Rolling bearing residual life prediction method based on quadratic feature optimization and improved lstm

Abstract

The application discloses a rolling bearing residual useful life prediction method based on secondary feature optimization and improved LSTM, and belongs to the technical field of mechanical and electrical equipment state monitoring and fault diagnosis. First, the time domain, frequency domain and time-frequency domain features of the rolling bearing full life cycle denoising signal are acquired to construct a high-dimensional feature set and perform normalization processing, and a secondary feature optimization method is proposed based on feature 3 constraint criterion and Pearson correlation coefficient. Then, the first prediction time is determined through 3σ criterion multi-feature comprehensive average, and the degradation stage is divided to reduce the data amount, and then a bidirectional long short-term memory model based on Bayesian optimization and self-attention mechanism is further proposed to realize rolling bearing RUL autonomous prediction, so that the algorithm efficiency is further improved. Finally, the prediction experiment verification is carried out based on the PRONOSTIA platform data set. The results show that the method has high prediction accuracy and can effectively improve the accuracy of the residual useful life of the rolling bearing.

Description

Technical Field

[0001] This invention belongs to the field of electromechanical equipment condition monitoring and fault diagnosis technology, specifically involving a method for predicting the remaining service life of rolling bearings based on secondary feature optimization and improved LSTM. Background Technology

[0002] Rotating machinery is widely used in industries such as manufacturing, aerospace, and military. Rolling bearings, as one of the key components of rotating machinery, play a vital role in its normal operation. Once a rolling bearing fails or is damaged during operation, it will cause the machine to stop and affect production, and in severe cases, it may even cause a major safety accident. Therefore, predicting its remaining useful life (RUL) is of great significance.

[0003] Currently, there are three main types of methods for predicting the remaining useful life (RUL) of rolling bearings: physical model-based, data-driven, and hybrid models. Physical model-based methods require extensive expert and experiential knowledge, are costly, and have limited versatility. Data-driven models are a current research hotspot. Data-driven methods primarily rely on data processing techniques, collecting rolling bearing operation data from various sensors and mining the mapping relationship between the data and RUL. Current methods are based on probability statistics and machine learning. Traditional probability statistics methods are heavily influenced by model selection, while machine learning can be divided into shallow machine learning and deep learning. In recent years, deep learning has shown great promise in predicting remaining useful life. Hybrid models combine physical and data-driven methods, resulting in complex predictions and less attention.

[0004] The data-driven model primarily performs lifetime prediction in two steps. First, the rolling bearing state information is characterized by features such as time domain, frequency domain, time-frequency domain, and entropy of the collected data. Next, the extracted features are input into the lifetime prediction model to obtain RUL information. During the feature extraction stage, different features have varying sensitivities to degradation and differing degrees of correlation with degradation. Using all features would affect the prediction accuracy; therefore, further screening of sensitive features is necessary.

[0005] In the life prediction model building stage, deep networks have gained increasing attention for their intelligence, accuracy, robustness, generalization, and optimization characteristics in predicting the life of rolling bearings. Among them, LSTM, with its long-term memory, is widely used in trend prediction. However, LSTM is a unidirectional information flow structure and cannot fully exploit the forward and backward connections in time series. Bidirectional LSTM (BiLSTM) combines the advantages of LSTM models and bidirectional networks, obtaining long-term dependencies within the data from both forward and backward directions, providing past and future information to the sequence and uncovering more hidden information in the time series. Building on this, RUL prediction, which incorporates an attention mechanism in deep networks, has also received considerable attention. It autonomously selects important information by assigning appropriate weights to different parts of the network, significantly improving the model's prediction performance. However, existing algorithms mainly rely on experience to determine the network's hyperparameters, which easily introduces subjective errors and reduces algorithm efficiency. Therefore, efficient selection of hyperparameters has become a crucial factor affecting model accuracy and quality.

[0006] In the entire life cycle evolution of rolling bearings, directly predicting the life from the initial performance degradation stage would save computational costs and improve algorithm efficiency and model accuracy. Therefore, the First Prediction Time (FPT) of rolling bearings has become a key concern, and finding a suitable and accurate FPT can improve the prediction accuracy and model accuracy of bearings. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a method for predicting the remaining service life (FPT) of rolling bearings based on quadratic feature optimization and an improved LSTM. This method can effectively optimize ideal features that accurately characterize rolling bearing degradation, fully considering the importance of each sensitive feature. After quadratic feature optimization, a 3σ criterion based on multiple features is proposed to determine the FPT, avoiding the instability risk of using a single index to determine the FPT. The dataset after FPT is used as the sensitive degradation feature set. While making full use of available data, this method avoids the time spent on manual parameter confirmation and repeated debugging, thus improving both the accuracy and execution efficiency of the life prediction model.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the remaining service life of rolling bearings based on quadratic feature optimization and improved LSTM, the specific steps of which are as follows:

[0009] Step 1: Extract the acceleration vibration signal from the bearing horizontal direction sensor;

[0010] Step 2: Perform wavelet threshold noise reduction on the acceleration vibration signal;

[0011] Step 3: Extract multi-domain features from the data points obtained by wavelet threshold denoising, including time domain, frequency domain and time-frequency domain features. After feature extraction for each sampled data throughout the entire lifetime, a high-dimensional feature set is obtained.

[0012] Step 4: Perform min-max normalization on the high-dimensional features in the high-dimensional feature set to reduce the difference in the order of magnitude of different features and the difference in their impact on the model.

[0013]

[0014] In the formula: K is the number of samples for each feature, i represents the i-th sampling, and z i Let maxz be the feature value at the i-th sampling time. i minz is the maximum value of the feature. i For the minimum value of the feature, m i These are the normalized eigenvalues;

[0015] Step 5: The extracted multi-domain features have high dimensionality and inconsistent quality, resulting in varying contributions to model prediction. Therefore, a secondary feature optimization is performed on the normalized high-dimensional features:

[0016] First feature optimization: The selection process involves calculating the correlation, monotonicity, and robustness values, then obtaining a comprehensive value through linear weighting as the screening criterion, and finally selecting features with high comprehensive indicators.

[0017] The features are divided into trend terms and residual terms using a smoothing method, as shown below:

[0018] X(t k ) = X T (t k )+X R (t k )

[0019] X(t k ) for t k Characteristics of time; X T (t k X is the trend term. R (t k ) is the residual term, t k Indicates the time of the kth sampling;

[0020] The formulas for calculating the time correlation, monotonicity, and robustness indices are as follows:

[0021]

[0022]

[0023]

[0024] In the formula: K is the total number of time points; δ() is the step function, and its calculation formula is:

[0025]

[0026] The linear weighting of indicators is used as the final feature selection criterion, and the calculation formula is as follows:

[0027] S=ω1Corr(X,T)+ω2Mon(X)+ω3Rob(X)

[0028]

[0029] In the formula, ω c Each has its own weight;

[0030] Second feature optimization: Select features with high absolute values ​​of Pearson correlation coefficients with RMS features, and remove features with low absolute values ​​of correlation coefficients, finally obtaining the feature set after the second optimization.

[0031]

[0032] In the formula, Cov represents the covariance between two vectors, X and Y are two eigenvectors, D represents the variance, and ρ X，Y The Pearson correlation coefficient represents the correlation between two vectors, with a value range of [-1, 1]. The closer the absolute value is to 1, the stronger the correlation.

[0033] Step 6: Based on the feature set selected by the secondary feature optimization, select the features with higher absolute values ​​of Pearson coefficients, and determine the FPT according to the 3σ criterion.

[0034] |x(t l -μ|≥3σ,l=1,2,...,5,

[0035] In the formula, μ and σ are the mean and variance of the characteristic during the steady-state operation period, respectively, and x(t) l ) is a feature in t l The value at time l represents the number of points that continuously satisfy the 3σ criterion. Since the feature may have spurious fluctuations, determining the value as FPT when a value satisfies 3σ may lead to inaccurate results. Therefore, it is set that when 5 consecutive points satisfy 3σ, the first of the 5 points is determined as FPT.

[0036] Step 7: Determine the FPT for each feature with a high Pearson coefficient using the 3σ criterion, then calculate the average of these values ​​to obtain the final FPT.

[0037]

[0038] In the formula, m is set to 4, and FPT(j) is the first prediction time point of the j-th feature;

[0039] Step 8, Define Failure Time Labels: Divide the difference between the failure time and the current time by the difference between the failure time and the degradation point time as the label data.

[0040]

[0041] In the formula: t i For the current time, t F Failure time point, t FPT This is the initial prediction time point; once the FPT is determined, the FPT point label is 1, and the failure point label is 0.

[0042] Step 9: Optimize the hyperparameters of SaBiLSTM using Bayesian methods, including the number of cells N in the first hidden layer. u1 The number of cells in the second hidden layer, N u2 batchsize, learning rate, and dropout rate;

[0043] Step 10: Input the optimized hyperparameters into the SaBiLSTM model and input the test set into the model for testing;

[0044] Step 11: Output the test results and calculate the mean squared error (RMSE) and mean absolute error (MAE) between the true value and the test results:

[0045]

[0046]

[0047] In the formula: n is the total number of samples in the test set, y i Authentic labels For labels generated for the test set, the smaller the evaluation metric, the smaller the difference between the predicted value and the true value, and the higher the prediction accuracy of the model.

[0048] The purpose of this invention is to predict the remaining service life of rolling bearings, comprising two parts: In the first part, vibration data is acquired and preprocessed. Experimental verification was conducted using vibration acceleration data collected from the PRONOSTIA rolling bearing accelerated life test bench provided by FEMTO-ST Research Institute. The preprocessing includes wavelet denoising, extraction of time-domain, frequency-domain, and time-frequency-domain features to construct a high-dimensional feature set, and normalization. Then, a second feature optimization is performed on the high-dimensional feature set. The first feature optimization calculates the correlation, robustness, and monotonicity values ​​of each high-dimensional feature, and then calculates the comprehensive value of each feature through linear weighting, selecting features with high comprehensive values. The second feature optimization uses RMS as the standard, selecting features with large absolute values ​​of RMS Pearson correlation coefficients to construct a sensitive feature set. The first prediction time point (FPT) is calculated, and the degraded features after the FPT following the second feature optimization are used as the sensitive feature set input to the network model. The second part is the network model. The model of this invention is based on the BSBiLSTM model. After optimizing the hyperparameters of the SaBiLSTM network through Bayesian methods, an attention mechanism is added to construct the SaBiLSTM model. The Bayesian-optimized hyperparameters are input into it. One bearing is used for training, and another bearing is used for testing. Finally, the RUL prediction of the bearing is obtained.

[0049] This invention addresses the challenge of identifying features closely related to rolling bearing degradation, particularly when determining the hyperparameters of the network model. It proposes a method for predicting the remaining service life of rolling bearings based on quadratic feature optimization and an improved LSTM. The model's predictive performance is validated using rolling bearing data from the PRONOSTIA platform. The main conclusions are as follows:

[0050] 1. After extracting the time domain, frequency domain, and time-frequency domain features of the bearing vibration signal through signal denoising, rich degradation information is obtained. A secondary feature optimization method based on 3 constraints and Person correlation coefficient is proposed to screen out more ideal degradation features. After feature selection, the complexity of network parameters can be further reduced.

[0051] 2. A comprehensive method for determining FPT based on the high absolute value of RMS Pearson correlation coefficient and the 3σ criterion is proposed. The data of the post-degradation stage of FPT is used as the input of the model to reduce the amount of data, which provides a basis for accurately predicting the remaining service life of rolling bearings.

[0052] 3. A BSBiLSTM model is proposed, which adds an attention mechanism to BiLSTM, assigns different weights to feature information, and optimizes the hyperparameters of the model through Bayesian self-optimization, reducing human error and further reducing the time cost of adjusting the network, so that the overall performance of the model reaches the optimal level. Attached Figure Description

[0053] Figure 1This is a flowchart of the rolling bearing prediction method of the present invention;

[0054] Figure 2 The platform for the dataset used in this invention;

[0055] Figure 3 Example 1 shows the time-domain plot of the original bearing vibration data signal extracted;

[0056] Figure 4 The image shows the time domain diagram of the vibration data after wavelet denoising in Example 1.

[0057] Figure 5 The feature map after the first feature optimization in Example 1;

[0058] Figure 6 The correlation coefficient between the first feature optimization in Example 1 and RMS is shown.

[0059] Figure 7 This is a feature map after the second feature optimization in Example 1;

[0060] Figure 8 The FPT diagram determined in Example 1;

[0061] Figure 9 Here is a diagram of the SaBiLSTM model structure;

[0062] Figure 10 This is a graph showing the prediction results for Example 1. Detailed Implementation

[0063] The invention will be further described below, but the scope of the invention is not limited to the description provided.

[0064] like Figure 1 The present invention consists of two main parts. The first part obtains the sensitive feature set after secondary feature optimization and determines the FPT through data preprocessing. The second part is to use a Bayesian optimized model to predict lifetime.

[0065] Step 1: In the data preprocessing section, vibration acceleration data were collected using the PRONOSTIA rolling bearing accelerated life test bench provided by FEMTO-ST Research Institute, and the acceleration vibration signal from the horizontal sensor was extracted; experiments were conducted as follows: Figure 2Two 3035B DYTRAN high-frequency accelerometers were installed on the platform to collect vibration signals in the horizontal and vertical directions. The sampling frequency was 25.6 kHz. In the experiment, a 0.1 s long signal was collected every 10 s, resulting in 2560 data points collected each time. The platform provided degradation data for 17 bearings from normal to failure under three operating conditions. There were 7 bearings under operating conditions 1 and 2, and 3 bearings under operating condition 3. Two bearings were used as the training set for each operating condition. The dataset contained data from both horizontal and vertical channels. In this example, the degradation data of horizontal acceleration was selected for the experiment. Bearing 1 was used for training under operating condition 1, and bearing 2 was used for testing. The acceleration vibration signal data of bearing 1 is shown below. Figure 3 As shown;

[0066] Step 2: Perform wavelet threshold denoising on the acquired acceleration vibration signal. Denoising reduces white noise in the signal and weakens the fluctuations in extracted features. Figure 4 .

[0067] Step 3: Extract 19 commonly used time-domain features and 5 frequency-domain features. The time-frequency domain features are used to analyze the signal.

[0068] Line 3

[0069] The energy values ​​of the eight frequency bands generated by the layer wavelet packet decomposition are used to extract features from each sampled data throughout the entire lifetime to obtain a high-dimensional feature set;

[0070] The time-domain characteristics are as follows:

[0071]

[0072] The frequency domain characteristics are as follows:

[0073]

[0074] Where n is the number of sampling points each time, f i X is the signal frequency. i x represents the spectral amplitude. i This represents the value of the i-th sampling point;

[0075] Step 4: To reduce the differences in the contributions of various features to the model and the magnitude of these differences, min-max normalization is performed on the high-dimensional features in the high-dimensional feature set, as shown below:

[0076]

[0077] In the formula: K is the number of samples for each feature, i represents the i-th sampling, and z i Let maxz be the feature value at the i-th sampling time. i minz is the maximum value of the feature.i For the minimum value of the feature, m i These are the normalized eigenvalues;

[0078] Step 5: The extracted multi-domain features have high dimensionality and inconsistent quality, resulting in varying contributions to model prediction. A secondary feature optimization process is then performed on the extracted high-dimensional features.

[0079] First feature selection: During the selection process, the correlation, monotonicity, and robustness values ​​are calculated, and then a comprehensive value is obtained through linear weighting as the screening criterion. Finally, features with high comprehensive indicators are selected.

[0080] The features are divided into trend terms and residual terms using a smoothing method, as shown below:

[0081] X(t k ) = X T (t k )+X R (t k )

[0082] X(t k (t) represents time t k Features; X T (t k X is the trend term. R (t k ) is the residual term, t k Indicates the time of the kth sampling;

[0083] The formulas for calculating the time correlation, monotonicity, and robustness indices are as follows:

[0084]

[0085]

[0086]

[0087] In the formula: K is the total number of time points; δ() is the step function, and its calculation formula is:

[0088]

[0089] The three indicators are linearly weighted and used as the final feature selection criterion. The calculation formula is as follows:

[0090] S=ω1Corr(X,T)+ω2Mon(X)+ω3Rob(X)

[0091]

[0092] In the formula, ω cFor their respective weights, the features after the first feature optimization are as follows: Figure 5 ;

[0093] Second feature optimization: Features with high absolute values ​​of Pearson correlation coefficients in RMS features are selected, while features with low absolute values ​​of correlation coefficients are removed. The final feature set after the second optimization is as follows:

[0094]

[0095] In the formula, Cov represents the covariance between two vectors, X and Y are two eigenvectors, D represents the variance, and ρ X,Y The Pearson correlation coefficient represents the correlation between two vectors, ranging from -1 to 1. The closer the absolute value is to 1, the stronger the correlation. Figure 6 , 7 ;

[0096] Step 6: Based on the feature set selected by the quadratic feature optimization, select the features with higher absolute values ​​of Pearson coefficients, and determine their FPT by averaging according to the 3σ criterion. The 3σ criterion is as follows:

[0097] |x(t l -μ|≥3σ,l=1,2,...,5,

[0098] In the formula, μ and σ are the mean and variance of the characteristic during the steady-state operation period, respectively, and x(t) l ) is a feature in t l The value at time l represents the number of points that continuously satisfy the 3σ criterion. Since the feature may have spurious fluctuations, determining the value as FPT when a value satisfies 3σ may lead to inaccurate results. Therefore, it is set that when 5 consecutive points satisfy 3σ, the first of the 5 points is determined as FPT.

[0099] Step 7: Determine the FPT for each feature with a high Pearson coefficient using the 3σ criterion.

[0100] We calculate their average to obtain the final FPT:

[0101]

[0102] In the formula, m is set to 4, and FPT(j) is the first prediction time point of the j-th feature, such as... Figure 8 ;

[0103] Step 8, Define Failure Time Labels: Divide the difference between the failure time and the current time by the difference between the failure time and the degradation point time as the label data.

[0104]

[0105] In the formula: t i For the current time, t F Failure time point, t FPT For the first prediction time point, once the FPT is determined, the FPT point label is 1, and the failure point label is 0;

[0106] Step 9: Optimize the hyperparameters of SaBiLSTM using Bayesian methods, including the number of cells N in the first hidden layer. u1 The number of cells in the second hidden layer, N u2 The optimization steps are as follows: batch size, learning rate, dropout rate;

[0107] a. Define the objective function: minimize the validation error, i.e. minimize the loss function. In this paper, the loss function is the mean squared error (MSE).

[0108] b. Define the range of hyperparameters: Define the range of each hyperparameter to be optimized;

[0109] c. Set initial observation points: Define random initial observation points and use them as initial hyperparameters for model training.

[0110] d. Proxy model construction: Construct a proxy probability model for the objective function to estimate the function distribution, using a Gaussian regression model as the proxy model;

[0111] e. Acquisition function construction: Construct the acquisition function to determine the next observation point, i.e., the next hyperparameter point; f. Obtain new results: Combine the new observation point to obtain the new MSE;

[0112] g. End optimization: The optimization process ends when the maximum number of iterations is reached, and the optimal hyperparameters are obtained based on the minimum objective function;

[0113] Step 10: Input the optimized hyperparameters into the SaBiLSTM model, and input the test set into the model for testing.

[0114] BiLSTM learns the unit information from both the forward and backward directions at each time step to obtain the output value at each time step. It learns the information inherent in the forward and backward directions of the time series, automatically acquiring comprehensive feature information while addressing issues such as time series dependencies and gradient explosion. An attention mechanism is added to the BiLSTM output layer. This attention layer further filters the time series data, selectively learning intermediate features during model training. By automatically assigning weights, these features are associated with the output sequence, filtering out useless information and highlighting key information, thus improving the accuracy of the model's prediction results. Finally, a fully connected layer performs dimensionality reduction to obtain the final remaining lifetime prediction result. The model is as follows: Figure 9 As shown;

[0115] Step 11: Output the test results and calculate the root mean square error (RMSE) and mean absolute error (MAE) between the true value and the test result, as follows:

[0116]

[0117]

[0118] In the formula: n is the total number of samples in the test set, y i Authentic labels For the labels generated for the test set, the smaller the evaluation metric, the smaller the difference between the predicted and true values, and the higher the prediction accuracy of the model. The prediction results are as follows: Figure 10 As can be seen, the predicted results have the same overall trend as the actual values, and the degree of agreement is good.

[0119] This invention addresses the challenge of identifying features closely related to rolling bearing degradation, particularly when determining the hyperparameters of the network model. It proposes a method for predicting the remaining service life of rolling bearings based on quadratic feature optimization and an improved LSTM. The model's predictive performance is validated using rolling bearing fatigue data from the PRONOSTIA platform. The main conclusions are as follows:

[0120] 1. After extracting the time domain, frequency domain, and time-frequency domain features of the bearing vibration signal through signal denoising, rich degradation information is obtained. A secondary feature optimization method based on 3 constraints and Person correlation coefficient is proposed to screen out more ideal degradation features. After feature selection, the complexity of network parameters can be further reduced.

[0121] 2. A comprehensive method for determining FPT based on the high absolute value of RMS Pearson correlation coefficient and the 3σ criterion is proposed. The data of the post-degradation stage of FPT is used as the input of the model to reduce the amount of data, which provides a basis for accurately predicting the remaining service life of rolling bearings.

[0122] 3. A BSBiLSTM model is proposed, which adds an attention mechanism to BiLSTM, assigns different weights to feature information, and optimizes the hyperparameters of the model through Bayesian self-optimization, reducing human error and further reducing the time cost of adjusting the network, so that the overall performance of the model reaches the optimal level.

Claims

1. A method for predicting the remaining service life of rolling bearings based on quadratic feature optimization and improved LSTM, the specific steps of which are as follows: Step 1: Extract the acceleration vibration signal from the bearing horizontal direction sensor; Step 2: Perform wavelet threshold noise reduction on the acceleration vibration signal; Step 3: Extract multi-domain features from the data points obtained by wavelet threshold denoising, including time domain, frequency domain and time-frequency domain features. After feature extraction for each sampled data throughout the entire lifetime, a high-dimensional feature set is obtained. Step 4: Perform min-max normalization on the high-dimensional features in the high-dimensional feature set to reduce the difference in the order of magnitude of different features and the difference in their impact on the model. In the formula: K is the number of samples for each feature, i represents the i-th sampling, and z i Let maxz be the feature value at the i-th sampling time. i minz is the maximum value of the feature. i For the minimum value of the feature, m i These are the normalized eigenvalues; Step 5: Perform secondary feature optimization on the normalized high-dimensional features: First feature optimization: The selection process involves calculating the correlation, monotonicity, and robustness values, then obtaining a comprehensive value through linear weighting as the screening criterion, and finally selecting features with high comprehensive indicators. The features are divided into trend terms and residual terms using a smoothing method, as shown below: X(t k )=X T (t k )+X R (t k ) X(t k ) for t k Characteristics of time; X T (t k X is the trend term. R (t k ) is the residual term, t k Indicates the time of the kth sampling; The formulas for calculating the time correlation, monotonicity, and robustness indices are as follows: In the formula: K is the total number of time points; δ() is the step function, and its calculation formula is: The linear weighting of indicators is used as the final feature selection criterion, and the calculation formula is as follows: S=ω1Corr(X,T)+ω2Mon(X)+ω3Rob(X) In the formula, ω c Each has its own weight; Second feature optimization: Select features with high absolute values ​​of Pearson correlation coefficients with RMS features, and remove features with low absolute values ​​of correlation coefficients, finally obtaining the feature set after the second optimization. In the formula, Cov represents the covariance between two vectors, X and Y are two eigenvectors, D represents the variance, and ρ X，Y The Pearson correlation coefficient represents the correlation between two vectors, with a value range of [-1, 1]. The closer the absolute value is to 1, the stronger the correlation. Step 6: Based on the feature set selected by the secondary feature optimization, select the features with higher absolute values ​​of Pearson coefficients, and determine the FPT according to the 3σ criterion. |x(t l )-μ|≥3σ,l=1,2,...,5, In the formula, μ and σ are the mean and variance of the characteristic during the steady-state operation period, respectively, and x(t) l ) is a feature in t l The value at time t, l represents the number of points that continuously satisfy the 3σ criterion; If five consecutive points satisfy 3σ, then the first of the five points is determined as FPT. Step 7: Determine the FPT for each feature with a high Pearson coefficient using the 3σ criterion, then calculate the average of these values ​​to obtain the final FPT. In the formula, m is set to 4, and FPT(j) is the first prediction time point of the j-th feature; Step 8: Define the fault time label: Divide the difference between the failure time value and the current time value by the difference between the failure time and the degradation point time as the label data. In the formula: t i For the current time, t F Failure time point, t FPT This is the initial prediction time point; once the FPT is determined, the FPT point label is 1, and the failure point label is 0. Step 9: Optimize the hyperparameters of SaBiLSTM using Bayesian methods, including the number of cells N in the first hidden layer. u1 The number of cells in the second hidden layer, N u2 batchsize, learning rate, and dropout rate; Step 10: Input the optimized hyperparameters into the SaBiLSTM model and input the test set into the model for testing. Step 11: Output the test results and calculate the mean square error and mean absolute error between the true value and the test results. In the formula: n is the total number of samples in the test set, y i Authentic labels For labels generated for the test set, the smaller the evaluation metric, the smaller the difference between the predicted value and the true value, and the higher the prediction accuracy of the model.

Images