PD identification method for gas switchgear based on cost-sensitive and dual-branch structure

Abstract

The present invention discloses a PD identification method for gas switchgear based on cost sensitivity and dual-branch structure, including the steps of: S1, constructing a one-dimensional convolutional neural network to extract the time series characteristics of the original ultra-high frequency data; S2, solving the number of effective samples of each type of sample one by one, and constructing a cost-sensitive mechanism for the number of effective samples; S3, integrating the cost-sensitive mechanism for effective samples into Res2Net to improve the model's attention to minority classes; S4, constructing a bilateral branch network structure to optimize the algorithm training process; S5, completing model training based on measured PD ultra-high frequency data to achieve PD pattern recognition. The present invention includes a cost-sensitive mechanism for the number of effective samples, an improvement on the Res2Net algorithm, and a bilateral branch network structure, which can well alleviate the problem of reduced model representation learning performance caused by conventional rebalancing learning methods, comprehensively improve the robustness of the model in handling unbalanced learning problems, and improve the accuracy of PD pattern recognition.

Description

Technical Field

[0001] The invention relates to the field of GIS PD pattern recognition, and in particular to a gas switchgear PD recognition method based on cost sensitivity and double-branch structure. Background Art

[0002] Gas insulated switchgear (GIS) has been widely used in new power systems and smart substations due to its small size, excellent insulation performance and easy installation. However, equipment defects are inevitable during the manufacturing, transportation, installation, and mechanical operation processes, which can induce partial discharge (PD) in GIS during operation. The PD mode type is an important factor in evaluating the degree of GIS insulation degradation. Accurately identifying the PD mode is a key step in sensing the insulation status of GIS, which is of great research significance for ensuring reliable power transmission.

[0003] With the maturity of ultra high frequency (UHF) discharge detection technology, PD feature construction and PD pattern recognition based on UHF data have become the main research directions. Among them, feature construction includes statistical features, fractal features, waveform features, etc. In recent years, with the growth of UHF data and the development of computer technology, the powerful learning ability of convolutional neural network (CNN) has effectively eliminated the process of manually compiling PD features and has become an important technical means for identifying PD patterns (GAN and CNN for imbalanced partial discharge pattern recognition inGISYanxin WangJing YanZhou YangQianzhen Jing Jianhua Wang lYingsan Geng). However, due to the large differences in the conditions for generating different PD patterns, PD data is unbalanced in PD experiments and actual application scenarios. At present, most methods only focus on improving the accuracy of the model, but ignore the improvement of the model's recognition performance for minority PD patterns. Common improvement strategies for imbalanced learning include rebalancing at the data level and improvements at the algorithm level. Y. Wang et al. and Y. Liu et al. used GAN and SMOTE algorithms to balance the data and then trained the model to enhance the model's learning performance for the minority class, but this method often introduces the risk of overfitting or underfitting the model and reduces the model's robustness. Cui Yu et al. used an ensemble learning algorithm from an algorithmic perspective to gradually strengthen the model's attention to the minority class, thereby improving the model's accuracy for the minority class, but this method will lead to a decrease in the model's representation learning performance and cause the majority class recognition effect to deteriorate (SVM SMOTE-based Power Transformer Fault Sample Balancing Method Liu Yunpeng, Jia Hui, Xu Ziqiang, Wang Quan, Li Zhe, Gao Shuguo Hebei Provincial Key Laboratory of Power Transmission and Transformation Equipment Security Defense, North China Electric Power University, Baoding 071003; 2. National Key Laboratory of New Energy Power System, North China Electric Power University, Beijing 102206; 3. State Grid Hebei Electric Power Co., Ltd. Electric Power Research Institute, Shijiazhuang 050021)). Summary of the invention

[0004] In order to overcome the shortcomings and deficiencies of the prior art, the present invention proposes a PD identification method for gas-insulated switchgear based on cost-sensitive Res2Net and dual-branch structure. The method includes constructing a cost-sensitive mechanism for the number of effective samples of a class, improving the Res2Net algorithm, and constructing a bilateral branch network structure. While improving the accuracy of PD pattern recognition, it will not reduce the recognition effect of most classes, greatly improving the robustness of the model in practical applications.

[0005] The present invention is achieved by at least one of the following technical solutions.

[0006] The PD identification method of gas switchgear based on cost sensitivity and dual-branch structure includes the following steps:

[0007] S1. Build a one-dimensional convolutional neural network, gradually extract the time series features of the original UHF data by smoothly moving the convolution kernel window, and construct a two-dimensional feature map by feature splicing to facilitate subsequent network calculations; the original UHF data includes surface discharge, tip discharge, particle discharge and insulator air gap discharge;

[0008] S2. Establish the relationship between the number of effective samples and the actual number of samples, solve the effective number of each type of samples, and construct a sensitive mechanism for the number of effective samples;

[0009] S3, integrating the cost-sensitive matrix in step S2 into the Res2Net network, constructing a Res2Net model with class-valid sample number constraints, and determining the theoretical number of samples in the sample space based on the test results of the Res2Net model with class-valid sample number constraints;

[0010] S4. Build a bilateral branch network structure containing regular learning and rebalancing learning to achieve a smooth transition from regular learning to rebalancing learning during algorithm training;

[0011] S5. Based on the measured PD UHF data, complete the bilateral branch network structure training, and use the trained bilateral branch network structure to identify the PD pattern.

[0012] Furthermore, in step S1, the one-dimensional convolutional neural network directly extracts the time series features of the original data, and the time series feature extraction is realized by sliding the convolution kernel.

[0013] Furthermore, the one-dimensional convolution kernel window size is 1×37, the sliding step is 1, and the number of convolution kernels is 64.

[0014] Furthermore, based on random coverage, the relationship between the actual number of samples and the effective sample volume is established, and then a cost-sensitive mechanism similar to effective samples is constructed:

[0015]

[0016] In the formula, λ represents the intermediate variable; M represents the theoretical number of effective samples in the sample space; n represents the actual number of samples; E n Indicates the number of valid samples.

[0017] Furthermore, the penalty weight matrix W of the Res2Net model is constrained based on the number of valid class samples:

[0018]

[0019] In the formula represents the effective number of samples of the jth class; w M Represents the Mth diagonal element in the W matrix.

[0020] Furthermore, the loss function of the Res2Net model based on the number of class effective samples is as follows:

[0021] Loss Res2Net =W·CE(p,y)

[0022] Where CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, and W represents the penalty weight matrix.

[0023] Furthermore, the bilateral branch network structure is constructed as follows:

[0024] A one-dimensional convolutional neural network is used to extract the time series features of the original data, and the one-dimensional time series features are converted into a two-dimensional time series feature map through feature splicing;

[0025] The multi-layer Res2Net structure is used to mine the feature information of the two-dimensional time series feature map, and the information is input into the conventional learning and rebalancing learning branches in sequence. Conventional learning relies on the cross entropy function to evaluate the learning degree, and rebalancing learning evaluates the learning degree by constraining the cross entropy function through a class cost-sensitive mechanism.

[0026] The features obtained from the bilateral branches are fused based on the cumulative learning mechanism, and the final PD pattern recognition is achieved using a fully connected network.

[0027] Furthermore, the number of valid class samples constrains the Res2Net model and the bilateral branch network structure loss function, as shown in the formula:

[0028] Loss BBN =βCE(p,y)+(1-β)·W·CE(p,y)

[0029] Where CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, β∈[0,1] represents the cumulative learning factor, and W represents the penalty weight matrix.

[0030] Furthermore, the training of the bilateral branch network structure includes:

[0031] In the representation learning part, a one-dimensional convolutional neural network is used to extract and expand the feature information and dimension of time-series UHF data, and deep features are learned through multiple groups of Res2Net networks;

[0032] Construct two branches, conventional learning and rebalancing learning, respectively. The features obtained by conventional learning are f a , the feature obtained by rebalancing learning is f b ;

[0033] In the classification learning part, bilateral features are adaptively fused based on the cumulative learning strategy to realize the training of the bilateral branch network structure model.

[0034] Furthermore, PD UHF data is obtained through GIS partial discharge simulation experiments, and model training is completed based on the measured PD UHF data.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The gas switchgear PD recognition method based on cost sensitivity and dual-branch structure of the present invention, on the one hand, deeply mines PD data features through Res2Net with a multi-layer convolution structure, and uses a class-effective cost-sensitive mechanism to improve Res2Net's attention to the PD mode of the gas-insulated switchgear; on the other hand, the constructed bilateral branch structure retains the excellent representation learning ability of Res2Net while taking into account the class-cost-sensitive mechanism to constrain the Res2Net model to accurately recognize the PD mode of the gas-insulated switchgear PD, thereby alleviating the problem of reduced model robustness caused by the commonly used rebalancing strategy. While improving the accuracy of PD pattern recognition of gas-insulated switchgear, the present invention will not reduce the recognition effect of most classes, greatly improving the robustness of the model in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 is a structural diagram of a method for identifying PD of a gas switchgear based on cost sensitivity and a dual-branch structure according to an embodiment of the present invention;

[0039] Figure 2 It is a data spectrum of the surface PD of the GIS pot-type insulator according to an embodiment of the present invention;

[0040] Figure 3 It is a data map of the cutting-edge PD in the GIS of an embodiment of the present invention;

[0041] Figure 4 It is a data map of GIS particle PD of an embodiment of the present invention;

[0042] Figure 5 It is a data map of the air gap PD of the GIS pot-type insulator according to an embodiment of the present invention;

[0043] Figure 6 is a one-dimensional convolutional neural network structure diagram of an embodiment of the present invention;

[0044] Figure 7 2 is a Res2Net structure diagram of an embodiment of the present invention.

[0045] Figure 8 is a performance curve diagram of the model under different theoretical effective sample numbers of the embodiment of the present invention;

[0046] Fig. 9 is a characteristic distribution diagram of an embodiment of the present invention;

[0047] Fig.10 is a training curve diagram of an embodiment of the present invention;

[0048] Fig.11 It is a PD pattern recognition effect diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the implementation methods of the present invention are not limited thereto.

[0050] In order to improve the recognition accuracy of the model for minority class samples, the present invention first proposes a Res2Net model (Ence-Res2Net) constrained by the number of class effective samples by incorporating a cost-sensitive mechanism based on the number of class effective samples into Res2Net; then, a dual-branch network structure with sample consistency is constructed, which consists of conventional learning and rebalancing learning branches. On the one hand, Res2Net is used as a learner for "conventional learning" to improve the model representation learning performance; on the other hand, Ence-Res2Net is used as a learner for rebalancing learning to improve the model classification performance. Considering the limited amount of PD data samples, the present invention improves the model convergence speed by maintaining sample consistency during dual-branch training. Through experimental verification and testing, the present invention can well alleviate the problem of decreased model representation learning performance caused by conventional rebalancing learning methods, comprehensively improve the robustness of the model in handling unbalanced learning problems, and improve the PD pattern recognition accuracy of gas insulated switchgear.

[0051] like Figure 1As shown, a PD identification method for a gas switchgear based on cost sensitivity and a dual-branch structure, specifically, includes using a PD pattern recognition model to realize PD pattern recognition, and building a PD pattern recognition model includes the following steps:

[0052] S1. Build a one-dimensional convolutional neural network, gradually extract the time series features of the original UHF data by smoothly moving the convolution kernel window, and use feature splicing to construct a two-dimensional feature map to facilitate subsequent network calculations; the one-dimensional convolution kernel window size is 1×37, the sliding step size is 1, and the number of convolution kernels is 64.

[0053] Figure 2 to Figure 5 The original PD data spectrum measured in the GIS partial discharge experimental platform of the present invention includes surface discharge, tip discharge, particle discharge and insulator air gap discharge. 0 , x 1 …, x m-1 As input to a one-dimensional convolutional neural network;

[0054] As one embodiment, the i-th convolution kernel window of size 1×37 is smoothed in a sliding manner with a step size of 1, and the extracted feature is y i，0 ,y i，1 ,…; Using feature splicing technology, the features extracted by each one-dimensional convolution kernel are spliced ​​to construct a feature map: {[y 0，0 ,y 0，1 ,…],[y 1，0 ,y 1，1 ,…],[y 2，0 ,y 2，1 ,…]}.like Figure 6 The one-dimensional convolutional neural network structure diagram in this embodiment is shown.

[0055] S2. Establish the relationship between the theoretical effective sample number and the actual sample number, solve the effective number of each type of sample in the following way, and construct a sensitive mechanism for the effective sample number of each type;

[0056] Based on the idea of ​​random coverage, the relationship between the actual number of samples and the effective volume of samples is established:

[0057]

[0058] In the formula, λ represents the intermediate variable; M represents the theoretical number of effective samples in the sample space; n represents the actual number of samples; E n represents the number of valid samples. The penalty weight matrix W of the four types of PD data (including surface discharge, tip discharge, particle discharge and insulator air gap discharge) is calculated as follows:

[0059]

[0060] In the formula represents the effective number of samples of the jth class; w M Represents the Mth diagonal element in the W matrix.

[0061] S3. Integrate the cost-sensitive matrix in S2 into Res2Net and construct a Res2Net model with class-valid sample number constraints.

[0062] As one example, Figure 7 The Res2Net model structure diagram of the present invention is shown as follows. First, n 1 A 1×1 two-dimensional convolution kernel is used to preprocess the features of the feature input and construct a convolution kernel containing n 1 dimensional feature map f; then divide f into k feature subsets, denoted as f 1 , …, f k ; Use the hierarchical convolution kernel to gradually extract the deep features of the feature subset to form the corresponding feature sub-map z 1 , …, z k ; Finally, these feature sub-maps are concatenated to form feature map z, and then n 2 A 1×1 two-dimensional convolution kernel is used to extract features from the feature map z again. The penalty weight in S2 is incorporated into the loss function of the Res2Net model, as shown in the following formula:

[0063] Loss Res2Net =W·CE(p,y)

[0064] In the formula, CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, and W represents the penalty weight matrix. The performance of the Res2Net model with the number of valid class samples constrained is tested. Figure 8 As shown in Figure 1, it is a curve diagram of the performance index parameters of the Res2Net model constrained by the number of class effective samples under different theoretical effective sample numbers. Figure 8 It shows that as the number of theoretical effective samples increases, the model pays more attention to the learning of minority samples, resulting in a fluctuating decrease in Acc, while the G-mean and Avg-acc curves fluctuate and continuously approach Acc. At M = 3000, the average accuracy (Avg-acc) and G-mean achieve local optimal values, and Acc here is close to the local optimal value. The theoretical number of samples when the model obtains the optimal performance is used as the theoretical effective sample number parameter value of the present invention.

[0065] S4. Build a bilateral branch network structure containing regular learning and rebalancing learning to achieve a smooth transition from regular learning to rebalancing learning during algorithm training, including the following steps:

[0066] In the representation learning part, a one-dimensional convolutional neural network is used to extract and expand the feature information and dimension of time-series UHF data, and deep features are learned through multiple groups of Res2Net networks;

[0067] Then, the conventional learning and rebalancing learning dual branches with the same two-dimensional convolutional network structure are constructed respectively, where the features obtained by conventional learning are f a , the feature obtained by rebalancing learning is f b ;

[0068] Finally, in the classification learning part, bilateral features are adaptively fused based on the cumulative learning strategy to construct a bilateral branch network structure loss function:

[0069] Loss BBN =β·CE(p,y)+(1-β)·W·CE(p,y)

[0070] Where CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, β∈[0,1] represents the cumulative learning factor, and W represents the penalty weight matrix.

[0071] S5. Carry out GIS partial discharge simulation experiment according to the steps to obtain reliable PD UHF data. Before starting the GIS partial discharge simulation experiment, turn off mobile phones and other instruments that generate UHF interference signals. Next, check the air pressure in the GIS air chamber, use the pulse current method to confirm that there is no partial discharge in the transformer and calibrate the UHF detector. Finally, put the artificial simulation defect module into the GIS air chamber in turn, adjust the voltage control console, and gradually increase the transformer output voltage so that the GIS generates stable PD. Complete the model training based on the measured PD UHF data.

[0072] Fig. 9 This is a sample distribution diagram of the feature f extracted in this embodiment in two-dimensional space. The figure shows that the features acquired by the model achieve obvious clustering within the class and reliable separation between classes. Fig.10 This is the training convergence diagram of the present invention. Fig.10 It shows that as the number of training times increases, conventional learning gradually transitions to the rebalancing learning branch, the model loss suddenly increases in the middle process and then converges to the minimum loss value, and Acc (accuracy) converges to the maximum value;

[0073] S6. Obtain a PD pattern recognition model for steps S2 to S5 to achieve PD pattern recognition. Fig.11 This is the accuracy of the PD model based on actual case sample analysis of the present invention.

[0074] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A PD identification method for gas switchgear based on cost sensitivity and dual-branch structure, characterized in that: The following steps are involved: S1. Build a one-dimensional convolutional neural network, gradually extract the time series features of the original UHF data by smoothly moving the convolution kernel window, and construct a two-dimensional feature map by feature splicing; the original UHF data includes four types of discharge data: surface discharge, tip discharge, particle discharge and insulator air gap discharge; S2. Establish the relationship between the number of effective samples and the actual number of samples, solve the effective number of each type of samples, and construct a sensitive mechanism for the number of effective samples; establish the relationship between the actual number of samples and the effective sample volume based on random coverage, and then construct a cost-sensitive mechanism for effective samples: In the formula, λ represents the intermediate variable; M represents the theoretical number of effective samples in the sample space; n represents the actual number of samples; E n Indicates the number of valid samples; S3, integrating the cost-sensitive matrix in step S2 into the Res2Net network, constructing a Res2Net model with class-valid sample number constraints, and determining the theoretical number of samples in the sample space based on the test results of the Res2Net model with class-valid sample number constraints; S4. Build a bilateral branch network structure containing regular learning and rebalancing learning to achieve a smooth transition from regular learning to rebalancing learning during algorithm training; S5. Based on the measured PD UHF data, complete the bilateral branch network structure training, and use the trained bilateral branch network structure to identify the PD pattern.

2. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: In step S1, the one-dimensional convolutional neural network directly extracts the time series features of the original data, and the time series feature extraction is realized by sliding the convolution kernel.

3. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: The one-dimensional convolution kernel window size is 1×37, the sliding step is 1, and the number of convolution kernels is 64.

4. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: The penalty weight matrix W of the Res2Net model is constrained based on the number of effective class samples: In the formula represents the effective number of samples of the jth class; w M Represents the Mth diagonal element in the W matrix.

5. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: The loss function of the Res2Net model based on the number of effective class samples is as follows: Loss Res2Net =W·CE(p,y) Where CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, and W represents the penalty weight matrix.

6. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: The bilateral branch network structure is constructed as follows: A one-dimensional convolutional neural network is used to extract the time series features of the original data, and the one-dimensional time series features are converted into a two-dimensional time series feature map through feature splicing; The multi-layer Res2Net structure is used to mine the feature information of the two-dimensional time series feature map, and the information is input into the conventional learning and rebalancing learning branches in sequence. Conventional learning relies on the cross entropy function to evaluate the learning degree, and rebalancing learning evaluates the learning degree by constraining the cross entropy function through a class cost-sensitive mechanism. The features obtained from the bilateral branches are fused based on the cumulative learning mechanism, and the final PD pattern recognition is achieved using a fully connected network.

7. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 1 is characterized in that: The number of effective samples of the class constrains the Res2Net model and the bilateral branch network structure loss function, as shown in the formula: Loss BBN =β·CE(p,y)+(1-β)·W·CE(p,y) Where CE(·) represents the cross entropy function, p represents the probability predicted by the model, y represents the actual probability of the sample, β∈[0,1] represents the cumulative learning factor, and W represents the penalty weight matrix.

8. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to any one of claims 1 to 7, characterized in that: The training of the bilateral branch network structure includes: In the representation learning part, a one-dimensional convolutional neural network is used to extract and expand the feature information and dimension of time-series UHF data, and deep features are learned through multiple groups of Res2Net networks; Construct two branches, conventional learning and rebalancing learning, respectively. The feature obtained by conventional learning is f a , the feature obtained by rebalancing learning is f b ; In the classification learning part, bilateral features are adaptively fused based on the cumulative learning strategy to realize the training of the bilateral branch network structure model.

9. The PD identification method for gas switchgear based on cost sensitivity and dual-branch structure according to claim 8, characterized in that: Through GIS partial discharge simulation experiments, PD UHF data is obtained, and model training is completed based on the measured PD UHF data.

Images