Adaptive stage division and residual life prediction method for multi-stage degraded bearing

Abstract

The invention provides a self-adaptive stage division and residual life prediction method for a multi-stage degraded bearing, and relates to the technical field of rotary machinery prediction and health management. Firstly, multi-domain degradation features are extracted from original vibration signals. And secondly, fusing a Chebyshev inequality and a Transform network to realize online health state automatic identification. A novel parallel residual life prediction model fusing an attention mechanism is provided, and the model automatically captures degradation characteristics of different dimensions through parallel construction of a time convolution network and a bidirectional gating circulation unit channel. And the prediction model of the corresponding stage is adaptively and dynamically matched according to the stage identification result so as to accurately predict the RUL of the stage. And finally, processing the correlation between the multi-stage identification and the residual life prediction by adopting smoothing operation so as to obtain a final residual life value. According to the method, the change point of the degradation stage can be automatically identified, the key feature extraction process is optimized, and the dependence on artificial experience is reduced, so that the precision and robustness of a prediction result are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery prediction and health management, and in particular to a method for adaptive stage division and remaining life prediction of a multi-stage degraded bearing. Background Art

[0002] With the rapid development of Industry 4.0, fault prediction has become the core of prognostics and health management (PHM) technology, and has received widespread attention and has been widely used in modern industrial equipment, such as aircraft turbofan engines, wind turbines, machine tools and other key equipment. As an important component of rotating machinery, the health status of rolling bearings directly affects the operating stability of the entire equipment. However, bearings are prone to failure in complex environments. Studies have shown that about 40-50% of motor failures are caused by bearing failures. Therefore, accurate prediction of the remaining useful life (RUL) of bearings has become a key means to achieve predictive maintenance, improve equipment operating reliability and reduce maintenance costs.

[0003] At present, RUL prediction methods can be roughly divided into model-based, data-driven and hybrid methods. With the rapid growth of industrial monitoring data, deep learning technology based on artificial intelligence has been widely used in the field of RUL prediction. Unlike traditional methods, deep learning technology can adaptively learn the complex nonlinear mapping relationship between monitoring data and RUL without relying on physical models or expert knowledge. Deep learning models such as convolutional neural network (CNN) and recurrent neural network (RNN) have been widely used in the RUL prediction task of rolling bearings. However, CNN has a weak ability to capture long-distance dependencies when processing time series data, while recurrent networks such as long short-term memory networks and gated recurrent units are more suitable for processing complex time series data due to their long-term memory function. Therefore, deep learning models based on RNN and its variants perform better than traditional methods in RUL prediction. Although these deep learning methods have achieved certain results, the multi-stage degradation characteristics of rolling bearings have brought new challenges to RUL prediction. As the equipment runs, bearings often go through multiple degradation stages, and most existing methods rely on a single fixed model and cannot fully capture the complex change trends over the entire life cycle. Therefore, the multi-stage prediction framework proposed in recent years has been widely used in the RUL prediction of rolling bearings, which can improve the accuracy of prediction by processing the degradation process in segments. Existing studies have shown that the accuracy of RUL prediction can be significantly improved by accurately identifying the change points of the degradation stage. In addition, the degradation state is closely related to RUL prediction. However, most existing methods deal with degradation assessment and RUL prediction problems separately, resulting in insufficient information sharing and reduced prediction efficiency. Summary of the invention

[0004] In view of the above technical problems, the present invention proposes a method for adaptive stage division and remaining life prediction of multi-stage degraded bearings. This method constructs temporal convolutional networks (TCN) and bidirectional gated recurrent unit (BiGRU) channels in parallel, and introduces a global attention mechanism (GAM) to enhance the prediction ability of the model, thereby realizing the comprehensive expression of long-term and short-term features. First, the degradation characteristics of the original vibration signal are extracted. The Chebyshev inequality and the Transformer network are integrated to realize the adaptive identification of the online health status. Then, based on TBiGAN (TCN-BiGRU-Global-Attention-Network), the staged RUL prediction is realized to ensure that the prediction of each degradation stage is more accurate. Finally, the online stage division result is combined with the staged prediction value to obtain the final RUL value of the rolling bearing.

[0005] Specifically, the present invention provides a method for adaptive stage division and remaining life prediction of a multi-stage degraded bearing, comprising:

[0006] Extracting degradation features of the original vibration signal of the rolling bearing, wherein the degradation features include time domain features, frequency domain features and time-frequency domain features;

[0007] The degradation features are divided into stages using an adaptive stage division method that integrates Chebyshev inequality and Transformer, generating corresponding classification labels and identifying healthy operation stages, slow degradation stages, and fast degradation stages;

[0008] The multi-domain degradation feature set is used as the input data of the model, and is input into a parallel prediction model based on a temporal convolutional network, a bidirectional gated recurrent unit, and a global attention module, and the parallel prediction model is trained in stages according to the stage division results;

[0009] The trained parallel prediction model is used to predict the remaining service life of the bearing in different stages, and the predicted values ​​of the remaining service life in different stages are smoothed to estimate the final remaining service life value.

[0010] As a further illustration of the present invention, the time domain features include root mean square and peak-to-peak values ​​in the time domain; and / or the frequency domain features include root mean square frequency and average frequency in the frequency domain.

[0011] As a further illustration of the present invention, the extraction process of the time-frequency domain features includes:

[0012] The signal is decomposed into several intrinsic mode functions IMF through empirical mode decomposition;

[0013] Perform Hilbert transform on each IMF to obtain the instantaneous frequency and instantaneous amplitude;

[0014] Calculating the Hilbert spectrum by using the instantaneous frequency and the instantaneous amplitude;

[0015] The marginal spectrum is obtained through the Hilbert spectrum, and finally the time-frequency domain characteristics of the signal are extracted.

[0016] As a further explanation of the present invention, the signal is decomposed into a number of intrinsic mode functions IMF by empirical mode decomposition, which is specifically expressed by the following formula:

[0017] (1)

[0018] In the formula, represents the original vibration signal; Indicates IMF components; is the residual.

[0019] As a further illustration of the present invention, when each IMF is subjected to a Hilbert transform, the analytical form of the IMF is expressed as It is defined as:

[0020] (2)

[0021] In the formula, yes The Hilbert transform of and instantaneous phase :

[0022] (3)

[0023] By instantaneous phase , the instantaneous frequency is obtained by the following formula :

[0024] (4).

[0025] As a further illustration of the present invention, the Hilbert spectrum is expressed as a three-dimensional function of time, instantaneous frequency and instantaneous amplitude:

[0026] (5).

[0027] As a further illustration of the present invention, the marginal spectrum is obtained by integrating the Hilbert spectrum along the time axis:

[0028] (6).

[0029] As a further explanation of the present invention, the adaptive stage division method integrating Chebyshev inequality and Transformer is used to divide the degradation features into stages, generate corresponding classification labels and identify the healthy operation stage, slow degradation stage and fast degradation stage, which specifically includes:

[0030] Standardizing the time domain features, frequency domain features and time-frequency domain features of the degradation features;

[0031] The Chebyshev method is used to detect the stage change points of the root mean square of the time domain feature, so that the full life cycle data of the bearing is divided into a healthy operation stage, a slow degradation stage and a fast degradation stage, and the classification label is obtained. ;

[0032] The degraded feature dataset with classification labels is input into the Transformer network for training, and finally the classification results and the probability distribution of samples falling into each stage are output through the fully connected layer and the Softmax layer respectively.

[0033] As a further illustration of the present invention, the temporal convolutional network is formed by stacking a plurality of residual blocks, each residual block consisting of a dilated causal convolution, a dropout layer, and a normalization layer, and the residual block transmits input information to output through an identity path;

[0034] The bidirectional gated recurrent unit consists of a forward gated recurrent unit layer and a backward gated recurrent unit layer, enabling the model to process input time series along two time directions.

[0035] As a further illustration of the present invention, the data processing process of the parallel prediction model based on the temporal convolutional network, the bidirectional gated recurrent unit and the global attention module includes:

[0036] Standardize the input multi-feature data and use the sliding time window technology for data preprocessing;

[0037] The preprocessed data are input into the temporal convolutional network and bidirectional gated recurrent unit modules in parallel. The temporal convolutional network conducts in-depth analysis of local temporal characteristics through the dilated convolution structure, so that the model can capture the complex changes in the sequence data in the short term; the bidirectional gated recurrent unit captures the global temporal dependency of the sequence through its bidirectional recurrent structure.

[0038] A global attention module is introduced into the temporal convolutional network and the bidirectional gated recurrent unit respectively, so as to assign different weights to different time steps, so that the model dynamically pays attention to the most important temporal features for the current prediction task;

[0039] The outputs of the temporal convolutional network and the bidirectional gated recurrent unit are integrated through a merging layer, and the integrated feature data are sent to a fully connected layer. The fully connected layer performs a final weighted summation on the feature data through weights and biases to obtain an accurate remaining useful life (RUL) prediction value.

[0040] As a further illustration of the present invention, the following formula is used to smooth the remaining useful life prediction values ​​at different stages to estimate the final remaining useful life (RUL) value:

[0041]

[0042]

[0043] In the formula, Indicates that the sample is classified into categories by the softmax layer of Transformer The probability of (slow degradation stage, fast degradation stage), Indicates that the categories are divided after removing the healthy operation stage The probability of Indicates The periodic remaining useful life value is predicted by a parallel prediction network based on a temporal convolutional network, a bidirectional gated recurrent unit, and a global attention module. Represents the predicted final remaining useful life value.

[0044] Compared with the prior art, the present invention has the following beneficial technical effects:

[0045] The present invention provides an efficient and accurate method for predicting the remaining useful life of rolling bearings to solve the problems of insufficient capture of complex degradation processes and low prediction accuracy in the prior art. By combining deep learning algorithms with a multi-stage prediction framework, the present invention can automatically identify the change points of the degradation stage, optimize the key feature extraction process, and reduce reliance on manual experience, thereby improving the accuracy and robustness of the prediction results and meeting the requirements for high reliability and low maintenance costs of equipment in actual industrial applications.

[0046] Other features and advantages of the technical solution will be described in the subsequent description, and partly become apparent from the description, or understood by implementing the technical solution. The purpose and other advantages of the technical solution can be achieved and obtained through the structures specifically pointed out in the written description and the drawings.

[0047] The technical solution of the present technical solution is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the technical solution and constitute a part of the specification. Together with the embodiments of the technical solution, they are used to explain the technical solution and do not constitute a limitation of the technical solution. In the accompanying drawings:

[0049] Figure 1 Comparison of the time domain signal and HHT marginal spectrum at the 400th, 2000th, and 2790th sampling points of PHM 2012 bearing 1-1.

[0050] Figure 2 It is a multi-stage degradation trend diagram based on RMS.

[0051] Figure 3 The TCN model structure, where (a) is the three dilated causal convolutional layers, (b) is the residual block of TCN; (c) is the structure of TCN.

[0052] Figure 4 It is the BiGRU network model structure.

[0053] Figure 5 It is the global attention GAM module structure.

[0054] Figure 6 The parallel prediction model structure proposed in the present invention is based on a temporal convolutional network, a bidirectional gated recurrent unit and a global attention module.

[0055] Figure 7 It is the staged prediction process of the prediction model.

[0056] Figure 8 This is the multi-stage division result of bearing 1-1.

[0057] Fig. 9 The comparison results of RUL prediction of bearings under stage division are shown in Figure 1. (a) is the RUL prediction result of bearing 1-1; (b) is the RUL prediction result of bearing 1-2.

[0058] Fig.10 Comparison of RMSE and MAE of different prediction methods, (a) is the RMSE comparison result, (b) is the MAE comparison result. DETAILED DESCRIPTION

[0059] The preferred embodiments of the present technical solution are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present technical solution, and are not used to limit the present technical solution.

[0060] The present invention provides a method for adaptive stage division and remaining life prediction of a multi-stage degraded bearing, comprising the following steps:

[0061] Step 1: Extract the degradation features of the original vibration signal of the rolling bearing. The degradation features include time domain features, frequency domain features and time-frequency domain features.

[0062] The vibration signal of rolling bearing contains rich degradation information, but there is a lot of redundancy and noise in the original data, which increases the complexity of degradation modeling. In order to solve this problem, the present invention compresses the high-dimensional original data into a low-dimensional space through feature extraction to reduce data complexity and improve the processing efficiency of subsequent models.

[0063] Specifically, the time domain features extracted by the present invention include the root mean square value and peak-to-peak value in the time domain; the frequency domain features include the root mean square frequency and average frequency in the frequency domain, so as to more comprehensively describe the degradation trend of the bearing.

[0064] Time domain features can only reflect the amplitude change of the vibration signal over time, and it is difficult to describe the frequency structure and energy distribution of the signal. In order to solve this problem, the present invention combines frequency domain and time-frequency domain features for extraction. Through the Hilbert-Huang Transform (HHT), the original vibration signal can be analyzed in time and frequency, overcoming the shortcomings of the traditional Fourier transform that cannot characterize the characteristics of transient signals. Specifically, the signal is first decomposed into several intrinsic mode functions (IMF) through empirical mode decomposition (EMD), and then each IMF is Hilbert transformed to obtain the instantaneous frequency and instantaneous amplitude, and then the time-frequency domain features of the signal are extracted. The combination of these features enables the model to more accurately capture the degradation information of rolling bearings.

[0065] Specifically, the extraction process of time-frequency domain features includes the following steps:

[0066] Step 101: Decompose the signal into several intrinsic mode functions (IMFs) by empirical mode decomposition. Specifically, the following formula is used to express it:

[0067] (1)

[0068] In the formula, represents the original vibration signal; Indicates IMF components; is the residual.

[0069] Step 102: Perform Hilbert transform on each IMF to obtain the instantaneous frequency and instantaneous amplitude. The analytical form of IMF is It is defined as:

[0070] (2)

[0071] In the formula, yes The Hilbert transform of and instantaneous phase :

[0072] (3)

[0073] By instantaneous phase , the instantaneous frequency is obtained by the following formula :

[0074] (4).

[0075] Step 103: Obtain the Hilbert spectrum by calculating the instantaneous frequency and instantaneous amplitude. Specifically, the Hilbert spectrum is represented as a three-dimensional function of time, instantaneous frequency and instantaneous amplitude:

[0076] (5).

[0077] Step 104: Obtain the marginal spectrum through the Hilbert spectrum, and finally extract the time-frequency domain features of the signal. Specifically, the marginal spectrum is obtained by integrating the Hilbert spectrum along the time axis:

[0078] (6).

[0079] Figure 1 The comparison results of the time domain signal and HHT marginal spectrum of the original vibration signal sample points of bearing 1-1 in the PHM 2012 data set at different stages are shown. It can be seen that by comparing the spectrum changes of sample points at different stages, the late fault signal has obvious changes in the high-frequency part. It shows that the detailed change trend of the signal can be better shown after the time-frequency transformation of the signal, which is conducive to the extraction of degradation features. At the same time, the differences in different states also prove the necessity of staged RUL prediction for bearings.

[0080] Step 2: The degradation features are divided into stages using the adaptive stage division method that integrates Chebyshev inequality and Transformer, generating corresponding classification labels and identifying the healthy operation stage, slow degradation stage, and fast degradation stage.

[0081] As the device continues to operate, the rolling bearing exhibits a multi-stage degradation characteristic from the start of operation to failure, and each stage represents a different health state. According to the degradation speed characteristics, the state changes of rolling bearings during operation are generally divided into three stages, namely the healthy operation stage, the slow degradation stage, and the rapid degradation stage. Figure 2 is a stage division based on RMS. However, existing studies often rely on empirical thresholds and expert knowledge to manually analyze the health status of bearings. In order to solve the problems of over-reliance on manual work and poor real-time performance in the analysis process, the present invention introduces an adaptive stage division method that integrates Chebyshev inequality and Transformer. Compared with traditional CNN, Transformer can adaptively adjust the attention weight according to the input data, so as to respond to different input signals more flexibly. This method can monitor signal changes in real time and automatically identify degradation stages, thus laying the foundation for subsequent multi-stage RUL prediction.

[0082] Specifically, the above step 2 includes the following process:

[0083] Step 201: Standardize the time domain features, frequency domain features and time-frequency domain features of the degradation features.

[0084] Step 202: Use the Chebyshev method to detect the stage change point of the root mean square (RMS) of the time domain feature, so as to divide the full life cycle data of the bearing into three stages: healthy operation stage, slow degradation stage and rapid degradation stage, and obtain classification labels .

[0085] Step 203: Input the degraded feature data set with classification labels into the Transformer network for training, and finally output the classification results and the probability distribution of the samples falling into each stage through the fully connected layer and the Softmax layer respectively.

[0086] Step 3: The multi-domain degradation feature set is used as the input data of the model, and is input into a parallel prediction model based on a temporal convolutional network, a bidirectional gated recurrent unit, and a global attention module, and the parallel prediction model is trained in stages according to the stage division results.

[0087] 1) Temporal Convolutional Network (TCN)

[0088] In recent years, TCN has been widely used in complex tasks such as time series processing, machine vision, and fault diagnosis. The core idea of ​​TCN is to increase the range of the receptive field by dilating the convolution without increasing the number of parameters, so that it can effectively capture the long-term correlation in the time series. Its structure is as follows Figure 3(a). TCN is formed by stacking multiple residual blocks, each of which consists of a dilated causal convolution, a dropout layer, and a normalization layer, as shown in Figure 3 (b) shows that the residual block transmits the input information to the output through the same path, which protects the integrity of the information to a certain extent. The TCN model captures the deep features of the input information by stacking multiple residual blocks to form a deep neural network. The structure is as follows Figure 3 (c) as shown.

[0089] For the input sequence and filter , for each element in the sequence The operation of dilated convolution can be defined as follows:

[0090] (7)

[0091] In the formula, represents the expansion factor, represents the filter size, Indicates the historical direction.

[0092] 2) Bidirectional Gated Recurrent Unit BiGRU

[0093] As an improved LSTM, GRU has fewer parameters and can effectively shorten the running time. At the same time, the GRU network inherits the ability of LSTM to overcome the gradient vanishing and gradient exploding problems. However, the traditional unidirectional network structure can only learn the forward degradation law, and the model expression ability is insufficient. Therefore, the present invention adopts an improved bidirectional structure GRU so that the model has better fitting ability and prediction robustness. The BiGRU model consists of a forward GRU layer and a backward GRU layer, so that the model can process the input time series along two time directions. This unique network structure enables each hidden layer to capture past (forward) and future (backward) information at a specific time step, so that more comprehensive bearing degradation features can be extracted to improve the prediction performance of the network. The network model structure of BiGRU is as follows Figure 4 shown.

[0094] 3) Global Attention Module

[0095] The attention mechanism in neural networks aims to enhance the network's focus on key information by assigning different weights to different input information, which helps to significantly improve performance. In recent years, it has been widely used in fields such as text translation and speech recognition. There are many types of attention mechanisms, including global attention, local attention, and self-attention.

[0096] GAM is used in the proposed RUL prediction model and integrated into both TCN and BiGRU network structures. The introduction of GAM enables the model to not only focus on each time step in the sequence, but also optimize the weight allocation of each step based on global information. In the TCN module, GAM enables the network to more accurately identify and emphasize features that have a significant impact on the prediction target while capturing long-term dependencies. In the BiGRU module, due to its bidirectional structure, GAM not only considers the connection between the current moment and the past period, but also considers the information of the future moment, making the hidden state of the current time point The hidden state at all other time points When calculating the attention score, the information of the entire time series can be comprehensively considered. The structure is as follows Figure 5 shown.

[0097] The calculation process is as follows:

[0098] (8)

[0099] in, Represents the calculated attention score, using the softmax function for all Normalization is performed to ensure that the sum of the results is 1, thereby obtaining the weight of the current hidden layer state in all source vectors. The calculation process is:

[0100] (9)

[0101] Multiply the obtained attention weight with the hidden layer vector to get the context vector , the context vector is the comprehensive representation of all elements in the input sequence after attention weighting, and then and Concatenate and use the tanh function to convert it into an output vector.

[0102] (10)

[0103] (11)

[0104] Through this GAM, the model can dynamically weight each part of the input sequence according to the correlation between the current hidden state and the entire input sequence, so as to better reflect the global characteristics and key information points of the sequence and improve the model's ability to process sequence data.

[0105] 4) TBiGAN prediction method

[0106] In order to accurately predict the RUL of bearings, this paper proposes an innovative parallel prediction model that effectively combines TCN, BiGRU, and GAM. The core reason for choosing the TCN and BiGRU parallel structure is that they complement each other in processing time series data, which enhances the model's ability to understand complex time series data. Through the parallel processing framework, the model can efficiently capture the local and global temporal dependencies in bearing data. The network structure of TBiGAN (a parallel prediction model based on temporal convolutional network, bidirectional gated recurrent unit, and global attention module) is as follows: Figure 6 The data processing process of the prediction model includes the following steps:

[0107] Step 301: Standardize the input multi-feature data and use the sliding time window technology to preprocess the data so that the time series data can retain local dependencies.

[0108] Step 302: The preprocessed data is input in parallel to the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU). TCN uses an expanded convolution structure to perform in-depth analysis of local temporal characteristics, allowing the model to capture complex changes in the short and medium term in sequence data. BiGRU captures the global temporal dependencies of the sequence through its bidirectional recurrent structure, which strengthens the model's recognition of long-term trends and periodic patterns, and plays an indispensable role in predicting the future state of the device. Through the combination of these two network structures, the model can not only gain insight into the details in the sequence, but also grasp the overall trend, thereby improving the accuracy of the prediction.

[0109] Step 303: A global attention GAM module is introduced into the temporal convolutional network TCN and the bidirectional gated recurrent unit BiGRU respectively. The GAM module allows the model to dynamically focus on the most important temporal features for the current prediction task. This mechanism enhances the model's ability to focus on and capture key information in time series data by assigning different weights to different time steps.

[0110] Step 304: The outputs of the temporal convolutional network TCN and the bidirectional gated recurrent unit BiGRU are integrated through the concatenate layer (CL), and the integrated feature data with comprehensive and rich temporal information is sent to the fully connected layer (FCL). The fully connected layer performs the final weighted summation on the feature data through weights and biases to obtain an accurate RUL prediction value.

[0111] The parallel use of TCN and BiGRU fully combines the advantages of both in time series analysis. It is suitable for tasks such as rolling bearing RUL prediction that require long-term dependence and sensitivity to the dynamic characteristics of the sequence. It improves the model's ability to understand time series data and enhances the accuracy and reliability of the prediction. In addition, in the healthy stage of bearing operation, there is minimal wear but it does not affect the normal operation and life loss of mechanical equipment. Therefore, it is not meaningful to predict the RUL of the bearing in the healthy operation stage, so the RUL is predicted from the initial degradation point. The present invention divides the bearing degradation stage into a slow degradation stage and a rapid degradation stage. Each stage contains different information, and different stages correspond to their own prediction models. The above prediction model is trained in stages, such as Figure 7 As shown, the remaining service life RUL of the bearing can be predicted more accurately.

[0112] Step 4: Use the trained parallel prediction model to predict the remaining service life of the bearing in different stages, and perform smoothing operations on the remaining service life prediction values ​​at different stages to estimate the final remaining service life value.

[0113] Combined with the degradation stage recognition results of the adaptive classification network, formula (13) is used to smooth the RUL prediction values ​​at different stages to estimate the final RUL value.

[0114] (12)

[0115] (13)

[0116] In the formula, Indicates that the sample is classified into categories by the softmax layer of Transformer The probability of (slow degradation stage, fast degradation stage), Indicates that the categories are divided after removing the healthy operation stage probability. Indicates The staged RUL value predicted by the TBiGAN network, Represents the predicted final RUL value.

[0117] Example Application

[0118] This paper uses the bearing degradation data published in the IEEE PHM 2012 prediction challenge to evaluate the effectiveness of the proposed method for RUL prediction. The data set is provided by the French FEMTO-ST Institute, and the vibration signal is collected using the PRONOSTIA platform. Two accelerometers are installed on the top and side of the outer ring of the rolling bearing to measure the vibration signals of the bearing in the horizontal and vertical directions. Since the vertical vibration signal is less sensitive to bearing degradation, this study only uses the vibration signal in the horizontal direction for RUL prediction.

[0119] The sampling frequency in the experiment is 25.6kHz, the sampling time is 0.1s, and the data is recorded every 10s, that is, each recorded data contains 2560 sampling points. When the amplitude exceeds 20g, the bearing fails and the experiment is stopped. The platform is designed to provide real information to describe the natural degradation of bearings throughout their service life. The data set includes three working condition data. This example selects 5 bearing data sets that have been fully run to failure under the working condition for RUL prediction. The working condition load is 5000N and the speed is 1500r / min.

[0120] Step 1: Degraded feature extraction

[0121] During the degradation process of rolling bearings, vibration signals are an important basis for characterizing the health status of bearings. In order to capture the degradation trend of bearings more comprehensively, the present invention extracts rich degradation features from multiple perspectives in the time domain, frequency domain, and time-frequency domain. Time domain features include statistical characteristics of bearing vibration signals, such as root mean square and peak-to-peak values. In this example, the most important time domain feature is the root mean square (RMS) value, which reflects the magnitude of signal energy and can better track the energy changes during bearing degradation. By calculating the RMS value in each time period, we can obtain a trend curve reflecting the bearing degradation process. Frequency domain feature extraction focuses on the energy distribution of vibration signals in the spectrum. In this example, the frequency domain root mean square frequency (RMS Frequency) and mean frequency (Mean Frequency) are selected to characterize the changes in the frequency components of the signal. As the bearing degradation intensifies, the high-frequency components in the signal will gradually increase, making the frequency feature an important indicator for evaluating the degradation state. Time-frequency domain analysis can reveal the joint changes of vibration signals in time and frequency. This example uses the Hilbert-Huang Transform (HHT) to evaluate the time-frequency characteristics of the bearing. HHT combines empirical mode decomposition (EMD) and Hilbert spectrum analysis to adaptively decompose the intrinsic mode functions in the vibration signal and obtain the instantaneous frequency and energy distribution of the signal. By extracting these time-frequency features, we can more accurately describe the dynamic changes in the signal characteristics during the degradation process.

[0122] Step 2: Multi-stage adaptive recognition

[0123] Since the degradation process of bearings usually includes multiple stages and the degradation characteristics of different stages are significantly different, directly modeling the entire life cycle will lead to large prediction errors. Therefore, this example proposes a multi-stage adaptive identification method based on the improved Chebyshev inequality.

[0124] The core of multi-stage division is to use RMS value for stage division. Since the original vibration signal usually contains a lot of noise, which will affect the stability of the RMS value, we use Savitzky-Golay filter to smooth the RMS curve. The filter window length is 11, and the polynomial order is 3. After filtering, the RMS value curve is smoother and can clearly reflect the different degradation stages of the bearing. For example, during the operation of bearing 1-1, we analyze its RMS value curve and use Chebyshev inequality to divide the degradation process into three stages (such as Figure 8 As shown in the figure). From the beginning of the experiment to 1426, the RMS value is close to 0, and the curve is relatively flat with little fluctuation, indicating that the bearing is in normal working condition; from 1426 to 2748, the RMS value gradually increases, indicating that the bearing defect expands and the surface wear intensifies; and after 2748, the RMS value rises sharply, indicating that the bearing has entered a serious fault stage. This division method can accurately identify different degradation stages of the bearing, and the division results are highly consistent with the bearing state changes, providing an effective stage classification label for subsequent adaptive stage identification.

[0125] Through the above stage division method, the classification labels of bearings in different time periods are obtained. We selected five bearing data sets that were fully operated until failure, and took one of the bearings as the target test object and the remaining four bearings as the training set to train the Transformer classifier. In order to further improve the robustness of the model, we took the average of 5 experiments and finally achieved a high recognition accuracy, with an average accuracy of more than 93% (see Table 1).

[0126] Table 1 Recognition evaluation results in the adaptive stage

[0127] Step 3: RUL prediction based on TBiGAN

[0128] The structure of the TBiGAN model is designed as a parallel combination of TCN and BiGRU, which can capture long-term and short-term dependencies in time series data at the same time. The channel configuration of the TCN module is [32, 64], and a 3×3 convolution kernel is used to capture the local features of the time series. The BiGRU module consists of two layers, with 32 and 64 hidden neurons, respectively. At the same time, the dimension of the GAM is the same as the last hidden layer of TCN and BiGRU to enhance the model's ability to focus on key time series features. The batch size is set to 64. The dropout regularization is set to 0.5 to prevent overfitting. The learning rate is set to 0.0003 to ensure stable and effective model training. In the training process of the proposed TBiGAN, it is worth mentioning that the above hyperparameters are mainly determined by the grid search algorithm and four-fold cross validation.

[0129] In order to verify the effectiveness of the TBiGAN model, we conducted experiments on the IEEE PHM 2012 bearing degradation dataset. We selected data from five bearings with the following working conditions, used the data from four bearings to train the model, and used the remaining one bearing for testing. In order to verify the effectiveness of the degradation stage division method, we took bearings 1-1 and 1-2 as examples to compare the proposed method with a single prediction model method that fits the entire degradation process without performing degradation stage division. The results are shown in Figure 2. Fig. 9 As shown. Compared with a single prediction model without stage division, the RUL predicted by the proposed stage division method is closer to the true value, the RUL prediction curve can track the actual value well, and almost all the predicted values ​​are within the error range of 30%, reflecting the stability of the algorithm. Especially in the late stage of degradation, TBiGAN can provide more accurate prediction results, which helps to provide a reliable basis for maintenance decisions. It is proved that the use of a staged prediction strategy can better capture the degradation characteristics of different stages and greatly improve the accuracy of RUL prediction. Table 2 gives the evaluation results of 5 bearings. Compared with the traditional prediction results that do not consider the degradation stage division, the method proposed in the present invention shows the lowest RMSE, MAE and the highest R2 and SCORE, which further verifies the superiority of the proposed method.

[0130] Table 2 Evaluation indicators of prediction results of different stage division methods

[0131] Table 3 Evaluation metrics of prediction results for different stagedivision methods

[0132] Choosing different network structures has an important impact on the RUL prediction performance. The model proposed in the present invention can give full play to the unique advantages of both in processing long-term and short-term dependencies of complex, multi-dimensional time series by introducing TCN and BiGRU in parallel. In order to verify the superiority of the parallel network proposed in this paper, the positive impact of each module of TBiGAN on the performance of the model is studied. The proposed model was subjected to ablation analysis and compared with the BiGRU-AN with the TCN module deleted and GAM fused, the TCN-AN with the BiGRU module deleted and GAM fused, the TCN-BiGRU with the GAM module deleted, the TCN with the BiGRU and GAM modules deleted, and the BiGRU with the TCN and GAM modules deleted.

[0133] Taking bearing 1-1 and bearing 1-3 as examples, all other prediction processes remain unchanged. Compare the prediction results of the proposed model with those of the above models. Fig.10 The prediction results of different network models for two bearings running to failure are shown. The results show that the prediction model using TBiGAN model is more accurate than the prediction model using either of the two combined techniques in the proposed method alone, showing the smallest difference from the actual RUL in both bearings. At the same time, the network model with attention mechanism shows better results than the model without attention mechanism.

[0134] To further evaluate the advantages of TBiGAN, we compared it with other common RUL prediction models, including Transformer, BiLSTM, TCN-BiLSTM, and CNN models. The comparison results show that TBiGAN has better prediction performance in terms of evaluation indicators such as RMSE, MAE, and R2 (see Table 3). This fully demonstrates the excellent performance of the TBiGAN model in processing bearing degradation data and performing RUL prediction.

[0135] Table 3 Comparison results between the proposed method and other prediction methods

[0136] Obviously, those skilled in the art can make various changes and modifications to the technical solution without departing from the spirit and scope of the technical solution. Thus, if these modifications and variations of the technical solution fall within the scope of the claims of the technical solution and their equivalents, the technical solution is also intended to include these modifications and variations.

Claims

1. A method for adaptive stage division and remaining life prediction of multi-stage degraded bearings, characterized in that: include: Extracting degradation features of the original vibration signal of the rolling bearing, wherein the degradation features include time domain features, frequency domain features and time-frequency domain features; The degradation features are divided into stages using an adaptive stage division method that integrates Chebyshev inequality and Transformer, generating corresponding classification labels and identifying healthy operation stages, slow degradation stages, and fast degradation stages; The multi-domain degradation feature set is used as the input data of the model, and is input into a parallel prediction model based on a temporal convolutional network, a bidirectional gated recurrent unit, and a global attention module, and the parallel prediction model is trained in stages according to the stage division results; The trained parallel prediction model is used to predict the remaining service life of the bearing in different stages, and the remaining service life prediction values ​​at different stages are smoothed to estimate the final remaining service life value.

2. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The time domain features include root mean square and peak-to-peak values ​​in the time domain; and / or the frequency domain features include root mean square frequency and average frequency in the frequency domain.

3. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The extraction process of the time-frequency domain features includes: The signal is decomposed into several intrinsic mode functions IMF through empirical mode decomposition; Perform Hilbert transform on each IMF to obtain the instantaneous frequency and instantaneous amplitude; Calculating the Hilbert spectrum by using the instantaneous frequency and the instantaneous amplitude; The marginal spectrum is obtained through the Hilbert spectrum, and finally the time-frequency domain characteristics of the signal are extracted.

4. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 3, characterized in that: The signal is decomposed into several intrinsic mode functions IMF by empirical mode decomposition, which are specifically expressed by the following formula: （1）； In the formula, represents the original vibration signal; Indicates IMF components; is the residual.

5. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 3, characterized in that: When the Hilbert transform is performed on each IMF, the analytical form of the IMF is expressed as It is defined as: （2）； In the formula, yes The Hilbert transform of and instantaneous phase : (3) By instantaneous phase , the instantaneous frequency is obtained by the following formula : (4)。 6. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 3, characterized in that: The Hilbert spectrum is expressed as a three-dimensional function of time, instantaneous frequency and instantaneous amplitude: (5); The marginal spectrum is obtained by integrating the Hilbert spectrum along the time axis: (6)。 7. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The adaptive stage division method integrating Chebyshev inequality and Transformer is used to divide the degradation features into stages, generate corresponding classification labels and identify the healthy operation stage, slow degradation stage and fast degradation stage, which specifically include: Standardizing the time domain features, frequency domain features and time-frequency domain features of the degradation features; The Chebyshev method is used to detect the stage change points of the root mean square of the time domain feature, so that the full life cycle data of the bearing is divided into a healthy operation stage, a slow degradation stage and a fast degradation stage, and the classification label is obtained. ; The degraded feature dataset with classification labels is input into the Transformer network for training, and finally the classification results and the probability distribution of samples falling into each stage are output through the fully connected layer and the Softmax layer respectively.

8. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The temporal convolutional network is formed by stacking multiple residual blocks, each residual block consists of a dilated causal convolution, a dropout layer, and a normalization layer, and the residual block transmits input information to output through an identity path; The bidirectional gated recurrent unit consists of a forward gated recurrent unit layer and a backward gated recurrent unit layer, enabling the model to process input time series along two time directions.

9. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The data processing process of the parallel prediction model based on the temporal convolutional network, the bidirectional gated recurrent unit and the global attention module includes: Standardize the input multi-feature data and use the sliding time window technology for data preprocessing; The preprocessed data are input into the temporal convolutional network and bidirectional gated recurrent unit modules in parallel. The temporal convolutional network conducts in-depth analysis of local temporal characteristics through the dilated convolution structure, so that the model can capture the complex changes in the sequence data in the short term; the bidirectional gated recurrent unit captures the global temporal dependency of the sequence through its bidirectional recurrent structure. A global attention module is introduced into the temporal convolutional network and the bidirectional gated recurrent unit respectively, so as to assign different weights to different time steps, so that the model dynamically pays attention to the most important temporal features for the current prediction task; The outputs of the temporal convolutional network and the bidirectional gated recurrent unit are integrated through a merging layer, and the integrated feature data are sent to a fully connected layer. The fully connected layer performs a final weighted summation on the feature data through weights and biases to obtain an accurate remaining useful life prediction value.

10. The method for adaptive stage division and remaining life prediction of multi-stage degraded bearings according to claim 1, characterized in that: The following formula is used to smooth the remaining useful life forecast values ​​at different stages and estimate the final remaining useful life value: ; ; In the formula, Indicates that the sample is classified into categories by the softmax layer of Transformer The probability of the category They represent the slow degradation stage and the fast degradation stage respectively. Indicates that the categories are divided after removing the healthy operation stage The probability of Indicates The periodic remaining useful life value is predicted by a parallel prediction network based on a temporal convolutional network, a bidirectional gated recurrent unit, and a global attention module. Represents the predicted final remaining useful life value.

Images