Rolling bearing residual life prediction method based on improved graph convolutional network and gating cycle unit
By combining an improved graph convolutional network with a gated recurrent unit, the problem of factor interference in the prediction of the remaining life of rolling bearings is solved, and high-precision and robust prediction results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUZHOU UNIV
- Filing Date
- 2024-06-07
- Publication Date
- 2026-04-21
AI Technical Summary
In complex industrial environments, existing data-driven methods for predicting the remaining life of rolling bearings are subject to interference from external and internal factors, resulting in poor prediction accuracy.
An improved combined model of graph convolutional networks and gated recurrent units is adopted. Through adaptive variational mode decomposition, time-domain and frequency-domain feature extraction, and the combination of graph convolutional networks and gated recurrent units, data processing is performed to extract hidden features and reduce the influence of interference factors.
It improves the robustness and accuracy of rolling bearing remaining life prediction, can effectively identify vibration signals from different experimental platforms, and has high prediction accuracy and generalization ability.
Smart Images
Figure CN121902308A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated cyclic unit. Background Technology
[0002] Currently, there are four main methods for predicting the remaining life of rolling bearings: (1) Statistical model-based prediction methods. These methods are easily affected by factors such as machining accuracy and working environment. (2) Physical model-based prediction methods. These methods rely on physical models to model and analyze the failure mechanism, but establishing an effective physical model requires relevant personnel to have a deep understanding of complex bearing structures, and these methods have low generalization ability. (3) Data-driven prediction methods. Due to the rapid development of sensor technology in recent years, data collection has become more convenient. Therefore, data-driven prediction methods have become the mainstream. Data-driven methods include machine learning-based methods and uncertainty-based methods. Machine learning methods eliminate human intervention and experience dependence during feature extraction.
[0003] Data-driven methods have played a significant role in the field of rolling bearing remaining life prediction in existing publicly available technologies. However, in complex industrial environments, due to the influence of external conditions and internal factors, these influencing factors may mask the characteristic information of newly collected data samples that require remaining life prediction, resulting in poor training and testing effects and consequently affecting the accuracy of rolling bearing remaining life prediction. Summary of the Invention
[0004] The purpose of this invention is to address the inconveniences and drawbacks of existing technologies and to provide a method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit. This method can extract hidden features from data through data processing, reduce interference from internal and external factors of the bearing, and achieve prediction of the remaining life of rolling bearings.
[0005] To achieve the above objectives, the technical solution of the present invention is: a method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit, which extracts hidden features through data processing, reduces interference from internal and external factors of the bearing, and realizes the prediction of the remaining life of rolling bearings.
[0006] In one embodiment of the present invention, the method includes:
[0007] Step S1: Obtain the rolling bearing accelerated life dataset;
[0008] Step S2: Perform Adaptive Variational Mode Decomposition (AVMD) on the obtained dataset;
[0009] Step S3: Extract time-domain and frequency-domain statistical features from the data samples obtained in step S2, and perform normalization processing.
[0010] Step S4: Set a sliding window with a predetermined time step for the normalized data, and use the data within the time step as a new sample, with the sample set X. s ={x1,x2,x3,…,x N}, where x i ∈R W×H This represents the sample feature input of a certain bearing under a certain working condition at time i, where N is the number of samples and Y is the label. s ={y1,y2,y3,…,y N}, where y i ∈[0,1] represents the percentage of the bearing's remaining life at time i;
[0011] Step S5: Take one bearing as the test set and use some or all of the remaining bearing data as the training set.
[0012] Step S6: Set the output dimension of the embedding operation, set the pooling kernel size and stride of the one-dimensional max pooling layer, and then set the sample X. s New samples are obtained by performing embedding operations and max pooling in batches.
[0013] Step S7: Output the new samples after max pooling to the graph convolutional network GCN improved by the compressed excitation network SEN to obtain new multidimensional feature data.
[0014] Step S8: Input the new multidimensional feature data into the gated recurrent unit (GRU) and the fully connected layer (FC) to predict the remaining life of the rolling bearing.
[0015] In one embodiment of the present invention, in step S1, the rolling bearing accelerated life dataset is the rolling bearing accelerated life dataset open sourced from the IEEE PHM 2012 challenge.
[0016] In one embodiment of the present invention, in step S2, adaptive variational mode decomposition (AVMD) is performed on the obtained dataset. The original data is decomposed into different intrinsic mode function (IMF) analytical signals by Hilbert transform. The estimated center frequency is mixed into the different IMF analytical signals, and each IMF is transferred to the corresponding baseband frequency.
[0017] In one embodiment of the present invention, in step S2, the adaptive variational mode decomposition (AVMD) uses a particle swarm optimization algorithm to optimize the penalty parameter α and the number of mode decompositions k in the variational mode decomposition; the particle velocity and particle position update calculation formulas of the particle swarm optimization algorithm are expressed as follows:
[0018]
[0019] in, This represents the d-th component of the velocity vector of particle i in the k-th iteration; The d-th component of the position vector of particle i in the k-th iteration is represented by c1 and c2, which represent acceleration constants; r1 and r2 represent two random numbers with values in the range [0, 1]; w represents the inertial weight; pbest id pbest represents the optimal solution for particle i. d This represents the globally optimal solution.
[0020] In one embodiment of the present invention, in step S7, the basic formula of the graph convolutional network GCN is expressed as:
[0021]
[0022] Among them, H (l) H represents the input features of the l-th layer; (l+1) Indicates output features; w (l) This represents a randomly initialized weight matrix; σ(·) represents a nonlinear activation function. This represents an adjacency matrix with self-connected nodes; This represents the degree matrix of the self-connection matrix.
[0023] In one embodiment of the present invention, in step S8, the fully connected layer FC consists of two layers.
[0024] In one embodiment of the present invention, in step S8, the gated loop unit GRU consists of a reset gate R and an update gate F. The reset gate determines whether to ignore the previous state information, and the update gate determines whether to save the previous state information.
[0025] In one embodiment of the present invention, in step S8, the gated recurrent unit (GRU) selectively stores state information through reset and update gates, while adding input information, and realizes model training and output through its internal loop mechanism. The specific calculation formula is as follows:
[0026] R t =σ(X) t W xr +H t-1 W hr +b r )
[0027] F t =σ(X) t W xf +H t-1 W hf +b f )
[0028]
[0029] Among them, R t Indicates the gate that controls the reset; F t This indicates a gating mechanism that controls updates; The candidate hidden state is obtained by resetting the gate; H t The current state is the output hidden state obtained by updating the gate; W j Let the weight matrix be j = xr, xf, xh, hr, hf, hh, b i Let i be the bias vector, i = r, f, h; σ(·) is the Sigmoid activation function, outputting an S-shaped curve in the range [0,1]; tanh(·) is the Tanh activation function, outputting an S-shaped curve in the range [-1,1]; X t Let t represent the input sample, where t = 1, 2, 3, ..., N.
[0030] In one embodiment of the present invention, the method constructs a model by selecting the Adam optimizer to optimize and update the parameters of each part. The initial learning rate of the model is set to 0.0005, the lower limit of the learning rate of the model is 0.0001, and the number of samples per batch sent to the model in batches during training is 64.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] 1. The method of the present invention is highly robust, highly accurate, and easy to implement. It can effectively identify the vibration signals of rolling bearings on different experimental platforms and has certain implications for the prediction of the remaining life of rolling bearings.
[0033] 2. The method of this invention utilizes the particle swarm optimization algorithm to adaptively optimize the parameters of the variational mode decomposition algorithm, and combined with the subsequent improved graph convolutional network, it can process vibration signals more flexibly.
[0034] 3. The network constructed by the method of the present invention can extract the dependency relationship between different time nodes of the bearing vibration signal and has high accuracy in predicting the remaining life of roller bearings under different working conditions. Attached Figure Description
[0035] Figure 1 This is a flowchart of the method of the present invention.
[0036] Figure 2 The PRONOSTIA testbed was used to collect and verify the dataset for this invention.
[0037] Figure 3 This is a schematic diagram of the network model framework of the present invention.
[0038] Figure 4 The present invention provides normalized feature maps of time-domain statistical features and frequency-domain statistical features.
[0039] Figure 5 The effect of different time steps on RMSE in this invention
[0040] Figure 6 The diagram shows the predicted effects of this invention on bearings B1-1, B1-3, B1-4, B2-1, B2-2, and B2-3.
[0041] Figure 7 This is a bar chart comparing the prediction metrics RMSE and MAE of this invention with those of different network models.
[0042] Figure 8 This is a comparison chart showing the prediction effects of the present invention and different network models on bearings B1-1, B1-4, and B1-7. Detailed Implementation
[0043] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0044] This invention provides a method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and gated recurrent units. Through data processing, it reduces interference from internal and external factors of the bearing and completes the prediction of the remaining life of rolling bearings by extracting hidden features from the data. The method includes the following steps:
[0045] Step S1: To verify the robustness and generalization of the proposed method, the rolling bearing accelerated life dataset from the IEEE PHM 2012 challenge was selected as the data for verification.
[0046] Step S2: Perform Adaptive Variational Mode Decomposition (AVMD) on the obtained dataset, that is, use the particle swarm optimization algorithm to optimize the penalty parameter α and the number of mode decompositions k in the variational mode decomposition;
[0047] Step S3: Extract 11 time-domain statistical features and 12 frequency-domain statistical features from the data sample, and normalize the features.
[0048] Step S4: Set a sliding window with a certain time step for the normalized data, and use the data within the time step as a new sample, with the sample set X. s ={x1,x2,x3,…,x N}, where x i ∈R W×H This represents the sample feature input of a specific bearing under a certain operating condition at time i, where N is the number of samples. Label Y s ={y1,y2,y3,…,y N}, where y i∈[0,1] represents the percentage of the remaining life of the bearing at time i.
[0049] Step S5: Take one bearing as the test set and use some or all of the remaining bearing data as the training set.
[0050] Step S6: Set the output dimension of the embedding operation, and set the pooling kernel size and stride of the one-dimensional max pooling layer. Then, process the sample X. s New samples are obtained by performing embedding operations and max pooling in batches.
[0051] Step S7: Output the new samples after max pooling to a graph convolution network (GCN) improved from the squeeze-and-excitation network (SEN) to obtain new multidimensional feature data.
[0052] Step S8: Finally, the new multidimensional features are input into the gated recurrent unit (GRU) and the fully connected layer (FC) to predict the remaining life of the rolling bearing.
[0053] In the above steps, the model used in this invention is an improved graph convolutional network with gated recurrent units, and its structure is as follows: Figure 3 As shown, the network model consists of three main parts: preprocessing, feature extraction, and remaining lifetime prediction. Preprocessing utilizes adaptive variational mode decomposition (ADD) with Hilbert transform to decompose the original data into different intrinsic mode function (EMF) analytical signals. The estimated center frequency is then mixed into the different EMF analytical signals, transferring each EMF to its corresponding baseband frequency. Feature extraction leverages an improved graph convolutional network (GRN), which utilizes a compression excitation mechanism to aggregate information from the original data and fully capture channel-related dependencies. The GRN can fuse data from different time points and extract the interrelationships between data, establishing new multidimensional features. Remaining lifetime prediction utilizes gated recurrent units (GRNs) to capture long-term sequence relationships and predict the bearing's remaining service life.
[0054] In adaptive variational mode decomposition (VMD), to prevent the penalty parameter α from being too large or too small, which could lead to missing or redundant frequency band information, and to prevent the number of mode decompositions k from being too large or too small, which could result in over-decomposition or under-decomposition of the signal, particle swarm optimization (PSO) is used to optimize the penalty parameter α and the number of mode decompositions k in VMD. The particle velocity and particle position update formulas of the PSO algorithm can be expressed as:
[0055]
[0056] in, This represents the d-th component of the velocity vector of particle i in the k-th iteration; The d-th component of the position vector of particle i in the k-th iteration is represented by c1 and c2, which represent acceleration constants; r1 and r2 represent two random numbers with values in the range [0, 1]; w represents the inertial weight; pbest id pbest represents the optimal solution for particle i. d This represents the globally optimal solution.
[0057] After adaptive variational mode decomposition, the data is input into an embedding layer and a one-dimensional max pooling layer. The embedding operation can capture key features of the data and improve efficiency during task execution. The one-dimensional max pooling layer can reduce the model size, improve the computational efficiency of the model, and enhance the robustness of the model.
[0058] In the improved graph convolutional network, the synergistic effect of the compressed activation network and the graph convolutional network is utilized to further extract the interrelationship features between samples at different time points from the data passing through the embedding layer and the one-dimensional max pooling layer. The basic formula of the graph convolutional network can be expressed as:
[0059]
[0060] Among them, H (l) H represents the input features of the l-th layer; (l+1) Indicates output features; w (l) This represents a randomly initialized weight matrix; σ(·) represents a nonlinear activation function. This represents an adjacency matrix with self-connected nodes; This represents the degree matrix of the self-connection matrix.
[0061] The output of the improved graph convolutional network is then input into the GRU. The output of the GRU network is then fed into two fully connected layers to output the final result. The GRU network model mainly consists of a reset gate (R) and an update gate (F). The reset gate determines whether the network chooses to ignore previous state information, while the update gate determines whether the network saves previous state information. The reset and update gates selectively store state information while simultaneously incorporating input information. An internal loop mechanism is used to train and output the model. The specific calculation formula is as follows:
[0062] R t =σ(X) t W xr +H t-1 W hr +b r )
[0063] F t =σ(X) t W xf +H t-1 W hf +b f )
[0064]
[0065] Among them, R t Indicates the gate that controls the reset; F t This indicates a gating mechanism that controls updates; The candidate hidden state is obtained by resetting the gate; H t The current state is the output hidden state obtained through the update gate. W is the weight matrix; b is the bias vector; σ(·) is the Sigmoid activation function, outputting an S-shaped curve in the range [0,1]; tanh(·) is the Tanh activation function, outputting an S-shaped curve in the range [-1,1]; X t Let t represent the input sample, where t = 1, 2, 3, ..., N.
[0066] The Adam optimizer was selected to optimize and update the parameters of each part of the model. The initial learning rate of the model was set to 0.0005, and the lower limit of the learning rate of the model was 0.0001. During the training process, the number of samples fed into the network model in batches was 64.
[0067] Example
[0068] Reference Figure 1 The present invention provides a method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated cyclic unit, which specifically includes the following steps.
[0069] Step S1: Obtain the rolling bearing accelerated life dataset from the IEEE PHM 2012 challenge. The experimental data used is horizontal vibration signal with a sampling frequency of 25.6 kHz, a sampling time of 0.1 s, and a sampling interval of 10 s. Use the bearing data from the first two operating conditions for verification experiments, as shown in Table 1.
[0070] Table 1. IEEE PHM2012 Dataset Information
[0071]
[0072] Step S2: Perform adaptive variational mode decomposition preprocessing on the sorted data samples, and extract time-domain and frequency-domain statistical features.
[0073] Step S3: Normalize the extracted statistical features to obtain deeper information, which will be used as input samples for model training, with the percentage of remaining lifespan as the label. The normalization formula is expressed as:
[0074]
[0075] Where x' is the normalized data and x is the data before normalization. Figure 4 Normalized feature maps for time-domain and frequency-domain statistical features.
[0076] Step S4: Set a sliding window of a certain time step for the normalized dataset, and treat the data within the time step as a new sample. The time step is selected as follows: Figure 5 As shown, the RMSE effect is better when the time step is set to 128.
[0077] Step S5: Select one bearing data point from the new dataset under the same working condition as the test set, and select some or all bearing data from the remaining bearings as the training set.
[0078] Step S6: Train the dataset samples in batches, with 64 samples in each batch. The output of the embedding layer is 512-dimensional. The pooling kernel size of the one-dimensional max pooling layer is 2, and the stride is 2. The output dimension of graph convolutional layer 1 is 64, the output dimension of graph convolutional layer 2 is 32, the output dimension of the GRU network is 32, the output dimension of fully connected layer 1 is 8, and the output dimension of fully connected layer 2 is 1.
[0079] Step S7: To reflect model performance and verify the model's generalization ability, root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics. The formulas for calculating RMSE and MAE are as follows:
[0080]
[0081] In the formula, N is the sample size; y i The actual value; These are predicted values.
[0082] Step S8: In the experimental dataset, three bearings were selected from each of operating conditions 1 and 2 as visualization examples. In operating condition 1, bearings 1-1, 1-3, and 1-4 were selected; in operating condition 2, bearings 2-2, 2-3, and 2-6 were selected. The prediction result curves were smoothed using the Savitzky-Golay algorithm with a window size of 20, enabling a more accurate judgment of the prediction results. The visualization is as follows: Figure 6 As shown.
[0083] Figure 2 The PRONOSTIA testbed was used to collect and verify the dataset for this invention.
[0084] To demonstrate the superior predictive performance of the SEGCN-GRU model, the CNN-GRU model and the Deep Residual Networks (DRSN) model were compared with the SEGCN-GRU model. The results are shown in Table 2. The proposed SEGCN-GRU model significantly outperforms the other two methods in predicting most bearings. Figure 7 As shown in Table 3, the proposed model has an overall mean RMSE of 0.1080 and an overall mean MAE of 0.0853, representing an improvement of approximately 32.5% in prediction performance compared to the CNN-GRU model and approximately 30% in prediction performance compared to the DRSN model. The results are illustrated in Table 3. The RUL prediction results of different models for bearings 1-1, 1-4, and 1-7 are compared and visualized. The visualization results are shown in Table 3. Figure 8 As shown in the figure, the fitting result of this method is optimal, thus further confirming the superior performance of the SEGCN-GRU model.
[0085] Table 2 Performance evaluation of different models on the IEEE PHM 2012 dataset
[0086]
[0087] Table 3 Average performance of different models under different operating conditions
[0088]
[0089] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit, characterized in that, By extracting hidden features from the data through data processing, interference from internal and external factors in bearings can be reduced, enabling the prediction of the remaining life of rolling bearings.
2. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit as described in claim 1, characterized in that, include: Step S1: Obtain the rolling bearing accelerated life dataset; Step S2: Perform Adaptive Variational Mode Decomposition (AVMD) on the obtained dataset; Step S3: Extract time-domain and frequency-domain statistical features from the data samples obtained in step S2, and perform normalization processing. Step S4: Set a sliding window with a predetermined time step for the normalized data, and use the data within the time step as a new sample, with the sample set X. s ={x1,x2,x3,…,x N }, where x i ∈R W×H This represents the sample feature input of a certain bearing under a certain working condition at time i, where N is the number of samples and Y is the label. s ={y1,y2,y3,…,y N }, where y i ∈[0,1] represents the percentage of the bearing's remaining life at time i; Step S5: Take one bearing as the test set and the remaining bearing data or all bearing data as the training set. Step S6: Set the output dimension of the embedding operation, set the pooling kernel size and stride of the one-dimensional max pooling layer, and then set the sample X. s New samples are obtained by performing embedding operations and max pooling in batches. Step S7: Output the new samples after max pooling to the graph convolutional network GCN improved by the compressed excitation network SEN to obtain new multidimensional feature data. Step S8: Input the new multidimensional feature data into the gated recurrent unit (GRU) and the fully connected layer (FC) to predict the remaining life of the rolling bearing.
3. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S1, the rolling bearing accelerated life dataset is the rolling bearing accelerated life dataset open sourced from the IEEE PHM 2012 challenge.
4. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S2, adaptive variational mode decomposition (AVMD) is performed on the obtained dataset. The original data is decomposed into different intrinsic mode function (IMF) analytical signals through Hilbert transform. The predicted center frequency is mixed into the different IMF analytical signals, and each IMF is transferred to the corresponding baseband frequency.
5. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S2, the Adaptive Variational Mode Decomposition (AVMD) is performed, and the particle swarm optimization algorithm is used to optimize the penalty parameter α and the number of mode decompositions k in the variational mode decomposition. The particle velocity and particle position update calculation formulas of the particle swarm optimization algorithm are expressed as follows: in, This represents the d-th component of the velocity vector of particle i in the k-th iteration; The d-th component of the position vector of particle i in the k-th iteration is represented by c1 and c2, which represent acceleration constants; r1 and r2 represent two random numbers with values in the range [0, 1]; w represents the inertial weight; pbest id pbest represents the optimal solution for particle i. d This represents the globally optimal solution.
6. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S7, the basic formula for the Graph Convolutional Network (GCN) is expressed as follows: Among them, H (l) H represents the input features of the l-th layer; (l+1) Indicates output features; w (l) This represents a randomly initialized weight matrix; σ(·) represents a nonlinear activation function. This represents an adjacency matrix with self-connected nodes; This represents the degree matrix of the self-connection matrix.
7. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S8, the fully connected layer FC consists of two layers.
8. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, In step S8, the gated loop unit GRU consists of a reset gate R and an update gate F. The reset gate determines whether to ignore the previous state information, and the update gate determines whether to save the previous state information.
9. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit as described in claim 8, characterized in that, In step S8, the gated recurrent unit (GRU) selectively stores state information through reset and update gates, while simultaneously adding input information. The model is trained and output through its internal loop mechanism. The specific calculation formula is as follows: R t =σ(X t W xr +H t-1 W hr +b r ) F t =σ(X t W xf +H t-1 W hf +b f ) Among them, R t Indicates the gate that controls the reset; F t This indicates a gating mechanism that controls updates; The candidate hidden state is obtained by resetting the gate; H t The current state is the output hidden state obtained by updating the gate; W j Let the weight matrix be j = xr, xf, xh, hr, hf, hh, b i Let i be the bias vector, i = r, f, h; σ(·) is the Sigmoid activation function, outputting an S-shaped curve in the range [0,1]; tanh(·) is the Tanh activation function, outputting an S-shaped curve in the range [-1,1]; X t Let t represent the input sample, where t = 1, 2, 3, ..., N.
10. The method for predicting the remaining life of rolling bearings based on an improved graph convolutional network and a gated recurrent unit according to claim 2, characterized in that, This method uses the Adam optimizer to optimize and update the parameters of each part of the model. The initial learning rate of the model is set to 0.0005, the lower limit of the learning rate of the model is 0.0001, and the number of samples per batch fed into the model during training is 64.