Bearing segmented life prediction method and system based on graph neural network

Abstract

The invention discloses a bearing segmented life prediction method and system based on a graph neural network, and relates to the technical field of rotating machinery health monitoring. The method comprises the following steps: acquiring a vibration signal of a to-be-detected bearing; based on the health index of the bearing, dividing the whole life process of the bearing into two stages of a stable stage and a degeneration stage by using a life division model; the service life prediction model is used for conducting service life prediction on the vibration signals in the stable period and the degradation period respectively, prediction results are combined, a final bearing service life prediction result is obtained, the service life prediction model constructs the vibration signals into multi-mode time scale features, and the service life prediction model is used for predicting the service life of the bearing. And carrying out mining analysis on time dependence in the multi-modal time scale features so as to realize life prediction. According to the method, a two-stage division strategy fused with deep learning is combined with an efficient mining analysis method of time scale information, and efficient prediction of the residual service life of the bearing is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rotating machinery health monitoring, and particularly relates to a bearing segmented life prediction method and system based on a graph neural network. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] As a core component of rotating machinery, the performance and state of bearings play a crucial role in the overall stability of the equipment. However, in actual working conditions, bearings often face complex environments such as high load, strong vibration, and high temperature, which makes bearings prone to wear and damage, thereby inducing faults and even causing equipment downtime, seriously affecting the stability and service life of the equipment. In order to overcome the problem of improper maintenance under the traditional periodic maintenance mode, modern industry generally uses sensors to collect bearing vibration signals in real time and combines signal analysis methods to realize the state monitoring and management of equipment. In recent years, with the rapid development of machine learning, especially deep learning technology, data-driven bearing health state monitoring and residual life prediction methods have gradually attracted widespread attention. Compared with traditional methods, deep learning can mine potential degradation patterns from complex and high-dimensional vibration data, thereby significantly improving the accuracy and robustness of bearing residual service life prediction, providing a reliable reference for intelligent maintenance decision-making.

[0004] However, there are still some problems to be solved in the current research on bearing residual service life prediction based on deep learning. First, due to the depth and modeling ability of neural network structure, the existing models have obvious differences in expressing bearing degradation characteristics at different degradation stages. A single model is difficult to fully fit the degradation process of bearings throughout their life cycle, which makes it prone to insufficient accuracy in the prediction process. Second, traditional handcrafted features have obvious shortcomings in describing the non-linear characteristics of signals and the implicit degradation information. These handcrafted features are difficult to fully reflect the complex changes of equipment operating state, thereby limiting the accuracy of life prediction. In addition, traditional deep learning research methods also have certain limitations in analyzing the complexity of time series. They fail to effectively capture the correlation and dependency between long time series features, resulting in unsatisfactory prediction results and difficulty in popularization and application in engineering practice.

[0005] The existence of the above problems makes the bearing residual service life prediction technology based on deep learning still face many challenges in practical application. Therefore, how to realize efficient and accurate prediction of bearing residual service life has become a technical problem to be solved in the prior art. SUMMARY

[0006] In view of the deficiencies of the prior art, the purpose of the present application is to provide a bearing segmented life prediction method and system based on a graph neural network, which adopts a two-stage division strategy combined with an efficient mining analysis method of time scale information based on deep learning, thereby realizing efficient prediction of the remaining service life of the bearing.

[0007] In order to achieve the above-mentioned purpose, the present application is realized by the following technical scheme: The present application provides a bearing segmented life prediction method based on a graph neural network in the first aspect, comprising the following steps: Obtaining a vibration signal of a bearing to be detected, the vibration signal comprising a horizontal vibration signal and a vertical vibration signal; Dividing the whole life process of the bearing into two stages of a stable period and a degradation period based on a health index of the bearing using a life division model; Using a life prediction model to respectively predict the life of the vibration signal in the stable period and the degradation period, and combining the prediction results to obtain the final bearing life prediction result, wherein the life prediction model constructs the vibration signal into a multi-modal time scale feature, and mines and analyzes the time dependence in the multi-modal time scale feature, thereby realizing life prediction.

[0008] Further, the specific steps of dividing the whole life process of the bearing based on the health index of the bearing using the life division model are as follows: Constructing a health index of the bearing according to influencing factors of the working state of the bearing; Determining a division condition according to the characteristics of the stable period and the degradation period; Using the life division model to process the vibration signal according to the health index and the division condition.

[0009] Further, the specific steps of constructing a health index of the bearing according to influencing factors of the working state of the bearing are as follows: Pretreating the vibration signal of the bearing; Introducing a trend consistency index and an expression ability vector as basic weight factors for constructing the health index, and dynamically allocating weights by using a game theory strategy.

[0010] Further, the specific steps of pretreating the vibration signal of the bearing are as follows: Performing Hilbert transform on the vibration signal of the bearing, and then performing fast Fourier transform; Then, performing filter processing on the key frequency band.

[0011] Further, the specific steps of using the life prediction model to respectively predict the life of the vibration signal in the stable period and the degradation period are as follows: Extracting features from the vibration signal; Modeling and fusing the extracted features from multiple time scales to obtain a fusion graph; Deep relationships between nodes in the fusion graph are mined using a graph multi-layer perception neural network, and the nodes are updated; The updated nodes are modeled using an encoder, and a regression algorithm is used to predict the remaining life of the bearing.

[0012] Furthermore, feature extraction is performed on the vibration signal to obtain hand-designed features and deep learning features, wherein a multi-scale large kernel attention module is used to extract deep features to obtain deep learning features.

[0013] Furthermore, the specific steps of modeling the extracted features from multiple time scales are as follows: A cosine similarity is used to construct a long-term feature graph to capture the global consistency structure showing the degradation trend in the vibration signal; Dot product similarity is used to construct a short-term feature graph on a finer-grained time scale to strengthen the dynamic change perception of local impact features and microscopic damage signals of the bearing.

[0014] The second aspect of the present application provides a bearing segmented life prediction system based on a graph neural network, comprising: A data acquisition module configured to acquire a vibration signal of a bearing to be detected; A stage division stage configured to divide the full life process of the bearing into a stable period and a degradation period based on a health index of the bearing using a life division model; A life prediction module configured to use a life prediction model to predict the life of the vibration signal in the stable period and the degradation period respectively, and combine the prediction results to obtain the final bearing life prediction result, wherein the life prediction model constructs the vibration signal into multi-modal time scale features, and mines and analyzes the time dependence in the multi-modal time scale features, thereby realizing life prediction.

[0015] The third aspect of the present application provides a computer readable storage medium storing a computer program, the computer program being adapted to be loaded and executed by a processor to perform the steps of the bearing segmented life prediction method based on the graph neural network as described in the first aspect of the present application.

[0016] The fourth aspect of the present application provides a computer device, comprising: A processor adapted to execute a computer program; A computer readable storage medium storing a computer program, the computer program being executed by the processor to implement the bearing segmented life prediction method based on the graph neural network as described in the first aspect of the present application.

[0017] The one or more technical solutions have the following beneficial effects: The application discloses a bearing segmented life prediction method and system based on a graph neural network.

[0018] The application discloses a bearing segmented life prediction method and system based on a graph neural network.

[0019] The application discloses a bearing segmented life prediction method and system based on a graph neural network.

[0020] Advantages of the additional aspects of the application will be partly given in the following description, partly will become obvious from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0022] Figure 1 Flow chart of bearing segmented life prediction method based on graph neural network in embodiment one of the present application; Figure 2 Structure diagram of ShaRNN in embodiment one of the present application; Figure 3 Structure diagram of MLKA-T in embodiment one of the present application; Figure 4 Stage division TRW-ShaRNN result presentation diagram in embodiment one of the present application. DETAILED DESCRIPTION

[0023] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0024] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof; The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Embodiment one: The embodiment one of the present application provides a bearing segmented life prediction method based on graph neural network. A TRW-ShaRNN model is constructed to divide the bearing life into two stages. Then, according to the stage division result, the life of the two stages of the bearing is predicted based on the MTNT-Net model, and finally the two stage prediction results are combined.

[0026] As Figure 1As shown, specifically comprising the following steps: Step 1: Obtain the vibration signal of the bearing to be detected, the vibration signal including horizontal vibration signal and vertical vibration signal.

[0027] In a specific embodiment, in the prior art, if only a single direction vibration signal is used in the signal collection process, some fault features may be missed, resulting in a decrease in prediction accuracy. Therefore, the embodiment simultaneously collects data in the horizontal and vertical directions, observes the same state from more perspectives, and can extract more comprehensive features (such as energy features, frequency domain features, time-frequency features, etc.). The subsequent steps of the embodiment are to extract the same time domain, frequency domain, and deep features from the horizontal and vertical original vibration signals, respectively, to avoid missing fault features, extract more comprehensive features for life prediction, and improve prediction accuracy.

[0028] In the embodiment, the vibration signal of the bearing to be detected is collected by using the experimental platform built. Since the bearing will be subjected to loads in multiple directions in actual work, the data set is subjected to radial force, rotational speed, and torque on the horizontal and vertical axes of the bearing during collection, and two columns of original signals are collected by an acceleration sensor.

[0029] Step 2: Divide the full life process of the bearing into two stages of stable period and degradation period based on the health index of the bearing using a life division model.

[0030] Step 2.1: Construct a health index of the bearing according to the influencing factors of the working state of the bearing.

[0031] Step 2.1.1: Preprocess the vibration signal of the bearing.

[0032] Step 2.1.1.1: Perform Hilbert Transform (HT) on the vibration signal of the bearing to obtain an envelope signal, and then perform Fast Fourier Transform (FFT) to extract overall frequency energy features. Then, the envelope spectrum is filtered through the following five key frequency bands to extract the energy of different key frequency bands. Finally, each vibration signal obtains 5-dimensional frequency band energy features, and a total of 10-dimensional frequency band energy features are obtained.

[0033] Step 2.1.1.2: Then, filter the key frequency band.

[0034] The following key frequency bands are used for filtering in this embodiment: Shaft Frequency (SF), Ball Pass Frequency Outer race (BPFO), Ball Pass Frequency Inner race (BPFI), Ball Spin Frequency (BSF), and Fundamental Train Frequency (FTF), as shown in equations (1) - (5), respectively: (1), (2), (3), (4), (5).

[0035] where, is the number of rolling elements, is the shaft speed, is the diameter of the rolling elements, is the average diameter of the bearing, is the contact angle in degrees with respect to the radial plane. Each frequency defines a critical band, which is key in the evolution of bearing failure. The set of five critical bands is represented as: .

[0036] where each frequency defines a critical band , representing the 5 bands, respectively.

[0037] After filtering through these 5 key frequency bands, the final spectral features are obtained.

[0038] The 10-dimensional spectral features obtained are then introduced into the trend consistency index and expression capacity vector, and the game theory strategy is used for fusion, and finally the health index TR is obtained.

[0039] Step 2.1.2: Introduce the trend consistency index and expression capacity vector as the basic weight factor for the construction of the health index, and use the game theory strategy for dynamic weight distribution.

[0040] In one specific implementation, since the frequency domain energy features extracted based on HT and FFT are more sensitive to shock and fault signals, and different frequency bands represent the contributions of different mechanical structural components or different excitation modes to vibration, the energy features of a single frequency band can only reflect the degradation of the corresponding fault frequency band and cannot comprehensively characterize the overall health status of the bearing. Therefore, this embodiment uses these frequency domain features to fuse and construct health indicators for stage division, which can detect local minor damage early, reduce oversensitivity to single anomalies, and improve overall robustness.

[0041] To mitigate the inconsistencies and uncertainties arising from multiple features under noise interference, this embodiment introduces a Trend Consistency Coefficient (TCC) to measure the trend correlation and relative consistency among spectral features. This coefficient comprehensively considers the numerical distance and trend correlation between features, achieving adaptive suppression of redundant or anomalous features from the perspectives of feature correlation and evolutionary consistency. TCC automatically identifies features with stronger expressive power and trend stability, assigning them higher fusion weights, while allocating lower weights to features with poor consistency, thereby improving the robustness and discriminative power of the fused features.

[0042] TCC is calculated as shown in equation (6): (6).

[0043] Where M is the Manhattan distance and S is the Spearman correlation coefficient. The Manhattan distance measures the deviation between the eigenvalue and the life curve, reflecting the overall degree of deviation; the smaller the value, the closer it is to the ideal degradation trend. The Spearman correlation coefficient measures the monotonicity of the eigenvalue over time. A eigenvalue with a small Manhattan distance and a high Spearman correlation coefficient indicates a stronger consistency with the bearing degradation trend.

[0044] Create the consistency matrix TCCM as shown in equation (6): (7).

[0045] in This represents the trend consistency calculation result between the i-th feature and the j-th feature.

[0046] Simultaneously, the root mean square error (RMSE) is introduced to construct a feature expressive power vector to measure the performance quality of each feature in the degradation prediction task. The smaller the RMSE, the stronger the feature expressive power; conversely, the larger the RMSE, the weaker the feature expressive power. The expressive power vector is shown in Equation (8): (8).

[0047] where J = 10, represents the 10-dimensional band energy feature to be fused. The expression ability vector reflects the interpretation ability of the feature in the real degradation process, and is directly related to the fitting degree of the actual bearing degradation. The embodiment can filter out features with large fluctuations, poor noise resistance, and poor sensitivity to degradation by introducing RMSE.

[0048] To realize the coordination of feature fusion, the trend consistency coefficient (TCC) and the expression ability vector (RMSE) are used as basic weight factors for constructing the health index. The optimal collaborative weight configuration is determined by introducing a game theory strategy to dynamically allocate the weights of the two. The specific steps are as follows: (1) Construct the basic weight vector.

[0049] For each frequency energy feature, first calculate TCC and RMSE to obtain two basic vectors, as shown in equation (9): (9).

[0050] wherein and represent the TCC vector and the RMSE vector, respectively, and J = 10 represents the 10-dimensional energy feature.

[0051] (2) Construct the weighted linear combination.

[0052] The comprehensive weight vector after fusion is shown in equation (10): (10).

[0053] wherein , are linear combination coefficients.

[0054] (3) Optimize the weights by game theory.

[0055] The linear weighting method based on game theory is used to optimize the coefficients of the linear combination, and a target function that minimizes the deviation is constructed, with the constraint condition being to reduce the difference between the comprehensive vector after fusion and the original vector . The difference is shown in equation (11): (11).

[0056] (4) Solve the linear equation to obtain the optimal coefficients.

[0057] Solve the matrix equation to obtain , as shown in equation (12): (12).

[0058] ​Renormalization to get the final comprehensive coefficient , and the final comprehensive weight vector. As shown in equation (13): (13).

[0059] (5) Generating a weighted fusion energy feature.

[0060] For each time point , the original band energy feature is weighted and summed according to to obtain the fusion feature , as shown in equation (14): (14).

[0061] Subsequently, the weighted fusion feature is obtained, and the Savitzky-Golay smoothing algorithm is introduced to denoise the initial fusion result, and finally a health indicator TR with good robustness and discriminability is constructed. This embodiment evaluates the monotonicity of the TCC evaluation feature over time and the consistency degree of the degradation trend, and quantifies the error of the feature fitting degradation curve through RMSE. Both of them are used to get the final comprehensive weight through weight fusion, which is used to weight multiple band energy features to generate a comprehensive health indicator TR. The fused TR continues to rise over time, which is consistent with the actual physical degradation process, can accurately and monotonically reflect the bearing degradation state, can clearly distinguish different life stages, can be used for health stage division, and finally realizes the comprehensive characterization of the bearing health state.

[0062] Step 2.2: Determine the division condition according to the characteristics of the stable period and the degradation period.

[0063] This embodiment adopts a sliding window method to realize health stage division by dynamically calculating the WD distance between the current window probability density function (PDF) and the previous window PDF of the TR indicator. By comparing the statistical distribution changes of adjacent windows, the slight shift of signal characteristics is detected in real time. When the bearing is in the stable period, the difference between the PDF of adjacent windows is small, and the WD distance remains at a low level, indicating that the feature distribution is stable; when entering the degradation period, due to defect expansion or wear intensification, the PDF between adjacent windows deviates significantly, resulting in a sharp increase in WD distance, reflecting the state mutation. This significant deviation is a typical feature of the degradation period, indicating that the performance of the bearing has undergone irreversible degradation.

[0064] To this end, the embodiment determines a threshold based on an empirical model, when the WD distance exceeds the threshold and the trend of change is irreversible, it is determined that the degradation phase is entered. By comparing with adjacent windows, rather than a static reference PDF, the dynamic evolution process of the bearing from health to degradation can be more sensitively captured, and the real-time performance and robustness are improved.

[0065] The algorithm divides the observed time series into W windows, and calculates the Wasserstein Distance (WD) relative to the previous window. Then, the algorithm tracks the changes in the difference between these values until they exceed a certain threshold. When the difference exceeds this threshold, the algorithm will detect an event.

[0066] Regarding the threshold that controls the divergence of WD Inspired by the exponential deterioration model, the threshold function is defined as an empirical model. As shown in equation (15): (15).

[0067] where, is the window index, and L is the end of life, is the estimated standard deviation of the measured difference in WD distance between windows. and are two parameters related to the natural degradation of the bearing and have a physical interpretation. The parameter reflects the sensitivity of the method to changes at the beginning of the record, which is related to the estimated standard deviation . On the other hand, ~1 controls the sensitivity of the algorithm as it approaches the end of the bearing's life, which varies depending on the specific conditions of the experiment.

[0068] Step 2.3: Process the vibration signal according to the health indicators and partition conditions using the life partition model.

[0069] In this embodiment, the life partition model is a ShaRNN network, the structure of which is shown in Figure 2 The two stages generated by the TRW method are used as labels, and the original horizontal and vertical vibration signals of the bearing are used as inputs, which are input into a shallow recurrent neural network (Shallow Recurrent Neural Network. Shallow RNN, ShaRNN). Figure 4For the TRW-ShaRNN stage division result on bearing B1_1, the solid line in the figure is the TR health indicator, and the dashed line is the first point of degradation (FPT). The period before the FPT is the stable period, and the period after the FPT is the degradation period. It can be seen that before the FPT, the TP curve is relatively stable, and after the FPT, the TR curve rises rapidly, indicating that the bearing has entered the degradation period. Therefore, this method can effectively identify the key stages of the bearing.

[0070] The specific steps of the degradation stage division method based on Trend-RMSE-WD indicators (TRW) are as follows: 1. First, use the TRW method to divide the vibration signal of the bearing into stages during the entire life process. Specifically, the WD distance between adjacent windows is calculated using a sliding window, and when the distance exceeds the threshold and shows a significant deviation from the trend, it is determined that the degradation stage has been entered; otherwise, it is determined to be a stable stage. Through this method, the entire degradation process can be preliminarily divided into two stages: stable stage and degradation stage, and each time window is assigned a corresponding stage label (0 for stable and 1 for degradation).

[0071] 2. The stage division result obtained is used as a supervision learning label to further train the Shallow RNN network. The original vibration signal of the bearing (horizontal and vertical directions) is used as the network input, and the two-stage label (0 / 1) generated by the TRW method is used to train the Shallow RNN model to learn the time sequence feature pattern of the vibration signal in different stages, thereby achieving automatic recognition and prediction of the degradation state of the bearing.

[0072] The ShaRNN network includes two layers. The first layer models each Brick (short time period) separately to extract local time dependence, and all time windows are developed in parallel to avoid the gradient vanishing problem in long sequences. The second layer integrates the high-order features of each Brick extracted by the first layer as time steps to capture long-term dependencies. This model can automatically learn the mapping between feature evolution and label changes without manual design of boundaries, improving the accuracy and robustness of stage division. Finally, the life division model outputs the stage classification result (0 / 1) of each sliding window, accurately identifying the health state (stable period or degradation period) of the vibration signal in the corresponding time period.

[0073] Step 3: Use the life prediction model to predict the life of the vibration signal in the stable period and the degradation period respectively, and combine the prediction results to obtain the final bearing life prediction result.

[0074] The life prediction model constructs the vibration signal as a multi-modal time scale feature, and analyzes the time dependence in the multi-modal time scale feature, so as to realize life prediction.

[0075] Step 3.1: Life prediction is performed on the vibration signals in the stable period and the degradation period respectively by using the life prediction model.

[0076] Step 3.1.1: Feature extraction is performed on the vibration signal.

[0077] In this embodiment, the vibration signal is subjected to feature extraction to obtain hand-designed features and deep learning features. The multi-scale large kernel attention module is used to extract deep features to obtain deep learning features.

[0078] Specifically, 46-dimensional time domain and frequency domain features are extracted.

[0079] The time domain features are 24-dimensional, as shown in Table 1. The frequency domain features are 11-dimensional: the original vibration signal is subjected to fast Fourier transform (FFT) to extract overall frequency energy features; the Hilbert envelope signal is subjected to fast Fourier transform (HT_FFT) to extract impact-sensitive modulation frequency features. Both of them are calculated on 5 key frequency bands to obtain average frequency band energy features (10-dimensional), which comprehensively represent different frequency domain modes. In addition, the overall energy mean of the original frequency spectrum is calculated as a noise energy feature (1-dimensional), which assists in distinguishing background noise and fault signals.

[0080] Table 1 Time domain statistical features

[0081] Then, in order to make up for the shortcomings of traditional artificial features in depicting signal nonlinearity and implicit degradation information, a multi-scale large kernel attention (MLKA-T) module suitable for time series is used to mine deep features, and 32-dimensional deep features are extracted. In this embodiment, a total of 78-dimensional features are extracted for life prediction.

[0082] LKA is originally applied to the field of image vision, has adaptability and long-range correlation, and the core idea is to decompose a large kernel convolution into a deep convolution with a dilation rate of and a deep convolution and point-wise convolution (where K is the size of the kernel and d is the dilation rate). Based on LKA, this embodiment proposes MLKA-T (Multi-scale Large Kernel Attention for Time series) to extract the multi-scale features in the bearing vibration signal. The bearing vibration signal has the following characteristics: first, the signal has a lot of non-stationarity and strong noise, and the early micro-fault information is easy to be submerged by the background noise; second, the signal contains obvious periodic impact characteristics, and these impact characteristics show a multi-scale evolution law in time and frequency as the fault develops; third, the fine-grained features of the signal at different stages (such as the stable period and the degradation period) are significantly different, and the dynamic combination of long-term dependence and short-term local patterns needs to be captured. The structural diagram of the MLKA-T module is shown in Figure 3

[0083] First, a number of simple convolution modules composed of convolution layers (Conv), maximum pooling layers (MaxPooling), layer normalization (LayerNorm) and Swish activation functions are used to extract local features of the vibration signal, to mine local impact features, transient high-energy features and small abnormal changes, thereby enhancing the sensitivity of the model to early degradation features and improving the overall expression ability. Subsequently, the multi-scale large kernel attention module (MLKA) approximates large kernel convolution by combining depthwise convolution, dilated depthwise convolution and point-wise convolution, which can effectively capture long-term time dependence features, especially the features of chronic changes in the degradation period, helping the model better understand the degradation trend and long-term evolution law of features. At the same time, the depthwise convolution with a small kernel size is used as a skip connection to further enhance the ability to capture local detail features, and to achieve dynamic balance extraction of local micro-fault features and global degradation features. The calculation process of the MLKA module is shown in equation (16): (16) By fusing the features generated by multiple MLKA modules with different kernel sizes and dilation rates, the multi-scale information (local transient anomalies and long-term degradation trends) of the vibration signal can be effectively extracted and comprehensively expressed. Finally, the convolution network is used to construct more nonlinear high-order feature combinations to more accurately represent the health state changes of the bearing, supporting subsequent tasks such as stage division and remaining useful life prediction.

[0084] Step 3.1.2: Model and fuse the features extracted from the multi-time scale to obtain a fusion graph.

[0085] ​To fully explore the feature evolution law of bearing vibration signals under multiple time scales, the extracted features are modeled and fused from multiple time scales to obtain the final fusion graph. The long-term trend reflects the overall performance degradation and long-term consistency of the features, while the short-term fluctuation reflects the local impact or the development of minor defects. Therefore, the cosine similarity is used to construct a long-term feature graph to capture the global consistency structure that shows the degradation trend in the vibration signal. At the same time, the dot product similarity is used to construct a short-term feature graph on a finer time scale to strengthen the dynamic change perception of local impact features and microscopic damage signals of the bearing. Through an adaptive fusion mechanism, the threshold is dynamically and adaptively adjusted so that when the short-term features significantly deviate from the global features, a larger weight is allocated, effectively combining the structures of the two types of graphs to realize the collaborative modeling of different time information and fully represent the bearing degradation characteristics.

[0086] Step 3.1.3: Use the graph multi-layer perception neural network to mine the deep relationships between nodes in the fusion graph and update the nodes.

[0087] The fusion graph is fed into the Message Passing Neural Network (MPNN) to mine the deep relationships between nodes in the feature graph and update the node representation.

[0088] During the degradation process of bearing vibration signals, there is a complex nonlinear correlation and dynamic dependence between different time domain, frequency domain, and depth features. MPNN can automatically learn and strengthen the internal relationship between different degradation mode features using global and local information of the fusion graph, thereby improving the discrimination and accuracy of node representation and more accurately depicting the evolution of bearing health status.

[0089] Step 3.1.4: Use the encoder to model the updated nodes in time series and use the regression algorithm to predict the remaining life of the bearing.

[0090] The updated node features are input into the DSC-Transformer encoder for time series modeling. Bearing vibration signals exhibit obvious stage changes and long-term dependencies throughout their life cycle. The DSC-Transformer encoder combines depthwise separable convolution (DSC) and multi-head attention mechanism to globally capture long-term degradation trends, enabling comprehensive learning of complex time series features throughout the bearing's life cycle. Finally, the regression prediction module outputs the remaining useful life (RUL) prediction results of the bearing, providing a basis for operation and maintenance strategy optimization and early maintenance. The performance of the model is verified and analyzed in combination with multiple evaluation indicators.

[0091] Step 3.2: Merge the prediction results of the stable period with the prediction results of the degradation period.

[0092] On the basis of stage division, the bearing life cycle is divided into stable period and degradation period in this embodiment. A targeted prediction model is established for each stage to obtain the stage-based residual life prediction result. Finally, the prediction result of the stable period is spliced with the prediction result of the subsequent degradation period in time to obtain the RUL prediction value of the entire life cycle of the bearing.

[0093] Due to the significant difference in degradation mechanism and signal characteristics of the bearing in different stages, direct overall modeling will lead to insufficient fitting of the model to the characteristics of the stable period or the degradation period, reducing the prediction accuracy. After stage division, a prediction model is established for different stages, which better fits the characteristic evolution law of each stage. Splicing the prediction results can more comprehensively and finely reflect the actual degradation process, significantly improving the overall prediction accuracy and robustness.

[0094] In order to better illustrate the excellent effect of the method of this embodiment, experimental verification is carried out, and the following is an experimental result description.

[0095] On the XJTU-SY dataset, the TRW-ShaRNN model can effectively identify the key stages of the bearing and accurately capture important features in the bearing degradation process. The network parameter settings on the public dataset XJTU-SY are shown in Table 2. The results before and after segmenting the XJTU-SY dataset for prediction are shown in Table 3. It can be seen that after segmentation, the RMSE and MAE of each bearing sample are effectively reduced, significantly improved, indicating that the model better captures the health state change trend of the bearing in different degradation stages. In addition, the prediction curve of this embodiment is highly consistent with the true life curve, and the absolute error is low, further verifying the accuracy of the segmented life prediction method.

[0096] Table 2. Network parameter settings

[0097] Table 3. Comparison of prediction results before and after segmentation of MTNT-Net

[0098] Table 4 shows the test results of MTNT-Net and other models. SAGCN-SA, T-GCN, and AGCN-TF are all based on graph neural networks, and RESFA-Bi-LSTM is based on Bi-LSTM. By comparing the results, it can be seen that MTNT-Net performs significantly better in most bearing tests. Specifically, the RMSE of SAGCN-SA, T-GCN, AGCN-TF, and RESFA-Bi-LSTM is generally higher, Generally lower than the MTNT-Net model.

[0099] Since stage 2 is closer to the end of life, the model is more critical to capture the bearing degradation in this stage, so the visualization results of stage 2 of bearing B1_3 of different prediction models further show that the MTNT-Net is superior to the existing model in bearing segmented life prediction, and can more effectively capture the key features in the bearing degradation process, significantly improving the prediction performance.

[0100] Table 4. Comparison of different segmented life prediction models

[0101] Embodiment two: The embodiment two of the present application provides a bearing segmented life prediction system based on a graph neural network, comprising: A data acquisition module configured to acquire a vibration signal of a bearing to be detected; A stage division stage configured to divide the full life process of the bearing into two stages of a stable period and a degradation period based on a health index of the bearing by using a life division model; A life prediction module configured to use a life prediction model to respectively predict the life of the vibration signal in the stable period and the degradation period, and combine the prediction results to obtain a final bearing life prediction result, wherein the life prediction model constructs the vibration signal into multi-modal time scale features, and mines and analyzes the time dependence in the multi-modal time scale features, thereby realizing life prediction.

[0102] Embodiment three: The embodiment three of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is suitable for being loaded and executed by a processor to perform the steps in the bearing segmented life prediction method based on the graph neural network as described in the embodiment one of the present application.

[0103] Embodiment four: The embodiment four of the present application provides a computer device, which comprises: A processor suitable for executing a computer program; A computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps in the bearing segmented life prediction method based on the graph neural network as described in the embodiment one of the present application.

[0104] The steps and methods involved in the above embodiments two, three and four correspond to the embodiment one, and the specific embodiments can refer to the related description part of the embodiment one.

[0105] Those skilled in the art can be aware that units and algorithm steps of each example described in combination with the embodiments disclosed in the application can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application. In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in or transmitted by a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data processing device such as a server, data center, etc. containing one or more available media sets. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, DVD) or semiconductor media (for example, solid state disk (SSD)) and the like. The above description is only a specific implementation of the application, but the protection scope of the application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the application, which should be covered within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the segmented life of bearings based on graph neural networks, characterized in that, Includes the following steps: The vibration signal of the bearing to be tested is acquired, including horizontal vibration signal and vertical vibration signal; Based on the health indicators of bearings, the entire life process of bearings is divided into two stages: the stable period and the degradation period using a life segmentation model. The lifetime prediction model is used to predict the lifetime of vibration signals during the stable period and the degradation period, and the prediction results are combined to obtain the final bearing lifetime prediction result. The lifetime prediction model constructs the vibration signal into multimodal time-scale features and mines and analyzes the time dependence in the multimodal time-scale features to achieve lifetime prediction.

2. The bearing segment life prediction method based on graph neural networks as described in claim 1, characterized in that, The specific steps for dividing the entire lifespan of a bearing using a lifespan segmentation model based on bearing health indicators are as follows: Construct bearing health indicators based on factors affecting bearing operating conditions; The criteria for classifying the period are determined based on the characteristics of the stable period and the degradation period; Vibration signals are processed using a lifespan classification model based on health indicators and classification conditions.

3. The bearing segment life prediction method based on graph neural networks as described in claim 2, characterized in that, The specific steps for constructing bearing health indicators based on the factors affecting bearing operating conditions are as follows: Preprocess the vibration signal of the bearing; Trend consistency index and expressive ability vector are introduced as the basic weighting factors for constructing health indicators, and game theory strategy is used for dynamic weight allocation.

4. The bearing segment life prediction method based on graph neural networks as described in claim 3, characterized in that, The specific steps for preprocessing the vibration signal of the bearing are as follows: The vibration signal of the bearing is first subjected to Hilbert transform, and then to fast Fourier transform. Then, the key frequency bands are filtered.

5. The bearing segment life prediction method based on graph neural networks as described in claim 1, characterized in that, The specific steps for using a lifetime prediction model to predict the lifetime of vibration signals during the stable and degradation phases are as follows: Feature extraction of vibration signals; The extracted features are modeled and fused from multiple time scales to obtain a fused map; Graph multilayer perceptron neural networks are used to mine and fuse deep relationships between nodes in a graph, and then node updates are performed. The updated nodes are modeled temporally using an encoder, and the remaining life of the bearing is predicted using a regression algorithm.

6. The bearing segment life prediction method based on graph neural networks as described in claim 5, characterized in that, Feature extraction is performed on the vibration signal to obtain both hand-designed features and deep learning features. In particular, a multi-scale large kernel attention module is used to extract deep features to obtain deep learning features.

7. The bearing segment life prediction method based on graph neural networks as described in claim 5, characterized in that, The specific steps for modeling extracted features across multiple time scales are as follows: Cosine similarity is used to construct long-term feature maps to capture the globally consistent structure in vibration signals that shows a degradation trend; By employing dot product similarity to construct short-term feature maps at a finer time scale, the dynamic changes in bearing local impact features and micro-damage signals are enhanced.

8. A bearing segment life prediction system based on graph neural networks, characterized in that, include: The data acquisition module is configured to acquire the vibration signal of the bearing to be tested; The phase division stage is configured to divide the entire life process of the bearing into two stages: the stable period and the degradation period, based on the bearing's health indicators and the life division model. The life prediction module is configured to use a life prediction model to predict the life of vibration signals during the stable period and the degradation period, and then merge the prediction results to obtain the final bearing life prediction result. The life prediction model constructs the vibration signal into multimodal time-scale features and mines and analyzes the time dependence in the multimodal time-scale features to achieve life prediction.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-7: the bearing segment life prediction method based on graph neural networks.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the bearing segment life prediction method based on a graph neural network as described in any one of claims 1-7.

Images