Circuit Breaker Remaining Life Prediction Method Based on Multi-Feature Fusion and Multi-Head Attention

Abstract

This invention presents a method for predicting the remaining life of circuit breakers based on multi-feature fusion and multi-head attention. First, vibration events in the vibration signal are segmented using a short-time energy dual-threshold method. Mechanical characteristic parameters are extracted from different vibration event intervals. Simultaneously, a convolutional variational autoencoder is used to mine mechanical degradation features in the vibration signal. The mechanical characteristic parameters and degradation features are then concatenated to obtain a comprehensive degradation feature for remaining life prediction. Finally, a remaining life prediction model is constructed based on a GRU network and a multi-head attention mechanism. The multi-head attention mechanism captures the dependencies of the vibration signal in multiple different representation subspaces, improving the accuracy of quantitative prediction of remaining life.

Description

Technical Field

[0001] This invention belongs to the field of circuit breaker remaining life prediction technology, specifically a circuit breaker remaining life prediction method based on multi-feature fusion and multi-head attention. Background Technology

[0002] As switching devices evolve towards intelligent operation, it is crucial to apply intelligent sensing and advanced detection technologies to their health management. Low-voltage universal circuit breakers, as key equipment in low-voltage power grid systems, rapidly disconnect circuits to ensure the performance and safety of electrical equipment when faults such as short circuits, overloads, and undervoltage occur. Therefore, predicting the remaining lifespan of circuit breakers is of great significance for improving their operational reliability.

[0003] Existing technologies have conducted a series of studies on the prediction of the remaining useful life (RUL) of circuit breakers. For example, Yang Qiuyu et al. (Yang Qiuyu, Peng Yanqing, Zhuang Zhijian, Xiong Lei. Assessment of the Remaining Mechanical Life of High Voltage Circuit Breakers Based on Stochastic Fuzzy Theory, 2016, 52(08):161-165+171.) collected the overtravel of circuit breaker contacts and the average speed of tripping for the prediction of the remaining mechanical life. The invention application with application number 202211045544.0 discloses a method for predicting the life of circuit breaker tripping accessories based on action time and driving energy. It uses the CEEMDAN-WD and short-time energy entropy ratio dual threshold method to accurately calibrate the action vibration event, and then extracts the mechanism action time and driving energy parameters corresponding to the vibration event as performance degradation indicators, and constructs a binary Wiener model to realize the quantitative prediction of the remaining life.

[0004] The aforementioned literature on circuit breaker mechanical life prediction has several shortcomings, mainly: 1) The parameterized life degradation information obtained through direct measurement or indirect detection using vibration signals fails to effectively mine the deep features of the degradation data; 2) The division of vibration event intervals is not detailed enough, resulting in limited feature extraction; 3) The life prediction model is relatively simple. Due to measurement factors and individual differences, the circuit breaker degradation process becomes uncertain. For complex nonlinear processes, statistical data-driven methods cannot provide sufficient expressive power, resulting in low computational efficiency and low prediction accuracy. Therefore, this invention proposes a circuit breaker remaining life prediction method based on multi-feature fusion and multi-head attention. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method for predicting the remaining life of circuit breakers based on multi-feature fusion and multi-head attention. First, the method segments the tripping vibration events in the vibration signal using a short-time energy dual-threshold method, extracting mechanical characteristic parameters from different event intervals. Simultaneously, CVAE is used to mine mechanical characteristic degradation features in the vibration signal. The mechanical characteristic parameters and degradation features are then concatenated to obtain a comprehensive degradation feature for remaining life prediction. Finally, a remaining life prediction model is constructed based on a GRU network and a multi-head attention mechanism. The multi-head attention mechanism captures the dependencies of the vibration signal in multiple different representation subspaces, improving the accuracy of quantitative prediction of remaining life.

[0006] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0007] A method for predicting the remaining life of circuit breakers based on multi-feature fusion and multi-head attention, characterized by the following steps:

[0008] The first step is to conduct a mechanical life test on the universal circuit breaker and collect vibration signals during the circuit breaker's opening process.

[0009] The second step is to divide the vibration signal into frames, calculate the short-time energy of all framed signals, and extract the vibration event interval T1 to T6 from the vibration signal based on the short-time energy; where T1 is the moment when the push rod of the circuit breaker tripping accessory collides with the triggering mechanism, and T6 is the moment when the amplitude of the moving contact decays to the minimum.

[0010] The third step is to extract the mechanical characteristic parameters of the circuit breaker from the interval T1 to T3, including the trigger mechanism action time and the average opening speed; and to extract the mechanical characteristic degradation features of the circuit breaker from the interval T2 to T6 using a convolutional variational automatic encoder; and to combine the two features to obtain the comprehensive degradation features; where T2 is the moment when the moving and stationary contacts of the circuit breaker separate, and T3 is the moment when the moving contact of the circuit breaker reaches its maximum displacement.

[0011] The fourth step is to embed a multi-head self-attention mechanism layer between the hidden layer and the output layer of the GRU network to obtain the remaining lifetime prediction model, and then use the trained model to predict the remaining lifetime of circuit breakers.

[0012] Furthermore, in the third step, during the training of the convolutional variational autoencoder, the total number of iterations is divided into three equal parts. The loss weight of the KL divergence in the first third of the total number of iterations is 0.001τ, and the loss weights of the KL divergence in the last two thirds are 0.01τ and 0.1τ, respectively, where τ is a constant.

[0013] Furthermore, the convolutional variational autoencoder includes two parts: encoding and decoding. The input of the encoding part passes through a reconstruction layer, a convolutional layer, a max pooling layer, another convolutional layer, another max pooling layer, another convolutional layer, a reconstruction layer, a fully connected layer, and a reparameterization layer in sequence to obtain the output of the encoding part. Each convolutional layer and fully connected layer is followed by an activation operation.

[0014] Furthermore, the number of attention heads in the remaining lifespan prediction model is 5 to 8, with the model achieving the best prediction performance when the number of attention heads is 6.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. Circuit breaker vibration signals contain rich mechanical state information. On one hand, the vibration signals contain vibration impact events, each corresponding to a specific mechanical action of the circuit breaker. Therefore, this allows for the detection of mechanical characteristic parameters with strong characterization capabilities of macroscopic degradation processes, including mechanism action time and average opening speed. On the other hand, the waveform of the vibration signal is also an effective representation of the mechanical degradation state. By extracting mechanical characteristic degradation features with strong characterization capabilities of microscopic degradation processes through convolutional variational autoencoders (CVAs), and then concatenating the mechanical characteristic parameter features and mechanical characteristic degradation features, a comprehensive degradation feature for remaining life prediction is obtained. The output of the CVA is the probability density distribution of parameter-constrained variables, rather than specific values. Therefore, the obtained latent variables are more meaningful. Furthermore, the CVA alleviates the overfitting problem by adding a regularization term to the loss function.

[0017] 2. During the parameter optimization of the convolutional variational autoencoder, the loss weight of KL divergence changes dynamically with the iteration process. The total number of iterations is evenly divided into three parts. The loss weight of KL divergence in the first third is one-tenth and one-hundredth of the total number of iterations in the last two thirds, respectively. This ensures the authenticity of the mechanical degradation features extracted by the convolutional variational autoencoder while guaranteeing the diversity of the generated data.

[0018] 3. This invention constructs a remaining lifetime prediction model based on GRU and a multi-head self-attention mechanism. GRU is highly suitable for handling problems highly correlated with time series data. Compared to traditional RNNs and LSTMs, GRU adds gating units to RNNs, increasing the network's ability to remember historical data. Simultaneously, GRU uses update and reset gates instead of the input, output, and forget gates in LSTMs, reducing the number of model parameters and complexity while maintaining model accuracy and making training more stable. Considering the complexity of the circuit breaker's mechanical mechanism and the uncertainty of its operation, a multi-head self-attention mechanism is placed between the hidden and output layers of the GRU network. This multi-head self-attention mechanism captures different dependencies of signals in multiple different representation subspaces, improving the accuracy of quantitative prediction of the circuit breaker's remaining lifetime.

[0019] 4. The vibration signal generated during the circuit breaker tripping process contains different vibration events, requiring interval segmentation to obtain the vibration event intervals directly related to the tripping action. Considering that the tripping vibration signal is mainly generated by collisions between mechanisms, and the intensity of different vibration events varies, this invention uses the short-time energy dual-threshold method to divide the vibration signal into different intervals. Compared with the short-time zero-crossing rate method, the short-time energy method is more suitable for detecting signals with large amplitude changes. Based on this, the dual-threshold method has high accuracy in detecting the two endpoints of each tripping event segment. This invention selects the vibration signal during the circuit breaker tripping process as the medium for life analysis, fully utilizing the non-invasive measurement advantages and rich mechanical state information contained in the vibration signal. In addition, tripping vibration is a manifestation of the circuit breaker's breaking capacity and is the most direct and powerful tool for reflecting the circuit breaker's functionality and reliability. This invention provides an integrated scheme for predicting the remaining mechanical life of a universal circuit breaker using vibration signals. Compared with fault diagnosis or condition assessment, which only evaluate the current state, this invention can provide technical support for the long-term use of circuit breakers in terms of prognosis and maintenance, helping to improve their reliability and safety, extend their life to a certain extent, and increase the economic benefits of the power system. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the present invention;

[0021] Figure 2 This is a structural diagram of the remaining lifetime prediction model;

[0022] Figure 3 The image shows the endpoint detection results of the short-time energy dual-threshold method.

[0023] Figure 4 This is a graph showing the trend of the mechanical characteristic parameters of circuit breaker #1 as a function of the number of operations.

[0024] Figure 5A comparison chart of decoding signals from different encoders;

[0025] Figure 6 Comparison diagram of CVAE encoded signals;

[0026] Figure 7 Prediction error diagrams for circuit breakers #2 and #3 with different attention head numbers;

[0027] Figure 8 Visualization results of the MSA weights for circuit breaker #1;

[0028] Figure 9 The results are the predicted results of the ablation test of circuit breaker #1. Detailed Implementation

[0029] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0030] This invention relates to a method for predicting the remaining life of circuit breakers based on multi-feature fusion and multi-head attention (hereinafter referred to as the method), comprising the following steps:

[0031] The first step is to conduct a mechanical life test on the universal circuit breaker and collect vibration signals during the circuit breaker's opening process.

[0032] The second step is to divide the vibration signal into frames, calculate the short-time energy of all framed signals, and extract the vibration event interval T1 to T6 from the vibration signal using the dual-threshold method based on the short-time energy envelope.

[0033] 2-1 The vibration signal x(t) is framed using a window function; to eliminate high-frequency interference and reduce energy leakage, a Hamming window is chosen, the expression of which is:

[0034] ω(v)=0.54-0.46cos(2πv / (L-1)) (1)

[0035] Where L is the window length, v represents the sequence number of the sampling point within the Hamming window, and 0≤v≤L-1;

[0036] The expression for the framing operation is:

[0037] z f (v)=ω(v)*x((f-1)*inc+v) (2)

[0038] In the formula, z f (v) represents the signal of the f-th frame, where f = 1, 2, ..., f n f n The total number of frames is denoted by 'inc', and the frame shift is 'inc'.

[0039] 2-2 Using equation (3), the short-time energy of all frame signals is calculated. The short-time energy of the f-th frame signal is expressed as:

[0040]

[0041] 2-3 The short-time energy of all framed signals forms a short-time energy envelope along the frame sequence direction. Based on this, the dual threshold method is used to detect the short-time energy of contact impact and transposition, and the opening vibration event interval T1 to T6 is extracted from the vibration signal.

[0042] The basic principle of the dual-threshold method is: select a higher threshold H2 on the short-time energy envelope for a rough judgment, that is, the interval AB above the threshold H2 is definitely the interval to be found. Then determine a lower threshold H1, and search left and right in the interval AB until the point intersecting with the threshold H1 is found. The new interval CD is the segment to be extracted.

[0043] Before and after the moving contact of the circuit breaker strikes the stationary contact, the vibration event range during the tripping process is T1 to T6, depending on the magnitude of the short-term energy. At time T1, the push rod of the tripping accessory collides with the triggering mechanism, generating the first vibration. At this time, since the vibration signal is collected at the contact point and the tripping accessory is far away, the vibration is relatively weak. At time T2, the constraint of the tripping spring is released, the spring releases energy, and under the action of the tripping spring, the moving and stationary contacts separate, and the short-term energy value is relatively large. At time T3, the moving contact reaches the maximum displacement, generating strong vibration, which is transmitted to the main shaft and the support and buffering mechanisms such as the stop pin through the connecting mechanism. Under the action of the reaction force, the moving contact rebounds, and at this time, the vibration signal amplitude reaches its peak value, and the short-term energy is also at its peak value. At times T4 and T5, the moving contact rebounds and strikes the stop pin again, and the second and third rebounds occur, with relatively large short-term energy values. At time T6, the amplitude of the moving contact gradually decays to the minimum. This time is taken as the end mark of the tripping action. Therefore, the tripping vibration event range is T1 to T6.

[0044] The third step is to extract the mechanical characteristic parameter features of the circuit breaker from the interval T1 to T3, and extract the mechanical characteristic degradation features of the circuit breaker from the interval T2 to T6. The mechanical characteristic parameter features and the mechanical characteristic degradation features are then combined to obtain the comprehensive degradation features.

[0045] The interval T1 to T3 in section 3-1 reflects the speed characteristics of the moving and stationary contacts before impact. Based on this, it is possible to detect mechanical characteristic parameters with strong characterization ability of macroscopic degradation process. The mechanical characteristic parameters include the trigger mechanism action time F1 and the average opening speed F2; the calculation formula is shown below:

[0046]

[0047] Where T1 is the moment when the push rod of the circuit breaker tripping accessory collides with the triggering mechanism, T2 is the moment when the moving and stationary contacts separate, T3 is the moment when the maximum displacement of the moving contact ends, and l is the maximum displacement of the moving contact;

[0048] 3-2 The vibration is stronger in the interval T2 to T6, and the signal components are richer, containing more information on mechanical property degradation. Therefore, a convolutional variational autoencoder (CVAE) is used to extract mechanical property degradation features from the vibration signal in the interval T2 to T6. Due to the influence of randomness, the interval T2 to T6 detected by the dual threshold method has inconsistent lengths, which is not conducive to the construction of the dataset. Therefore, the first 580 sampling points of this interval are truncated to make the sample length uniform and their amplitudes are normalized.

[0049] CVAE consists of two parts: encoding and decoding. The encoding part includes two reconstruction layers, three convolutional layers, two max pooling layers, one fully connected layer, and one reparameterization layer. The decoding part includes one fully connected layer, two reconstruction layers, three convolutional layers, and two max pooling layers.

[0050] The reconstruction layer transforms the dimensions of the input:

[0051] G k =g(S k (5)

[0052] In the formula, g represents the dimension transformation operation, G k This represents the output vector of the reconstruction layer, and the size of the output vector is 1×d;

[0053] The convolutional layer performs horizontal sliding convolution on the input data across multiple time steps. The expression for the convolution operation is:

[0054] H k =δ(W k *G k +B k (6)

[0055] In the formula, H k The output of the convolutional layer is represented by *, where * indicates a one-dimensional convolution operation with depth, and W represents the output of the convolutional layer. k and B k These represent the weight matrix and bias matrix of the convolutional layer, respectively, and δ is the activation function, which uses the ReLU activation function.

[0056] The output H of the max pooling layer to the convolutional layer k The expression for max pooling operation, used for feature selection and dimensionality reduction, is:

[0057] I k =p(Hk (7)

[0058] In the formula, p represents the max pooling operation, and I k This is the output of the max pooling layer;

[0059] The output of the max pooling layer is passed through another reconstruction layer to obtain vector J. k Vector J k After passing through a fully connected layer, we obtain the mean vector μ and the standard deviation vector σ. The calculation formula for the fully connected layer is:

[0060]

[0061] In the formula, W a and B a W represents the weight matrix and bias matrix of the fully connected layer that yields the mean vector, respectively. b and B b Let represent the weight matrix and bias matrix of the fully connected layer, respectively, to obtain the standard deviation vector. φ is the activation function, which is the ReLU activation function.

[0062] The output of the fully connected layer passes through a reparameterization layer to obtain the output of the encoded part, namely the latent variable Z;

[0063] Z=μ+ε·σ (9)

[0064] In the formula, ε represents the weighting factor, which is randomly sampled from the standard normal distribution N(0,I);

[0065] Finally, the latent variable Z is decoded and restored to the same dimension as the CVAE input to obtain the CVAE output, namely the mechanical property degradation feature M = (M1, M2, ..., M...). A A represents the number of mechanical property degradation features;

[0066] 3-3 By concatenating the mechanical property degradation features with the normalized mechanical property parameter features F=(F1',F2'), the comprehensive degradation feature Y=(M1,M2,...,M A F1', F2') = (Y1, Y2, ..., Y D ); where D represents the number of comprehensive degradation features;

[0067] The fourth step is to build a remaining lifetime prediction model based on the GRU network and the multi-head self-attention mechanism, train the model, and use the trained model for circuit breaker remaining lifetime prediction.

[0068] The GRU network consists of an input layer, a GRU layer, a hidden layer, and an output layer. Considering the complexity of the circuit breaker's mechanical mechanism and the uncertainties during its operation, the multi-head self-attention mechanism layer is placed between the hidden layer and the output layer. Figure 2 As shown; the multi-head self-attention mechanism can compute multiple relationships between each time step and other time steps in parallel, resulting in higher computational efficiency and more accurate capture of features in the input sequence; the integrated degenerate features obtained in the third step are input into the GRU network, and the information learned at different time steps, i.e., the output of the hidden layer, is represented as O = [O1, O2, ..., O]. t ] T , where t represents the time step for returning. The multi-head self-attention mechanism, based on self-attention, learns separately for different features, requiring the computation of matrix Q. m K m and V m The formula is as follows:

[0069]

[0070] In the formula, and Let represent the query matrix, key matrix, and value matrix of the m-th attention head, respectively;

[0071] Next, calculate the importance score for each attention point, using the following formula:

[0072]

[0073] In the formula, d(K) m ) represents matrix K m The dimension;

[0074] The importance scores of each attention head are concatenated to obtain the feature attention; the feature attention is then passed through a fully connected layer to obtain the multi-head self-attention output Y. out The formula is as follows:

[0075] Y out =Concact(head1,head2,...,head h W P (12)

[0076] In the formula, h represents the number of attention heads, and W P The weight matrix represents the multi-head self-attention mechanism. The results of multi-head attention can be weighted and merged to obtain the remaining lifetime prediction result.

[0077] Example

[0078] The circuit breaker remaining life prediction method based on multi-feature fusion and multi-head attention in this embodiment includes the following steps:

[0079] Step 1: Vibration signals during the circuit breaker's opening process were collected using a circuit breaker mechanical life test system. The test operation frequency was set to 20 times / hour according to relevant circuit breaker standards. For each opening operation, contact vibration signals were collected using a USB-7648A data acquisition card at a sampling frequency of 20kHz, with a sampling duration of 0.15s and 3000 data points sampled per operation. Life tests were conducted on a certain number of DW15-1600 circuit breakers. The mechanical life prediction method for circuit breakers was tested using vibration signals collected from three test units. Considering that an excessively low average opening speed would directly lead to circuit breaker failure, the average opening speed was selected as the mechanical life failure criterion, and the average opening speed threshold was set to 5m / s. The mechanical life operation counts for circuit breakers #1, #2, and #3 were 4354, 3522, and 3946, respectively.

[0080] The second step involves selecting a Hamming window function to frame the vibration signal, resulting in multiple frames. Adjusting the window length and frame shift allows the energy entropy ratio envelope curve to comprehensively reflect the impact of each vibration event during the tripping process. Appropriate window length and frame shift not only allow for the representation of minor impacts during vibration but also prevent unnecessary interference or excessive energy leakage due to excessive computational load. After debugging, the window length was determined to be 50 and the frame shift inc = 10. The vibration signal is then divided into intervals reflecting different tripping vibration events using the short-time energy dual-threshold method to enhance feature representation. By comparing the short-time energy values ​​at different times, a threshold value for the dual-threshold method is set, and the tripping vibration event interval T1 to T6 is extracted from the vibration signal. Since the vibration is relatively weak at time T1, the corresponding threshold value is set to 500g based on multiple measurement results. 2 At time T2, the moving and stationary contacts separate, and the short-term energy value is relatively small, resulting in a 5×10⁻⁶ energy level. 4 g 2 Set as the threshold value; and the peak values ​​of short-time energy appear sequentially at times T3, T4, and T5. Set the maximum value of the short-time energy as the threshold value at time T3, and set 6.7 × 10 5 g 2 and 4.4×10 5 g 2 The threshold values ​​are set for times T4 and T5 respectively. At time T6, the amplitude gradually decreases, and the corresponding threshold value is set to 1.6 × 10⁻⁶. 5 g 2 , Figure 3 (a)~(f) show the positions at times T1, T2, T3, T4 and T5; thus, the vibration event range directly related to the tripping action can be obtained.

[0081] The third step is to extract the mechanical characteristic parameters of the circuit breaker from the interval T1 to T3, including the triggering mechanism action time F1 and the average opening speed F2. These two parameters are more significant in engineering and are a direct reflection of the opening action performance. Figure 4 The graph shows the trend of mechanical characteristic parameters of circuit breaker #1 with the number of operations. It can be seen that the triggering mechanism action time F1 increases with the number of operations, which is due to the gradual increase in jamming of the shunt trip triggering mechanism and the tripping half-shaft under long-term wear. The average tripping speed F2 decreases. Vibration is strong and the signal components are richer in the interval T2 to T6, containing more information on mechanical characteristic degradation. Therefore, mechanical characteristic degradation features are extracted from the interval T2 to T6. Due to the influence of randomness, the DTM detected an inconsistent length in the interval T2 to T6, which is not conducive to the construction of the dataset. Therefore, this example extracts the first 580 sampling points of this interval to unify the length of the vibration signal in this segment and normalizes its amplitude. CVAE is used to extract mechanical characteristic degradation features from the interval T2 to T6. The parameters of each layer of CVAE are shown in Table 1; where k represents the kernel size, d represents the number of kernels, and s is the stride.

[0082] Table 1. Structure and parameters of convolutional variational autoencoders

[0083]

[0084] The Adam algorithm was chosen to optimize the parameters of the convolutional variational autoencoder. The learning rate was set to 0.001, the batch size for each training iteration was set to 50, and the total number of training iterations was set to 150. The loss function for CVAE was:

[0085]

[0086] Where, x i For the data of the i-th sampling point, x i ′ represents the mechanical property degradation feature obtained by CVAE of the data at the i-th sampling point, N is the length of the vibration signal, α is the loss weight of the KL divergence, and μ (j) σ (j) These are the j-th elements of the mean vector and the standard deviation vector, respectively, where J is the dimension of the latent variable Z;

[0087] The first term on the right-hand side of the loss function equation represents the reconstruction loss, and the second term represents the KL divergence loss. The KL divergence loss weight α plays a role in adjusting the reconstruction loss and the KL divergence loss to ensure that the model finds a suitable balance between the diversity and realism of the generated signals. When the KL divergence loss is small, the model pays more attention to the reconstruction loss, and vice versa, it will pay more attention to the KL divergence loss, making the generated signals more diverse. In this embodiment, the KL divergence loss weight α is set to 0.001 for the first 50 iterations, α is set to 0.01 for the next 50 iterations, and α is set to 0.1 for the last 50 iterations.

[0088] To visually demonstrate the effectiveness of CVAE in extracting mechanical property degradation features, the CVAE decoded signal is compared with the decoded signals of an autoencoder (AE) and a variational autoencoder (VAE). The inputs and hyperparameters of AE and VAE are identical to those of CVAE, and the total number of training iterations is 150. The decoded signals of the three models are compared as follows: Figure 5 As shown; from Figure 5 It can be seen that CVAE can decode and reconstruct the original vibration signal with good consistency. Furthermore, root mean square error (RMSE) and mean absolute error (MAE) are introduced to evaluate the performance of the three models in reconstructing the vibration signal; the definitions of RMSE and MAE are as follows:

[0089]

[0090]

[0091] In the formula, Q represents the total number of circuit breaker operations. Characteristic points representing vibration signal decoding Characteristic points representing the original vibration signal;

[0092] The reconstruction error comparison results of the three datasets are shown in Table 2. It can be seen that the RMSE and MAE of the CVAE model reconstruction are the smallest for each dataset, and its reconstruction ability is better. Therefore, CVAE is more robust and the extracted mechanical property degradation features are more representative.

[0093] Table 2 Comparison of Reconstruction Errors

[0094]

[0095] To further visualize the mechanical property degradation characteristics extracted by CVAE, vibration signals from the circuit breaker sample at three different time periods were input into the CVAE system. Operation counts 201-300 were designated as period 1, 1301-1400 as period 2, and 3401-3500 as period 3. The visualization results are shown below. Figure 6 (a) and Figure 6 As shown in (b), with a mechanical characteristic degradation feature dimension of 2, it can be seen from the figure that the mechanical characteristic degradation features of the same circuit breaker specimen differ at different times. Vibration signals of three circuit breaker specimens at time 2 are then input into CVAE, and the encoded visualization results are shown below. Figure 6As shown in (c), it can be seen that the mechanical characteristic degradation characteristics of different circuit breaker test specimens also differ at the same time, which realizes the effective characterization of degradation micro information, makes up for the problem that the mechanical characteristic parameter characteristics alone are insufficient to characterize the degradation process, and also proves the necessity of extracting mechanical characteristic degradation characteristics.

[0096] The mechanical property degradation features extracted by CVAE are M=(M1,M2), and the mechanical property parameter features are F=(F1',F2'). The mechanical property degradation features and the mechanical property parameter features are concatenated to obtain the comprehensive degradation features Y=(M1,M2,F1',F2').

[0097] Step 4: Construct a remaining lifetime prediction model based on GRU network and multi-head attention mechanism;

[0098] To delve deeper into the patterns of time series samples, a sliding window method is employed to generate the dataset samples. This method preserves the correlation between adjacent elements in the time series while overcoming the low accuracy of single-step prediction. Let the sliding window size be W, and the step size be S. Each input sample can be represented as X = W × D, where D is the number of elements in the comprehensive degenerate feature set. The RUL label of the sample can be represented as R... true =QW-(c-1)*S, where c is the sample number; to obtain more samples, set W=5, S=1, then the total number of samples is C=Q-4; the input data for training the model can be described as The loss metric is the mean squared error (MSE), which can be expressed as follows:

[0099]

[0100] in, These are the predicted and actual remaining lifetime values ​​for the c-th sample, respectively.

[0101] The remaining lifespan prediction model comprises two GRUs, one multi-head attention mechanism layer, and two fully connected layers. The number of hidden units in the GRUs is set to 64 and 32, respectively, while the number of hidden units in the two fully connected layers is set to 50 and 1, respectively. Furthermore, batch normalization (BN) is added after each network layer to reduce internal covariate bias in the neural network, making training deep network models easier and more stable, enhancing the model's nonlinear expressive power, and ultimately accelerating training speed. During training, predicted values ​​are calculated through forward propagation. and the true value The MSE between layers was calculated, and the back propagation (BP) of each layer parameters was performed using the optimizer Adam. The optimization learning rate was set to 0.001, the batch size was set to 50, and the epochs were set to 100.

[0102] RMSE and MAE are used to evaluate the performance of the remaining mechanical life prediction model. Both of these evaluation indicators reflect the accuracy of the predicted remaining mechanical life of the circuit breaker. The smaller the RMSE and MAE, the higher the prediction accuracy. RMSE better reflects abnormal fluctuations, while MAE better reflects the actual predicted situation. The definitions of RMSE and MAE are as follows:

[0103]

[0104]

[0105] Furthermore, the number of attention heads is an important parameter that significantly impacts the accuracy of the prediction model. Using circuit breaker #1 as the training set and circuit breakers #2 and #3 as the test set, a Remaining Life (RUL) prediction model was used to predict RUL. The number of attention heads was set to 1-10, and the model input was the comprehensive degradation feature Y. The results of the prediction model's accuracy changing with the number of attention heads are as follows: Figure 7 As shown in Table 3, when the number of attention heads is 1, the model is a single-head self-attention mechanism. The model can only learn the single relationship between each time step and other time steps, resulting in poor expressive ability. However, when the number of attention heads is too large, the computational burden is high, and it can lead to overfitting, causing the performance of the prediction model to decline. The quantitative results in Table 3 show that the model has the best prediction effect when the number of attention heads is 6. Therefore, the number of heads in the multi-head attention mechanism is set to 6.

[0106] Table 3. Error results for different numbers of attention heads.

[0107]

[0108]

[0109] To further illustrate the effect of multi-head self-attention (MSA), taking the 200th vibration signal of circuit breaker #1 as an example, the weight relationship between each time step calculated by MSA and other time steps is visualized as follows: Figure 8 As shown, the attention of each head number pays different attention to different time steps. The first and sixth attentions pay more attention to the fifth time step, the second and third attentions pay more attention to the third and fourth time steps, and the fourth and fifth attentions pay more attention to the second time step. This shows that MSA captures the different dependencies of vibration signals, solves the problem of insufficient representation ability of self-attention mechanism, and also proves the necessity of using MSA.

[0110] Then, cross-validation was used to conduct RUL tests on three circuit breaker samples. To verify the effectiveness of extracting mechanical characteristic parameter features, mechanical characteristic degradation features, and MSA, ablation experiments were performed. The prediction effects of five models—GRU, mechanical characteristic parameter features-GRU, mechanical characteristic degradation features-GRU, GRU-MSA, and the method of this invention—were compared. The parameters of GRU were consistent, and the input of the model was the vibration signal in the interval T2-T6. Figure 9 The prediction performance of the above models was compared on circuit breaker #1, with circuit breaker #2 as the training set. Compared with the actual remaining life curve, GRU showed the worst fit. The prediction results of parametric feature-GRU were affected by degradation uncertainty, exhibiting large fluctuations and numerous outliers. Deep feature-GRU mitigated the impact of uncertainty to some extent, showing less fluctuation. GRU-MSA showed a better fit, but still exhibited some fluctuation. The proposed method showed the best fit, with minimal fluctuations and few outliers. The quantitative results in Table 4 show that the proposed method reduced the RMSE by 63.42%, 55.13%, 49.92%, and 30.40% respectively, and the MAE by 62.62%, 54.39%, 49.58%, and 30.18% for circuit breaker #1 compared to other prediction methods. This demonstrates that each module played a crucial role.

[0111] Table 4 Ablation Experiment Results

[0112]

[0113] In this embodiment, the data acquisition software is implemented using LabVIEW, the data processing software using Matlab, the lifetime prediction software using Python, the deep learning environment using Tensorflow+Keras, and the software platform using PyCharm. The LabVIEW, Matlab, and Python languages ​​used in this embodiment are well-known to those skilled in the art. All percentages in this embodiment are numerical percentages.

[0114] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A circuit breaker residual life prediction method based on multi-feature fusion and multi-head attention, characterized by, The method comprises the following steps: Firstly, the mechanical life test of the universal circuit breaker is carried out, and the vibration signal in the breaking process of the circuit breaker is collected; Second step, frame the vibration signal, calculate the short-time energy of all frame signals, and extract the vibration event interval from the vibration signal based on the short-time energy ; wherein, is the moment when the push rod of the circuit breaker opening accessory collides with the trigger mechanism, is the moment when the moving contact amplitude decays to the minimum Thirdly, the mechanical characteristic parameters of the circuit breaker are extracted from the interval , including the operating time of the triggering mechanism and the average opening speed; the mechanical characteristic degradation features of the circuit breaker are extracted from the interval using the convolutional variational autoencoder; the two kinds of features are spliced to obtain comprehensive degradation features; wherein, is the moment when the moving and static contacts of the circuit breaker are separated, is the moment when the moving contact of the circuit breaker reaches the maximum displacement; and is the moment when the moving contact rebounds to hit the stop pin again and the second and third rebounds occur. Fourthly, a multi-head self-attention mechanism layer is embedded between the hidden layer and the output layer of the GRU network to obtain a residual life prediction model, and the trained model is used for residual life prediction of the circuit breaker.

2. The method of claim 1, wherein the method is based on multi-feature fusion and multi-head attention. In the third step, in the convolutional variational autoencoder training process, the total number of iterations is evenly divided into three parts, and the loss weight of KL divergence in the first third of the total number of iterations is 0.001 , and the loss weights of KL divergence in the last two thirds are 0.01 and 0.1 , respectively, and c is a constant. 3.The multi-feature fusion and multi-head attention based residual life prediction method of a circuit breaker according to claim 1 or 2, characterized in that, The convolutional variational autoencoder comprises an encoding part and a decoding part; wherein the input of the encoding part sequentially passes through a reconstruction layer, a convolutional layer, a max-pooling layer, a convolutional layer, a max-pooling layer, a convolutional layer, a reconstruction layer, a full connection layer and a reparameterization layer to obtain the output of the encoding part; and an activation operation is followed after each convolutional layer and full connection layer.

4. The method of claim 1, wherein the method is based on multi-feature fusion and multi-head attention. The number of attention heads of the residual life prediction model is 5-8, and the model prediction effect is best when the number of attention heads is 6.

Images