A multi-cycle feature alignment-based bearing residual life prediction method across working conditions

Abstract

The application discloses a cross-condition bearing residual life prediction method based on multi-period feature alignment, which first divides the whole life cycle vibration signal of a rolling bearing into signal samples with the same length, selects the bearing signal under an arbitrary condition W i as a source domain, and selects the bearing signal under the remaining arbitrary conditions as a target domain. Secondly, the autocorrelation function is used for periodic detection of each sample of the source domain and the target domain, a clustering algorithm is used for clustering of the signal sample periods, K periodic lengths are obtained, and the signal samples are reshaped. Then, a convolutional neural network is constructed, feature extraction and feature fusion are performed on the reshaped signal samples, and multi-scale periodic degradation features are obtained. Finally, a regression network is constructed, the multi-scale periodic degradation features are used for residual service life prediction, and a loss function based on multi-period feature correlation alignment is constructed for training and testing. The application improves the extraction of bearing degradation features and accurately predicts the residual life of the bearing.

Description

Technical Field

[0001] The present invention belongs to the field of rotating machinery health management, and in particular relates to a method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment. Background Art

[0002] Rotating machinery is widely used in industrial manufacturing and often operates for extended periods under high intensity, high speed, and variable operating conditions. This accelerates wear and aging, increasing the risk of failure. With the development of artificial intelligence (AI), health management technologies based on big data and AI are gaining increasing attention. Among them, prognostics and health management (PHM) has become a mainstream intelligent maintenance model. PHM monitors and analyzes equipment operating data in real time, integrating advanced algorithms such as sensor technology, big data analytics, and machine learning to dynamically assess equipment health, predict potential failures, and implement targeted maintenance measures. The increasing maturity of PHM applications has not only improved production efficiency and extended equipment life, but also provided solid support for enterprises' transformation and upgrading in the era of intelligent manufacturing. Real-time monitoring and prediction of the remaining useful life (RUL) of critical machinery is particularly important, ensuring safe and reliable operation.

[0003] In recent years, researchers at home and abroad have conducted extensive research on rolling bearing life prediction, with the main approaches categorized as physical model-driven and data-driven. Traditional physical models rely on precise mathematical or physical formulas. While they can provide theoretically accurate predictions, they are difficult to apply to complex nonlinear systems and require high levels of researcher expertise, limiting their widespread application. With the advent of the big data era, advances in sensor technology and data processing capabilities have driven the rapid development of data-driven approaches. These approaches exploit the potential information in monitoring data, reducing reliance on specialized knowledge and offering greater adaptability and transferability. In particular, the application of deep learning technology, which automatically extracts fault features through the construction of deep neural networks, has significantly improved the accuracy of life prediction, reduced costs, and promoted the development of intelligent and reliable life prediction for rotating machinery. However, in real-world industrial applications, the diversity of operating conditions and the complexity of the operating environment result in different distribution characteristics of data across the entire bearing lifecycle. Factors such as operating conditions, loads, and speeds affect the consistency of data distribution, making it difficult for deep learning models to generalize between training and test sets, thereby reducing the performance of deep learning-based rolling bearing remaining service life prediction. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this paper proposes a cross-operating-condition bearing remaining life prediction method based on multi-cycle feature alignment. This method incorporates multi-cycle convolutional neural network feature fusion technology to effectively extract multi-level fault features from the signal. Furthermore, a multi-cycle feature correlation alignment strategy is introduced to align the covariance matrices of source and target domain features at the individual cycle feature and fusion feature levels, significantly reducing the inconsistency of feature distributions under different operating conditions. This method not only enhances the model's generalization capabilities but also improves prediction accuracy and reliability under variable operating conditions.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] A method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment includes the following steps:

[0007] S1, based on the sampling frequency, the vibration signal of the rolling bearing throughout its life cycle is divided into signal samples of the same length, and according to different working conditions W={W1,W2,…,W p}, where l is the number of working conditions, select any working condition W i The bearing signal under the condition W is used as the source domain. j , and the bearing signal under j≠i is taken as the target domain.

[0008] S2. Use the autocorrelation function (ACF) to perform periodic detection on each sample in the source domain and the target domain to identify the signal sample period.

[0009] S3. Use the K-Means algorithm to cluster the signal sample periods identified in the previous step, obtain the K most significant period lengths by finding the cluster centers, and reshape the signal samples based on these significant signal sample periods to prepare for feature extraction by the convolutional neural network.

[0010] S4. Convolutional neural network (CNN) is constructed, consisting of convolutional layers, pooling layers, and nonlinear activation functions. Feature extraction is performed on reshaped signal samples with different significant periods to generate multi-scale period degradation features. An average pooling layer is used to align the multi-scale period features and perform feature fusion to obtain the multi-scale period degradation features.

[0011] S5. Construct a regression network and use multi-scale periodic degradation features to predict the remaining service life, completing the transformation process from feature extraction to rolling bearing remaining service life prediction.

[0012] S6. Based on the multi-scale periodic degradation features extracted by convolutional neural networks at different periods, a multi-period feature correlation alignment loss function is constructed between the source domain and the target domain. By minimizing this loss function, the feature distributions of the source and target domains are aligned, reducing the inconsistency of feature distributions under different working conditions.

[0013] S7. Train the model constructed in S4 and S5 on the training set using the loss function constructed in S6. Test the trained model on the test set to obtain the predicted life of the bearing.

[0014] Furthermore, the step S1 divides the horizontal vibration signal of the rolling bearing throughout its life cycle into samples of equal length and divides them into source domain and target domain according to different working conditions. Specifically, the source domain signal sample is The target domain signal sample is Where n is the number of source domain signal samples, and m is the number of target domain signal samples.

[0015] Furthermore, the step S2 performs periodic detection on each sample in the source domain and the target domain. First, the autocorrelation function is used to quantify the similarity of the signal at different lag times, and its formula is as follows:

[0016]

[0017] Where R(τ) represents the similarity of the signal x(t) at the lag time τ, τ represents the lag time, that is, the offset of the time series relative to itself, and N represents the length of the time series. Then, based on the significant peaks in the autocorrelation function, the period T = {T1, T2, ..., T n+m}.

[0018] Furthermore, the step S3 obtains the significant period of the signal sample and performs signal reshaping. First, the K-means algorithm is used to reshape the sample period T of the source domain and the target domain. n+m} to cluster and obtain K cluster centers C={C1,C2,…,C K} as a significant cycle. Then, the source domain sample S S Based on the significant cycles C={C1,C2,…,C K} is reshaped according to the following formula, for example, for any significant period C k,k＝1,2,,K , convert the source domain sample from length 1×l to length

[0019]

[0020] Similarly, the target domain samples It is also reshaped according to the cycle of significance.

[0021] Furthermore, the step S4 constructs a convolutional neural network (CNN) to extract degradation features of the sample signal in a multi-cycle situation. The K-cycle samples reshaped in the source domain and the target domain are respectively subjected to feature extraction using the constructed convolutional neural network to obtain features {f1, f2, ..., f k}, then the average pooling layer performs feature size alignment and vertical splicing, and the convolutional layer is used for feature fusion.

[0022] Furthermore, in step S5, a regression network is constructed using the fused multi-scale periodic degradation features to complete the transformation from feature extraction to life prediction. The constructed regression network includes a fully connected layer with 100 hidden nodes, a Dropout layer, and an output layer.

[0023] Furthermore, step S6 constructs a multi-cycle feature correlation alignment loss function between the source domain and the target domain. Specifically, the CORAL loss function achieves feature alignment by minimizing the difference in the covariance matrix between the source domain and the target domain, and its expression is as follows:

[0024]

[0025] where f s represents the learned features of the source domain, f s represents the learned features of the target domain, ‖·‖ F represents the Frobenius norm, which is used to calculate the sum of squares of matrix elements. d is the feature dimension. Dividing by 4d² is to normalize the loss value to ensure that the loss magnitude is consistent under different feature dimensions. Based on the CORAL loss function, a loss function is constructed based on multi-cycle features. Its expression is as follows:

[0026]

[0027] where MSE pred is the minimum mean square error, which is used to measure the predicted lifespan The difference between the actual lifespan value y is expressed as follows:

[0028]

[0029] is the source domain feature after periodic feature fusion and target domain features The loss between , is expressed as:

[0030]

[0031] α is the parameter that controls the contribution of this part of the loss;

[0032] For each period feature {f1,f2,…,f k The loss between the source domain features and the target domain features under} is expressed as follows:

[0033]

[0034] Among them, i=1,2,…,K, {λ1,λ2,…,λ K} is the coefficient for adjusting the weight of each loss item.

[0035] Furthermore, in step S7, a training set is constructed using the full lifecycle data of multiple bearings under one operating condition as the source domain and the full lifecycle data of a bearing under another operating condition as the target domain. The models constructed in steps S4 and S5 are trained on this training set using the loss function constructed in step S6, with Adam being the optimal optimization method. The trained model is then tested on multiple bearings under the other operating condition to obtain their predicted lifespans.

[0036] The beneficial effects of the present invention are:

[0037] The present invention analyzes the periodicity of bearing vibration signals based on the autocorrelation function, combines K-Means clustering to extract significant period lengths, and constructs a multi-period extraction and fusion method based on a convolutional neural network, thereby improving the extraction of bearing degradation characteristics.

[0038] The present invention constructs a multi-cycle feature correlation alignment loss function, which aligns the feature distributions of the source domain and the target domain at different cycle scales, reduces the distribution differences between different working conditions, and improves the generalization ability of the model across working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method of the present invention;

[0040] Figure 2 Schematic diagram of the multi-scale periodic feature fusion method based on convolutional neural network of the present invention;

[0041] Figure 3 This is a framework of a cross-operating-condition bearing remaining life prediction method based on multi-cycle feature alignment. DETAILED DESCRIPTION

[0042] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0043] like Figure 1 As shown, the present invention provides a method for predicting the remaining life of bearings across working conditions based on multi-cycle feature alignment. First, the vibration signal of the rolling bearing throughout its life cycle is divided into samples of equal length based on the sampling frequency, and then divided into a source domain and a target domain according to different working conditions. Next, the autocorrelation function is used to analyze the periodicity of the sample signal, and the K-Means algorithm is used to obtain the significant period length. Then, the signal samples are reshaped based on these significant periods to prepare for subsequent feature extraction. Subsequently, a convolutional neural network is constructed to extract features from the reshaped signals under multiple periods, and these features are aligned and fused through an average pooling layer. Finally, a regression network is constructed using the fused multi-scale periodic degradation features to achieve the prediction of the remaining service life of the rolling bearing. In order to ensure the consistency of feature distribution under different working conditions, the present invention constructs a multi-cycle feature correlation alignment loss function between the source domain and the target domain, and trains and tests the model based on the loss function.

[0044] In order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are described in detail below using the XJTU-SY rolling bearing full life cycle dataset.

[0045] Example 1

[0046] A method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment includes the following steps:

[0047] S1. Example 1 uses the XJTU-SY dataset for experimental verification. Five rolling bearings of model LDK UER204 are experimented under three different operating conditions. Acceleration sensors are placed in the vertical and horizontal directions of the bearing rotation plane to collect the vibration acceleration signals of the rolling bearing throughout its life cycle.

[0048] During operation of the experimental platform, the sampling frequency was 25.6 kHz. The sensors collected data every minute, with each sampling period lasting 1.28 seconds, resulting in 32,768 data points. To reduce redundancy and computational complexity, each 32,768 data points were uniformly downsampled to 4,096 data points. Taking Bearing1_1 as an example, a total of 4,030,464 data points were collected. After partitioning the data by degradation unit time, 123 signal samples were obtained, each with a length of 4,096 points. The bearing datasets under different operating conditions were divided into source and target domains, as shown in Table 1.

[0049] Table 1 XJTU-SY cross-operating condition remaining service life prediction tasks

[0050]

[0051] Taking task T1 as an example, the source domain Bearing2-2 and Bearing2-3 have a total of 694 samples. 161 samples were obtained in the target domain Bearing1-2

[0052] S2. Use the autocorrelation function shown in formula (1) to perform periodicity detection on each sample in the source domain and the target domain. Taking task T1 as an example, the period T of each sample in the source domain and the target domain is {T1, T2, ..., T 855}.

[0053] S3, use K-means algorithm to analyze the sample period T of source domain and target domain = {T1, T2, ..., T 855} for clustering, and obtain K = 3 cluster centers C = {15, 216, 91} as significant cycles. Then, based on the significant cycles, the source domain samples and target domain samples are reshaped according to the following formula:

[0054]

[0055] S4 Figure 2 As shown, a three-layer convolutional neural network is constructed to extract the degradation features of the reshaped sample signal. First, the specific structure of the constructed three-layer convolutional neural network is as follows:

[0056] The first layer includes a convolution layer, a maximum pooling layer, and a nonlinear activation function. The number of input channels is 1, the number of output channels is 16, the convolution kernel size is 2*2, the pooling window size is 2*2, and the activation function is ReLU.

[0057] The second layer includes a convolution layer, a maximum pooling layer, and a nonlinear activation function. The number of input channels is 16, the number of output channels is 32, the convolution kernel size is 2*2, the pooling window size is 2*2, and the activation function is ReLU.

[0058] The third layer includes a convolution layer, a maximum pooling layer, and a nonlinear activation function. The number of input channels is 32, the number of output channels is 64, the convolution kernel size is 2*2, the pooling window size is 2*2, and the activation function is ReLU.

[0059] The samples of the three cycles of source domain and target domain reshaping are respectively extracted using the constructed convolutional neural network to obtain the features {f1, f2, f3}. Then, an average pooling layer with a window size of 4*9 is used to align the feature sizes and vertically splice them, and a convolution layer with a convolution kernel size of 2*2 is used for feature fusion.

[0060] S5. Use the fused multi-cycle degradation features to construct a regression network to complete the transformation from feature extraction to life prediction. The constructed regression network includes a fully connected layer with 100 hidden nodes, a Dropout layer, and an output layer.

[0061] S6. Based on the CORAL loss function, a multi-cycle feature correlation alignment loss function is constructed between the source domain and the target domain. Its expression is as follows:

[0062]

[0063] Among them, the weight of the fusion feature is α∈(0,1), and the weight of the feature in each cycle is MSE pred is the minimum mean square error, which measures the predicted lifespan The difference between the actual life value y, Loss CORALfusion is the source domain feature after periodic feature fusion and target domain features CORAL loss between For each period feature {f1,f2,…,f k}CORAL loss between source domain features and target domain features.

[0064] S7. Taking task T1 as an example, the bearings Bearing2-2 and Bearing2-3 under one working condition are used as the source domain, and the bearing Bearing1-2 under another working condition is used as the target domain to construct the training set, such as Figure 3 As shown in the figure, the source and target domains are trained, and the CORAL loss function is used to align the source and target domain features. Finally, the trained model is tested on Bearing 1-3 and Bearing 1-5 bearings to obtain their predicted life curves.

[0065] Table 2 shows the remaining service life prediction results of 6 test bearings in the XJTU-SY cross-operating condition remaining service life prediction task. The root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation indicators to compare the method proposed in the present invention with the convolutional neural network life prediction method without domain adaptation. The results in the table show that after feature alignment, the RMSE and MAE of the remaining service life prediction of all test bearings were reduced, indicating that the method of the present invention has shown excellent results in narrowing the feature differences between the source domain and the target domain and improving the prediction performance of the target domain. This verifies the effectiveness of this method in bearing life prediction and helps to enhance the generalization ability of the model under different operating conditions.

[0066] Table 2 Prediction error of remaining service life of XJTU-SY test bearings

[0067]

[0068] The proposed method was also compared with classic feature alignment methods such as MECA, MMD, and CMMD. Table 3 shows the prediction error comparison for test bearings 1-5. As can be seen from Table 3, the proposed method achieves lower RMSE and MAE values, outperforming MECA, MMD, and CMMD.

[0069] Table 3 Comparison results of different feature alignment methods for XJTU-SY

[0070]

Claims

1. A method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment, characterized in that: The following steps are involved: S1. Divide the vibration signal of the rolling bearing throughout its life cycle into signal samples of the same length, and select any working condition W. i The bearing signal under the condition W is used as the source domain. j , and the bearing signal under j≠i is taken as the target domain; S2. Use the autocorrelation function to perform periodic detection on each sample in the source domain and the target domain to identify the signal sample period; S3. Use the K-Means algorithm to cluster the signal sample periods, obtain the K most significant period lengths by finding the cluster centers, and reshape the signal samples based on the period lengths; S4. Construct a convolutional neural network to extract and fuse features of the reshaped signal samples with different period lengths to obtain multi-scale period degradation features; S5. Construct a regression network and use multi-scale periodic degradation characteristics to predict the remaining service life; S6. Based on the multi-scale periodic degradation features extracted by the convolutional neural network at different periods, a multi-period feature correlation alignment loss function is constructed between the source and target domains for training, and the predicted life of the bearing is obtained by testing on the test set. The specific process of constructing the multi-cycle feature correlation alignment loss function is as follows: The CORAL loss function achieves feature alignment by minimizing the difference in the covariance matrix between the source domain and the target domain. Its expression is as follows: where f s represents the learned features of the source domain, f s represents the learned features of the target domain, ||·|| F represents the Frobenius norm, d is the feature dimension, and based on the CORAL loss function, a loss function is constructed based on multi-cycle features. Its expression is as follows: where MSE pred is the minimum mean square error, which measures the predicted lifespan The difference between the actual life span value y, is the source domain feature after periodic feature fusion and target domain features CORAL loss between For each period feature {f1,f2,…,f k }, α is the parameter controlling the contribution of this part of the loss; {λ1,λ2,…,λ K } is the coefficient for adjusting the weight of each loss item.

2. The method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment according to claim 1 is characterized in that: The step S1 is specifically implemented as follows: based on the sampling frequency, the vibration signal of the rolling bearing throughout its life cycle is divided into signal samples of the same length, and at the same time, according to different working conditions W={W1, W2, ..., W p }, where l is the number of working conditions, select any working condition W i The bearing signal under the condition W is used as the source domain. j The bearing signal under j≠i is taken as the target domain; the source domain signal sample is obtained as The target domain signal sample is Where n is the number of source domain signal samples, and m is the number of target domain signal samples.

3. The method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment according to claim 2 is characterized in that: The specific process of performing periodic detection in step S2 is as follows: First, the autocorrelation function (ACF) is used to quantify the similarity of signals at different lag times. The formula is as follows: Where R(τ) represents the similarity of the signal x(t) at the lag time τ, τ represents the lag time, that is, the offset of the time series relative to itself, and N represents the length of the time series; Then, based on the significant peaks in the autocorrelation function, the period T = {T1, T2, ..., T n+m }.

4. The method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment according to claim 3 is characterized in that: The specific implementation process of step S3 is as follows: First, the K-means algorithm is used to perform a random sampling of the source and target domains over a period of T = {T1, T2, ..., T n+m } to cluster and obtain K cluster centers C={C1,C2,…,C K } as a significant cycle; Then, the source domain sample S S Based on the significant cycles C={C1,C2,…,C K }Reshape according to the following formula, Reshape into Similarly, the target domain samples It is also reshaped according to the cycle of significance.

5. The method for predicting the remaining life of a bearing across working conditions based on multi-cycle feature alignment according to claim 4 is characterized in that: The specific process of step S4 is as follows: the K period samples reshaped in the source domain and the target domain are respectively subjected to feature extraction using the constructed convolutional neural network to obtain features {f1, f2, ..., f k }, then the average pooling layer is used to align the feature sizes and vertically splice them, and the convolution layer is used for feature fusion.

6. The method for predicting the remaining life of a bearing across operating conditions based on multi-cycle feature alignment according to claim 5 is characterized in that: The regression network constructed in step S5 includes a fully connected layer, a Dropout layer, and an output layer constructed sequentially.

Images