the A-type test set to test the trained ViT model to evaluate the effectiveness of recognition and classification; S6. By comparing the training results of the A and B training sets, the superiority of this method is verified.

2. The method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 1, characterized in that: In step S1, a UHF sensor is used to collect partial discharge phenomena of GIS equipment; The original data collected in step S1 when the GIS equipment has a partial discharge phenomenon includes: suspended discharge, tip corona discharge, insulating discharge, and free particle discharge.

3. The method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 1 is characterized in that: In step S1, the PRPD spectrum is formed into a two-dimensional spectrum by combining the local discharge pulse with the power frequency phase, wherein one coordinate axis represents the power frequency phase φ and the other coordinate axis represents the discharge amount q; This kind of map can clearly show the discharge intensity and occurrence number n in different phases, and then identify and classify the type of partial discharge and evaluate the insulation status of electrical equipment.

4. The method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 1, characterized in that: The step S2 specifically includes: S201, manually remove unclear or erroneous maps and retain clear and high-quality PRPD maps; S202, preprocessing the image and normalizing it into a grayscale image of uniform size; S203. Construct a class A training set and a class A test set according to a certain ratio.

5. A method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 4, characterized in that: In step S203, after preprocessing, a class A training set and a class A test set are constructed in a certain ratio, wherein the ratio of class A training set to class A test set is 8:

2.

6. A method for optimizing partial discharge identification of gas-insulated switchgear for a small sample set according to claim 1, characterized in that: The step S3 of applying the K nearest neighbor interpolation method to the class A training set includes: S301, select a PRPD spectrum as a sample, and record the sample as A(x,y) ,in x and y Represent the rows and columns of the image respectively. This sample is the reference atlas and is used to generate new samples including; S302, calculating the K nearest neighbor samples of the selected sample, using the Euclidean distance in the K nearest neighbor algorithm for calculation; For each sample A i (x,y) , calculate its difference with the sample A(x,y) By calculating the distance of each sample, we can find the K neighbor samples with the smallest distance, which are recorded as A n1 (x,y) , A n2 (x,y) ,……, A nK (x,y) , the distance formula is as follows: ； in, d(i,s) Representation sample A i (x,y) With the benchmark sample A(x,y) The distance between A(x,y) is the pixel value of the selected benchmark sample; A i (x,y) is the pixel value of the current sample; S303, generate new PRPD map data by interpolating the selected sample and its neighbors; generate new samples by weighted average method B(x,y) , and its calculation formula is as follows: ； in: B(x,y) is the pixel value of the generated new PRPD map; A nj (x,y) It is j The pixel value of neighbor samples; K is the number of neighbors selected; Weight w j According to the definition of the distance between the neighbor and the benchmark sample, it is usually set to the inverse of the distance. The specific calculation formula is as follows: ； in: w j It is j The weight of the neighbors; d(nj,s) Indicates j Neighborhood samples A nj (x,y) With the benchmark sample A(x,y) The distance between is a small constant that prevents the denominator from being zero; S304: Combine the class A training set and the class B sample set to construct a class C training set.

7. A method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 6, characterized in that: When generating new PRPD spectrum data in step S3, the number of samples actually generated is dynamically adjusted according to the number of sample sets of the four partial discharge types to ensure that the final number of samples of each type is balanced; for types with a small number of samples, more samples are generated; and for types with a large number of samples, the number of samples generated is reduced.

8. A method for optimizing partial discharge identification of gas-insulated switchgear for a small sample set according to claim 7, characterized in that: In step S4, the ViT model, which stands for Vision Transformer, is an emerging deep learning model that applies the Transformer architecture, which has been very successful in the field of natural language processing, to computer vision tasks; In step S4, the ViT model is used to train the C-type training set to obtain an optimized recognition model. The training process includes selecting appropriate hyperparameters to improve the diversity of training data and performing multiple rounds of iterative training to ensure that the final model has sufficient adaptability and recognition effect.

9. A method for optimizing partial discharge identification of gas-insulated switchgear for small sample sets according to claim 8, characterized in that: In step S5, the trained ViT model is tested with a Class A test set to evaluate the effectiveness of recognition and classification, specifically including generating a loss curve and an accuracy curve by training the Class A test set to visualize the performance changes during the training process and analyze the recognition effect of the model on different partial discharge types.

10. A method for optimizing partial discharge identification of gas-insulated switchgear for a small sample set according to claim 9, characterized in that: The step S6 specifically includes: S601, putting the A and B training sets into the ViT model for training; S602, using the class A test set to test the two ViT models trained in step S601 respectively; S603, taking the best accuracy of each test as the best model and comparing them to verify the superiority of this method.

Images