A Deep Transfer Learning-Based Method for Predicting the Remaining Life of Rolling Bearings After Multi-Stage Degradation

Abstract

A method for predicting the remaining life (RUL) of rolling bearings across multiple stages based on deep transfer learning is proposed. First, raw vibration signal data of the rolling bearing is collected and segmented using a sliding window method to obtain multi-stage vibration signal data. Next, the change in the health index curve is obtained by calculating the maximum mean difference between the self-health data and the multi-stage vibration signal data. Then, the bearing degradation stage is determined by applying statistical process analysis to the health index curve. The raw vibration signal data is preprocessed, partitioned, and labeled before being fed into a residual network and an Adaptive-Transformer network for training. Finally, the test set is fed into the Adaptive-Transformer network for prediction, outputting the final predicted RUL of the rolling bearing. This method considers the RUL of the bearing at each stage, reducing the prediction loss RMSE by approximately 0.1 and improving the prediction accuracy by 4%–6%.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and bearing life prediction technology, and in particular to a method for predicting the remaining life of rolling bearings after multi-stage degradation based on deep transfer learning. Background Technology

[0002] As the most basic component of industrial rotating machinery, rolling bearings can cause unplanned economic losses and equipment downtime if they fail during industrial production. When bearings degrade or fail during operation, they directly affect the reliability and stability of the entire equipment. Therefore, bearing remaining life (RUL) prediction plays a crucial role in avoiding dangerous situations before operational risks arise, and has received increasing attention in predictive and health management (PHM) technologies. In recent years, transfer learning (TL) has rapidly developed as an effective tool for addressing differences in domain distribution features. As a branch of TL, domain adaptation (DA) maps features from different domains to a new common feature space to solve the problem. To address the domain shifting problem, DA selects appropriate feature representations to make the distributions of the source and target domains in the new feature space as consistent as possible. DA provides a new solution for RUL prediction, and the model shows better performance under different operating conditions. However, these TL methods often require a large amount of cross-condition data to improve RUL prediction performance. Although the above-mentioned RUL prediction methods have made great progress, the following problems still exist:

[0003] (1) Existing prediction methods often use a single time-domain feature as a health indicator and divide the health indicator curve into degradation stages by setting a threshold. However, since the threshold cannot automatically adapt to the degradation process of the bearing, it is impossible to accurately obtain the multi-level degradation information of the bearing.

[0004] (2) Existing prediction methods do not take into account the correlation between multiple degradation stages in predicting remaining lifespan, and cannot accurately identify health indicators and degradation stages based on the polygon of degradation trends.

[0005] (3) The prediction accuracy of existing transfer learning prediction models largely depends on the amount of data in the source domain and the differences in data distribution under multiple operating conditions. However, in real-world scenarios, it is often difficult to collect enough data to obtain the differences in feature distribution and train the transfer learning network, resulting in poor performance of bearing remaining life prediction results under multiple operating conditions.

[0006] (4) Existing prediction methods only consider the overall prediction accuracy of the model, while ignoring that each degradation stage has similar degradation characteristics that can be used to adjust the model accuracy, thus reducing the prediction accuracy of a single degradation stage.

[0007] (5) Existing prediction methods cannot overcome the problem of decreased accuracy of bearing RUL prediction across operating conditions due to small data volume and data imbalance in actual industrial scenarios. Summary of the Invention

[0008] The purpose of this invention is to improve the low prediction accuracy and poor generalization of bearing cross-condition remaining life prediction models, and to provide a method for predicting the remaining life of rolling bearings at different degradation stages based on deep transfer learning, so as to achieve accurate prediction of the remaining life of bearings at different degradation stages under cross-condition conditions.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for predicting the remaining life of rolling bearings after multi-stage degradation based on deep transfer learning includes the following steps:

[0011] Step 1: Collect the raw vibration signal data of the rolling bearing and use the sliding window segmentation method to obtain multiple vibration signal data segments;

[0012] Step 2: Calculate self-health data S H The maximum mean difference between the vibration signal data in step 1 and the mean difference is measured by the MMD value, and the change of the health index curve is obtained by normalizing the MMD value.

[0013] Step 3: Using the Statistical Process Analysis (SPA) method, the health index curves obtained in Step 2 are used to identify the first prediction time point (FPT) and the sharp transition point (STP), thus obtaining the various degradation stages of the bearing.

[0014] Step 4: Perform data preprocessing on the raw vibration signal data obtained in Step 1, and extract features to obtain feature parameters;

[0015] Step 5: Divide the feature parameters obtained in Step 4 into training set and test set, and label the training data of each degradation stage as in Step 3.

[0016] Step 6: Feed the training set from Step 5 into the classification network and the time series prediction network for training;

[0017] Step 7: Input the test set obtained in Step 5 into the time series prediction network trained in Step 6 for prediction, and output the final rolling bearing remaining life prediction result.

[0018] Further preferred, in step 1, the original vibration signal data of the rolling bearing includes two types: horizontal vibration signal data and vertical vibration signal data. These are collected by an accelerometer installed on the bearing base to obtain the full life cycle data of the bearing operation. The threshold for ending the collection is set when the bearing vibration amplitude exceeds 20g.

[0019] Further preferred, in step 1, the specific method of the sliding window segmentation method is to divide the entire original vibration signal data into equal-distance segments according to the step size l, where l is taken as 64.

[0020] Further preferred, in step 2, the self-health data S H This refers to the first segment of vibration signal data divided into time periods using the sliding window segmentation method described in step 1.

[0021] Further preferably, in step 2, the maximum mean difference (MMD) value is the distance between two distributions in the regenerated Hilbert space, which is the self-health data S. H The degree of difference between each segment of vibration signal data.

[0022] Further preferably, in step 3, the statistical process analysis method involves analyzing the average μ and variance σ of a set of multiple points on the curve. 2 The standard deviation σ is used to locate the point where the curve slope changes abruptly, thereby finding the first prediction time point (FPT) and the sharp transition point (STP) to divide the health indicator curve into multiple stages of degradation.

[0023] Further preferred, in step 4, the data preprocessing of the original vibration signal data includes standardization and normalization, and the feature parameters obtained by feature extraction of the preprocessed data include root mean square value, standard deviation, peak-to-peak value and waveform factor.

[0024] Further preferably, in step 5, the specific method for dividing the feature parameters obtained in step 4 into a training set and a test set is to divide the training set and the test set according to a 7:3 ratio, wherein the training set is represented as: X train = (x1, x2, ..., x K ), where x i =(x 1 x 2 , ..., x k ), where K and k represent the number of samples and the number of feature parameters, respectively, and the test set is similar.

[0025] Further preferred, in step 6, the classification network is a residual network, and the time series prediction network is an Adaptive-Transformer network.

[0026] The advantages and beneficial effects of this invention are as follows:

[0027] (1) This invention proposes a novel method for determining health indicators and classifying multi-stage degradation. Compared with traditional statistical models, this method can accurately obtain health indicators that change with multi-stage degradation of bearings without setting thresholds, reducing computational costs by 65%.

[0028] (2) This invention proposes a multi-degradation stage adaptive identification method that combines residual network and feature extraction. It can accurately identify health indicators and degradation stages based on the variability of bearing degradation trends, and shorten the identification time of degradation stages to less than 10 milliseconds.

[0029] (3) This invention proposes a hierarchical adaptive RUL model based on Adaptive-Transformer, which transfers the learned multi-stage degradation features from one working condition to multiple working conditions, which is beneficial for predicting bearing remaining life across working conditions using small sample datasets.

[0030] (4) This invention proposes a hierarchical adaptive tuning (HAT) method, which reduces the RMSE loss of the adaptive RUL model by about 0.1 and improves the prediction accuracy by 4%-6%.

[0031] (5) This invention combines the bearing degradation process and life prediction to achieve prediction of the remaining life of bearings across operating conditions, and solves the problem that the prediction accuracy is less than 60% when the amount of industrial data is less than 1,500 and the tag data is unbalanced. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the overall process of a method for predicting the remaining life of rolling bearings based on deep transfer learning.

[0033] Figure 2 This is a detailed flowchart of a method for predicting the remaining life of rolling bearings based on deep transfer learning.

[0034] Figure 3 Here is a flowchart of the method for constructing health indicators based on MMD.

[0035] Figure 4 This is a schematic diagram of the residual network structure.

[0036] Figure 5 This is a schematic diagram of the Adaptive-Transformer network structure.

[0037] Figure 6 This is a schematic diagram comparing the results of the present invention with the Adaptive-Transformer without the HAT module. Figures (a) and (b) show the RUL prediction results of test bearing 1_3, and Figures (c) and (d) show the RUL prediction results of test bearing 2_1. Detailed Implementation

[0038] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0039] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific examples. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0040] like Figure 1 and Figure 2 As shown, a method for predicting the remaining life of rolling bearings under multi-stage degradation based on deep transfer learning includes the following steps:

[0041] 1) Acquire the raw vibration signal of the rolling bearing.

[0042] Raw vibration signal data of the rolling bearing was collected, and then segmented into multiple vibration signal data segments using a sliding window method. The raw vibration signal data of the rolling bearing included both horizontal and vertical vibration signal data, which were collected separately using an accelerometer mounted on the bearing base to obtain the bearing's entire life cycle data. A vibration amplitude exceeding 20g was used as the threshold for ending the acquisition. The sliding window segmentation method involved dividing the entire raw vibration signal data into equal-distance segments based on a step size l; in this invention, l was set to 64.

[0043] 2) Construct a health index based on MMD using the maximum mean difference measure.

[0044] Calculate self-health data S H The maximum mean difference (MMD) value between the vibration signal data and the data from step 1 is used, and the change in the health index curve is obtained by normalizing the MMD value. Among these, the change in the health index curve is derived from the health data S. H This refers to the first segment of vibration signal data divided into time periods using the sliding window segmentation method described in step 1). MMD is the most commonly used metric in transfer learning; it measures the distance between two distributions in the reproducing Hilbert space (RHS), and the difficulty of solving it is reduced by introducing kernel learning. Furthermore, MMD can also be used to test whether two samples come from two different distributions. The specific formula for MMD is as follows:

[0045]

[0046] Where S and T are two distinct distributions, ES [·] represents the expectation of distribution S, and f∈F indicates that a set of functions belonging to F is defined in the RHS. In practice, the function set F cannot be arbitrarily defined. In MMD, the function domain is defined as any vector within the unit sphere of the RHS, i.e., ||f||<1. The dot product of vectors f in the RHS of φ(·) in this space can be calculated by the following formula:

[0047] f(x) =<f,φ(x)> H (2)

[0048] The variables are mapped to vectors in the RHS through a kernel function. Substituting equation (2) into equation (1), the MMD formula can be rewritten as:

[0049]

[0050] Where μ S and μ T E represents S [ <f,φ(d S )> H ] and E T [ <f,φ(d T )> H ].

[0051] To address the difficulty in choosing the mapping function, a kernel function method is introduced to represent the squared MMD as follows:

[0052]

[0053]

[0054] Where D S and D T For the characteristics of the source and target domains, ||·|| H The reproducing kernel Hilbert space (RKHS) of the feature map, n S and n T These represent the number of samples in the source and target domains, respectively. The function φ(·) maps the original data to RKHS. This represents the inner product of the mapping process. If dist mmd (D S D T If ) = 0, then D S =D T then dist mmd (D S D T ) represents the average embedding distance between the two feature distributions in RKHS. For ease of calculation, equation (4) can be simplified to the matrix form shown in equation (5):

[0055] MMD = tr(KL)

[0056]

[0057]

[0058] Based on this, the present invention proposes a health index construction method based on MMD to obtain the degree of difference between self-health data s1 and the original vibration data spliced ​​by the sliding window. Typically, measurement data from the initial operating state of a bearing without significant degradation can be considered as self-health data. Taking self-health data s1 as the benchmark, s... i This represents degradation data as the bearing degrades throughout its lifecycle. As the bearing degrades, s1 and s... i The differences between them will gradually increase until they are completely different. To determine the health indicators as the original signal sample changes, s1 and s are calculated. i The difference is assessed using MMD values, thereby providing real-time health status evaluation. The process for constructing health indicators based on MMD is as follows: Figure 3 As shown.

[0059] 3) The health indicator curves obtained in step 2 are identified by statistical process analysis (SPA) to obtain the various stages of degradation.

[0060] Statistical process analysis methods involve analyzing the mean (μ) and variance (σ) of a set of multiple points on a curve. 2 The curve slope abruptly changes using the standard deviation (σ) and standard deviation (FPT) to identify the FPT and STP, thus classifying the health index curve into multi-stage degradation phases. First, a piecewise averaging method is used to process the MMD-based health index curve, resulting in a smooth curve and avoiding data redundancy and unclear degradation trends. Then, by calculating three monitoring benchmarks, abrupt changes are effectively monitored, and detected abrupt changes are recorded as suspicious points. Finally, by finding multiple suspicious points and ranking their abrupt changes, multiple stages of bearing degradation can be accurately obtained. To classify the degradation stages during bearing system operation, this invention proposes a four-stage model for bearing degradation. The specific degradation stages are as follows:

[0061] (1) Healthy stage: During this stage, the bearing operates smoothly and there is no obvious downward trend in health indicators.

[0062] (2) Mild degradation stage: The bearing begins to degrade, and the health indicators show a slight downward trend.

[0063] (3) Slow degradation stage: Compared with the mild degradation stage, health indicators often show a lot of irregular fluctuations, which indicates that the bearing has begun to degrade significantly.

[0064] (4) Accelerated degradation stage: In this stage, health indicators begin to decline at a faster rate, and the bearing will eventually fail in this stage.

[0065] This invention selects three monitoring benchmarks: mean (μ), variance (σ) 2 The formulas for σ and σ are as follows:

[0066]

[0067]

[0068]

[0069] The specific implementation steps for dividing bearings into multi-stage degradation stages using statistical process analysis methods are as follows:

[0070] Step (3.1) yields a new smoothed health index curve by calculating the average data for each step size l′, and is expressed as follows:

[0071] Step (3.2) uses the sliding window method to obtain a subset of the sample point data. Two consecutive subsets can be represented as... and Where W represents the sliding window length, l represents the single movement step size, and p is the starting position of the sliding window movement. The subset set can be represented as Y = (y1, y2, ..., y...). n′ ), where n′ can be obtained from the following formula:

[0072]

[0073] Step (3.3) Calculate the values ​​of two adjacent subsets y i and y i+1 Three baseline statistics are then compared. If y i+1 The baseline value is greater than y i A number of suspicious points will be recorded as mutation points if the number of suspicious points is M times that of mutation points.

[0074] Step (3.4) sorts the increase factors of the suspicious point benchmark values ​​in descending order, and regards the identified suspicious points as abrupt change points, thus accurately realizing the division of bearing degradation stages.

[0075] 4) By preprocessing the original vibration signal obtained in step 1) and extracting features, feature parameters are obtained.

[0076] Step 3) involves preprocessing the original vibration signal, including standardization and normalization. Feature parameters obtained by feature extraction of the preprocessed data include root mean square value, standard deviation, peak-to-peak value, and waveform factor.

[0077] 5) Divide the feature parameters obtained in step 4 into training set and test set, and label the data of each degradation stage through step 3.

[0078] The training set in step 5 is represented as X in the following form. train = (x1, x2, ..., x K ), where x i =(x 1 x 2 , ..., x k ), and K and k represent the number of samples and the number of feature parameters, respectively, and the test set is similar. Since the original vibration signal does not contain a label set, we create a label set using the following formula:

[0079]

[0080] in, Let T represent the label of the i-th sample. FPT The result obtained from step 3 is the sequence number of FPT points in the sample, i.e., the bearing at T FPT Degeneration begins.

[0081] 6) Feed the training set from step 5 into the residual neural network and the adaptive-transformer for training.

[0082] Training the residual network is to achieve adaptive identification of health indicators and multi-level degradation stages, laying the foundation for the remaining lifetime prediction of the Adaptive-Transformer. The Adaptive-Transformer network for cross-condition bearing RUL prediction uses different bearing degradation stages as source and target domains for domain adaptation, trains remaining lifetime prediction models for different degradation stages, including mild degradation prediction models, slow degradation prediction models and accelerated degradation prediction models, and reduces the model loss for each degradation stage through the hierarchical adaptive tuning (HAT) method.

[0083] 6.1) Training of Residual Neural Networks

[0084] The application of residual networks aims to alleviate the vanishing gradient problem in deep convolutional neural networks (CNNs) due to increased network depth. Therefore, to address this issue, the concept of residual blocks is introduced, which can be specifically expressed as follows:

[0085] x l+1 =h(x l )+F(x l w l (11)

[0086] The residual block is divided into two parts: the direct mapping part h(x)l ) and the remaining part F(x) l w l ).like Figure 4 As shown, F(x) l W l It consists of two or three convolution operations.

[0087] To address the issue of excessive reliance on manual methods for classifying bearing degradation stages, a multi-degradation stage adaptive identification method combining residual networks and feature extraction is proposed. This method can monitor changes in health indicators in real time and identify degradation stages, providing a foundation for domain adaptation between different stages in subsequent RUL prediction. The specific steps of the proposed degradation stage adaptive identification method are as follows:

[0088] Step (6.1) Data Processing. Extract the original vibration signal T = (t1, t2, ..., t... N By utilizing the time and frequency domain characteristics of data, we can reduce the amount of data without losing any information, and enable the network to train data more effectively.

[0089] Step (6.2) involves labeling the data at different stages. This is a labeled dataset obtained using the SPA method, where n0 is the number of labels. For the four degradation stages, there are different categories of labels as follows:

[0090] Category 0: Healthy Stage

[0091] Category 1: Mild Degeneration Stage

[0092] Category 2: Slow Degradation Stage

[0093] Category 3: Accelerated Degradation Stage

[0094] Step (6.3) Network Training. The labeled data is input into the network, and then convolution and pooling are performed for feature extraction and downsampling. The learnable parameters are fully trained using residual blocks, and finally the classification result is output through a fully connected layer.

[0095] The adaptive identification method proposed in this invention combines residual networks with feature extraction, which can effectively shorten the optimization path and accelerate network convergence. Compared with traditional methods, this method can adaptively identify health indicators and degradation stages based on changes in degradation trends.

[0096] 6.2) Training of Adaptive-Transformer

[0097] The Transformer, based on an attention mechanism, can convert the distance between any two positions in a sequence into a constant. This allows for better capture of semantic information with longer intervals when analyzing and predicting long sequences, avoiding the weakening of semantic information carried in sequence computation. Furthermore, the Transformer abandons the traditional sequential structure, thus achieving better parallelism. In particular, positional encoding enables it to remember the position of each piece of information, making it easier to use distributed GPUs for parallel training, thereby improving the efficiency of model training. The Transformer mainly includes the following structure:

[0098] Input layer: This includes a source domain data embedding layer, a target domain data embedding layer, and a position encoder layer. The embedding layer maps the data into vectors of dimension d_model, according to the network's computational requirements. Since the Transformer does not contain recursion or convolution, this model adds position encoding to represent the absolute or relative position of information within the sequence, allowing the network to utilize sequence information. The position encoding uses sine and cosine functions, as shown in the following formula:

[0099]

[0100]

[0101] Where pos is the position and i is the dimension. Therefore, each dimension of the position encoding corresponds to a sine signal, and experiments have shown that this function allows the model to easily learn the pattern of position.

[0102] The encoder consists of N stacked encoder layers, each connected sequentially to two sub-layers, a multi-head attention mechanism (MHA), and a position feedforward network (PWFFN). Each sub-layer uses residual jumpers and layer normalization to ensure sufficient training of sub-layer parameters, accelerating convergence. MHA addresses the limitation of self-attention mechanisms overemphasizing their own position when encoding current location information. It learns different queries (Q), keys (K), and values ​​(V) by using h different attention heads to perform different linear projections on the data. Then, h groups of different Q, K, and V values ​​are merged in parallel. Finally, the pooled outputs of h groups are concatenated and transformed through another learnable linear projection to produce the final output. The calculation formula is shown below:

[0103] MH(Q,K,V)=Concat(head1,head2,…,head h W O (14)

[0104] where head i =Attention(QWi Q KW i K VW i V )

[0105] The learnable parameters include and Based on this design, each head can focus on different parts of the input data.

[0106] The decoder consists of N stacked decoder layers. Each encoder has three sequentially connected sub-layers: a masked multi-head attention layer (masked MHA), MHA, and PWFFN. The masked MHA is identical to the MHA in the encoder block except for the addition of a mask. The mask indicates that certain values ​​are masked so that they do not take effect during parameter updates. In time series prediction, for a time step t, the decoder's output only needs to depend on the outputs before time t. Therefore, applying a masked MHA to the decoder's input yields information about previously predicted outputs, essentially recording information between the current input and the current time step.

[0107] Output layer: After passing through N layers of encoders and decoders, the input features undergo a linear transformation and are activated by an activation function to obtain the final output. In bearing RUL prediction, sigmoid is typically chosen as the activation function for the output layer.

[0108] Based on the traditional Transformer, this invention proposes an Adaptive-Transformer that uses each degradation stage of a single bearing as the source and target domains respectively, dynamically reducing the distribution difference between the source and target domains. The structure of the proposed Adaptive-Transformer network is as follows: Figure 5 As shown. In forward propagation, the data from the two degradation stages are used as the source and target domains, respectively. Then, the data is shuffled within each domain, and these domains are used as training data input. In backpropagation, to effectively update the network weights and reduce the distribution differences between domains, a new objective loss function is proposed, which consists of two parts:

[0109] (1) Regression loss L regression :L regression To minimize the prediction error of the training data, the network training parameters can be adaptively adjusted to make the predicted RUL closer to the actual result. This invention patent selects the radial basis function (RBF) as the kernel function and calculates L using the mean squared error (MSE). regression as follows:

[0110]

[0111] Where y i and Let represent the true RUL label and the model-predicted RUL label, respectively, and n represent the batch size during training. (2) Distribution difference loss L MMD (D S D T ):L MMD (D S D T The distribution difference (MMD) represents the difference in the distribution of high-dimensional features between the source and target domains, enabling the network to better extract domain-invariant features. As a non-parametric metric, MMD measures the first-order distribution divergence between two domains. Therefore, the distribution difference loss formula after introducing kernel learning can be expressed as:

[0112]

[0113] in Represents the i-th element of the source field, n S and n T Let represent the size of each batch in the source and target domains, respectively, and k(·, ·) represent the kernel function. Therefore, the network loss function can be defined as:

[0114] L total (D S D T ;θ)=L regression (θ)+L MMD (D S D T (17)

[0115] Where θ represents the learnable parameters in the network.

[0116] To better utilize the self-attention sublayer, this invention proposes the Attention Layer Importance Evaluation (ALIE) method. The ALIE method adaptively matches two degradation stages (D) while capturing temporal relevance. S D T The distribution of self-attention layers. Furthermore, an importance evaluation vector is introduced. To evaluate the relative importance of each self-attention layer, where L represents the number of self-attention layers. For the input degradation stage dataset (D... S D T The loss formula for distribution matching can be calculated as follows:

[0117]

[0118] in and These represent the l-th self-attention layer D. S and D T Data, This represents the l-th self-attention layer D. S and D T The distribution importance between them, d cosine This represents the cosine similarity.

[0119] Finally, considering the network optimization loss of ALIE, it can be re-expressed as:

[0120]

[0121] 7) Input the test set obtained in step 5 into the Adaptive-Transformer trained in step 6 for prediction, and output the final rolling bearing remaining life prediction result.

[0122] To verify the invention, extensive experiments were conducted on the XJTU-SY dataset under two operating conditions and seven cross-domain cases to predict the RUL of rolling bearings.

[0123] Experimental results are as follows Figure 6 As shown in the figures, Figures (a) and (b) present the RUL prediction results for test bearing 1_3, and Figures (c) and (d) present the RUL prediction results for test bearing 2_1. The experiments used data from one bearing under each of the two operating conditions, and the prediction results represent the mild degradation stage, slow degradation stage, and accelerated degradation stage, respectively. The red curve represents the prediction result of the proposed HAT method applied to the Adaptive-Transformer model, while the blue curve represents the prediction result without applying the HAT method. It can be seen from the figures that the red curve has a better fit to the true remaining life and better generalization than the blue curve. This method not only has better accuracy and advantages over other traditional methods, but also can identify multi-stage degradation of bearings, predict the RUL of each stage, and realize the transfer of multi-level domain-invariant features from one operating condition to a cross-condition, significantly improving prediction accuracy.

[0124] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other modifications based on the guidance of this specification and without departing from the scope of the claims of the present invention, and all of these modifications are within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of rolling bearings after multi-stage degradation based on deep transfer learning, characterized in that, Includes the following steps: Step 1: Collect the raw vibration signal data of the rolling bearing and use the sliding window segmentation method to obtain multiple vibration signal data segments; Step 2: Calculate self-health data S H The maximum mean difference (MMD) value is measured between the MMD values ​​and the multiple vibration signal data from step 1, and the change in the health index curve is obtained by normalizing the MMD values. MMD is a metric in transfer learning used to measure the distance between two distributions in the reconstructed Hilbert space and to test whether two samples come from two different distributions. The specific formula for MMD is: ； In the formula, S and T are two different distributions; Let S represent the expected value of the distribution S; This indicates that a set of functions belonging to F is defined in the RHS, and the domain of these functions is defined as any vector within the unit sphere of the RHS, i.e. In this space The dot product of vector f in the RHS is: ； Here, variables are mapped to vectors in the RHS through kernel functions; Step 3: Using the Statistical Process Analysis (SPA) method, the health index curves obtained in Step 2 are used to identify the first prediction time point (FPT) and the sharp transition point (STP), thus obtaining the various degradation stages of the bearing. Step 4: Perform data preprocessing on the raw vibration signal data obtained in Step 1, and extract features to obtain feature parameters; Step 5: Divide the feature parameters obtained in Step 4 into training set and test set, and label the training data of each degradation stage as in Step 3. Step 6: Feed the training set from Step 5 into the classification network and the time series prediction network for training; Step 7: Input the test set obtained in Step 5 into the time series prediction network trained in Step 6 for prediction, and output the final rolling bearing remaining life prediction result.

2. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 1, the original vibration signal data of the rolling bearing includes two types: horizontal vibration signal data and vertical vibration signal data. These are collected by an accelerometer installed on the bearing base to obtain the full life cycle data of the bearing. The threshold for ending the collection is set when the bearing vibration amplitude exceeds 20g.

3. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 1, the specific method of the sliding window segmentation method is to divide the entire original vibration signal data into equal-distance segments according to the step size l, where l is taken as 64.

4. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 2, the self-health data S H This refers to the first segment of vibration signal data divided into time periods using the sliding window segmentation method described in step 1.

5. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 2, the maximum mean difference (MMD) value is the distance between two distributions in the regenerated Hilbert space, which is the self-health data. The degree of difference between each segment of vibration signal data.

6. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 3, the statistical process analysis SPA method is performed by analyzing the average value of a set of multiple points on the curve. ,variance and standard deviation This allows us to pinpoint the location of abrupt changes in the curve's slope, thereby identifying the first prediction point (FPT) and the sharp transition point (STP) to divide the health indicator curve into multiple stages of degradation.

7. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 4, the original vibration signal data is preprocessed, including standardization and normalization. Feature parameters are extracted from the preprocessed data, including root mean square value, standard deviation, peak-to-peak value, and waveform factor.

8. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 5, the specific method for dividing the feature parameters obtained in step 4 into training and test sets is to divide the training and test sets according to a 7:3 ratio, wherein the training set is represented as follows: ,in And K and k represent the number of samples and the number of feature parameters, respectively, and the test set is the same.

9. The method for predicting the remaining life of rolling bearings based on deep transfer learning according to claim 1, characterized in that, In step 6, the classification network is a residual network, and the time series prediction network is an Adaptive-Transformer network.

Images