Remaining life prediction method for circuit breakers based on stage attention mechanism network model

Through the method based on the stage attention mechanism network model, combined with one-dimensional convolutional neural network and GRU network, the vibration signal characteristics of the circuit breaker are extracted, and the problem of universal circuit breaker life prediction is solved, and more accurate mechanical life prediction is achieved. Improve the stability and economic benefits of the power system.

CN114519305BActive Publication Date: 2024-11-22HEBEI UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210136953.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-15
Publication Date
2024-11-22
Estimated Expiration
2042-02-15

AI Technical Summary

Technical Problem

The existing technology has shortcomings in the mechanical life prediction of the circuit breaker field, especially the technical difficulty of accurate life prediction of universal circuit breakers. The existing methods are mainly aimed at high-voltage circuit breakers, and there is a lack of effective method for predicting the remaining life of universal circuit breakers.

Method used

Using a network model based on the stage attention mechanism network model, vibration signals are collected through the universal circuit breaker mechanical life test system, and modal component decomposition and reconstruction are used to combine the prediction of one-dimensional convolutional neural network and GRU network. The model, fuses the distributed attention mechanism and the time-step attention mechanism, extracts the contact breaking vibration segments, and constructs the data set for prediction.

Benefits of technology

Improve the accuracy and reliability of the remaining life prediction of the circuit breaker, reduce the amount of network parameters, improve the training rate, can more accurately reflect the mechanical state and functionality of the circuit breaker, extend the service life of the circuit breaker, and reduce resource waste. , improves the economic benefits of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114519305B_ABST
    Figure CN114519305B_ABST
Patent Text Reader

Abstract

This invention presents a method for predicting the remaining life of circuit breakers based on a stage-based attention mechanism network model. First, vibration signals during the opening process are collected. Then, the VMD algorithm is optimized and used to decompose the vibration signals, selecting modal components with high kurtosis for reconstruction. Next, contact segmentation vibrations are extracted from the reconstructed vibration signals based on the energy entropy ratio. Finally, a prediction model incorporating the stage-based attention mechanism is established. This model uses a one-dimensional convolutional neural network and a GRU network as its backbone. The stage-based attention mechanism consists of two stages: the first stage utilizes a distributed attention mechanism applied to the one-dimensional convolutional neural network, weighting the input samples in both time and feature dimensions; the second stage utilizes a time-step attention mechanism applied to the GRU network, further weighting the input samples in the time dimension. This method enhances the contribution of important information in both the time and feature dimensions to the prediction results, thereby improving prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit breaker remaining life prediction, and specifically is a circuit breaker remaining life prediction method based on a stage attention mechanism network model. Background Art

[0002] The low-voltage power distribution system is a key part of the power transmission and distribution network system, and plays an important role in transmitting electrical energy to the load. A momentary interruption or abnormality in the power supply may cause the power transmission and distribution network system to be paralyzed or disordered, and even cause safety accidents and huge economic losses. As an important electrical switching device in the low-voltage power distribution system, the universal circuit breaker has a wide range of applications. On the one hand, as a dispatching and control device for the low-voltage power grid, it can connect or disconnect specific distribution lines according to the power distribution needs; on the other hand, when a short circuit, overload, undervoltage or other faults occur in the transmission line, the faulty line can be cut off in time, which can not only protect the low-voltage power distribution system and electrical equipment from being damaged by the fault current, but also prevent safety accidents. For the mechanical life of universal circuit breakers, although engineers will specify the factory life based on experience, this specified life is only an estimate of the ideal life and has limited reference value in actual use. Therefore, it is necessary to combine the changes in the mechanical state in actual use and use reliable methods to predict the remaining life of the circuit breaker. On the one hand, it can provide predictive support for the maintenance of circuit breakers, and on the other hand, it can avoid the waste of resources caused by blindly replacing equipment, and at the same time it is conducive to ensuring the stability and safety of the low-voltage distribution network.

[0003] Remaining useful life (RUL) prediction technology has been maturely applied in many fields such as bearings, aircraft engines and lithium batteries. However, in the field of circuit breakers, especially in the field of mechanical life prediction, there are few studies and obvious deficiencies. This is mainly because the circuit breaker is a complex mechanical system and it is technically difficult to achieve a relatively accurate life prediction. For example, Yang Qiuyu et al. (Yang Qiuyu, Peng Yanqing, Zhuang Zhijian, Xiong Lei. Evaluation of the remaining mechanical life of high-voltage circuit breakers based on random fuzzy theory [J]. High Voltage Electrical Appliances, 2016, 52(08): 161-165+171.) applied random fuzzy theory to the mechanical life evaluation of high-voltage circuit breakers. The mechanical life model was established through the random fuzzy distribution function of the contact overtravel and the average opening speed, and the mechanical life of the high-voltage circuit breaker was objectively and effectively evaluated. However, this method is only for high-voltage circuit breakers. At present, in the field of universal circuit breakers, some scholars have conducted life prediction research on operating accessories or contact systems. For example, Sun Shuguang et al. (Sun Shuguang, Li Qin, Wang Jiaxing, Du Taihang, Wang Jingqin. Real-time remaining life prediction of operating accessories of universal circuit breakers based on performance degradation model [J]. Chinese Journal of Scientific Instrument, 2019, 40(10): 120-129. DOI: 10.19650 / j.cnki.cjsi.J1905286.) established an operating accessory degradation model based on the Wiener process and a RUL prediction model based on the concept of first arrival time, realizing real-time life prediction of operating accessories. Summary of the invention

[0004] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide a method for predicting the remaining life of a circuit breaker based on a stage attention mechanism network model.

[0005] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0006] A method for predicting the remaining life of a circuit breaker based on a stage attention mechanism network model, characterized in that the method comprises the following steps:

[0007] The first step is to use the universal circuit breaker mechanical life test system to collect the vibration signal during the circuit breaker opening process;

[0008] The second step is to optimize the VMD algorithm based on energy loss to obtain the optimal number of modal components and quadratic penalty factor; decompose the vibration signal using the optimized VMD algorithm to obtain multiple modal components; calculate the kurtosis of each modal component, select the modal component with larger kurtosis for reconstruction, and obtain the reconstructed vibration signal;

[0009] The third step is to divide the reconstructed vibration signal into frames to obtain a multi-frame reconstructed signal; calculate the energy entropy ratio of all frame reconstructed signals, and extract the contact break vibration segment from the reconstructed vibration signal according to the energy entropy ratio; construct a data set based on the contact break vibration segment, the data set includes multiple two-dimensional input samples, each input sample includes multiple contact break vibration segments, and each contact break vibration segment includes multiple sampling points;

[0010] The fourth step is to establish a prediction model of the fusion stage attention mechanism. The prediction model uses a one-dimensional convolutional neural network and a GRU network as the backbone network. The stage attention mechanism is divided into two stages. The first stage is a distributed attention mechanism applied to a one-dimensional convolutional neural network, and the second stage is a time step attention mechanism applied to a GRU network.

[0011] The one-dimensional convolutional neural network contains four convolutional layers, each of which contains three convolutional kernels, and each convolutional kernel performs horizontal sliding convolution on multiple time-step data of the input sample;

[0012] The output matrix of the second convolutional layer is used as the input matrix of the distributed attention mechanism. First, the input matrix is ​​weighted in the time dimension to obtain the time-weighted weights of all row vectors in the input matrix. The time-weighted weights of all row vectors constitute a time-weighted column vector. The time-weighted column vector is multiplied by the corresponding elements of the original input matrix to obtain a time-weighted matrix.

[0013] First, assuming that the input matrix size of the distributed attention mechanism is M'×L', the i-th row vector of the input matrix is ​​represented as x i =[x i1 ,x i2 ,...,x iL' ], i = 1, 2, ..., M', then according to equations (11) and (12) the row vector x i Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling time weight coefficient a i and the global average pooling time weight coefficient b i Similarly, after performing the same operation on all row vectors of the input matrix, we obtain M' groups of one-to-one corresponding global maximum pooling time weight coefficients and global average pooling time weight coefficients. Finally, we calculate the row vector x through the dense layer and softmax activation function of formula (13). i The time weighted weight α i; Similarly, calculate the time-weighted weights of all row vectors in the input matrix to obtain the time-weighted weights of M' row vectors, and the time-weighted weights of these M' row vectors constitute the time-weighted column vector; multiply the corresponding elements of the time-weighted column vector with the input matrix of the distributed attention mechanism to obtain the time-weighted matrix, that is, to realize the time-weighted input matrix of the distributed attention;

[0014] a i =globalmax(x i )=globalmax([x i1 ,x i2 ,...,x iL' ]) (11)

[0015] b i =globalaverage(x i )=globalaverage([x i1 ,x i2 ,...,x iL' ]) (12)

[0016] α i =softmax((w a *a i +bias a )+(w b *b i +bias b )) (13)

[0017] In the formula, w a 、w b are the convolution kernel matrices of the global maximum pooling and global average pooling in the time-weighted process, bias a 、bias b They are the bias matrices of global maximum pooling and global average pooling in the time weighted process respectively;

[0018] Then, the time weighting matrix is ​​further weighted in the feature dimension. Assume that the e-th column vector of the time weighting matrix is ​​x e =[x 1e ,x 2e ,...,x M'e ] T , e=1,2,...,L', then according to equations (14) and (15) the column vector x e Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling feature weight coefficient c e and the global average pooling feature weight coefficient d eSimilarly, after performing the same operation on all column vectors of the time weighted matrix, we obtain the global maximum pooling feature weight coefficient and the global average pooling feature weight coefficient corresponding to each group L'. Finally, we calculate the column vector x by the dense layer and softmax activation function shown in formula (16): e The feature weight β e ; Similarly, calculate the feature weighted weights of all column vectors in the time weighted matrix to obtain the feature weighted weights of L' column vectors, and the feature weighted weights of these L' column vectors constitute the feature weight row vector; multiply the feature weight row vector by the corresponding elements of the original time weighted matrix to obtain the feature weight matrix;

[0019] c e =globalmax([x 1e ,x 2e ,...,x M'e ] T ) (14)

[0020] d e =gobalaverage([x 1e ,x 2e ,...,x M'e ] T ) (15)

[0021] β e =softmax(w e *[c e ;d e ]+bias e ) (16)

[0022] Among them, w e is the convolution kernel matrix of the feature weighting process, bias e is the bias matrix for the feature weighting process;

[0023] Finally, the feature weighted matrix is ​​added to the input matrix of the distributed attention mechanism, and the added matrix is ​​used as the input matrix of the third convolutional layer. After passing through the third and fourth convolutional layers in sequence, the output matrix of the one-dimensional convolutional neural network is obtained;

[0024] The output matrix of the one-dimensional convolutional neural network is used as the input matrix of the GRU network. The scoring function is used to score the time step information corresponding to all moments in the output matrix of the one-dimensional convolutional neural network and calculate the weight probability. The weight probabilities at all moments constitute a weight column vector. The weight column vector is multiplied by the corresponding elements of the output matrix of the GRU network to obtain the time step attention weighted result. Finally, the time step attention weighted result passes through the fully connected layer to obtain the output of the prediction model, that is, the prediction value, thus completing the construction of the prediction model integrating the attention mechanism.

[0025] The prediction model of the attention mechanism in the fusion stage is trained, and the pre-trained prediction model is used to predict the remaining mechanical life of the circuit breaker.

[0026] In the second step, the specific process of VMD algorithm optimization is as follows: set the screening range of modal component number and quadratic penalty factor, combine the modal component number and quadratic penalty factor in sequence, increase the modal component number and quadratic penalty factor by a certain amount each time, obtain all possible parameter combinations within the screening range, substitute each set of parameters into VMD algorithm to decompose the vibration signal, and calculate the energy loss of the vibration signal before and after decomposition. When the energy loss is close to 0, the modal component number and quadratic penalty factor have optimal values, and the optimization of VMD algorithm is completed.

[0027] In the third step, the energy entropy ratio of all frame reconstructed signals forms an energy entropy ratio envelope curve along the frame sequence direction, and the lower threshold and upper threshold for two-level judgment are set. First, the upper threshold is used to perform the first-level judgment, and the energy entropy ratio of each point on the energy entropy ratio envelope curve is compared with the upper threshold. It is considered that the signal whose energy entropy ratio is not less than the upper threshold must belong to the contact break vibration segment, so as to obtain the signal interval determined by the first-level judgment; then the lower threshold is used to continue to judge outside the interval determined by the first-level judgment, and the left and right signal points whose energy entropy ratio is equal to the lower threshold are found. These two signal points are regarded as the starting point and end point of the contact break vibration segment, and the vibration signal between the two signal points is the contact break vibration segment, and the contact break vibration segment is used as the signal source for predicting the remaining life of the circuit breaker.

[0028] The third step also includes: extracting the peak-to-peak characteristics of the vibration signal from the contact breaking vibration segment, and regarding the point where the peak-to-peak characteristics change significantly as the inflection point of the circuit breaker performance degradation, that is, the dividing point between the stable period and the degradation period. Through this inflection point, the circuit breaker degradation process is divided into two stages: the stable period and the degradation period, and the change trends of the stable period and the degradation period are divided into two linear RUL label functions.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The prediction model of the fusion stage attention mechanism of the present invention uses a one-dimensional convolutional neural network and a GRU network as the main network, giving full play to the advantages and self-learning ability of the two neural networks. Compared with the conventional LSTM network, the GRU network can reduce the number of parameters of the operation by one third, thereby improving the training rate of the network. The distributed attention mechanism is integrated into the one-dimensional convolutional neural network, and the input samples are weighted in the time dimension and feature dimension respectively, which strengthens the contribution of important information in the time dimension and feature dimension to the prediction results; the time step attention mechanism is integrated after the GRU network, and the time dimension is weighted again, which further strengthens the contribution of important information in the time dimension to the prediction results, improves the prediction accuracy of the model, and is very important for time series prediction tasks.

[0031] 2. The VMD algorithm based on energy loss optimization solves the parameter selection problem of the VMD algorithm from an energy perspective, so that the VMD algorithm can achieve the best decomposition effect and retain the original information of the vibration signal to the greatest extent; the modal components with strong impact characteristics are screened out through the kurtosis index, so that the reconstructed vibration signal has a higher signal-to-noise ratio, which can effectively highlight the impact characteristics caused by the collision of the mechanism.

[0032] 3. In the vibration generated during the entire opening process, except for the vibration of the operating accessories with a small impact in the early stage, most of the energy and impact characteristics are concentrated in the contact breaking vibration segment. The contact breaking vibration segment is an integrated embodiment of the mechanical state of multiple mechanisms. Therefore, the present invention uses the short-time energy entropy ratio double threshold method to extract the contact breaking vibration segment from the reconstructed vibration signal, and uses the contact breaking vibration segment as a signal source to predict the mechanical life of the circuit breaker from an overall perspective. The short-time energy entropy ratio algorithm is an improved short-time analysis method. Compared with the short-time energy method, it is more sensitive to small impacts in vibration. Therefore, the short-time energy entropy ratio envelope curve can more comprehensively reflect the impact event of the entire vibration signal. On this basis, the double threshold method is used to detect the two endpoints of the contact breaking vibration segment with higher accuracy.

[0033] 4. The present invention selects the vibration signal during the circuit breaker opening process as the medium for life analysis, making full use of the non-invasive measurement advantages of the vibration signal and the rich mechanical state information advantages contained therein. In addition, the opening vibration is a reflection of the circuit breaker's breaking capacity and is the most direct and powerful tool to reflect the functionality and reliability of the circuit breaker. The present invention provides an integrated solution for predicting the mechanical remaining life of a universal circuit breaker using vibration signals. Compared with fault diagnosis or state assessment, which only evaluate the current state, the present invention can provide technical support for the long-term use of circuit breakers in terms of prognosis and maintenance, which helps to improve their reliability and safety, extend their service life to a certain extent, and increase the economic benefits of the power system.

[0034] 5. The RUL label function in the data set of the present invention is a piecewise linear function. According to the peak-to-peak value characteristics, the degradation process of the circuit breaker throughout its life is divided into two stages: the stable period and the degradation period. Compared with the conventional one-piece linear RUL label function, the present invention is more adaptable to the nonlinear degradation characteristics of the circuit breaker as a complex mechanical system, which helps to fundamentally improve the prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention;

[0036] Figure 2 is a flow chart of the distributed attention mechanism of the present invention;

[0037] Figure 3 It is a structural diagram of the mechanical life test system of the universal circuit breaker of the present invention;

[0038] Figure 4 Heat map of energy loss when optimizing the VMD algorithm;

[0039] Figure 5 This is the decomposition result diagram of the optimized VMD algorithm;

[0040] Figure 6 is the histogram of the kurtosis value of each modal component;

[0041] Figure 7 is a comparison diagram of the original vibration signal and the reconstructed vibration signal;

[0042] Figure 8 This is the endpoint detection result diagram of the short-time energy entropy ratio double threshold method;

[0043] Fig. 9 The peak-to-peak value trend and RUL label function graph extracted from the three datasets;

[0044] Fig.10 It is the operation diagram of the sliding window method;

[0045] Fig.11 This is a comparison chart of the evaluation indicators of the four prediction models CNN-GRU, DA-CNN-GRU, TA-CNN-GRU and SA-CNN-GRU in the #1 dataset;

[0046] Fig.12 This is a comparison chart of the evaluation indicators of the four prediction models CNN-GRU, DA-CNN-GRU, TA-CNN-GRU and SA-CNN-GRU in the #2 dataset;

[0047] Fig.13This is a comparison chart of the evaluation indicators of the four prediction models CNN-GRU, DA-CNN-GRU, TA-CNN-GRU and SA-CNN-GRU under the #3 dataset. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] The present invention is a circuit breaker remaining life prediction method (hereinafter referred to as method) based on a stage attention mechanism network model, comprising the following steps:

[0050] The first step is to use the universal circuit breaker mechanical life test system to collect the vibration signal during the circuit breaker opening process;

[0051] The second step is to optimize the VMD algorithm to obtain the optimal number of modal components and quadratic penalty factors; use the optimized VMD algorithm to decompose the vibration signal x(t) to obtain multiple modal components; calculate the kurtosis of each modal component, select the modal component with larger kurtosis for reconstruction, and obtain the reconstructed vibration signal x'(t);

[0052] 2-1 The VMD algorithm is optimized based on the energy loss standard to determine the number of modal components J and the quadratic penalty factor η that make the algorithm optimal; According to the Basseval theorem of formula (1), the time domain energy and E of the continuous signal t Equal to the frequency domain energy and E f , which is the basis for the optimization to be implemented;

[0053]

[0054] Among them, s represents a differential operator, that is, some continuous signal;

[0055] 2-2 Set the screening range of the modal component number J and the quadratic penalty factor η, combine the modal component number and the quadratic penalty factor in order, obtain all possible parameter combinations within the screening range, substitute each set of parameters into the VMD algorithm to decompose the vibration signal, and calculate the energy loss of the vibration signal before and after decomposition;

[0056] Assume that the discrete time sequence of the vibration signal x(t) is x(t) = [x(1), x(2), ..., x(N)], then the energy E of the vibration signal x(t) before decomposition is x for:

[0057]

[0058] The vibration signal is decomposed by the VMD algorithm to obtain J modal components. The energy E of the jth modal component is calculated according to formula (3): j ;

[0059]

[0060] Where x(n) is the value of the nth sampling point of the vibration signal x(t), n = 1, 2, ..., N, N is the number of sampling points; u j (n) is the value of the nth sampling point of the jth modal component, j = 1, 2, ..., J;

[0061] The energy loss E before and after the vibration signal decomposition is calculated by formula (4): loss :

[0062]

[0063] Ideally, there is no energy loss in the vibration signal before and after the VMD algorithm decomposition. In fact, due to the algorithm itself, there will be energy leakage when there is no over-decomposition, and the energy loss is greater than 0 at this time; when over-decomposition occurs, a certain frequency component will appear in multiple modal components, and the energy loss is less than 0 at this time; therefore, when the energy loss is close to 0, the VMD algorithm reaches the optimal value, and the corresponding modal component number J and the quadratic penalty factor η are selected as the optimal parameters of the VMD algorithm to complete the optimization of the VMD algorithm;

[0064] 2-3 The vibration signal x(t) is decomposed into multiple modal components using the optimized VMD algorithm, and the kurtosis of each modal component is calculated. The kurtosis can reflect the impact characteristics of the vibration signal and is suitable for vibration signals generated by collisions of various mechanisms. The modal components with larger kurtosis are selected for reconstruction, which plays a role in improving the signal-to-noise ratio of the vibration signal and highlighting the effective impact information, and the reconstructed vibration signal x'(t) is obtained.

[0065] The third step is to extract the contact breaking vibration segment from the reconstructed vibration signal; first, the reconstructed vibration signal is framed to obtain a multi-frame reconstructed signal; then, the energy entropy ratio of all frame reconstructed signals is calculated, and the short-time energy entropy ratio double threshold method is used to extract the contact breaking vibration segment;

[0066] 3-1 Use the window function to divide the reconstructed vibration signal x'(t) into frames to obtain a multi-frame reconstructed signal; in order to eliminate high-frequency interference and reduce leakage energy, the Hamming window is selected, and its expression is:

[0067] ω(v)=0.54-0.46cos(2πv / (λ-1)) (5)

[0068] Where λ is the window length, v represents the sequence number of the sampling point in the Hamming window, v = 0, 1, ..., λ-1;

[0069] The expression of the framing operation is:

[0070] z f (v)=ω(v)*x'((f-1)*inc+v) (6)

[0071] In the formula, z f (v) represents the value of the reconstructed signal of any frame, f = 1, 2, ..., F, F is the total number of frames, inc is the frame shift;

[0072] 3-2 Calculate the energy entropy ratio of the reconstructed signal of all frames, where the energy entropy ratio G of the reconstructed signal of the fth frame is calculated according to formula (7): f ;

[0073]

[0074] In the formula, E f , H f are the energy and spectral entropy of the reconstructed signal of the f-th frame respectively;

[0075] 3-3 The energy entropy ratio of all frame reconstructed signals forms an energy entropy ratio envelope curve along the frame sequence direction. On this basis, the double threshold method is used to detect the two endpoints of the contact break vibration segment, and the contact break vibration segment is extracted from the entire reconstructed vibration signal; the specific process is: set the lower threshold Q1 and upper threshold Q2 for two-level judgment, first use the upper threshold Q2 to make the first-level judgment, compare the energy entropy ratio of each point on the energy entropy ratio envelope curve with the upper threshold Q2, and it is considered that the signal with an energy entropy ratio not less than the upper threshold Q2 must belong to the contact break vibration segment, and the first-level judgment is determined. signal interval; then use the lower limit threshold Q1 to continue judging outside the interval determined by the first-level judgment, find the left and right signal points whose energy entropy ratio is equal to the lower limit threshold Q1, and regard these two signal points as the starting point and end point of the contact breaking vibration segment. The signal between these two signal points is the contact breaking vibration segment; the vibration signal before the starting point is mainly generated by the vibration of the opening switch accessories, and the vibration signal after the end point is mainly some invalid noise. These two parts of vibration cannot effectively reflect the overall mechanical state of the circuit breaker, so they are discarded, and the contact breaking vibration segment is separated as the signal source for life prediction;

[0076] 3-4 constructing a data set, the data set includes multiple input samples in a two-dimensional format, each input sample includes multiple contact break vibration segments, that is, the number of rows of the input sample; each contact break vibration segment includes multiple sampling points, that is, the number of columns of the input sample;

[0077] Firstly, the peak-to-peak characteristics of the vibration signal are extracted from the contact breaking vibration segment. The peak-to-peak trend reflects the nonlinear degradation trend of the circuit breaker throughout its life. Therefore, the point where the peak-to-peak characteristic changes significantly is regarded as the inflection point of the circuit breaker performance degradation, that is, the dividing point between the stable period and the degradation period. Through this inflection point, the circuit breaker degradation process is divided into two stages: the stable period and the degradation period. The change trends of the stable period and the degradation period are divided into two linear RUL label functions.

[0078] Then, using the contact-broken vibration fragments, a two-dimensional input sample is constructed by a sliding window method, wherein the window length is the number of time steps contained in an input sample, which is also the number of contact-broken vibration fragments, the window width is the number of sampling points contained in each contact-broken vibration fragment, and the window sliding direction is the time step direction;

[0079] Suppose any input sample is S k , then S k The expression is as follows:

[0080]

[0081] Wherein, m represents the mth contact break vibration segment in the input sample, that is, the contact break vibration segment at the mth moment, l represents the lth sampling point of the mth contact break vibration segment; M represents the number of contact break vibration segments or the number of time steps contained in the input sample, and L represents the number of sampling points contained in each contact break vibration segment, which is also the characteristic number;

[0082] Step 4: Establish a prediction model that integrates the stage attention mechanism. The prediction model uses a one-dimensional convolutional neural network and a GRU network as the backbone network. The stage attention mechanism is divided into two stages. The first stage is a distributed attention mechanism applied to the one-dimensional convolutional neural network, and the second stage is a time step attention mechanism applied to the GRU network. Train the prediction model and use the pre-trained prediction model to predict the remaining life of the circuit breaker.

[0083] 4-1 Using a deep one-dimensional convolutional neural network to perform convolution processing on the input sample to achieve dimensionality reduction and feature extraction of the input sample; in this embodiment, the one-dimensional convolutional neural network includes four convolutional layers, each convolutional layer includes three convolutional kernels, and each convolutional kernel performs horizontal sliding convolution on multiple time step data of the input sample;

[0084] The expression of convolution operation is:

[0085] H k =σ1(W k *S k +b k ) (9)

[0086] In the formula, H k represents the output of the convolutional layer, * represents a one-dimensional convolution operation with depth, and W k and b k They represent the weight matrix and bias matrix of the convolutional layer respectively, σ1 is the activation function, and the Relu activation function is used;

[0087] According to formula (10), the output H of the convolutional layer k Perform feature screening and dimensionality reduction:

[0088] V k = p(H k ) (10)

[0089] In the formula, p represents the maximum pooling operation, V k is the feature matrix after dimensionality reduction;

[0090] 4-2 The output matrix of the second convolutional layer is used as the input matrix of the distributed attention mechanism. First, the input matrix is ​​weighted in the time dimension to obtain the time-weighted weights of all row vectors in the input matrix. The time-weighted weights of all row vectors constitute a time-weighted column vector. The time-weighted column vector is multiplied by the corresponding elements of the original input matrix to obtain a time-weighted matrix.

[0091] First, assuming that the input matrix size of the distributed attention mechanism is M'×L', the i-th row vector of the input matrix is ​​represented as x i =[x i1 ,x i2 ,...,x iL' ], i = 1, 2, ..., M', then according to equations (11) and (12) the row vector x i Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling time weight coefficient a i and the global average pooling time weight coefficient b i Similarly, after performing the same operation on all row vectors of the input matrix, we obtain M' groups of one-to-one corresponding global maximum pooling time weight coefficients and global average pooling time weight coefficients. Finally, we calculate the row vector x through the dense layer and softmax activation function of formula (13). i The time weighted weight α i Similarly, the time-weighted weights of all row vectors in the input matrix are calculated to obtain the time-weighted weights of M' row vectors. The time-weighted weights of these M' row vectors constitute the time weight column vector α, α = [α1, α2, ..., α M' ] T; Multiply the time weight column vector α by the corresponding elements of the input matrix of the distributed attention mechanism to obtain the time weighted matrix, that is, to achieve the time weighting of the input matrix of the distributed attention;

[0092] a i =globalmax(x i )=globalmax([x i1 ,x i2 ,...,x iL' ]) (11)

[0093] b i =globalaverage(x i )=globalaverage([x i1 ,x i2 ,...,x iL' ]) (12)

[0094] α i =softmax((w a *a i +bias a )+(w b *b i +bias b )) (13)

[0095] In the formula, w a 、w b are the convolution kernel matrices of the global maximum pooling and global average pooling in the time-weighted process, bias a 、bias b They are the bias matrices of global maximum pooling and global average pooling in the time weighted process respectively;

[0096] Then, the time weighting matrix is ​​further weighted in the feature dimension. Assume that the e-th column vector of the time weighting matrix is ​​x e =[x 1e ,x 2e ,...,x M'e ] T , e=1,2,...,L', then according to equations (14) and (15) the column vector x e Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling feature weight coefficient c e and the global average pooling feature weight coefficient d eSimilarly, after performing the same operation on all column vectors of the time weighted matrix, we obtain the global maximum pooling feature weight coefficient and the global average pooling feature weight coefficient corresponding to each group L'. Finally, we calculate the column vector x by the dense layer and softmax activation function shown in formula (16): e The feature weight β e ; Similarly, calculate the feature weighted weights of all column vectors in the time weighted matrix to obtain the feature weighted weights of L' column vectors, and the feature weighted weights of these L' column vectors constitute the feature weight row vector β; multiply the feature weight row vector by the corresponding elements of the original time weighted matrix to obtain the feature weighted matrix;

[0097] c e =globalmax([x 1e ,x 2e ,...,x M'e ] T ) (14)

[0098] d e =gobalaverage([x 1e ,x 2e ,...,x M'e ] T ) (15)

[0099] β e =softmax(w e *[c e ;d e ]+bias e ) (16)

[0100] Among them, w e is the convolution kernel matrix of the feature weighting process, bias e is the bias matrix for the feature weighting process;

[0101] The output matrix of the second convolutional layer is weighted in the time dimension and feature dimension in turn, thus achieving the distribution of weights in two dimensions.

[0102] Finally, the feature weight matrix is ​​added to the input matrix of the distributed attention mechanism to complete the entire input sample S k The weighted matrix is ​​used as the input matrix of the third convolutional layer. After passing through the third and fourth convolutional layers, the output matrix of the one-dimensional convolutional neural network is obtained. The input sample S is subjected to distributed attention. k The important information is given a higher weight, taking into account the importance of the time dimension and feature dimension of the vibration signal;

[0103] 4-3 The output matrix of the one-dimensional convolutional neural network is used as the input matrix of the GRU network, and the GRU network is used to learn time series features and predictions; the output matrix of the GRU network is further weighted in the time step using the time step attention mechanism to enhance the importance of the time step information;

[0104] First, according to formula (17), the scoring function Φ is used to score the time step information corresponding to all moments in the output matrix of the one-dimensional convolutional neural network and calculate the weight probability; where the time step information h corresponding to the m'th moment in the output matrix of the one-dimensional convolutional neural network is m' score m' and weighted probability δ m' They are:

[0105] score m' =Φ(W T h m' +bias Φ ) (17)

[0106] Where W T is the weight matrix of the score function, bias Φ is the bias matrix of the score function;

[0107] δ m' =softmax(score m' ) (18)

[0108] The weight probabilities at all times constitute the weight column vector δ, which is combined with the output matrix X of the GRU network GRU According to formula (19), the corresponding elements are multiplied to obtain the time step attention weighted result g, realizing the secondary weighting of the time step and completing the second stage of the stage attention mechanism;

[0109]

[0110] in, Indicates the corresponding element multiplication operation;

[0111] Finally, the time-step attention weighted result g passes through the fully connected layer to obtain the output of the prediction model, that is, the predicted value y. This completes the construction of the prediction model integrating the attention mechanism. The expression of the fully connected layer is:

[0112] y=w D *g+bias D (20)

[0113] In the formula, w D is the weight matrix of the fully connected layer, bias D is the bias matrix of the fully connected layer;

[0114] 4-4 Train and optimize the prediction model of the attention mechanism in the fusion stage, and use the pre-trained prediction model to predict the remaining mechanical life of the circuit breaker.

[0115] Example

[0116] The circuit breaker remaining life prediction method based on the stage attention mechanism network model of this embodiment includes the following steps:

[0117] The first step is to use the universal circuit breaker mechanical life test system to collect the vibration signal during the circuit breaker opening process; Figure 3 It is a structural diagram of the mechanical life test system for universal circuit breakers. In this embodiment, three DW15-1600 universal circuit breakers are used as test objects, and the vibration signals collected at the crossbeams of the circuit breakers are used to predict the remaining mechanical life of the circuit breakers. According to the relevant standards of the circuit breakers, the test operation frequency is set to 20 times / h, and the sampling frequency of the USB-7648A data acquisition card is set to 20kHz. Considering that the specified service life of the circuit breaker is 4500 times, the number of vibration signal groups collected during the opening process of the three circuit breakers are 4447, 4466 and 4459 respectively, including three data sets #1, #2 and #3 in total. Since there is a running-in stage in the early stage of the circuit breaker, the early vibration signals are partially discarded, and the subsequent 4180 groups of vibration signals are retained.

[0118] Step 2: First, optimize the VMD algorithm, set the screening range of the modal component number J to [3, 10] and the screening range of the quadratic penalty factor η to [50, 3000], and combine the modal component number and the quadratic penalty factor in sequence. Each time the modal component is combined, the increment is 1, and the increment of the quadratic penalty factor is 50, so as to obtain all possible parameter combinations within the screening range. Substitute each set of parameters into the VMD algorithm in turn to decompose the vibration signal, and calculate the energy loss of the vibration signal before and after decomposition, as shown in Figure 4 As shown in Figure 1, when the number of modal components is 6 and the quadratic penalty factor is 1800, the minimum energy loss is 0.127, the VMD algorithm reaches the optimal state, and the optimization of the VMD algorithm is completed; secondly, taking the #1 data set as an example, the vibration signal is decomposed into 6 modal components using the optimized VMD algorithm, see Figure 5 ; Finally, taking the 600th, 2200th and 3800th as examples, the kurtosis of each modal component is calculated, and the three modal components with larger kurtosis are selected for reconstruction to obtain the reconstructed vibration signal, see Figure 7 ,It can be seen from the figure that the reconstructed vibration signal shows more significant impact characteristics.

[0119] The third step is to select the Hamming window function to frame the reconstructed vibration signal to obtain a multi-frame reconstructed signal; by adjusting the window length and frame shift, the energy entropy ratio envelope curve can fully reflect the impact of each vibration event in the opening process. The appropriate window length and frame shift can not only reflect the tiny impact in the vibration, but also prevent the introduction of unnecessary interference or excessive energy leakage due to excessive calculation. After debugging, the window length is determined to be 40 and the frame shift inc=20; the energy entropy ratio of all frame reconstructed signals is calculated, and the short-time energy entropy ratio double threshold method is used to extract the contact breaking vibration segment; wherein, the lower limit threshold Q1=1.017, the upper limit threshold Q2=1.021; the length of the contact breaking vibration segment extracted in this embodiment is 680, that is, the number of feature points of each segment is 680, and the obtained Figure 8 Results shown;

[0120] Extract the peak-to-peak characteristics of the contact breaking vibration segment and obtain Fig. 9 Based on the results, the point corresponding to the obvious change of the peak-to-peak characteristic is regarded as the inflection point of circuit breaker performance degradation, that is, the dividing point between the stable period and the degradation period. The circuit breaker degradation process is divided into two stages: the stable period and the degradation period through this dividing point.

[0121] The sliding window method is used to construct the input samples in two-dimensional format. The schematic diagram of the sliding window can be found in Fig.10 The time step of each input sample is 12, that is, it contains 12 contact break vibration segments.

[0122] Step 4: Establish a prediction model that integrates the stage attention mechanism. The prediction model uses a one-dimensional convolutional neural network and a GRU network as the backbone network. The stage attention mechanism is divided into two stages. The first stage is a distributed attention mechanism applied to a one-dimensional convolutional neural network, and the second stage is a time step attention mechanism applied to a GRU network. Train the prediction model and use the trained prediction model to predict the remaining life of the circuit breaker.

[0123] The one-dimensional convolutional neural network contains four convolutional layers, and the number of neurons in each convolutional layer is 64, 32, 16, and 12 respectively; each convolutional layer contains three convolution kernels, the convolution kernel size is 3×1, and the convolution kernel movement step is 1; the second and fourth convolutional layers are followed by maximum pooling layers, and the pooling window sizes are 4×1 and 3×1 respectively;

[0124] The GRU network contains two GRU layers, with 32 and 16 neurons in each GRU layer, respectively. To prevent overfitting, the output of the GRU layer is dropout, with the dropout rate parameter set to 0.2 and the activation function as Relu. Since the GRU network is lightweight, it is more conducive to expansion into a complex hybrid network than the LSTM network. It is used as the backend of the entire prediction model to learn the prominent features in the output matrix of the one-dimensional convolutional neural network.

[0125] The prediction model is trained and optimized. The Adam algorithm is used for parameter optimization. The learning rate is set to 0.001, the batch size is 128, the epoch is 50, and the loss function is MSE. The pre-trained prediction model is used to predict the remaining mechanical life of the circuit breaker.

[0126] The performance of the prediction model is evaluated by the root mean square error (RMSE) and the mean absolute error (MAE). Both evaluation indicators can reflect the accuracy of the prediction results of the remaining mechanical life of the circuit breaker. The smaller the RMSE and MAE, the higher the prediction accuracy. The RMSE can better reflect abnormal fluctuations, while the MAE can better reflect the actual situation of the prediction. The prediction model (SA-CNN-GRU) of the present invention and the common DCNN, LETM, GRU, BiLSTM and CNN-LSTM models in the prior art are used to predict the remaining mechanical life of the circuit breaker for three data sets, and the results shown in Table 1 are obtained. As can be seen from the table, the RMSE and MAE of the prediction model of the present invention on the three data sets are smaller than those of other models, indicating that the prediction model of the present invention has higher prediction accuracy and stronger generalization ability. Compared with the prediction models of single structures such as DCNN and GRU, the prediction effect of the SA-CNN-GRU model is greatly improved, which is due to the better information mining and learning ability of the attention mechanism in the fusion stage of the present invention.

[0127] Table 1 Performance evaluation results of different models on three datasets

[0128]

[0129]

[0130] The RMSE and MAE of the SA-CNN-GRU model were quantitatively compared with those of the DCNN, LSTM, and CNN-LSTM models, and the RMSE and MAE of the four models on the three data sets were averaged; compared with the DCNN, LSTM, and CNN-LSTM models, the average RMSE of the SA-CNN-GRU model was reduced by 40.91%, 31.11%, and 26.75%, respectively, and the average MAE was reduced by 41.54%, 36.10%, and 26.63%, respectively, indicating that the prediction model of the present invention has high prediction accuracy and is suitable for solving the problem of remaining life prediction of complex mechanical systems such as circuit breakers. Compared with the above conventional prediction models, its prediction accuracy has been greatly improved, indicating that the prediction model of the present invention has significant advantages in life prediction.

[0131] In order to further verify the weight contribution effect of the stage attention mechanism proposed in the present invention, the distributed attention mechanism in the stage attention mechanism is added to the one-dimensional convolutional neural network of the CNN-GRU model to obtain the DA-CNN-GRU model; the time step attention mechanism is added to the GRU network of the CNN-GRU model to obtain the TA-CNN-GRU model, and these two models are subjected to ablation tests with the CNN-GRU model and the SA-CNN-GRU model of the present invention, and the results shown in Table 2 are obtained.

[0132] Table 2 Ablation test results of the stage attention mechanism on three datasets

[0133]

[0134] The test results in Table 2 show that the addition of the distributed attention mechanism and the time-step attention mechanism both play a role in improving the prediction accuracy of the model. The prediction results of the two models integrating the single attention mechanism are relatively close, indicating that the weight contribution of the two attention mechanisms to the prediction results is similar. However, the SA-CNN-GRU model proposed in the present invention has the smallest prediction error. This is because the model integrates the advantages of the above two attention mechanisms and further strengthens the contribution of important information to the prediction results, so the improvement of the prediction effect is more obvious.

[0135] To further compare the prediction performance of the four models, CNN-GRU, DA-CNN-GRU, TA-CNN-GRU, and SA-CNN-GRU, continuous experiments were conducted on each model. The RMSE and MAE of 10 consecutive normal experiments on three data sets were obtained, and violin plots were drawn. Figures 11 to 13These are the violin plots corresponding to data set #1, data set #2, and data set #3. The violin plots are used to represent the RMSE and MAE distribution of the four models. The rectangle in the middle of each graph represents the range of the quartile data, the horizontal line in the rectangle represents the median of the data, and the thickness of the graph outline reflects the density of the data distribution. As can be seen from the figure, after multiple experiments, the SA-CNN-GRU model still showed the best prediction accuracy. The graph outline corresponding to the SA-CNN-GRU model is short and wide, indicating that the distribution of the prediction results is more concentrated. This feature shows that the model has a high prediction stability and the probability of abnormal prediction in the application is small.

[0136] The data acquisition software of this embodiment is partially implemented based on LabVIEW language, the data processing part is implemented based on Matlab language, the life prediction part is implemented based on Python language, the deep learning environment is built based on Tensorflow+Keras, and the software platform is Pycharm. The LabVIEW, Matlab and Python languages ​​used in this embodiment are well known to those skilled in the art. The percentages in this embodiment are all digital percentages.

[0137] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A circuit breaker remaining life prediction method based on a stage attention mechanism network model, characterized in that: The method comprises the following steps: The first step is to use the universal circuit breaker mechanical life test system to collect the vibration signal during the circuit breaker opening process; The second step is to optimize the VMD algorithm based on energy loss to obtain the optimal number of modal components and quadratic penalty factor; use the optimized VMD algorithm to decompose the vibration signal to obtain multiple modal components; calculate the kurtosis of each modal component, select the modal component with larger kurtosis for reconstruction, and obtain the reconstructed vibration signal; The third step is to divide the reconstructed vibration signal into frames to obtain a multi-frame reconstructed signal; calculate the energy entropy ratio of all frame reconstructed signals, and extract the contact break vibration segment from the entire reconstructed vibration signal according to the energy entropy ratio; construct a data set based on the contact break vibration segment, the data set includes multiple two-dimensional input samples, each input sample includes multiple contact break vibration segments, and each contact break vibration segment includes multiple sampling points; The fourth step is to establish a prediction model of the fusion stage attention mechanism. The prediction model uses a one-dimensional convolutional neural network and a GRU network as the backbone network. The stage attention mechanism is divided into two stages. The first stage is a distributed attention mechanism applied to a one-dimensional convolutional neural network, and the second stage is a time step attention mechanism applied to a GRU network. The one-dimensional convolutional neural network contains four convolutional layers, each of which contains three convolutional kernels, and each convolutional kernel performs horizontal sliding convolution on multiple time-step data of the input sample; The output matrix of the second convolutional layer is used as the input matrix of the distributed attention mechanism. First, the input matrix is ​​weighted in the time dimension to obtain the time-weighted weights of all row vectors in the input matrix. The time-weighted weights of all row vectors constitute a time-weighted column vector. The time-weighted column vector is multiplied by the corresponding elements of the original input matrix to obtain a time-weighted matrix. First, assuming that the input matrix size of the distributed attention mechanism is M'×L', the i-th row vector of the input matrix is ​​represented as x i =[x i1 ,x i2 ,...,x iL' ], i = 1, 2, ..., M', then according to equations (11) and (12) the row vector x i Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling time weight coefficient a i and the global average pooling time weight coefficient b i Similarly, after performing the same operation on all row vectors of the input matrix, we obtain M' groups of one-to-one corresponding global maximum pooling time weight coefficients and global average pooling time weight coefficients. Finally, we calculate the row vector x through the dense layer and softmax activation function of formula (13): i The time weighted weight α i ; Similarly, calculate the time-weighted weights of all row vectors in the input matrix to obtain the time-weighted weights of M' row vectors, and the time-weighted weights of these M' row vectors constitute the time-weighted column vector; multiply the corresponding elements of the time-weighted column vector with the input matrix of the distributed attention mechanism to obtain the time-weighted matrix, that is, to realize the time-weighted input matrix of the distributed attention; a i =globalmax(x i )=globalmax([x i1 ,x i2 ,...,x iL' ]) (11) b i =globalaverage(x i )=globalaverage([x i1 ,x i2 ,...,x iL' ]) (12) a i =softmax((w a *a i +bias a )+(w b *b i +bias b )) (13) In the formula, w a 、w b are the convolution kernel matrices of the global maximum pooling and the global average pooling in the time weighted process, bias a 、bias b They are the bias matrices of global maximum pooling and global average pooling in the time weighted process respectively; Then, the time weighting matrix is ​​further weighted in the feature dimension. Assume that the e-th column vector of the time weighting matrix is ​​x e =[x 1e ,x 2e ,...,x M'e ] T , e=1,2,...,L', then according to equations (14) and (15) the column vector x e Perform global maximum pooling and global average pooling respectively to obtain the global maximum pooling feature weight coefficient c e and the global average pooling feature weight coefficient d e Similarly, after performing the same operation on all column vectors of the time weighted matrix, we obtain the global maximum pooling feature weight coefficient and the global average pooling feature weight coefficient corresponding to each other in the L' group. Finally, we calculate the column vector x by the dense layer and softmax activation function shown in formula (16): e The feature weight β e ; Similarly, calculate the feature weighted weights of all column vectors in the time weighted matrix to obtain the feature weighted weights of L' column vectors, and the feature weighted weights of these L' column vectors constitute the feature weight row vector; multiply the feature weight row vector by the corresponding elements of the original time weighted matrix to obtain the feature weight matrix; c e =globalmax([x 1e ,x 2e ,...,x M'e ] T ) (14) d e =gobalaverage([x 1e ,x 2e ,...,x M'e ] T ) (15) b e =softmax(w e *[c e ;d e ]+bias e ) (16) Among them, w e is the convolution kernel matrix of the feature weighting process, bias e is the bias matrix for the feature weighting process; Finally, the feature weighted matrix is ​​added to the input matrix of the distributed attention mechanism, and the added matrix is ​​used as the input matrix of the third convolutional layer. After passing through the third and fourth convolutional layers in sequence, the output matrix of the one-dimensional convolutional neural network is obtained; The output matrix of the one-dimensional convolutional neural network is used as the input matrix of the GRU network. The scoring function is used to score the time step information corresponding to all moments in the output matrix of the one-dimensional convolutional neural network and calculate the weight probability. The weight probabilities at all moments constitute a weight column vector. The weight column vector is multiplied by the corresponding elements of the output matrix of the GRU network to obtain the time step attention weighted result. Finally, the time step attention weighted result passes through the fully connected layer to obtain the output of the prediction model, that is, the prediction value, thus completing the construction of the prediction model integrating the attention mechanism. The prediction model of the attention mechanism in the fusion stage is trained, and the pre-trained prediction model is used to predict the remaining mechanical life of the circuit breaker.

2. The circuit breaker remaining life prediction method based on the stage attention mechanism network model according to claim 1 is characterized in that: The specific process of VMD algorithm optimization in the second step is: set the screening range of the number of modal components and the quadratic penalty factor, combine the number of modal components and the quadratic penalty factor in sequence, increase the number of modal components and the quadratic penalty factor by a certain amount each time, obtain all possible parameter combinations within the screening range, substitute each set of parameters into the VMD algorithm to decompose the vibration signal, and calculate the energy loss of the vibration signal before and after decomposition. When the energy loss is close to 0, the number of modal components and the quadratic penalty factor have optimal values, and the optimization of the VMD algorithm is completed.

3. The circuit breaker remaining life prediction method based on the stage attention mechanism network model according to claim 1 is characterized in that: In the third step, the energy entropy ratio of all frame reconstructed signals forms an energy entropy ratio envelope curve along the frame sequence direction, and the lower threshold and upper threshold for two-level judgment are set. First, the upper threshold is used to perform the first-level judgment, and the energy entropy ratio of each point on the energy entropy ratio envelope curve is compared with the upper threshold. It is considered that the signal whose energy entropy ratio is not less than the upper threshold must belong to the contact break vibration segment, so as to obtain the signal interval determined by the first-level judgment; then the lower threshold is used to continue to judge outside the interval determined by the first-level judgment, and the left and right signal points whose energy entropy ratio is equal to the lower threshold are found. These two signal points are regarded as the starting point and end point of the contact break vibration segment, and the vibration signal between the two signal points is the contact break vibration segment, and the contact break vibration segment is used as the signal source for predicting the remaining life of the circuit breaker.

4. The circuit breaker remaining life prediction method based on the stage attention mechanism network model according to claim 1 or 3 is characterized in that: The third step also includes: extracting the peak-to-peak characteristics of the vibration signal from the contact breaking vibration segment, and regarding the point where the peak-to-peak characteristics change significantly as the inflection point of the circuit breaker performance degradation, that is, the dividing point between the stable period and the degradation period. Through this inflection point, the circuit breaker degradation process is divided into two stages: the stable period and the degradation period, and the changing trends of the stable period and the degradation period are divided into two linear RUL label functions.