A bearing residual life prediction method based on cross-domain adaptive reconstruction

Abstract

The application discloses a bearing residual life prediction method based on cross-domain adaptive reconstruction, acquires full life cycle vibration signals of bearings under multiple working conditions, and constructs source domain training samples and target domain test samples; a joint reconstruction model is trained by using source domain normal stage vibration signals; a source domain health index is constructed according to reconstruction errors between source domain vibration signal segments and reconstructed signal segments, and a failure threshold is determined, a target domain health index sequence is generated according to target domain reconstruction errors; a local health baseline is established by using a front health index of a bearing to be predicted, a degradation starting point is determined by means of continuous threshold exceeding, local slope and window head and tail change, a trend of a degradation segment is fitted, and the predicted failure time and residual life are obtained by extrapolating to the failure threshold. The application can alleviate the distribution difference of cross-condition data, reduce the dependence on fault labels and artificial thresholds, and improve the accuracy and robustness of degradation identification and residual life prediction.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault prediction and health management technology, specifically relating to a bearing remaining life prediction method based on cross-domain adaptive reconstruction. Background Technology

[0002] As a core supporting component of rotating machinery, the health status of rolling bearings directly affects the safety and reliability of equipment operation. Accurately constructing a Health Indicator (HI) that can characterize the bearing degradation process is a core prerequisite for achieving early fault warning and Remaining Useful Life (RUL) prediction. However, in actual industrial scenarios, bearings often operate under varying operating conditions and loads, resulting in significant differences in data distribution between different bearings and at different operating stages of the same bearing. Traditional methods are insufficient to effectively address such cross-domain issues.

[0003] Traditional methods for constructing health indicators can be mainly divided into two categories: physical indicators and virtual indicators. Physical indicators are directly derived from vibration signals (such as root mean square and kurtosis), which are simple to calculate but can only reflect a single degradation feature and are difficult to comprehensively describe the performance degradation process of a bearing throughout its entire life cycle. Virtual indicators are constructed through multi-feature fusion (such as principal component analysis and local preservation projection), but the fusion process relies on expert experience for feature selection, resulting in strong subjectivity and weak generalization ability. In addition, most traditional methods assume that the training data and test data are homogenized, and in application scenarios that span multiple operating conditions and bearings, the domain bias problem leads to a sharp decline in model performance.

[0004] In recent years, machine learning, especially deep learning, has demonstrated significant advantages in the field of mechanical fault diagnosis. Deep neural networks can automatically learn hierarchical features from raw vibration signals, avoiding the drawbacks of traditional methods that rely on manual feature extraction, and possess stronger feature representation capabilities and generalization performance. Among them, Convolutional Neural Networks (CNNs) can effectively extract local intrinsic structural information from signals through local connections and weight sharing mechanisms; Recurrent Neural Networks (RNNs) and their variant, Long Short-Term Memory (LSTM), excel at processing time-series data and can capture the long-term dependencies of degradation processes. Existing research has shown that hybrid models combining CNNs and LSTMs can achieve good results in predicting the remaining life of rolling bearings.

[0005] However, the aforementioned deep learning methods still have two limitations: first, most adopt a supervised training paradigm, requiring a large amount of labeled fault data, which is difficult to obtain in actual engineering; second, model training depends on source domain data, and when applied to new operating conditions and new bearings, model performance significantly degrades due to differences in data distribution. Domain Adaptation technology offers a new approach to solving this problem. By using unsupervised domain adaptation methods such as Covariance Alignment (CORAL), the feature extractor is forced to learn domain-invariant representations, effectively mitigating the domain bias problem. Meanwhile, Variational Autoencoders (VAEs), as unsupervised generative models, can measure the degree of anomalies through reconstruction errors, achieving health status assessment without fault labels. Combining VAEs with CORAL allows for both leveraging the reconstruction capabilities of VAEs to characterize health status and eliminating inter-domain differences through covariance alignment, thereby constructing a generalized health index applicable to multiple operating conditions and bearings.

[0006] Chinese invention patent application CN119047116A discloses a method and system for predicting the life of rolling bearings across operating conditions. This method extracts time-domain, frequency-domain, and wavelet packet time-frequency-domain features of vibration signals, classifies bearing degradation states through feature filtering and fuzzy C-means clustering, and uses maximum mean difference and domain discrimination to predict life across operating conditions. However, this method still relies on pre-defined signal features and their filtering results, making it difficult to fully characterize complex nonlinear degradation information. Its degradation starting point depends on changes in clustering states, making it susceptible to vibration noise, short-term fluctuations, and operating condition switching. Furthermore, this method requires source domain life labels and target domain samples for training; when target operating condition samples are difficult to obtain in advance or new operating conditions not covered during the training phase occur, the model's applicability is limited.

[0007] Furthermore, determining the failure threshold is a crucial step in remaining life prediction. Existing methods often rely on experiments or experience to set fixed thresholds, but the failure thresholds for different bearings often vary, and blindly setting a threshold can lead to increased prediction errors. Adaptively determining the failure threshold based on the statistical characteristics of reconstruction errors from normal data can effectively solve this problem and improve the practicality and robustness of the method.

[0008] In summary, existing methods face challenges in cross-domain scenarios, such as feature extraction relying on human experience, poor generalization ability due to domain bias, and difficulty in determining the failure threshold. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a bearing remaining life prediction method based on cross-domain adaptive reconstruction.

[0010] The objective of this invention can be achieved through the following technical solutions: This invention provides a bearing remaining life prediction method based on cross-domain adaptive reconstruction, comprising the following steps: The vibration signals of the bearing throughout its entire life cycle under multiple operating conditions are acquired. The bearing vibration signals used for model training are divided into source domain vibration signals, and the vibration signals of the bearing to be predicted are divided into target domain vibration signals. The source domain vibration signals and the target domain vibration signals are preprocessed to obtain source domain vibration signal segments and target domain vibration signal segments. Vibration signal segments during the normal bearing phase are selected from the source domain vibration signal segments as training samples. The joint reconstruction model is trained using the training samples, so that the joint reconstruction model can reconstruct the vibration signal segments and perform covariance alignment on the implicit representations corresponding to different working conditions. The source domain vibration signal segments corresponding to the entire life cycle of the source domain are input into the trained joint reconstruction model. The source domain health index value is generated based on the reconstruction error between each source domain vibration signal segment and its reconstructed signal segment. The failure threshold is determined based on the source domain health index value. The target domain vibration signal segments are input into the trained joint reconstruction model in chronological order, and a target domain health index sequence is generated based on the reconstruction error between each target domain vibration signal segment and its reconstructed signal segment. A local health baseline is established based on the initial health index values ​​of the target domain health index sequence. The degradation start point is determined based on the change of the target domain health index sequence relative to the local health baseline. The degradation trend is fitted to the target domain health index values ​​after the degradation start point, and the fitted degradation trend is extrapolated to the failure threshold to obtain the predicted failure time. The remaining life of the bearing to be predicted is determined based on the predicted failure time.

[0011] Furthermore, the preprocessing of the source domain vibration signal and the target domain vibration signal to obtain source domain vibration signal segments and target domain vibration signal segments, and the selection of vibration signal segments during the normal bearing phase from the source domain vibration signal segments as training samples, specifically includes: The source domain vibration signal and the target domain vibration signal are decomposed into three levels using the db5 wavelet, and the wavelet coefficients obtained by the decomposition are processed using a soft thresholding method to obtain the denoised source domain vibration signal and the denoised target domain vibration signal. The source domain vibration signal and the target domain vibration signal are segmented without overlap according to the method that each vibration signal segment contains 2560 sampling points, so as to obtain the source domain vibration signal segment and the target domain vibration signal segment. Robust normalization is performed on each source domain vibration signal segment and each target domain vibration signal segment. The robust normalization process includes subtracting the median of the vibration signal segment from each sampled value in the vibration signal segment and dividing by the proportionally scaled absolute median difference. The vibration signal segment corresponding to the first 40% of the vibration signal at the beginning of the entire life cycle of each source domain bearing is determined as the vibration signal segment of the bearing in the normal stage, and the vibration signal segments of the bearing in the normal stage corresponding to multiple operating conditions are used together to form the training sample.

[0012] Furthermore, the joint reconstruction model includes a variational autoencoder and a covariance alignment module; The variational autoencoder includes a long short-term memory encoder, a reparameterized sampling module, and a decoder; The long short-term memory encoder is used to receive vibration signal segments and extract the temporal features of the vibration signal segments, and output the mean and logarithmic variance of the latent representation based on the temporal features; The reparameterized sampling module is used to generate the latent representation based on the mean and the log-variance of the latent representation; The decoder is a fully connected network used to generate a reconstructed signal segment corresponding to the input vibration signal segment based on the implicit representation; The covariance alignment module is used to receive the implicit representations of vibration signal segments under different working conditions obtained by the long short-term memory encoder and the reparameterized sampling module, and to determine the covariance difference between different working conditions based on the implicit representations corresponding to different working conditions.

[0013] Furthermore, the covariance alignment of the implicit representations corresponding to different operating conditions specifically includes: Two different working conditions are selected from multiple working conditions, and the two different working conditions are respectively determined as working condition A and working condition B. The first implicit representation set corresponding to working condition A and the second implicit representation set corresponding to working condition B are obtained, and the covariance matrix of the first implicit representation set and the second implicit representation set are calculated respectively. The covariance alignment loss is calculated based on the difference between the two covariance matrices, and the covariance alignment loss is expressed as follows: in, This represents the covariance alignment loss between operating conditions A and B; The dimension of the implicit representation is represented; Let represent the covariance matrix of the first implicit representation set; Denotes the covariance matrix of the second implicit representation set; Denotes the Frobenius norm; By reducing the covariance alignment loss, the implicit representations corresponding to operating condition A and operating condition B are aligned in terms of covariance statistics, and the covariance alignment is performed for different combinations of operating conditions among multiple operating conditions.

[0014] Furthermore, the step of training a joint reconstruction model including a variational autoencoder and a covariance alignment module using the training samples specifically includes: The training samples corresponding to different working conditions are input into the joint reconstruction model. The reconstruction loss between the vibration signal segment and its reconstructed signal segment, the KL divergence loss corresponding to the hidden representation, and the covariance alignment loss between the hidden representations corresponding to different working conditions are calculated. A total loss function is constructed based on the reconstruction loss, the KL divergence loss, and the covariance alignment loss. The total loss function is expressed as follows: in, Represents the total loss function; This represents the reconstruction loss, which is the Huber loss; This represents the KL divergence loss; This represents the covariance alignment loss; This represents the weighting coefficient of the KL divergence loss; This represents the weighting coefficient of the covariance alignment loss; The model parameters of the joint reconstruction model are updated with the goal of reducing the total loss function until the training of the joint reconstruction model is completed.

[0015] Furthermore, the step of generating source domain health index values ​​based on the reconstruction error between each source domain vibration signal segment and its reconstructed signal segment, and determining the failure threshold based on the source domain health index values, specifically includes: Calculate the average absolute error between the sampled values ​​in each source domain vibration signal segment and the corresponding sampled values ​​in its reconstructed signal segment, and use the average absolute error as the reconstruction error of the corresponding source domain vibration signal segment; The reconstruction error corresponding to the normal stage of the source domain bearing is subjected to standardization, median filtering and exponential moving average smoothing in sequence to obtain the source domain health index value corresponding to the normal stage of the source domain, and the 80th percentile of the source domain health index value corresponding to the normal stage of the source domain is determined as the health reference threshold. The reconstruction error corresponding to the entire life cycle of the source domain is subjected to the same standardization, median filtering and exponential moving average smoothing as the normal stage of the source domain bearing, to obtain the source domain health index value corresponding to the entire life cycle of the source domain. The source domain health index values ​​corresponding to the entire life cycle of bearings from different source domains are merged, and the 99th percentile of the merged source domain health index values ​​is determined as the failure threshold.

[0016] Furthermore, the step of generating a target domain health index sequence based on the reconstruction error between each target domain vibration signal segment and its reconstructed signal segment specifically includes: According to the acquisition sequence of the target domain vibration signal segments during the entire life cycle of the bearing to be predicted, each target domain vibration signal segment is sequentially input into the trained joint reconstruction model to obtain a reconstructed signal segment that corresponds one-to-one with each target domain vibration signal segment. Calculate the average absolute error between the sampled value in each target domain vibration signal segment and the corresponding sampled value in its reconstructed signal segment to obtain the reconstruction error corresponding to each target domain vibration signal segment; The reconstruction error corresponding to each target domain vibration signal segment is subjected to standardization, median filtering and exponential moving average smoothing in sequence to obtain the target domain health index value corresponding to each target domain vibration signal segment. The target domain health index values ​​are arranged according to the acquisition order of the vibration signal segments in the target domain to obtain the target domain health index sequence.

[0017] Furthermore, establishing a local health baseline based on the initial health indicator values ​​of the target domain health indicator sequence specifically includes: The first N% of the target domain health index values ​​are selected from the beginning of the target domain health index sequence to form a local health sample set corresponding to the bearing to be predicted, where N represents the preset proportion of the first data selection. The distribution characteristics of target domain health index values ​​in the local health sample set are statistically analyzed. A local health baseline corresponding to the bearing to be predicted is established based on the distribution characteristics. A local health baseline threshold is determined based on the local health baseline so that the local health baseline threshold adapts to the distribution of front health index values ​​of different bearings to be predicted.

[0018] Furthermore, determining the degradation initiation point based on the change in the target domain health indicator sequence relative to the local health baseline specifically includes: The target domain health indicator sequence is scanned in chronological order using a preset detection window, and for each preset detection window, it is determined whether the following conditions are met simultaneously: Within the preset detection window, there are at least 5 consecutive target domain health indicator values ​​that exceed the local health baseline threshold determined by the local health baseline; The local slope determined by the change of the target domain health index value over time within the preset detection window is greater than the preset minimum slope threshold. The target domain health index value at the end of the preset detection window is higher than the target domain health index value at the beginning of the preset detection window. The first preset detection window that simultaneously meets the above conditions in chronological order is determined as the degradation detection window, and the moment corresponding to the target domain health index value that first exceeds the local health baseline threshold within the degradation detection window is determined as the degradation start point.

[0019] Furthermore, the step of fitting a degradation trend to the target domain health index values ​​after the degradation initiation point, and extrapolating the fitted degradation trend to the failure threshold to obtain the predicted failure time, and determining the remaining life of the bearing to be predicted based on the predicted failure time, specifically includes: From the degradation segment starting from the degradation initiation point, a predetermined proportion of target domain health index values ​​located at the beginning of the degradation segment are selected as fitting data, and the time corresponding to the last target domain health index value in the fitting data is determined as the current prediction time. Using the time corresponding to the fitted data as the independent variable and the target domain health index value in the fitted data as the dependent variable, a cubic polynomial fitting is performed to obtain a degradation trend model used to characterize the degradation trend of the bearing to be predicted. The degradation trend model is extrapolated along the direction of time increase, and the time when the degradation trend model first reaches the failure threshold is determined as the predicted failure time. The time interval between the predicted failure time and the current prediction time is determined as the remaining life of the bearing to be predicted.

[0020] Compared with the prior art, the present invention has the following advantages: (1) Existing bearing remaining life prediction models typically rely on data distribution under specific operating conditions. When the rotational speed, load, or individual bearing changes, the model is prone to misidentifying the differences in operating conditions as bearing degradation, leading to baseline drift of health indicators and a decrease in prediction accuracy. This invention introduces covariance alignment constraints into the implicit representation space of the variational autoencoder, aligning the statistical characteristics of the implicit representations corresponding to different operating conditions. While preserving vibration signal degradation information and signal reconstruction capabilities, it reduces information related to specific operating conditions. This can alleviate the data distribution differences between the source and target domains, improve the comparability and generalization ability of health indicators across different operating conditions and different bearings, and reduce false alarms and remaining life prediction biases caused by changes in operating conditions.

[0021] (2) Existing supervised methods require a large amount of full lifecycle data with fault status labels or remaining life labels. However, complete failure data is difficult to obtain in actual engineering, and artificially constructed life labels may deviate from the actual degradation process of bearings. This invention uses vibration signal segments of the bearing in the normal stage of the source domain to train a joint reconstruction model, and generates health indicators based on the reconstruction error between the vibration signal segments and their reconstructed signal segments, without the need to pre-label fault categories, degradation degrees, or remaining life. As the bearing state gradually deviates from the normal data distribution, the reconstruction error increases accordingly, thus continuously characterizing the process of bearing change from a normal state to a degraded state, and maintaining high sensitivity to early degradation.

[0022] (3) Existing methods for constructing health indicators usually require manual selection of time-domain, frequency-domain, or time-frequency-domain features, which can easily overlook nonlinear degradation information not covered by preset features. This invention extracts the temporal features of vibration signals through a long short-term memory encoder and trains the model jointly with reconstruction loss, KL divergence loss, and covariance alignment loss. This can take into account the signal reconstruction quality, the continuity of the latent representation distribution, and the consistency of features across working conditions, reduce the dependence on manual feature selection, and avoid excessive loss of effective degradation information during domain alignment.

[0023] (4) Existing methods often set uniform health or failure thresholds based on experience, which is difficult to adapt to the normal fluctuation range and failure level of different working conditions and different individual bearings. This invention determines the health reference threshold based on the distribution of health index values ​​in the normal stage of the source domain, and determines the failure threshold based on the distribution of health index values ​​throughout the entire life cycle of the source domain. This makes the health status judgment benchmark and the failure judgment benchmark match the corresponding data distribution, thereby reducing the subjectivity of manually setting thresholds and reducing early alarms caused by excessively low thresholds and lag in degradation recognition caused by excessively high thresholds.

[0024] (5) Existing degradation initiation detection methods often directly use the moment when the health index first exceeds a fixed threshold as the degradation initiation point, which is easily affected by impact noise, abnormal points, and short-term operating condition changes. This invention establishes a local health baseline based on the front health index value of the bearing to be predicted, and determines the degradation initiation point by combining the triple judgment conditions of the health index value continuously exceeding the local health baseline threshold, the local slope being greater than the preset minimum slope threshold, and the health index value at the end of the detection window being higher than the health index value at the beginning of the detection window. Among them, the local health baseline can adapt to the differences in the initial state and normal fluctuation level of different bearings, the continuous exceeding threshold condition is used to exclude isolated abnormal points, and the local slope and the change conditions at the beginning and end of the window are used to confirm the continuity and directionality of the degradation trend, thus improving the noise resistance, robustness, and accuracy of degradation initiation detection.

[0025] (6) Existing degradation trend predictions are easily affected by sudden changes in data in the later stages, and using fixed-length data makes it difficult to adapt to the problem of inconsistent degradation stages in different bearings. This invention selects a predetermined proportion of health index values ​​from the beginning of the degradation segment after the degradation start point and performs polynomial fitting. The degradation trend obtained by fitting is then extrapolated to the failure threshold. This allows for the prediction of the failure time in advance using the already established and relatively stable degradation trend, reducing the interference of drastic fluctuations in the near-failure stage on the fitting results, and establishing a continuous correspondence between health indexes, failure thresholds, and remaining life calculations.

[0026] (7) The present invention employs consistent noise reduction, segmentation and robust standardization processing for source domain vibration signals and target domain vibration signals, and generates source domain health index values ​​and target domain health index sequences in the same way. This can reduce the impact of noise, abnormal amplitude and data scale differences on model training and reconstruction errors, and ensure that the failure threshold and target domain health index have a unified data caliber. Attached Figure Description

[0027] Figure 1 This is a flowchart of the bearing health index construction and remaining life prediction method based on cross-domain adaptive reconstruction according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the joint reconstruction model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the training set reconstruction error distribution and threshold determination in an embodiment of the present invention. Figure 4 This is a graph showing the test bearing health index curve and remaining life prediction results according to an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] Example 1: like Figure 1 , Figure 2 As shown, this embodiment provides a bearing remaining life prediction method based on cross-domain adaptive reconstruction, which can be used for health status assessment and remaining life prediction of rolling bearings throughout their entire life cycle. This method uses vibration signals from multiple bearings that have completed their entire life cycle as the data source for model training, and the vibration signal of the bearing to be predicted as the prediction object. It establishes a consistent health status representation across operating conditions through a joint reconstruction model, and determines the degradation initiation point and predicted failure time based on the health index sequence.

[0030] This embodiment applies to the PHM 2012 bearing dataset, which contains full-lifecycle vibration signals of multiple bearings under three operating conditions. This embodiment uses the first two bearings in operating conditions 2 and 3 as source domain bearings, namely Bearing2_1 and Bearing2_2 in operating condition 2, and Bearing3_1 and Bearing3_2 in operating condition 3; and uses the bearing in operating condition 1 as the target domain bearing. Specifically, the speed in operating condition 2 is 1650 rpm and the load is 4200 N, the speed in operating condition 3 is 1500 rpm and the load is 5000 N, and the speed in operating condition 1 is 1800 rpm and the load is 4000 N. The differences in speed and load among these operating conditions can be used to verify the applicability of the joint reconstruction model under cross-operating condition conditions.

[0031] It should be noted that in this embodiment, the source domain represents one or more sets of known operating condition data used to establish normal reconstruction patterns and determine failure judgment criteria; the target domain represents the data set of the operating condition of the bearing to be predicted; the vibration signal segment represents a fixed-length signal extracted from a single acquisition record or continuous vibration signal; the latent representation represents the potential feature vector obtained after the vibration signal segment is processed by the encoder; and the health index value represents the bearing health status characterization value formed based on the reconstruction error between the vibration signal segment and its reconstructed signal segment. The source domain and target domain can be divided according to operating condition, bearing number, equipment number, or operating batch, as long as the data used for training and the data of the bearing to be predicted can be distinguished.

[0032] The following combination Figure 1 and Figure 2 The implementation process of this embodiment will be described.

[0033] Step S1: Obtain the full life cycle vibration signals of the bearing under multiple operating conditions, and preprocess the source domain vibration signals and target domain vibration signals to obtain source domain vibration signal segments and target domain vibration signal segments.

[0034] In this embodiment, the acceleration vibration signals of each bearing are read from the PHM 2012 bearing dataset, with a sampling frequency of 25.6 kHz. Each acquisition record contains 2560 sampling points, corresponding to a 0.1-second vibration waveform; the actual time interval between adjacent acquisition records is determined based on the timestamps in the data file, which is 10 seconds in this embodiment. Therefore, when calculating the degradation initiation point, predicting the failure time, and remaining lifetime, the actual timestamps corresponding to the acquisition records are used, instead of directly using the duration of each 0.1-second waveform as the time interval between adjacent health indicator values, thus ensuring that the time unit of remaining lifetime is consistent with the actual acquisition process of the dataset.

[0035] To denoise the original vibration signal, a three-level wavelet decomposition using the db5 wavelet is employed, and soft thresholding is used to process the detail coefficients at each level. Then, an inverse wavelet transform is performed to obtain the denoised vibration signal. The soft threshold can be determined based on the noise standard deviation and signal length; symmetrical extension can be used at boundary locations to reduce wavelet decomposition boundary effects. Subsequently, the signal is segmented without overlap, with each vibration signal segment containing 2560 sampling points, ensuring a one-to-one correspondence between each vibration signal segment and a single acquisition record.

[0036] In another specific embodiment, when the acquisition device outputs a continuous vibration signal, it can be segmented using a sliding window with a preset overlap rate, such as 25% to 75%. Using non-overlapping segments can reduce redundant calculations and maintain a direct correspondence between health index values ​​and acquisition records; using overlapping segments can improve the temporal resolution near changes in degradation state, and is suitable for scenarios with low acquisition frequency or rapid changes in degradation process.

[0037] Step S2: Select vibration signal segments of the bearing during the normal stage from the source domain vibration signal segments as training samples, and standardize the source domain vibration signal segments and the target domain vibration signal segments.

[0038] In this embodiment, the vibration signal segment of the first 40% of the bearing's entire life cycle in each source domain is defined as the vibration signal segment of the bearing's normal stage. The vibration signal segments of the normal stage corresponding to different source domain bearings and different source domain operating conditions are used together to form the training sample. By using early data as normal stage data, the joint reconstruction model mainly learns the vibration signal distribution when the bearing has not undergone significant degradation, avoiding the reduction of reconstruction error sensitivity to abnormal states after degradation stage samples are included in the training set.

[0039] Robust normalization is performed on each vibration signal segment. For the first... The first vibration signal segment The robust normalization result of each sample value is expressed as follows: in, Indicates the first The j-th sampled value in a vibration signal segment; This represents the corresponding robust standardization result; Indicates the first The median of all sampled values ​​in each vibration signal segment; Indicates the first The median of the absolute deviations of each sampled value relative to the median in each vibration signal segment; 1.4826 is a scaling factor that makes the absolute median difference correspond to the standard deviation scale of the Gaussian distribution; This represents the stability constant to prevent the denominator from being zero.

[0040] When the absolute median difference of a certain vibration signal segment is less than a preset lower limit, the preset lower limit can be used to replace the absolute median difference to avoid numerical amplification when the signal is approximately constant. In another specific embodiment, a uniform median and absolute median difference can also be calculated based on all normal training samples in the source domain, and this uniform statistic can be used to process the source and target domain data. Segment-wise robust standardization is beneficial for suppressing amplitude scale differences between different acquisition records; using a uniform statistic is beneficial for preserving the degradation information contained in the absolute change of vibration amplitude, and can be selected according to the consistency of sensors and the degree of difference in operating conditions.

[0041] Step S3: Construct a joint reconstruction model including a variational autoencoder and a covariance alignment module.

[0042] like Figure 2 As shown, the joint reconstruction model includes a long short-term memory encoder, a reparameterized sampling module, a decoder, and a covariance alignment module. The long short-term memory encoder receives the standardized vibration signal segment, extracts temporal features sequentially along the sampling points, and outputs the mean vector and log-variance vector of the latent representation based on the temporal features at the encoder end. The reparameterized sampling module generates the latent representation based on the mean vector and log-variance vector, and the decoder uses a fully connected network to map the latent representation into a reconstructed signal segment with the same length as the input vibration signal segment.

[0043] During the training phase, the process of generating the hidden representation is represented as follows: in, This represents the implicit representation; This represents the mean vector output by the Long Short-Term Memory encoder; This represents the log-variance vector of the output of the Long Short-Term Memory encoder; The variance vector representing the implicit representation; This represents a random vector sampled from a standard normal distribution; ⊙ represents the identity matrix; ⊙ represents element-wise multiplication. The mean vector can be used in the prediction phase. As a deterministic implicit representation, it can reduce the impact of random sampling on the reconstruction error of the same vibration signal segment; it can also perform multiple samplings and average the results of multiple reconstructions to characterize the uncertainty of the model.

[0044] The covariance alignment module receives the set of latent representations corresponding to different source domain conditions. Taking condition A and condition B as examples, the latent representations belonging to condition A and condition B within the same training batch are arranged row-wise to form the first latent representation matrix. Second implicit representation matrix And calculate the corresponding covariance matrix. Operating conditions The covariance matrix can be expressed as: in, Indicates working conditions The corresponding implicit representation is the covariance matrix; Indicates working conditions The corresponding implicit representation matrix; This indicates that the current training batch belongs to the operating condition. The number of samples; Indicates length is A column vector whose elements are all 1; Indicates working conditions The mean vector corresponding to the implicit representation; the superscript T indicates matrix transpose.

[0045] Calculate the covariance alignment loss based on the covariance matrices corresponding to operating conditions A and B: in, This represents the covariance alignment loss between operating conditions A and B; The dimension of the implicit representation is represented; Let represent the covariance matrix of the first implicit representation set; Denotes the covariance matrix of the second implicit representation set; Denotes the Frobenius norm. By reducing... This makes the second-order statistical characteristics of the implicit representations corresponding to different working conditions tend to be consistent, thereby reducing the influence of speed and load differences in the implicit space.

[0046] When the source domain contains more than two work cases, the covariance alignment loss of each pairwise combination of work cases can be calculated, and the average of all combinations can be taken as the covariance alignment loss for the current training batch. Alternatively, a source domain work case with a larger sample size can be selected as the reference work case, and the remaining work cases can be aligned with the reference work case. Pairwise alignment reduces the dependence on the selection of the reference work case, while alignment with the reference work case reduces the computational cost. To ensure the stability of the covariance estimation, each work case participating in the alignment should preferably include at least two training samples in the same training batch.

[0047] Step S4: Train the joint reconstruction model using training samples.

[0048] The training dataset consists of vibration signal segments from the normal operating stages of each bearing in the source domain, identified by the operating condition number. During training, batches are constructed according to a domain balancing method, ensuring that each batch contains training samples from different source domain operating conditions as much as possible. The variational autoencoder uses Huber loss for reconstruction. The Huber function represents the error e between the input vibration signal segment and the reconstructed signal segment at the corresponding sampling point: in, This represents the Huber loss corresponding to a single sampling point; This represents the difference between the input sample value and the reconstructed sample value; This represents the absolute value of the difference e; The transition parameter represents the Huber loss. It is applied to all vibration signal segments and all sampling points within a training batch. Taking the average, we obtain the reconstruction loss. Huber loss employs a squared penalty when the error is small and a linear penalty when the error is large, which can reduce the impact of random shock values ​​on the training process while ensuring reconstruction accuracy.

[0049] The KL divergence loss corresponding to the implicit representation is expressed as: in, Indicates the KL divergence loss; Indicates the number of samples in the training batch; The dimension of the implicit representation; Let r be the r-th component in the latent representation mean vector of the b-th training sample. Let represent the r-th component in the variance vector of the latent representation of the b-th training sample. KL divergence loss makes the latent representation distribution closer to the standard normal distribution, thereby improving the continuity of the latent space and reducing the possibility that the model only remembers a few training samples.

[0050] The total loss function is a weighted average of the reconstruction loss, KL divergence loss, and covariance alignment loss. in, Represents the total loss function; This represents the reconstruction loss, which is the Huber loss; This represents the KL divergence loss; This represents the covariance alignment loss; This represents the weighting coefficient of the KL divergence loss; The weighting coefficients of the covariance alignment loss are represented; the covariance alignment term is added to the total loss with a positive weight as a penalty term, so that the optimization direction of reducing the total loss also reduces the covariance difference between the implicit representations of different operating conditions.

[0051] In this embodiment, the Adam optimizer is used to update the model parameters, with an initial learning rate of 0.0003, a batch size of 64, and 30 training rounds. The value can be gradually increased from 0 to 0.001 in the early stages of training to avoid the KL divergence constraint suppressing the reconstruction capability too early. This can be determined using source domain validation data. During training, the total loss and the changes in each loss component are recorded. Training ends when the number of training rounds reaches 30 or the total loss improvement is lower than the preset value for several consecutive rounds. The model parameters with lower validation reconstruction errors are saved to obtain the trained joint reconstruction model.

[0052] Step S5: Input the source domain vibration signal segment corresponding to the entire life cycle of the source domain into the trained joint reconstruction model, generate source domain health index values ​​based on the reconstruction error, and determine the health reference threshold and failure threshold.

[0053] First, the vibration signal segments of the normal phase of the source domain are input into the trained joint reconstruction model to obtain reconstructed signal segments that correspond one-to-one with the vibration signal segments of each source domain. The average absolute reconstruction error of each vibration signal segment is expressed as: in, Indicates the first The average absolute reconstruction error of each vibration signal segment; This indicates the number of sampling points in a vibration signal segment; in this embodiment, it is 2560. Indicates the first The first input vibration signal segment The mean absolute reconstruction error is the absolute value of the difference between a sampled value and its corresponding reconstructed sampled value. It provides a relatively intuitive response to local anomalies and is compatible with the Huber loss used during the training phase.

[0054] The total average absolute reconstruction error during the normal phase of the source domain is used as the reference error set. To ensure consistent data scale for health indicators corresponding to different bearings, each reconstruction error can be standardized based on the median and absolute median difference of the reference error set, followed by median filtering and exponential moving average smoothing. The exponential moving average process is expressed as follows: in, Indicates the first The health indicator values ​​corresponding to each vibration signal segment; Indicates the first The result of median filtering of the standardized reconstruction error; Represents the coefficient of the exponential moving average. In this embodiment, the median filter window can take five health indicator values. The median filter is used to remove isolated spikes, and the exponential moving average is used to suppress high-frequency fluctuations and preserve the degradation trend over time.

[0055] Collect all health indicator values ​​corresponding to the normal stage of the source domain, and determine their 80th percentile as the health reference threshold, such as... Figure 3As shown on the left. The health reference threshold is used to describe the main fluctuation range of health index values ​​under normal conditions in the source domain. Subsequently, the full life cycle vibration signal segments of each training bearing in the source domain are input into the trained joint reconstruction model, and the full life cycle health index sequence of the source domain is generated according to the same mean absolute error, standardization, median filtering and exponential moving average processing procedure; the full life cycle health index values ​​of all bearings in the source domain are merged, and the 99th percentile of the merged health index value is determined as the failure threshold, such as... Figure 3 As shown on the right.

[0056] The health reference threshold and the failure threshold can be expressed as follows: in, Indicates a health reference threshold; Indicates the failure threshold; Represents 80th percentile operations; Represents 99th percentile operations; This represents the set of health indicator values ​​during the normal phase of the source domain; This represents the set of health indicator values ​​for the entire lifecycle of the source domain.

[0057] In another specific implementation, the health reference threshold can be determined by adding a preset multiple of the standard deviation to the mean of health indicators during the normal stage, and the failure threshold can also be determined by adding a preset multiple of the standard deviation to the mean of health indicators throughout the entire life cycle. Alternatively, candidate thresholds can be determined separately based on different source domain operating conditions, and then the median or weighted average of the candidate thresholds can be taken. Using quantiles is insensitive to extreme outliers and is suitable for situations where the distribution of health indicators is skewed; using the mean and standard deviation facilitates statistical interpretation when health indicators are approximately symmetrically distributed.

[0058] Step S6: Input the target domain vibration signal segments into the trained joint reconstruction model in chronological order to generate a target domain health index sequence.

[0059] For the bearing to be predicted in the target domain, the vibration signal segments are sequentially input into the trained joint reconstruction model according to the acquisition time sequence of each vibration signal segment throughout its entire life cycle, obtaining reconstructed signal segments that correspond one-to-one with the vibration signal segments in each target domain. Using the same mean absolute reconstruction error calculation method and the same standardization, median filtering, and exponential moving average processing procedures as the source domain, the target domain health index values ​​corresponding to each vibration signal segment are obtained and arranged in chronological order to form a target domain health index sequence. The source and target domains use the same data processing caliber, ensuring that the failure threshold determined in the source domain can be used in the target domain health index sequence.

[0060] Figure 4The health index curve and remaining life prediction results for the target domain bearing Bearing1_3 are shown. The health index curve remains relatively stable in the early stage, gradually rises in the middle stage, and increases rapidly in the later stage, corresponding to the bearing's state change from normal operation and slow degradation to accelerated degradation. It should be noted that the horizontal axis of the health index sequence can use either the vibration signal segment number or the actual acquisition time corresponding to each vibration signal segment; the actual acquisition time is preferred for calculating the remaining life.

[0061] Step S7: Establish a local health baseline based on the initial health index values ​​of the target domain health index sequence, determine the starting point of degradation, and fit and extrapolate the degradation trend of the target domain health index values ​​after the starting point of degradation.

[0062] First, the top N% of target domain health index values ​​are selected from the beginning of the target domain health index sequence to form a local health sample set corresponding to the bearing to be predicted. In this embodiment, N=10; when the early health stage of the bearing is relatively long, N can be set between 5 and 20. A local health baseline is established based on the median and absolute median difference of the local health sample set, and the local health baseline threshold is determined. in, Indicates the local health baseline threshold; Represents a local set of healthy samples; It represents the median of the local healthy sample set and serves as the local healthy baseline; Represents the absolute median difference of a local healthy sample set; This represents the local threshold coefficient. This local health baseline originates from the bearing being predicted and can reflect the differences in the initial state and normal fluctuation range of different bearings.

[0063] In another specific implementation, the preset high quantile of the local healthy sample set can be directly determined as the local health baseline threshold, for example, using the 90th to 99th quantile; alternatively, the mean of the local healthy sample set plus three times the standard deviation can be used. The median and absolute median deviation methods are highly robust to a small number of outliers in the preceding data, while the high quantile method is simple to calculate and does not require the health indicators to approximately follow a normal distribution.

[0064] The target domain health indicator sequence is scanned sequentially using a preset detection window. In this embodiment, the detection window length W can be 10 health indicator values, and the scanning step size is 1. For each detection window, continuous threshold exceeding condition, local slope condition, and window end rising condition are determined respectively. The local slope is the univariate linear least squares slope of the health indicator value within the detection window relative to the actual acquisition time, which is expressed as: in, This represents the local slope within the current detection window; W represents the number of health indicator values ​​within the detection window. Indicates the number of detection windows. The data collection time corresponding to each health indicator value; This represents the average value at each acquisition time within the detection window; Indicates the number of detection windows. Health indicator values ​​for each target domain; This represents the average value of the health indicators for each target domain within the detection window.

[0065] A detection window is identified as a candidate degradation detection window when it meets all three of the following conditions: First, there are at least five consecutive target domain health index values ​​within the window that exceed the local health baseline threshold; Second, the local slope k is positive and greater than the preset minimum slope threshold; Third, the target domain health index value at the end of the detection window is higher than the target domain health index value at the beginning of the detection window.

[0066] The target domain health indicator sequence is scanned from front to back. The first detection window that simultaneously meets all three conditions mentioned above is identified as the degradation detection window. The moment when the target domain health indicator value within this degradation detection window first exceeds the local health baseline threshold is identified as the degradation start point. The continuous exceedance threshold condition is used to exclude isolated outliers, the local slope condition is used to confirm that the health indicator has a continuous upward trend, and the window end rise condition is used to exclude fluctuations where the value rises briefly and then falls back. If no detection window that meets all three conditions is found, it can be concluded that the bearing has not yet entered the identifiable degradation stage, and the scanning continues after a new target domain vibration signal segment is obtained.

[0067] After determining the degradation initiation point, a predetermined proportion of target domain health index values ​​located at the beginning of the degradation segment are selected as fitting data from the degradation segment starting from the degradation initiation point. In this embodiment, the first 30% of the observed degradation segment data can be selected, and the number of fitting data is not less than the minimum sample size used to determine the coefficients of the cubic polynomial. The time corresponding to the last target domain health index value in the fitting data is determined as the current prediction time. .

[0068] Using the actual data collection time corresponding to the fitted data as the independent variable and the target domain health index value as the dependent variable, a cubic polynomial fitting was performed to obtain the degradation trend model: in, This represents the fitted health index value corresponding to time t; , , and This represents the coefficients of the cubic polynomial obtained using the least squares method. To reduce ill-conditioned problems caused by excessively large absolute time values, the time corresponding to the degradation start point can be used as the time origin, and the time can be normalized according to the acquisition interval before fitting; after extrapolation, it can be converted back to the actual time.

[0069] Extrapolating the degradation trend model along the direction of increasing time, and solving for the model that satisfies... and The real solutions are obtained, and the smallest real solution is determined as the predicted failure time. If multiple real solutions exist that are greater than the current prediction time, the solution that reaches the failure threshold first is selected. If no real solutions exist within the preset maximum prediction range, the observed degradation data can be added and the model refitted, or extrapolation can be performed using a quadratic polynomial, exponential function, or regression model with monotonic constraints that guarantees monotonic increase. Cubic polynomials can describe the nonlinear trend from slow degradation to accelerated degradation, while monotonic constraint models can reduce the nonphysical explanations caused by local fluctuations in health indicators.

[0070] The remaining life of the bearing to be predicted is expressed as: in, This indicates the remaining life of the bearing to be predicted; Indicates the predicted failure time; This indicates the current prediction time. When the vibration signal segment number is used as the time variable for fitting and extrapolation, the difference in the number needs to be multiplied by the actual time interval between adjacent acquisition records to obtain the remaining lifespan expressed in seconds, minutes, or hours.

[0071] Step S8: Output the target domain health index curve, degradation start point, predicted failure time and remaining lifetime prediction results, and verify the prediction results.

[0072] Repeat steps S6 and S7 for each bearing in the target domain of operating condition 1 to obtain the health index curves and remaining life prediction results for each bearing. The prediction error percentage (PE) and mean absolute percentage error (MAPE) are expressed as follows: in, Indicates the projected remaining lifespan; Indicates the actual remaining lifespan; This represents the percentage of prediction error for a single target domain bearing. This indicates the number of bearings in the target domain included in the statistics; Indicates the first The absolute value of the prediction error percentage for each target domain bearing; MAPE represents the mean absolute percentage error.

[0073] Table 1 shows the predicted remaining life of the bearing under certain operating conditions in this embodiment.

[0074] Table 1 As shown in Table 1, the algebraic sum of the percentage errors of the five bearings is -2.80%, the average percentage error is -0.56%, and the average absolute percentage error is 1.42%. Figure 4 The health index curves shown can reflect the degradation process of early stability, slow increase in the middle stage and rapid increase in the later stage, indicating that cross-condition covariance alignment, health index construction based on reconstruction error, and degradation initiation detection combining local health baseline and triple judgment can jointly support the prediction of remaining life.

[0075] The overall workflow for constructing health indicators and predicting remaining life consists of steps S1 to S8. A joint reconstruction model learns consistent normal reconstruction patterns across operating conditions from normal source domain data, using reconstruction errors to form health indicators. Health reference thresholds and failure thresholds are determined using the normal source domain stage and full lifecycle health indicator values, respectively, reducing reliance on manual fault labels and fixed empirical thresholds. A local health baseline is established using the initial health indicator values ​​of the bearing to be predicted, and the degradation starting point is determined by combining three conditions: continuous exceeding of thresholds, local slope, and upward movement at the end of the window. This reduces misjudgments caused by isolated anomalies and short-term operating condition fluctuations.

[0076] The specific bearing number, data ratio, filtering window, smoothing coefficient, detection window, loss weight, and fitted data ratio in this embodiment are used to illustrate one possible implementation and do not constitute a limitation on the application scenario. While maintaining the normal training of source domain data, the alignment of implicit representation covariance under different operating conditions, the reconstruction error health index, the source domain statistical threshold, the local health baseline of the target bearing, and the extrapolation relationship of the degradation trend unchanged, those skilled in the art can adjust the corresponding parameters according to the bearing model, sampling frequency, number of operating conditions, and degradation rate.

[0077] In another specific implementation, the source domain may include bearing data from different devices or sensor locations, with the device number or sensor location number used as the domain identifier; the target domain may also contain only partial lifecycle data of the bearing to be predicted up to the current moment. This method can be used for online remaining life prediction. Each time a new vibration signal segment of the target domain is obtained, the target domain health index sequence is updated, and the degradation start point and predicted failure time are reassessed, thus forming a rolling prediction result that is updated with operational data.

[0078] The above embodiments are used to further illustrate the technical solution of the present invention and are not intended to limit the present invention. For those skilled in the art, equivalent substitutions made to the signal segmentation method, encoder structure, threshold statistics method, smoothing method, or degradation trend model without departing from the technical concept of the present invention can all be used to achieve bearing cross-operating condition health status assessment and remaining life prediction.

[0079] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bearing remaining life prediction method based on cross-domain adaptive reconstruction, characterized in that, Includes the following steps: The vibration signals of the bearing throughout its entire life cycle under multiple operating conditions are acquired. The bearing vibration signals used for model training are divided into source domain vibration signals, and the vibration signals of the bearing to be predicted are divided into target domain vibration signals. The source domain vibration signals and the target domain vibration signals are preprocessed to obtain source domain vibration signal segments and target domain vibration signal segments. Vibration signal segments during the normal bearing phase are selected from the source domain vibration signal segments as training samples. The joint reconstruction model is trained using the training samples, so that the joint reconstruction model can reconstruct the vibration signal segments and perform covariance alignment on the implicit representations corresponding to different working conditions. The source domain vibration signal segments corresponding to the entire life cycle of the source domain are input into the trained joint reconstruction model. The source domain health index value is generated based on the reconstruction error between each source domain vibration signal segment and its reconstructed signal segment. The failure threshold is determined based on the source domain health index value. The target domain vibration signal segments are input into the trained joint reconstruction model in chronological order, and a target domain health index sequence is generated based on the reconstruction error between each target domain vibration signal segment and its reconstructed signal segment. A local health baseline is established based on the initial health index values ​​of the target domain health index sequence. The degradation start point is determined based on the change of the target domain health index sequence relative to the local health baseline. The degradation trend is fitted to the target domain health index values ​​after the degradation start point, and the fitted degradation trend is extrapolated to the failure threshold to obtain the predicted failure time. The remaining life of the bearing to be predicted is determined based on the predicted failure time.

2. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The preprocessing of the source domain vibration signal and the target domain vibration signal to obtain source domain vibration signal segments and target domain vibration signal segments, and the selection of vibration signal segments during the normal bearing phase from the source domain vibration signal segments as training samples, specifically includes: The source domain vibration signal and the target domain vibration signal are decomposed into three levels using the db5 wavelet, and the wavelet coefficients obtained by the decomposition are processed using a soft thresholding method to obtain the denoised source domain vibration signal and the denoised target domain vibration signal. The source domain vibration signal and the target domain vibration signal are segmented without overlap according to the method that each vibration signal segment contains 2560 sampling points, so as to obtain the source domain vibration signal segment and the target domain vibration signal segment. Robust normalization is performed on each source domain vibration signal segment and each target domain vibration signal segment. The robust normalization process includes subtracting the median of the vibration signal segment from each sampled value in the vibration signal segment and dividing by the proportionally scaled absolute median difference. The vibration signal segment corresponding to the first 40% of the vibration signal at the beginning of the entire life cycle of each source domain bearing is determined as the vibration signal segment of the bearing in the normal stage, and the vibration signal segments of the bearing in the normal stage corresponding to multiple operating conditions are used together to form the training sample.

3. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The joint reconstruction model includes a variational autoencoder and a covariance alignment module; The variational autoencoder includes a long short-term memory encoder, a reparameterized sampling module, and a decoder; The long short-term memory encoder is used to receive vibration signal segments and extract the temporal features of the vibration signal segments, and output the mean and logarithmic variance of the latent representation based on the temporal features; The reparameterized sampling module is used to generate the latent representation based on the mean and the log-variance of the latent representation; The decoder is a fully connected network used to generate a reconstructed signal segment corresponding to the input vibration signal segment based on the implicit representation; The covariance alignment module is used to receive the implicit representations of vibration signal segments under different working conditions obtained by the long short-term memory encoder and the reparameterized sampling module, and to determine the covariance difference between different working conditions based on the implicit representations corresponding to different working conditions.

4. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The covariance alignment of the implicit representations corresponding to different operating conditions specifically includes: Two different working conditions are selected from multiple working conditions, and the two different working conditions are respectively determined as working condition A and working condition B. The first implicit representation set corresponding to working condition A and the second implicit representation set corresponding to working condition B are obtained, and the covariance matrix of the first implicit representation set and the second implicit representation set are calculated respectively. The covariance alignment loss is calculated based on the difference between the two covariance matrices, and the covariance alignment loss is expressed as follows: in, This represents the covariance alignment loss between operating conditions A and B; The dimension of the implicit representation is represented; Let represent the covariance matrix of the first implicit representation set; Denotes the covariance matrix of the second implicit representation set; Denotes the Frobenius norm; By reducing the covariance alignment loss, the implicit representations corresponding to operating condition A and operating condition B are aligned in terms of covariance statistics, and the covariance alignment is performed for different combinations of operating conditions among multiple operating conditions.

5. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The training of the joint reconstruction model, including the variational autoencoder and the covariance alignment module, using the training samples specifically includes: The training samples corresponding to different working conditions are input into the joint reconstruction model. The reconstruction loss between the vibration signal segment and its reconstructed signal segment, the KL divergence loss corresponding to the hidden representation, and the covariance alignment loss between the hidden representations corresponding to different working conditions are calculated. A total loss function is constructed based on the reconstruction loss, the KL divergence loss, and the covariance alignment loss. The total loss function is expressed as follows: in, Represents the total loss function; This represents the reconstruction loss, which is the Huber loss; This represents the KL divergence loss; This represents the covariance alignment loss; This represents the weighting coefficient of the KL divergence loss; This represents the weighting coefficient of the covariance alignment loss; The model parameters of the joint reconstruction model are updated with the goal of reducing the total loss function until the training of the joint reconstruction model is completed.

6. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The process of generating source domain health index values ​​based on the reconstruction error between each source domain vibration signal segment and its reconstructed signal segment, and determining the failure threshold based on the source domain health index values, specifically includes: Calculate the average absolute error between the sampled values ​​in each source domain vibration signal segment and the corresponding sampled values ​​in its reconstructed signal segment, and use the average absolute error as the reconstruction error of the corresponding source domain vibration signal segment; The reconstruction error corresponding to the normal stage of the source domain bearing is subjected to standardization, median filtering and exponential moving average smoothing in sequence to obtain the source domain health index value corresponding to the normal stage of the source domain, and the 80th percentile of the source domain health index value corresponding to the normal stage of the source domain is determined as the health reference threshold. The reconstruction error corresponding to the entire life cycle of the source domain is subjected to the same standardization, median filtering and exponential moving average smoothing as the normal stage of the source domain bearing, to obtain the source domain health index value corresponding to the entire life cycle of the source domain. The source domain health index values ​​corresponding to the entire life cycle of bearings from different source domains are merged, and the 99th percentile of the merged source domain health index values ​​is determined as the failure threshold.

7. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The process of generating a target domain health index sequence based on the reconstruction error between each target domain vibration signal segment and its reconstructed signal segment specifically includes: According to the acquisition sequence of the target domain vibration signal segments during the entire life cycle of the bearing to be predicted, each target domain vibration signal segment is sequentially input into the trained joint reconstruction model to obtain a reconstructed signal segment that corresponds one-to-one with each target domain vibration signal segment. Calculate the average absolute error between the sampled value in each target domain vibration signal segment and the corresponding sampled value in its reconstructed signal segment to obtain the reconstruction error corresponding to each target domain vibration signal segment; The reconstruction error corresponding to each target domain vibration signal segment is subjected to standardization, median filtering and exponential moving average smoothing in sequence to obtain the target domain health index value corresponding to each target domain vibration signal segment. The target domain health index values ​​are arranged according to the acquisition order of the vibration signal segments in the target domain to obtain the target domain health index sequence.

8. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The step of establishing a local health baseline based on the leading health indicator values ​​of the target domain health indicator sequence specifically includes: The first N% of the target domain health index values ​​are selected from the beginning of the target domain health index sequence to form a local health sample set corresponding to the bearing to be predicted, where N represents the preset proportion of the first data selection. The distribution characteristics of target domain health index values ​​in the local health sample set are statistically analyzed. A local health baseline corresponding to the bearing to be predicted is established based on the distribution characteristics. A local health baseline threshold is determined based on the local health baseline so that the local health baseline threshold adapts to the distribution of front health index values ​​of different bearings to be predicted.

9. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The step of determining the degradation initiation point based on the change of the target domain health indicator sequence relative to the local health baseline specifically includes: The target domain health indicator sequence is scanned in chronological order using a preset detection window, and for each preset detection window, it is determined whether the following conditions are met simultaneously: Within the preset detection window, there are at least 5 consecutive target domain health indicator values ​​that exceed the local health baseline threshold determined by the local health baseline; The local slope determined by the change of the target domain health index value over time within the preset detection window is greater than the preset minimum slope threshold. The target domain health index value at the end of the preset detection window is higher than the target domain health index value at the beginning of the preset detection window. The first preset detection window that simultaneously meets the above conditions in chronological order is determined as the degradation detection window, and the moment corresponding to the target domain health index value that first exceeds the local health baseline threshold within the degradation detection window is determined as the degradation start point.

10. The bearing remaining life prediction method based on cross-domain adaptive reconstruction according to claim 1, characterized in that, The process of fitting a degradation trend to the target domain health index values ​​after the degradation initiation point, extrapolating the fitted degradation trend to the failure threshold to obtain the predicted failure time, and determining the remaining life of the bearing to be predicted based on the predicted failure time specifically includes: From the degradation segment starting from the degradation initiation point, a predetermined proportion of target domain health index values ​​located at the beginning of the degradation segment are selected as fitting data, and the time corresponding to the last target domain health index value in the fitting data is determined as the current prediction time. Using the time corresponding to the fitted data as the independent variable and the target domain health index value in the fitted data as the dependent variable, a cubic polynomial fitting is performed to obtain a degradation trend model used to characterize the degradation trend of the bearing to be predicted. The degradation trend model is extrapolated along the direction of time increase, and the time when the degradation trend model first reaches the failure threshold is determined as the predicted failure time. The time interval between the predicted failure time and the current prediction time is determined as the remaining life of the bearing to be predicted.

Images