A Remaining Useful Life Prediction Method for Rolling Bearings Based on MS-ALSTM

Abstract

The present invention proposes a method for predicting the remaining useful life of rolling bearings based on MS-ALSTM, which includes the following steps: S1. Obtain the full-cycle vibration signals of the rolling bearings and extract the time-domain features of the vibration signals; S2. After normalization, reduce the dimension through the principal component analysis method; S3. Perform discrete wavelet decomposition on the dimension-reduced comprehensive index to obtain inputs of different scales, and obtain the training set and the test set; S4. Introduce the attention mechanism, construct the ALSTM model based on the original LSTM unit, and input the training set for training; S5. Use the trained model for testing to obtain the output result, that is, realize the prediction of the remaining useful life. The experimental results show that the method of the present invention has high prediction accuracy, accurate prediction results and small errors.

Description

Technical Field

[0001] The present invention relates to the field of rolling bearing detection technology, in particular to a rolling bearing remaining life prediction method based on MS-ALSTM (Multi-Scale-Attention Long Short-Term Memory Network) Background Art

[0002] Rolling bearings are key components of rotating machinery. During operation, bearings often experience various types of failures, including pitting and wear, due to factors such as overload and insufficient lubrication. Once these failures occur, the equipment cannot function properly, and can even cause major accidents, resulting in significant losses to production and life. Therefore, predicting the remaining life of rolling bearings has become a key technology for providing prognostic measures and maintenance decisions, thereby enabling effective maintenance and improving equipment reliability, availability, and safety.

[0003] Currently, the remaining useful life (RUL) prediction of rolling bearings has become a hot research topic. Mainstream research approaches include physical model-based methods and data-driven methods. Physical model-based methods involve developing corresponding degradation models by studying the physical structure and failure mechanisms of the equipment. However, these methods are complex to construct and are only applicable to certain types of equipment, making them difficult to handle sudden or non-monotonic degradation processes. Data-driven methods rely on statistical or machine learning techniques to extract hidden degradation information from massive amounts of data. With the development of deep learning, recurrent neural networks (RNNs) have been widely applied to the RUL prediction of rolling bearings due to their significant advantages in processing time series data. However, traditional LSTM (Long Short-Term Memory) models are susceptible to interference from bearing noise signals when predicting bearing life, resulting in poor prediction results. Therefore, by introducing an attention mechanism to enhance model performance, a multi-scale feature set representing bearing trend changes can be constructed to mine information at different depths, thereby improving prediction accuracy. Summary of the Invention

[0004] Aiming at the problem that the existing model for predicting the remaining life is easily disturbed by bearing noise and has low prediction accuracy, the present invention provides a method for predicting the remaining life of rolling bearings based on MS-ALSTM.

[0005] The technical solution adopted in the present invention is as follows:

[0006] A method for predicting the remaining life of a rolling bearing based on MS-ALSTM, characterized by comprising the following steps:

[0007] S1. Obtain the full-cycle vibration signal of the rolling bearing, extract the time domain features of the vibration signal, and construct a time domain feature set that is highly correlated with the bearing degradation trend;

[0008] S2. Normalize the time-domain features, reduce the dimension of the processed data using the principal component analysis method, and construct a comprehensive feature index;

[0009] S3. Perform discrete wavelet decomposition on the comprehensive feature index to obtain the comprehensive feature distributions at different scales, and construct a multi-scale training set and a test set;

[0010] S4. Introduce an attention mechanism, construct an ALSTM model based on the original LSTM unit, input the training set for training, and input the test set for testing;

[0011] S5. Use the trained and tested model to predict the remaining useful life.

[0012] The further technical solution is as follows:

[0013] Step S1 includes the following steps: Select 11-dimensional time-domain features to construct a time-domain feature set.

[0014] Step S2 includes the following steps: Normalize the time-domain features, and limit the numerical range to between [0, 1].

[0015] Step S2 includes the following steps: Calculate the covariance matrix of the normalized feature matrix, perform eigenvalue decomposition on the covariance matrix, and sort the obtained eigenvalues λ1, λ2,..., λ d Arrange them in order, and take the eigenvectors corresponding to the larger eigenvalues to form an eigenvector matrix, which is the reconstructed feature set after dimensionality reduction.

[0016] In step S2, the PCA method is used for dimensionality reduction. First, calculate the covariance matrix of the normalized feature matrix, then perform eigenvalue decomposition on the covariance matrix, and sort the obtained eigenvalues λ1, λ2,..., λ d Arrange them in order, and take the larger eigenvalues λ1, λ2,..., λ d′ The corresponding eigenvectors w1, w2,..., w d′ Form an eigenvector matrix, which is the reconstructed feature set after dimensionality reduction. Here, the reconstruction threshold is selected as 90%, that is, the eigenvalues satisfying the following formula (1) are selected.

[0017]

[0018] Where 0 < d′ < d, after selecting the corresponding eigenvalues λ1, λ2,..., λ d′ Then, the corresponding eigenvectors w1, w2,..., w d′ Construct a comprehensive feature index CF (comprehensive feature indicator) according to formula (2)

[0019]

[0020] Step S3 specifically includes the following steps:

[0021] S31. Perform discrete wavelet decomposition on the comprehensive feature index to obtain trend data at different scales;

[0022] S32. Define the bearing life label. The point where the amplitude of the bearing starts to change significantly is used as the degradation starting point. The process from the degradation starting point to complete failure is the process in which the bearing life gradually increases until it reaches the highest point and then fails, and at the same time, it is also the process in which the remaining useful life RUL gradually decreases.

[0023] S33. After dividing the training set and the test set, convert the data into a single time window sequence that conforms to the input of the LSTM unit to obtain a complete data set.

[0024] The beneficial effects of the present invention are as follows:

[0025] 1. Aiming at the problem that the original vibration signal of the rolling bearing has a high noise content and a large interference to prediction, this paper selects time-domain feature indicators, and obtains comprehensive feature indicators after principal component dimensionality reduction for subsequent remaining useful life prediction. Compared with the original signal, the amount of calculation is reduced, and the stability and trend are improved.

[0026] 2. Aiming at the problem of low accuracy of the original LSTM, an attention mechanism is introduced to improve the original model, and at the same time, multi-scale transformation is performed for life prediction. Compared with the original model, more comprehensive information is obtained, and both the prediction accuracy and the accuracy are improved. Brief Description of the Drawings

[0027] Figure 1 It is a schematic flowchart of an embodiment of the present invention.

[0028] Figure 2 It is a full-cycle original vibration signal diagram extracted in an embodiment of the present invention.

[0029] Figure 3 It is a trend change diagram of the bearing comprehensive feature index in an embodiment of the present invention.

[0030] Figure 4 It is a comparison diagram of bearing remaining useful life prediction in an embodiment of the present invention. Specific Embodiments

[0031] The following describes the specific embodiments of the present invention with reference to the accompanying drawings and specific embodiments.

[0032] Embodiment 1

[0033] As shown in the appendix Figure 1 , in this embodiment, the IEEE PHM2012 competition data set is selected to implement a rolling bearing remaining useful life prediction method based on MS-ALSTM, including the following steps:

[0034] S1. Obtain the full-cycle vibration signal of the rolling bearing:

[0035] Conduct an accelerated degradation experiment on the rolling bearing on the PRONOSTIA test bench to obtain full-life cycle data. Collect the vibration signals in the horizontal and vertical directions through an acceleration sensor. The sampling frequency of the rolling bearing signal on the test bench is 25.6KHZ, the sampling time is 0.1s, and the sampling interval is 10s. The full-cycle vibration signal of the original signal is as shown in the appendix Figure 2 as follows.

[0036] S2. Extract the time-domain features of the vibration signal: [[ID=ll]]

[0037] Extract the corresponding time-domain features within each sampling time period. Since the change trend of the time-domain features is highly similar to the bearing degradation process, a key time-domain feature set is constructed, including the mean, standard deviation, variance, root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, and pulse index, to comprehensively characterize the information of the signal.

[0038] The selected 11-dimensional time-domain features are represented as follows:

[0039]

[0040]

[0041] where X i is the amplitude of the single-point vibration signal, and n is the number of signal points.

[0042] S3. Use PCA for dimensionality reduction to construct a comprehensive feature index:

[0043] Normalize the 19-dimensional features to obtain a normalized feature set with an amplitude range of 0-1. Due to computational problems, PCA (Principal Component Analysis) is used here to reduce the feature dimensionality, calculate the importance of each principal component, and construct a comprehensive feature containing 90% of the information, as shown in the appendix Figure 3 as follows. During this process, the trend of the comprehensive feature index can roughly represent the degradation process of the rolling bearing, and the remaining life of the bearing gradually decreases. At this time, the normalized life is used as the data label. It is assumed here that there is no degradation wear before the critical point where the bearing vibration signal does not increase significantly, and the remaining life remains unchanged. The corresponding life is obtained by linearly dividing the intermediate failure process points, and a complete label set is constructed;

[0044] S4. Construct multi-scale training sets and test sets:

[0045] First, the comprehensive features are decomposed by three-layer discrete wavelet transform and then reconstructed at each time layer to obtain the comprehensive feature distributions at different scales. Taking Bearing 1-1 as an example, after multiple experiments, it is decomposed into three scales: small, medium, and large. If the number of divided layers is too large, it will cause an increase in data volume, complex calculations, and a large loss of contained information. As the time scale decreases, the trend of this feature becomes more and more obvious, the detailed fluctuations gradually decrease, and the curve tends to be smooth. For bearings that are often in a high-noise state, comprehensively using the information at different scales can capture the time-varying characteristics of the bearings and reduce the interference of high-frequency fluctuations to a certain extent. Divide the training set and the test set.

[0046] Discrete wavelet transform is a typical multi-resolution analysis method. In each layer, the signal is approximately divided into a high-level signal and a detailed signal. The high-level signal can display trend information, while the detailed signal can show more local information.

[0047] S5. Remaining useful life prediction:

[0048] To reduce the influence of the time dependence existing between the learned representations, an Attention layer is introduced to capture the importance of historical information representations, and an improved ALSTM model is formed based on the original LSTM unit. Then, the multi-scale comprehensive features are respectively input into the ALSTM model for training, the parameters of the Adam optimizer are adjusted to make the model loss function converge, the trained model is input into the test set for remaining useful life prediction and an error curve is plotted, and the hidden layer results output by the model are upsampled and then subjected to mean pooling processing to obtain the final prediction result.

[0049] LSTM (Long Short-Term Memory neural network) is a special recurrent neural network. Compared with ordinary RNN networks, the LSTM network introduces gating units, overcoming some problems of gradient explosion and gradient disappearance in long-distance propagation. The LSTM unit includes an input layer, a recurrent layer, and an output layer. The recurrent layer contains a forget gate, an input gate, and an output gate. The input gate imports input information, the forget gate determines the influence weight of the unit state at the previous moment on the unit state at this moment, and the output gate is used to output the unit information. The update process of the recurrent unit information is shown as follows:

[0050] f t =σ(W f ·[h t-1 ,x t +b f )

[0051] i t =σ(W i ·[h t-1 ,x t +b i )

[0052] Ct = f t * C t-1 + i t * tanh(W C · [h t-1 , x t + b C )

[0053] o t = σ(W o · [h t-1 , x t + b0)

[0054] h t = o t * tanh(C t )

[0055] In the above formula, C t represents the long-term state at time t, and h t is the short-term state. [h t-1 , x t represents the concatenated vector of the output at time t-1 and the input at time t. W f and b f represent the weight matrix and bias term of the forget gate. W i and b i represent the weight matrix and bias term of the input gate. W c and b c represent the weight matrix and bias term of the basic cell state. W o and b o represent the weight matrix and bias term of the output gate. σ() and tanh() represent the sigmoid activation function and the tanh activation function respectively.

[0056] The Attention mechanism essentially assigns different weights to different inputs to obtain a more targeted output. Reasonable addition of the Attention mechanism can improve the accuracy and efficiency of the model.

[0057] The prediction result in the original LSTM model only depends on the output of the last recurrent unit, which has limitations. Therefore, the Attention mechanism is introduced to consider the output results of the intermediate units, assign different weighting coefficients, calculate the weighted average information, and obtain the optimized prediction result.

[0058] To measure the effect of the model, the mean absolute error (MAE) and the root mean square error (RMSE) are used as indicators for result comparison. The calculation formulas of the mean absolute error and the root mean square error are as follows.

[0059]

[0060]

[0061] The above is the single-point prediction value, and y i is the true value of the prediction point, and n represents the number of prediction points.

[0062] Input the test set into the trained model to obtain the final prediction result. As shown in the appendix Figure 4 shown, it can be seen that after the original index is scaled and segmented, the prediction results of the large-scale input are generally low and volatile; the medium-scale prediction results are higher than the actual values, with large errors and obvious overall trends; the small-scale has good trends and is overall stable. The final prediction result synthesizes the values of each scale, can effectively characterize the life change trend, and has a better prediction effect. Therefore, it can be seen that after different scale transformation processes, the accuracy and accuracy of the model prediction effect can be improved.

Claims

1. A method for predicting the remaining useful life of a rolling bearing based on MS-ALSTM, characterized in that, It includes the following steps: S1. Obtain the full-cycle vibration signal of the rolling bearing, extract the time-domain features of the vibration signal, and select 11-dimensional time-domain features to construct a time-domain feature set with a high degree of correlation with the bearing degradation trend. The 11-dimensional time-domain features are: mean value, standard deviation, root mean square, root mean square of the squares, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, kurtosis factor, and waveform factor; S2. Normalize the time-domain features, and perform dimensionality reduction on the processed data using the principal component analysis method to construct a comprehensive feature index; Step S2 includes the following steps: Calculate the covariance matrix of the normalized feature matrix, perform eigenvalue decomposition on the covariance matrix, and sort the obtained eigenvalues λ1, λ2, ..., λ d Sort them, and take the eigenvectors corresponding to the larger eigenvalues to form an eigenvector matrix, which is the reconstructed feature set after dimensionality reduction; Select the reconstruction threshold as 90%, that is, select the eigenvalue that satisfies the following formula (1): Where 0 < d′ < d, corresponding eigenvalues λ1, λ2,..., λ are selected according to Equation (1). d′ After that, the corresponding eigenvectors w1, w2,..., w d′ A comprehensive feature index CF is constructed according to Equation (2): S3. Perform discrete wavelet decomposition on the comprehensive feature index to obtain the comprehensive feature distributions at different scales, and construct a multi-scale training set and a test set; S4. Introduce the attention mechanism, construct an ALSTM model based on the original LSTM unit, input the training set for training, and input the test set for testing; S5. Use the trained and tested model to predict the remaining life.

2. The method for predicting the remaining life of a rolling bearing based on MS-ALSTM according to claim 1, wherein Step S2 includes the following steps: Normalize the time-domain features, and limit the numerical range to between [0, 1].

3. A remaining useful life prediction method for rolling bearings based on MS-ALSTM according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Perform discrete wavelet decomposition on the comprehensive feature index to obtain the comprehensive feature distribution data at different scales; S32. Define the bearing life label. The starting point of degradation is the point where the amplitude of the bearing starts to change significantly. The process from the starting point of degradation to complete failure is the process in which the bearing life gradually increases until it reaches the highest point and fails, and at the same time, it is also the process in which the remaining useful life (RUL) gradually decreases; S33. After dividing the training set and the test set, convert the data into a single time window sequence that conforms to the input of the LSTM unit to obtain a complete data set.

Images