Contactor electrical life multi-stage prediction method based on clustering analysis and time series network

Abstract

The application is a multi-stage prediction method for contactor electrical life based on clustering analysis and time sequence network, comprising: performing AC contactor life test, and obtaining key features through feature importance and correlation analysis screening; taking the key features of single action of the contactor as a feature mode, and constructing a contact wear factor based on mode similarity measurement; performing fuzzy C-means clustering on the contact wear factor, dividing the contactor life test process into multiple degradation stages, taking the divided degradation stages as training labels, and taking the key features as inputs to train the degradation stage online evaluation model; taking the key features of each degradation stage and the corresponding residual life to form a stage data set, training a parallel time sequence network, and obtaining a residual life prediction model of each stage to predict the residual life of the contactor. The characteristics of the multi-stage performance degradation process of the contactor and the problem of the relatively one-sided feature extraction of a single model are comprehensively considered, and the accuracy of the life prediction is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of alternating current contactor residual life prediction, and specifically relates to a contactor electrical life multi-stage prediction method based on clustering analysis and time sequence network. BACKGROUND

[0002] As a common switch in the power distribution system, the contactor plays a role in frequently turning on and off the main circuit and controlling the large capacity circuit. In recent years, with the rapid development of direct current power supply related industries such as electric vehicles, charging piles, energy storage devices and photovoltaic power generation, the demand for contactors has also increased. As an indispensable key component in the power supply system, the contactor needs to meet high reliability and safety. With the increase of the number of actions, the contacts of the switch will be worn to varying degrees, resulting in gradual reduction of its performance and reliability. Once the contactor fails, it will cause huge economic losses and even casualties. Therefore, it is particularly important to deeply study the performance degradation law of the contactor and then build a more accurate and practical working condition fitting life prediction model to predict the residual life of the contactor.

[0003] In the aspect of switch residual life prediction, most of the studies on the life prediction of the contactor are full life prediction. For example, Xing Chaoken et al. (Xing Chaoken, Liu Shuxin, Gao Shuyu, Liu Yang, Li Jing, Cao Yundong. Residual electrical life prediction of alternating current contactor based on data enhancement SDAE-BiGRU[J]. High Voltage Technology, 2024, 50(11): 4990-5004.) regard the life cycle of the contactor as a long time sequence, build a contactor life prediction model based on BiGRU, and deeply mine the bidirectional dependence relationship of multi-dimensional feature parameters in the time sequence dimension. Although the above method realizes relatively accurate prediction of the residual life of the alternating current contactor, it regards the degradation process as a single stage, and uses the same model to predict from the initial stage to the end stage, ignoring the fact that the degradation process of the switch presents a staged characteristic. For the switch with a long service life such as the alternating current contactor, the above method obviously has certain defects. SUMMARY

[0004] In view of the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a contactor electrical life multi-stage prediction method based on clustering analysis and time sequence network. The method comprehensively considers the multi-stage characteristics of the performance degradation process of the contactor and the problem of one-sidedness of feature extraction of a single model, builds a residual life multi-stage prediction model based on parallel time sequence network on the basis of reasonable division of the degradation stages, and makes targeted prediction of the residual life of the contactor in different degradation stages, so as to improve the accuracy of life prediction.

[0005] The technical problem solved by the present application adopts the following technical scheme:

[0006] A contactor electrical life multi-stage prediction method based on clustering analysis and time sequence network, comprising the following contents:

[0007] Performing AC contactor life test, obtaining electrical signals in the contactor life action process and extracting original features and permutation entropy of the original features, and screening key features capable of effectively representing contactor performance degradation through feature importance and correlation analysis;

[0008] Taking the key features of single action of the contactor as a feature mode, constructing a contact wear factor based on mode similarity measurement: normalizing all key features, and taking the mean value of the first 10 normalized feature modes as a base mode M b ; calculating the weighted Euclidean distance between the normalized feature mode and the base mode, and forming a distance sequence in the order of action times; I : dividing the distance sequence into multiple subsequences with a length of 10, calculating the mean value of the weighted Euclidean distance in each subsequence, and taking the mean value of the weighted Euclidean distance of all subsequences as the contact wear factor;

[0009] Performing fuzzy C-means clustering on the contact wear factor, dividing the contactor life test process into multiple degradation stages, taking the divided degradation stages as training labels, and training a degradation stage online evaluation model with the key features as inputs, the degradation stage online evaluation model being used for predicting the degradation stage of the key features of single action;

[0010] Taking the key features of each degradation stage and the corresponding remaining life as a stage data set, training a parallel time sequence network with the stage data set, and obtaining a remaining life prediction model of each stage;

[0011] Predicting the degradation stage to which the key features belong by the degradation stage online evaluation model, and predicting the remaining life of the contactor by the remaining life prediction model of the corresponding stage.

[0012] Further, the parallel time sequence network takes Bi-LSTM and Bi-TCN as a double-branch network, and realizes dynamic weighted fusion of output features of the double-branch network by using an attention-based feature fusion module.

[0013] Further, the attention-based feature fusion module includes using average pooling to compress the dimensions of the output features of the Bi-TCN network and the Bi-LSTM network, calculating feature weights for feature fusion, and mapping the fused features to RUL values through a fully connected layer.

[0014] Further, the key features are contact pressure drop A1, release voltage A2, cumulative arcing energy A7, permutation entropy of contact pressure drop A8, permutation entropy of release voltage A9, permutation entropy of attraction time A10 and permutation entropy of contact resistance A11.

[0015] Further, the feature importance analysis is realized by an XGBoost algorithm, in the process of constructing a tree model of the XGBoost algorithm, the XGBoost algorithm adopts a greedy strategy to search for an optimal split point in a feature space, if a certain feature is used more times in the tree model splitting, it is indicated that the importance of the feature in the modeling process is higher, and the splitting times of each feature are recorded.

[0016] In the weighted Euclidean distance calculation, the weight is calculated according to the splitting times of the feature.

[0017] Further, the degradation stage online evaluation model is realized by using a Transformer architecture, and one-hot encoding is used to encode the degradation stage.

[0018] Compared with the prior art, the present application has the following beneficial effects:

[0019] (1) In view of the problem that the contactor performance degradation process is difficult to intuitively and accurately measure, a contact wear factor is constructed based on modal similarity measurement to represent the contactor performance degradation process. Permutation entropy is introduced to nonlinearly reconstruct the original features, and the original features and their permutation entropy are used as full information features. Through feature importance and correlation analysis (XGBoost algorithm and Spearman algorithm) on the full information features, irrelevant and redundant features are removed, and key features that can effectively represent the contactor performance degradation are obtained. Then, the key features under single action of the contactor are used as a feature mode, and the weighted Euclidean distance between the feature mode and the base mode is calculated to construct the contact wear factor. The results show that the contact wear factor is monotonous in the whole life cycle, and the contact wear factor can clearly show that the contactor performance degradation presents a multi-stage characteristic.

[0020] (2) In the present application, the contactor performance degradation process has the characteristics of fuzzy boundary, multi-state and multi-factor interaction. Based on unsupervised clustering and a Transformer model, the degradation stage is evaluated. The FCM algorithm is used to unsupervisedly divide the contact wear factor into multiple sections, i.e. multiple stages of the contactor performance degradation, and a Transformer model is used to online evaluate the contactor degradation stage, realizing the mapping between the key features and the degradation stage. The average accuracy of the model in the online evaluation of the alternating current contactor degradation stage is 90.23%, and the average F1-score is 88.92%.

[0021] (3) To address the multi-stage nature of contactor performance degradation, a multi-stage remaining lifetime prediction model based on parallel time-series networks was established. This model includes two parallel branch networks, Bi-LSTM and Bi-TCN, and also introduces an attention-based feature fusion module to achieve dynamic weighted fusion of the output features of the two branch networks, solving the problems that traditional TCN cannot utilize future time features and LSTM cannot effectively learn long-term sequence features. In the multi-stage prediction of AC contactor remaining lifetime, the minimum RMSE and MAE were only 153.16 and 127.24, respectively. Ablation experiments showed that the multi-stage remaining lifetime prediction model based on parallel time-series networks reduced RMSE by 167.06 and 160.71, respectively, and MAE by 149.5 and 108.54, respectively, in the most critical stage 3, compared to the Bi-LSTM network model and the Bi-TCN network model used alone. Furthermore, compared to the prediction results obtained by the feature fusion mechanism using element-wise addition, the attention-based feature fusion module in this invention reduces the RMSE and MAE by 31.51 and 42.33 times, respectively. Comparing the multi-stage prediction results of the contactor in this application with the overall prediction results, the results show that the RMSE and MAE of the multi-stage prediction for AC contactors are reduced by 29.81% and 31.31%, respectively, compared to the overall prediction. Attached Figure Description

[0022] Figure 1 This is a waveform diagram of voltage and current during a single switching operation of an AC contactor, including contact voltage, contact current, and coil current.

[0023] Figure 2 This is a graph showing the trend of the original characteristics and corresponding permutation entropy as a function of the number of operations during the full life test of AC contactor #AC1.

[0024] Figure 3 The feature importance score for AC contactor #AC1.

[0025] Figure 4 The characteristic correlation coefficient is for AC contactor #AC1.

[0026] Figure 5 The trend of contact wear factor of the constructed AC contactor #AC1 with the number of operations.

[0027] Figure 6 This is a flowchart illustrating an embodiment of the multi-stage prediction method for contactor electrical lifetime based on cluster analysis and temporal networks according to the present invention.

[0028] Figure 7 The degradation stage classification results for AC contactor #AC1.

[0029] Figure 8 This is the confusion matrix.

[0030] Figure 9 For AC1 as the training set, AC2, AC3, AC4, and AC5 are respectively the degradation phase evaluation results of the test set, wherein (a) is the degradation phase evaluation result of AC2 as the test set; (b) is the degradation phase evaluation result of AC3 as the test set; (c) is the degradation phase evaluation result of AC4 as the test set; and (d) is the degradation phase evaluation result of AC5 as the test set.

[0031] Figure 10 The overall structure schematic diagram of the parallel timing network according to an embodiment of the present application.

[0032] Figure 11 For AC1 as the training set, the multi-stage life prediction result of AC2 as the test set, wherein the horizontal coordinate CNA represents the current number of actions, and the vertical coordinate represents the remaining life number; wherein (a) is the life prediction result of stage 0; (b) is the life prediction result of stage 1; (c) is the life prediction result of stage 2; and (d) is the life prediction result of stage 3.

[0033] Figure 12 For AC1 as the training set, the multi-stage life prediction result of AC3 as the test set; wherein (a) is the life prediction result of stage 0; (b) is the life prediction result of stage 1; (c) is the life prediction result of stage 2; and (d) is the life prediction result of stage 3.

[0034] Figure 13 For AC1 as the training set, the multi-stage life prediction result of AC4 as the test set; wherein (a) is the life prediction result of stage 0; (b) is the life prediction result of stage 1; (c) is the life prediction result of stage 2; and (d) is the life prediction result of stage 3.

[0035] Figure 14 For AC1 as the training set, the multi-stage life prediction result of AC5 as the test set; wherein (a) is the life prediction result of stage 0; (b) is the life prediction result of stage 1; (c) is the life prediction result of stage 2; and (d) is the life prediction result of stage 3.

[0036] Figure 15 For AC2 as the training set and AC3 as the test set, the ablation experiment result, wherein (a) is the life prediction result of AC3 at stage 0; (b) is the life prediction result of AC3 at stage 1; (c) is the life prediction result of AC3 at stage 2; and (d) is the life prediction result of AC3 at stage 3.

[0037] Figure 16 The prediction results of the multi-stage prediction and the overall prediction of the AC contactor life are compared, and #AC2 is taken as the training set to predict the residual life of #AC4. DETAILED DESCRIPTION

[0038] The technical solutions of the application will be further described below with reference to the drawings and the embodiments, but this is not intended to limit the protection scope of the present application.

[0039] The contactor electrical life multi-stage prediction method based on clustering analysis and time sequence network includes the following steps:

[0040] First step: Perform AC contactor life test, and calculate original features according to electrical signals in the contactor life action process;

[0041] The collected contactor coil current, contact voltage and contact current data are processed to calculate original features, and the original features are contact pressure drop, release voltage, contact resistance, arc time, arc energy and cumulative arc energy.

[0042] Second step: Extract permutation entropy of the original features as the full information features together with the original features, and then screen out key features capable of effectively representing contactor performance degradation through feature importance and correlation analysis.

[0043] Considering the individuality of the performance degradation of each contactor, the gradual microscopic features should be fully excavated when studying the contact wear process. In order to comprehensively consider the information contained in different features, the permutation entropy of the original features except the cumulative arc energy is also extracted for the analysis of the contactor performance degradation in the embodiments of the application, and the final selected full information features include the original features (contact pressure drop, release voltage, contact resistance, arc time, arc energy, cumulative arc energy) and the permutation entropy of the first six original features. Since different features have different importance in representing the contactor performance degradation process, and there is information redundancy between the features, the feature selection is performed through the feature importance and correlation analysis, so as to obtain the key features capable of effectively representing the contactor performance degradation, which is beneficial to improve the prediction accuracy.

[0044] Third step: Take the key features of a single action of the contactor as a feature mode, and construct a contact wear factor based on the mode similarity measurement. Specifically, first, normalize all the key features, and take the mean value of the first 10 normalized feature modes as the base mode M b; calculate the weighted Euclidean distance between the normalized feature mode and the base mode, and arrange the action times in order to form a distance sequence I; then divide the distance sequence into multiple subsequences with a length of 10, calculate the mean value of the weighted Euclidean distance in each subsequence, and form a contact wear factor by combining the mean values of the weighted Euclidean distances of all subsequences.

[0045] The weighted Euclidean distance is used as the similarity measurement value, the similarity measurement quantifies the proximity between objects through the distance index, and is used for grouping and difference analysis of complex data. The definition of the similarity measurement is that when the sample mode is a feature vector, each sample corresponds to a point in the feature space, and if the feature difference between samples is significant, the distance between them in the feature space will change accordingly. In the present application, the key features of each action of the contactor are regarded as a sample, and as the number of actions increases, the contact wear degree increases, and the distance between samples gradually increases. Through the similarity measurement, a contact wear factor representing the contact wear degree can be constructed, and the contact wear factor also reflects the performance degradation of the contactor.

[0046] Fourth step: offline division of the degradation stage and online evaluation of the degradation stage:

[0047] Offline division of the degradation stage: the fuzzy C-means (FCM) algorithm is used to process the contact wear factor, and the performance degradation of the contactor is divided into multiple degradation stages to provide training labels for the online evaluation model of the degradation stage.

[0048] Online evaluation of the degradation stage: the degradation stage is used as the training label, and the key features of the failed contactor are used as the input to train the online evaluation model of the degradation stage, which is used to predict the degradation stage of the key features of a single action. In this embodiment, the Transformer model is used for online evaluation of the degradation stage of the contactor.

[0049] The offline division of the degradation stage is to use the FCM algorithm to process the contact wear factor of the contactor to divide the performance degradation of the contactor into multiple stages to provide training labels for the subsequent part. The online evaluation of the degradation stage is to train the online evaluation model of the degradation stage through historical failure data and the offline division of the degradation stage label, and after multiple training, the online evaluation model of the degradation stage is obtained. According to the key feature sequence obtained online, the current degradation stage can be identified through the online evaluation model of the degradation stage.

[0050] Fifth step: multi-stage life prediction of RUL (remaining useful life): the key features of each degradation stage and the corresponding remaining life form a stage data set, and a parallel time series network is trained using the stage data set to obtain a remaining life prediction model for each stage; the remaining life prediction models for each stage form a multi-stage prediction model of the remaining life based on the parallel time series network.

[0051] The degradation stage to which the key feature belongs is predicted by the online degradation stage assessment model, and the remaining life of the contactor is predicted by the remaining life prediction model of the corresponding stage.

[0052] To address the multi-stage characteristics of contactor performance degradation and the long-term time-series nature of failure data, a multi-stage remaining lifetime prediction model based on a parallel time-series network is constructed. The model is trained using historical failure data from different stages to predict the remaining lifetime for each degradation stage, and the prediction results provide multi-level support for operation and maintenance decisions. Extensive experiments demonstrate that the multi-stage remaining lifetime prediction model based on the parallel time-series network can achieve more accurate remaining lifetime predictions while maintaining high computational efficiency.

[0053] Example 1:

[0054] This embodiment presents a multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks (hereinafter referred to as the method, see [link]). Figures 1-16 ), including the following steps:

[0055] Step 1: Conduct a full life test of the AC contactor and extract the original characteristics that can characterize contact degradation by using the voltage and current data during the contactor's opening and closing process.

[0056] An AC contactor was selected as the test object. The contactor's rated voltage was 220V, rated current was 12A, and the main circuit load was an inductive load. During the test, a multi-function data acquisition card was used to collect voltage and current signals during the contactor's opening and closing processes at a sampling frequency of 100kHz. The test frequency was 300 times / hour, and the test temperature was maintained at room temperature (20~25℃). Finally, five representative contactors of the same model that failed were selected for further study. The lifespan of each contactor is shown in Table 1.

[0057] Table 1 Contactor lifespan information

[0058]

[0059] Among them, the lifespan is the total number of times the contactor operates from brand new to failure.

[0060] Figure 1 This refers to waveform data acquired during a single on / off operation cycle of an AC contactor. From... T At time 1, the coil of the AC contactor is energized and excited, and the contacts begin to move under the action of electromagnetic force; T 1 to T Time 2 represents the dynamic contact engagement process; from T 2 to T At time 3, the contacts of the AC contactor are in the energized state; T 3 toT The dynamic breaking process begins at time 4, with the contacts separating under the action of the reaction spring; until... T At time 4, the contacts are completely disconnected, and the AC contactor completes one closing and opening operation.

[0061] Figure 2 This figure shows the original characteristics of AC contactor #AC1 and the trend of its corresponding permutation entropy as a function of the number of operations. The horizontal axis of the figure represents the number of operations of the contactor. The first column shows the trend of the original characteristics representing the degradation of the contactor as a function of the number of operations, extracted by processing the contactor voltage and current data. The original characteristics include contact voltage drop (A1), release voltage (A2), pull-in time (A3), contact resistance (A4), arcing time (A5), arcing energy (A6), and cumulative arcing energy (A7). Among them, A7 is the cumulative arcing energy, which is the sum of the arcing energy of each operation. Its trend is strong and is not shown in the figure. Figure 2 The second column shows the trend of permutation entropy feature PE as the number of actions varies for different original features.

[0062] Step 2: Screening key features of the contactor;

[0063] Since the degradation of contact performance due to electrical wear is a nonlinear process, the various features characterizing the degree of wear also exhibit nonlinear characteristics. Entropy is a good nonlinear parameter for characterizing feature changes. Permutation entropy, by comparing and analyzing adjacent sample points in a time series, can effectively extract dynamic feature information contained in the time-series signal. It also has advantages such as simple calculation and immunity to noise, making it suitable for online monitoring. The basic principle of permutation entropy is as follows:

[0064] Let the length be One-dimensional feature sequence for:

[0065]

[0066] Where y is an element in Y.

[0067] For one-dimensional feature sequences Reconstructing the phase space yields the matrix. for:

[0068]

[0069] matrix middle, and These represent the embedding dimension and latency, respectively. Each row is considered a reconstruction component, and the number of reconstruction components is proportional to the number of data points in the system. The length, embedding dimension, and latency are related, and satisfy the following conditions: ,but K is the number of reconstructed components.

[0070] Each reconstructed component is rearranged in ascending order according to the size of its element values, and then the index of each element in the column it belonged to in the reconstructed component before sorting is extracted. This forms a symbol sequence. Any reconstructed component can yield a symbol sequence. G for:

[0071]

[0072] For all reconstructed components, the probability of the symbol sequence appearing is: The probability of a symbol sequence occurring is denoted as . ,and , According to the expression for Shannon's information entropy, the calculation method for the entropy of a feature sequence permutation is defined as follows:

[0073]

[0074] in, For the first The probability of a sequence of symbols appearing; The permutation entropy;

[0075] After normalization, the permutation entropy ranges from [0, 1]. The larger the permutation entropy, the stronger the randomness of the feature sequence in that state.

[0076] The original features (A1-A7) and the permutation entropy (A8-A13) of the original features excluding the accumulated arc energy are extracted as full-information features to analyze the performance degradation of the contactor.

[0077] Regarding feature importance, the analysis primarily examines the correlation strength between features and contactor electrical life, using feature importance to characterize this correlation strength. Feature importance scores are positively correlated with the information value contained in the feature. Regarding feature relevance, the analysis focuses on the redundancy of information between different features. Feature importance analysis is implemented using the XGBoost algorithm, whose core idea is to build a powerful classifier by integrating multiple classification and regression tree models. In constructing the XGBoost tree model, the algorithm employs a greedy strategy to search for the optimal split point in the feature space. The more times a feature is used during tree model splitting, the higher its importance in the modeling process. The number of splits for each feature is recorded, and this number is used as a quantitative indicator of feature importance. Representation. Each feature's The values ​​are recorded not only to indicate the importance of the features, but also for subsequent weighted Euclidean distance calculations.

[0078] Figure 3 Rank the features of AC contactor #AC1 by importance. For #AC1, the pull-in time A3 is... T score The values ​​are relatively small, indicating that this feature has minimal correlation with contactor performance degradation and can be directly discarded. The feature importance of the permutation entropy A12 of arcing time and the permutation entropy A13 of arcing energy is also less than 35, and they are also discarded.

[0079] Taking into account the characteristics of different contactors The scores vary, and to account for individual differences as much as possible and fully preserve the degenerative information in the features, a comprehensive analysis of the five contactor features is conducted. Scoring Results. The top eight features in each contactor are selected. If the selected top eight features appear three or more times in five contactors, then that feature is considered to reflect contactor performance degradation. In this embodiment, the features obtained through feature importance screening are {Contact Voltage Drop A1, Release Voltage A2, Arcing Time A5, Arcing Energy A6, Cumulative Arcing Energy A7, Arrangement Entropy of Contact Voltage Drop A8, Arrangement Entropy of Release Voltage A9, Arrangement Entropy of Pull-in Time A10, Arrangement Entropy of Contact Resistance A11}.

[0080] However, even after feature importance screening, information overlap still exists among the features. Therefore, the Spearman correlation coefficient method is used to measure the feature relevance of the selected features. Spearman can effectively quantify the degree of non-linear correlation between features during feature selection. For features... and Spearman correlation coefficient The calculation formula is:

[0081]

[0082] in, It is the first of features A and B i element and The difference in rank, n It is the length of the characteristic sequence, which is the number of times the contactor operates throughout its entire lifespan. The range is [0, 1]. The larger the value, the higher the degree of information overlap between the two features.

[0083] The Spearman correlation coefficient is used to measure the correlation between features. #AC1 The correlation coefficient results between features are as follows: Figure 4 As shown, there are varying degrees of correlation among the features, which is generally considered... A value greater than 0.6 indicates a strong correlation. Figure 4It can be seen that there is significant information redundancy among contact voltage drop A1, arcing time A5, and arcing energy A6. Analysis of the original characteristics of the five contactors with the number of operations reveals that as the contactors enter the later stages of degradation, the contact voltage drop A1 of all contactors shows a significant increasing trend. Figure 2 The figure shows the original characteristics of contactor #AC1 and the trend of its permutation entropy with the number of operations. It can be seen from the figure that the contact voltage drop A1 increases significantly in the later stages of contactor degradation, and some studies use the contact voltage drop A1 to determine whether a contactor has entered the later stages of its lifespan. Therefore, the information contained in the contact voltage drop A1 is more important. Ultimately, the key features that can effectively characterize contactor performance degradation are {contact voltage drop A1, release voltage A2, accumulated arc energy A7, permutation entropy of contact voltage drop A8, permutation entropy of release voltage A9, permutation entropy of pull-in time A10, and permutation entropy of contact resistance A11}. This includes four permutation entropy features, further demonstrating the effectiveness of the proposed permutation entropy features.

[0084] Step 3: Take the key features of a single contactor action as a feature mode, and construct the contact wear factor based on the modal similarity metric:

[0085] In similarity metrics, Euclidean distance is the most commonly used method, defined as follows: If the key features of two contactor actions are known as follows: and ,in This refers to the number of key features, i.e., the dimension of the feature space. and Euclidean distance for:

[0086]

[0087] As can be seen from the above equation, using Euclidean distance for similarity measurement does not consider the influence weights of key features in the feature modes. Each key feature has different representational meanings and levels of importance for contactor performance degradation. Therefore, this embodiment calculates the weighted Euclidean distance by weighting each key feature. :

[0088]

[0089] In the formula, For the first The weight coefficients of each key feature For the first The first action under the next action One key feature For the first The first action under the next action Key features. Number of splits recorded in eigenvalue importance analysis Calculate according to the following formula:

[0090] ;

[0091] in, Indicates the first The number of splits for each key feature.

[0092] When constructing the contact wear factor, this invention uses the key features of a single contactor operation as a feature mode, and constructs the contact wear factor based on a modal similarity metric. First, all key features are normalized individually. Then, the mean of the top 10 normalized feature modes is taken as the base mode. ,Right now , here For the normalized first Each feature mode, at this time The value is an integer between 1 and 10. The normalization method uses characteristic range normalization. The normalized feature sequence is expressed as follows:

[0093]

[0094] in, A feature sequence among the key features of a contactor The One value, and These are the maximum and minimum values ​​of the feature sequence, respectively. After the above transformation, the feature values ​​are mapped to the range [0, 1].

[0095] Then, the weighted Euclidean distance between each feature mode and the base mode is calculated based on the weighted Euclidean distance, and a distance sequence is formed in order of the number of actions. :

[0096]

[0097] In the formula, Indicates the first The similarity metric between the eigenmode and the base mode under each action is: This represents the difference between them, namely the first The degree of wear of the contactor contacts during the next operation; here n This represents the total number of times the contactor operates.

[0098] The distance sequence is then divided into multiple subsequences of length 10. The mean of the weighted Euclidean distance in each subsequence is calculated, and the mean of the weighted Euclidean distance of all subsequences is used to form the contact wear factor.

[0099] The contact wear factor of AC contactor #AC1 is as follows: Figure 5 As shown, the contact wear factor gradually increases with the number of operations, indicating that contact wear becomes increasingly severe. The fitted curve reveals that the contact wear degradation process exhibits multi-stage characteristics. Specifically, in the first 1000 operations, the contact transitions from brand new to slightly worn, with the contact wear factor rising sharply. During the middle of the lifespan, the contact wear factor generally shows a slow upward trend. A sharp increase occurs around 23,000 operations, indicating that the contact wear has further intensified.

[0100] Step 4: Use clustering algorithms to perform offline segmentation of the contactor's degradation stages;

[0101] Offline degradation stage segmentation is beneficial for better describing the slow change process of a contactor from a healthy state to a failure state. Unsupervised classification of features using clustering algorithms is an effective way to segment degradation stages. The FCM algorithm, based on the similarity between sample data, optimizes the objective function through multiple iterations, grouping samples with high similarity into the same stage. In this embodiment, the contact wear factor is used as the input to the FCM algorithm, and the algorithm process is as follows:

[0102] Assuming the contact wear factor sequence is For a one-dimensional feature sequence, the number of states is set. The contact wear factor sequence Divided into Each stage. The objective function of the FCM algorithm. for:

[0103]

[0104] In the formula, Let be the membership matrix, and ,in Indicates the first The wear factor of the first contact on the first Membership degree of each degenerate stage. This is the cluster center matrix. As a weighted index, it is generally taken as ; is the length of the contact wear factor sequence. For the first Cluster centers at each stage of degradation; For the first Individual contact wear factor.

[0105] Set termination tolerance ,satisfy Random initialization and Iteratively seek the objective function J The minimum value of .

[0106] Update cluster center :

[0107]

[0108] For iteration sequence Cluster centers in a degenerate stage, μ go(l-1) For iteration The first The wear factor of the first contact on the first Membership degree of each degradation stage For iteration l The first The wear factor of the first contact on the first Membership degree of each degenerate stage.

[0109] Update membership :

[0110]

[0111] Repeat the iteration until Terminate when satisfied. For iteration Membership matrix of the order For iteration Membership matrix of the order represent and Euclidean distance, represent and The Euclidean distance.

[0112] Figure 7 The diagram shows the degradation stage classification of contact wear factor for contactor #AC1. The performance degradation process of the AC contactor is divided into four stages: normal operation (stage 0), slight degradation (stage 1), moderate degradation (stage 2), and severe degradation (stage 3). In the normal operation stage, the AC contactor is first put into operation, and the contact surface is relatively smooth with no obvious wear. As the number of switching operations increases, the contact surface gradually wears down, entering the slight and moderate degradation stages. These two stages account for the majority of the overall degradation, indicating that the AC contactor experiences some degree of contact wear during normal operation, but this does not lead to a significant performance decrease. However, with further increases in wear, the contactor enters the severe degradation stage. In this stage, the contact surface is severely ablated, and its operating performance is severely degraded.

[0113] Step 5: Online assessment of contactor degradation stage;

[0114] This embodiment uses the Transformer model for online evaluation of the contactor degradation stage. The Transformer model, by introducing multi-head self-attention, directly models the dependencies between any elements in the sequence data, and its computation is highly parallelized, giving it high computational efficiency when processing long sequence data. These characteristics make the Transformer model advantageous in capturing the dynamic relationship between key input features and the degradation stage, enabling more accurate evaluation of the contactor performance degradation stage.

[0115] Five AC contactor datasets were constructed by using the key features and degradation stage (label) of each contactor as one dataset. Training and test sets were selected from each dataset for online evaluation of contactor degradation stages. To eliminate the influence of dimensions between features, range normalization was used to normalize the key features in the AC contactor datasets, mapping feature values ​​to the range [0, 1]. Furthermore, since degradation stages are not numerical values, mapping them to [0, 1, 2, 3] results in a partial order, meaning the degradation stages are not independent. Therefore, one-hot encoding was used to encode the degradation stages, representing them as 1000, 0100, 0010, 0001.

[0116] In this embodiment, the number of heads in the multi-head self-attention mechanism of the Transformer model is set to 8. A confusion matrix is ​​introduced to count the number of correctly classified and incorrectly classified observations, respectively. The confusion matrix is ​​shown below. Figure 8 Where TP and FN represent the number of actual positive values ​​correctly or incorrectly evaluated, and FP and TN represent the number of actual negative values ​​correctly or incorrectly evaluated. Based on the confusion matrix, the performance of the Transformer model is evaluated using accuracy and F1-Score metrics.

[0117] To verify the generalization ability of the proposed Transformer model, five AC contactors were trained and tested using cross-validation. The key features and degradation stages of the entire lifespan of each AC contactor were used as the training set, while the key features and degradation stages of the entire lifespan of the remaining AC contactors were used as the test set. To avoid the randomness of the experimental results, the prediction result for each contactor was the average of ten tests to ensure the accuracy of the test results.

[0118] When #AC1 is the training set, the remaining contactors are the test set. The evaluation results of the contactor degradation phase are as follows: Figure 9As shown in the figure, the horizontal axis represents the predicted label, and the vertical axis represents the true label. The analysis results indicate that the Transformer model exhibits good classification stability. In the online evaluation of degradation stages, the model has few misclassifications, and the misclassifications mainly occur between adjacent degradation stages. The recognition accuracy for different contactor degradation stages reaches over 90%.

[0119] The evaluation metrics for all experiments are shown in Table 2, where Acc and F1 represent accuracy and F1-score, respectively, Null indicates that the test and validation sets are the same, and Avg represents the mean of the evaluation metrics. The Transformer model constructed in this embodiment achieves an average accuracy and F1-score of over 80% on different training and test sets, validating its good classification performance and generalization ability, and also demonstrating the effectiveness of the degradation stage segmentation.

[0120] Table 2 Evaluation Indicators for AC Contactor Performance Degradation Stage

[0121]

[0122] Step 6: Phased prediction of contactor remaining life;

[0123] In the field of contactor lifetime prediction, both the TCN and Bi-LSTM network models have been widely used and achieved good results. However, these two network models still have some problems in practical applications. On the one hand, the traditional TCN network model, when predicting the remaining lifetime of contactors, relies on unidirectional convolution, which means the model can only learn from historical data and cannot extract feature information from future time dimensions. On the other hand, while the Bi-LSTM network model can process sequences through forward and backward LSTM units, it struggles to extract feature information from long-term sequences. Therefore, this embodiment constructs a multi-stage remaining lifetime prediction model based on a parallel temporal network. The parallel temporal network includes two branches: Bi-LSTM and Bi-TCN, which operate in parallel. Furthermore, to achieve deep fusion of the output features from the parallel Bi-TCN and Bi-LSTM network branches, the feature weighting issue is fully considered. An attention-based feature fusion module is adopted to adaptively and dynamically adjust the feature weights, ensuring that the final fused feature vector accurately and efficiently reflects the complex relationship between key features and remaining lifetime. The specific structure of the parallel temporal network is as follows: Figure 10 As shown.

[0124] This invention employs an attention-based feature fusion module to achieve a weighted fusion process of the output features of Bi-TCN and Bi-LSTM networks. This includes using average pooling to compress the dimensionality of the output features of the Bi-TCN and Bi-LSTM networks, calculating feature weights for feature fusion, and mapping the fused features to RUL values ​​through a fully connected layer.

[0125] Step 1: Use average pooling to compress the dimensionality of the output features of the Bi-TCN network and the Bi-LSTM network;

[0126] Outputs of Bi-TCN and Bi-LSTM and They can be represented as follows:

[0127]

[0128] in, For the number of channels, For key feature length, z The value is 1- Integers.

[0129] Because Bi-TCN and Bi-LSTM networks are limited by their receptive fields, their output features often only contain local information. Therefore, average pooling is used to compress the dimensionality of the output features of Bi-TCN and Bi-LSTM networks, converting the local receptive features of each branch into a channel-dimensional global representation vector. This ensures that the features processed by average pooling fully reflect the input data.

[0130]

[0131] in, This represents the channel vectors of the output features of the Bi-TCN network and the Bi-LSTM network after average pooling. and Channel vectors and The The vector factor of the channel; and These represent the first and second generations of the Bi-TCN network and the Bi-LSTM network, respectively. The output of the first channel One characteristic.

[0132] Step 2: Calculate feature weights and perform feature fusion;

[0133] The two channel vectors processed by average pooling are added element-wise and then fused to obtain the vector. ,at this time Feature information containing two branches:

[0134]

[0135] Composed of two cascaded linear transformation layers The process yields attention coefficients in two dimensions. :

[0136]

[0137] in, and These are the first and second linear transformation layers, respectively. and These are the first-level and second-level activation functions, respectively. ,in The attention coefficient of Bi-TCN, Let be the attention coefficients of the Bi-LSTM; then calculate the weighting parameters for the output features of the Bi-TCN network and the Bi-LSTM network, respectively. and :

[0138]

[0139] Will and The output features applied to the Bi-TCN network and the Bi-LSTM network, respectively, are given by the following formulas:

[0140]

[0141] in, This represents element-wise multiplication. and This represents the weighted output of the Bi-TCN network and the Bi-LSTM network.

[0142] Will and Perform an addition operation to obtain the fused features. for:

[0143]

[0144] Step 3: Finally, the RUL value is obtained through processing with a fully connected layer.

[0145] Model parameter settings: The parallel temporal network consists of a feature input layer, a Bi-TCN network, a Bi-LSTM network, an attention-based feature fusion module, and a RUL output layer. Specific parameters are shown in Table 3. The number of input features in the feature input layer is determined by the key feature dimension. The Bi-TCN network contains three cascaded bidirectional TCN modules, where the kernel size is 3, and the dilation factor increases exponentially by a power of 2. Dropout within each Bi-TCN module is set to 0.2. The Bi-LSTM network contains two cascaded bidirectional LSTM modules with gate units set to 32 and 16 respectively. Dropout is also used to address overfitting in deep learning models, and its value is also set to 0.2. A fully connected layer is used in the RUL output layer to process the fused features. The predicted RUL value is obtained.

[0146] The parallel Bi-TCN and Bi-LSTM networks were trained using multiple batches, with batch size and number of training iterations set to 128 and 50, respectively, and learning rate set to 0.01. The Adam optimizer was used for iterative parameter updates during training.

[0147] Table 3 Parallel Temporal Network Model Parameters

[0148]

[0149] Stage Dataset Construction and Evaluation Metrics: A stage dataset is constructed by combining the key features of each degradation stage of each contactor with its corresponding remaining lifetime label. A parallel temporal network is trained using this stage dataset to obtain the remaining lifetime prediction model for each stage. This embodiment uses the sliding window method to construct the feature samples of the stage dataset, assuming the contactor has a remaining number of actions. Window width is The sliding step size is Considering the importance of maximizing the number of samples for improving model training performance, this study chooses to increase the sliding step size. The value is 1. This represents the total number of feature samples in the constructed stage dataset. Each feature sample corresponds to a remaining lifetime label (RUL label), and the size of a single feature sample and its corresponding RUL label are respectively... and , Indicates the number of key features.

[0150] The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to assess the effectiveness of contactor remaining life prediction. The calculation formulas are as follows:

[0151]

[0152]

[0153] in, and These represent the predicted and actual values ​​of RUL, respectively. The smaller the evaluation index, the higher the prediction accuracy.

[0154] Analysis of multi-stage prediction results for contactor remaining life: Figures 11-14 With #AC1 as the training set, and #AC2, #AC3, #AC4, and #AC5 as the test sets, the multi-stage prediction results of the remaining lifetime are shown. In the figure, the horizontal axis CNA represents the number of current actions, and the vertical axis RUL represents the number of remaining lifetime actions. It can be seen that the overall prediction results are quite close to the true values, especially with #AC2 as the test set, where good prediction results are achieved at each stage. Table 4 shows the evaluation metrics for the multi-stage prediction results of the remaining lifetime of each contactor when #AC1 is the training set.

[0155] In the early stages of contactor operation, its performance is relatively good, and RUL (Remaining Life Scale) serves as a reference at this time. In the later stages of operation, RUL prediction can provide accurate information for formulating maintenance or replacement plans. In the experiment, the evaluation index of the remaining life prediction results of all contactors in stage 3 was the smallest among all stages, proving that the constructed parallel timing network has good predictive performance.

[0156] Table 4 Evaluation Indicators for AC Contactor Life Prediction Results

[0157]

[0158] Ablation Experiment: To verify the effectiveness of the multi-stage remaining lifetime prediction model based on parallel temporal networks used in this embodiment, an ablation experiment was conducted. #AC2 was used as the training set, and #AC3 as the test set to test the prediction performance of different models. In the ablation experiment, the Bi-TCN network model, the Bi-LSTM network model, and the multi-stage remaining lifetime prediction model based on parallel temporal networks were used to predict the remaining lifetime of the contactor in stages. The prediction results were then compared.

[0159] Furthermore, to verify the effectiveness of using an attention-based feature fusion module to fuse the output features of parallel Bi-TCN and Bi-LSTM networks in the multi-stage remaining lifetime prediction model based on parallel temporal networks, the prediction results were compared with those obtained by fusing the output features of parallel Bi-TCN and Bi-LSTM networks using an element-wise addition method, forming a control experiment. PTN represents the prediction result obtained by fusing the output features of the two-branch network using an element-wise addition method, and FA-PTN represents the prediction result obtained by fusing the output features of the two-branch network using the attention-based feature fusion module used in this invention. The contactor remaining lifetime prediction results for each stage are as follows: Figure 15 As shown in Table 5, the evaluation metrics for the prediction results are as follows. It can be seen that, at each stage, the prediction results of the multi-stage remaining lifetime prediction model based on parallel temporal networks used in this invention are closer to the true values, and the evaluation metrics are all lower than those of other models. This is because the parallel temporal network integrates the characteristics of Bi-TCN and Bi-LSTM networks, and their parallel operation allows the parallel temporal network to learn more comprehensive and effective feature information. Among different feature fusion methods, the evaluation metrics of the prediction results based on attention are lower than those of the element-wise addition method at each stage. Specifically: in stage 1, the RMSE and MAE of FA-PTN are reduced by 455.24 and 358.15 times respectively compared to PTN; in the most critical stage 3, the RMSE and MAE of FA-PTN are reduced by 31.51 and 42.33 times respectively compared to PTN, indicating that the attention-based feature fusion module can effectively improve the prediction performance by dynamically weighting different features.

[0160] Table 5 Evaluation Indicators for Ablation Experiment Prediction Results

[0161]

[0162] Comparative experiment:

[0163] Different prediction mechanisms: Considering the multi-stage nature of contactor performance degradation, this invention constructs a multi-stage remaining lifetime prediction model based on a parallel time-series network to predict the remaining lifetime of the contactor at different degradation stages. To further verify its prediction effectiveness, the staged prediction results of the contactor's remaining lifetime are compared with the overall prediction results. Figure 16The figure shows the results of multi-stage remaining life prediction for #AC4 using #AC2 as the training set. As can be seen from the figure, the multi-stage prediction results are closer to the actual values, especially in stages 1 and 2. Table 6 shows the evaluation metrics for the prediction results of different prediction mechanisms. By comparing the evaluation metrics of the overall prediction results (overall prediction is a full life prediction of the contactor, without distinguishing stages) and the multi-stage prediction results, the RMSE and MAE of the multi-stage prediction results for AC contactors are reduced by 29.81% and 31.31% respectively compared to the overall prediction. The results indicate that multi-stage life prediction of contactors better reflects the actual degradation process of contactors and can significantly improve prediction accuracy. The reason for this is that the degradation process of contactors is inherently multi-stage, and the data characteristics of different stages differ significantly. The overall prediction method, because it fails to fully consider this characteristic, is prone to large deviations between the prediction results and the actual situation.

[0164] Table 6 Evaluation Indicators for Prediction Results under Different Prediction Mechanisms

[0165]

[0166] Different Prediction Models: To verify the advantages of the multi-stage remaining lifetime prediction model based on parallel temporal networks used in this embodiment compared to other models, it was compared with some classic machine learning algorithms, including BP neural networks, random forests (RF), support vector regression (SVR), CNNs, and GRUs. The inputs to each model were key features, and the output was the RUL value. When #AC1 was the training set, staged remaining lifetime predictions were performed for #AC2 and #AC3 respectively. Evaluation metrics for the prediction results of each model are shown in Table 7. The evaluation metrics in the table are obtained by summing and averaging the prediction results of each stage, representing the overall prediction performance of the model.

[0167] Experimental results show that the evaluation index of the remaining lifetime multi-stage prediction model based on parallel temporal networks proposed in this embodiment is the smallest compared with the prediction results of other models.

[0168] Table 7 Evaluation Indicators for Prediction Results of Different Models

[0169]

[0170] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A multi-stage prediction method for the electrical life of contactors based on cluster analysis and time-series networks, characterized in that, Includes the following: A full-life test of the AC contactor was conducted to obtain the electrical signals during the entire life operation of the contactor and extract the original features and the permutation entropy of the original features. Key features that can effectively characterize the performance degradation of the contactor were screened through feature importance and correlation analysis. The key features of a single contactor action are taken as a feature mode, and a contact wear factor is constructed based on the modal similarity metric: all key features are normalized, and the mean of the top 10 normalized feature modes is used to form the base mode. M b Calculate the weighted Euclidean distance between the normalized eigenmodes and the base modes, and form a distance sequence in chronological order of the number of actions. I The distance sequence is then divided into multiple subsequences of length 10. The mean of the weighted Euclidean distance in each subsequence is calculated, and the mean of the weighted Euclidean distance of all subsequences is used to form the contact wear factor. Fuzzy C-means clustering is performed on the contact wear factor to divide the contactor's full life test process into multiple degradation stages. The degradation stages are used as training labels, and key features are used as input to train an online degradation stage evaluation model. The online degradation stage evaluation model is used to predict which degradation stage the key features of a single action are in. A stage dataset is formed by combining the key features of each degradation stage with the corresponding remaining lifetime. A parallel temporal network is trained using the stage dataset to obtain the remaining lifetime prediction model for each stage. The degradation stage to which the key feature belongs is predicted by the online degradation stage assessment model, and the remaining life of the contactor is predicted by the remaining life prediction model of the corresponding stage.

2. The multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks according to claim 1, characterized in that, The parallel temporal network uses Bi-LSTM and Bi-TCN as dual-branch networks, and utilizes an attention-based feature fusion module to achieve dynamic weighted fusion of the output features of the dual-branch networks.

3. The multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks according to claim 2, characterized in that, The attention-based feature fusion module includes using average pooling to compress the dimensionality of the output features of the Bi-TCN network and the Bi-LSTM network, calculating feature weights for feature fusion, and mapping the fused features to RUL values ​​through a fully connected layer.

4. The multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks according to claim 1, characterized in that, The key features are contact voltage drop, release voltage, accumulated arc energy, arrangement entropy of contact voltage drop, arrangement entropy of release voltage, arrangement entropy of pull-in time, and arrangement entropy of contact resistance.

5. The multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks according to claim 1, characterized in that, Feature importance analysis is implemented using the XGBoost algorithm. In the process of constructing the tree model of the XGBoost algorithm, the XGBoost algorithm adopts a greedy strategy to search for the optimal split point in the feature space. If a certain feature is used more often when splitting the tree model, it indicates that the feature is more important in the modeling process. The number of splits for each feature is recorded. In weighted Euclidean distance calculation, the weights are calculated based on the number of splits of the feature.

6. The multi-stage prediction method for contactor electrical lifetime based on cluster analysis and time-series networks according to claim 1, characterized in that, The online evaluation model for the degradation stage is implemented using the Transformer architecture and employs one-hot encoding to encode the degradation stage.

Images