Turntable bearing residual life prediction method based on graph neural network-physical information neural network cooperation

Abstract

The invention relates to the field of turntable bearing life prediction, in particular to a turntable bearing residual life prediction method based on graph neural network-physical information neural network collaboration, which comprises the following steps: carrying out multi-class sensor acquisition on a large turntable bearing, and carrying out noise reduction on the acquired signal to obtain a smooth signal; a physical information embedded graph neural network (Physics-Attention GNNs) related to turntable bearing equipment is constructed and trained, nodes of the graph neural network represent operation data of different modes of the acquisition equipment, three times of graph convolution processing spliced with the physical information is carried out in a task, and the physical information is acquired; global physical information such as temperature and torque is spliced into vibration acceleration node features before convolution of each layer of graph, various space-time graph structures are constructed, and space-time relevance is enhanced. According to the method, the service life of the large turntable bearing can be predicted, and heavy economic losses caused by failure faults can be effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of slewing bearing life prediction, specifically a method for predicting the remaining life of slewing bearings based on the collaboration of graph neural network and physical information neural network. Background Technology

[0002] Slewing bearings, as key rotating components in industrial equipment, are widely used in large-scale equipment fields such as construction machinery, wind power equipment, and port machinery. Their operating status directly affects the overall reliability of the equipment. Therefore, predicting the remaining life of slewing bearings can anticipate their failures or malfunctions, effectively avoiding significant economic losses caused by such failures.

[0003] In the development of bearing life prediction research, early methods relied mainly on empirical formulas and physical models for life estimation. However, these traditional failure mechanism-based prediction methods struggle to capture the nonlinear characteristics of bearing degradation and are ill-suited for complex operating conditions. With the rise of data-driven approaches, recurrent neural networks such as Long Short-Term Memory (LSTM) networks can capture the temporal characteristics of life degradation, and Transformers further enhance long-sequence modeling capabilities. Predictive models combining these two approaches have also performed well in bearing life prediction, but their interpretability of physical mechanisms is limited. In recent years, Physical Information Neural Networks (PINs) have emerged, demonstrating excellent performance in life prediction due to their high correlation with physical mechanisms. Combining them with digital twin technology to quantify uncertainty and incorporate physical laws into model training improves prediction reliability. However, they are heavily reliant on prior physical knowledge, and their predictive performance significantly degrades when the physical model is inaccurate or incomplete.

[0004] Graph neural networks are suitable for modeling complex relationships between multiple components of a bearing, and can mine potential information in structured data to extract features and predict bearing life. However, their prediction accuracy is highly dependent on the accuracy of graph construction. Furthermore, slewing bearings operate at low speeds, are subjected to complex loads, and have weak fault signals. Multimodal methods improve the comprehensiveness of life prediction by fusing information from multiple sources such as vibration, temperature, and torque. However, existing methods still face challenges in feature fusion and heterogeneous data processing, including difficulties in feature alignment and uneven modal weight distribution. Therefore, this invention proposes a slewing bearing remaining life prediction method based on a synergistic approach of graph neural networks and physical information neural networks. This method combines physical information neural networks with graph neural networks, embedding multimodal features into the graph construction of the graph neural network. This more accurately constructs the graph structure of multi-source heterogeneous data, uses fully connected layers to solve the modal weight distribution problem, and improves the interpretability of physical mechanisms in data-driven prediction methods, effectively improving life prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the remaining life of turntable bearings based on the collaboration of graph neural networks and physical information neural networks, in order to solve the problems of lack of interpretability of physical mechanisms and difficulty in extracting structured data information between multiple modes, thereby improving the accuracy of life prediction.

[0006] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0007] A method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network includes the following steps:

[0008] Step S1: Obtain the vibration acceleration, temperature, and torque dataset of the large turntable bearing; extract the sample size synchronized with the torque and temperature modes from the vibration acceleration dataset; and perform data processing on the sample size.

[0009] Step S2: After processing the data, input it into the Physics-Attention GNNs model for training through a fully connected method. The model includes: a three-layer GCNConv to extract spatial relationships and a multi-head attention mechanism for the node features of each graph.

[0010] Step S3: Introduce the output RUL into the physical information loss function. If the condition is satisfied, perform forward propagation; otherwise, perform backward optimization.

[0011] Step S4: Finally, the RUL prediction value is output through global average pooling and dimensionality reduction of the fully connected layer.

[0012] Step S1 specifically includes:

[0013] Step S1.1: Obtain vibration acceleration, temperature and torque datasets during the operation of the turntable bearing using different sensors, and perform wavelet decomposition and threshold noise reduction on the original vibration dataset;

[0014] Step S1.2: Ten-dimensional statistical and frequency domain features were extracted from the denoised signal, namely the time-series features of mean, standard deviation, maximum value, minimum value, root mean square, peak value, skewness, kurtosis, frequency domain mean, and frequency domain maximum value.

[0015] Step S1.3: Normalize vibration, temperature and torque using minmaxscaler, and use zero padding to keep the feature dimensions of different modes consistent so that multiple sensors can interact fully.

[0016] Step S2 specifically includes:

[0017] Step S2.1: By introducing vibration acceleration, temperature and torque as three nodes into the graph neural network and using fully connected processing, feature vectors are formed for each node, and edges represent the physical connection relationships between modes, thus demonstrating multimodal fusion;

[0018] Step S2.2: The task involves three graph convolution processes concatenating physical information. After three layers of GCNConv extracting spatial relationships, higher-order interaction features between nodes are extracted layer by layer. Physical information is concatenated at each layer, strengthening the spatiotemporal correlation. The computational function can be simplified as follows: In the formula, It is a node A set of randomly sampled neighbor nodes, (·) represents an aggregation operation. This represents the physical information splicing operation. It is a vector dot product. It is the network layer parameter matrix. (·) is the sigmoid activation function.

[0019] Step S3 specifically includes:

[0020] Step S3.1: The first type of interval monotonicity loss function is used as the physical information loss function, with specific parameters including topk, margin, and physical loss coefficient λ. After parameter tuning, a comparison is performed. In the formula, The loss function of a neural network; This represents the error between the predicted value and the actual value. for The weights; For physical constraints; for The weights; N is the total number of physical constraint terms;

[0021] Step S3.2: Perform forward propagation of information during data training. The forward propagation process is to first aggregate the physical information of neighboring nodes layer by layer through graph convolution, then calculate the weights through the attention mechanism, and finally propagate the error backward by taking the partial derivative of the loss function.

[0022] Step S4 specifically includes:

[0023] Step S4.1: The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) between the predicted and actual RUL values ​​reflect the prediction performance. , , In the formula, n is the total number of samples; For the first The true value of each sample; For the first Predicted values ​​for each sample;

[0024] Step S4.2: Dimensionality reduction is achieved through global average pooling and fully connected layers. Partial derivatives of the loss function are then backpropagated to continuously decrease the loss function. Training ends when the loss function reaches convergence. The final output is the RUL prediction value.

[0025] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0026] This invention combines graph neural networks with physical information neural networks. This method can effectively model the relationships between various modes in a complex turntable bearing system, fully explore the potential information of structured data, has strong feature representation capabilities, improves prediction accuracy, and can handle dynamically changing working condition data, making it suitable for complex environments in actual engineering. It ensures that the prediction results conform to actual physical laws and has stronger interpretability and reliability. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the method of this approach;

[0028] Figure 2 This is a physical image of the turntable bearing test bench of the present invention;

[0029] Figure 3 Comparison of preprocessed waveforms for vibration acceleration;

[0030] Figure 4 This is a diagram of the overall network architecture.

[0031] Figure 5 For lifetime prediction process;

[0032] Figure 6 For physical information loss module and reverse optimization;

[0033] Figure 7 Comparison of prediction results with parameter adjustments (in a bar chart);

[0034] Figure 8 A scatter plot comparing the specific prediction effects of different parameters;

[0035] Figure 9 Here is a detailed diagram of the prediction results from this model;

[0036] Figure 10 A bar chart comparing the prediction results of multiple models;

[0037] Figure 11 A scatter plot showing the specific prediction results of multiple models;

[0038] Figure 12 The connection method of graph nodes;

[0039] Figure 13 A bar chart comparing the prediction results for different connection methods;

[0040] Figure 14 A scatter plot comparing the specific prediction effects of different connection methods;

[0041] Figure 15 A bar chart comparing prediction results for datasets divided at different proportions;

[0042] Figure 16 A scatter plot comparing the specific prediction results for different proportions of the dataset. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0044] Example 1

[0045] This embodiment discloses a method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network, such as... Figure 1 This includes the following steps:

[0046] Step S1: Obtain the vibration acceleration, temperature, and torque dataset of the large turntable bearing; extract the sample size synchronized with the torque and temperature modes from the vibration acceleration dataset; and perform data processing on the sample size.

[0047] Step S2: After processing the data, input it into the Physics-Attention GNNs model for training through a fully connected method. The model includes: a three-layer GCNConv to extract spatial relationships and a multi-head attention mechanism for the node features of each graph.

[0048] Step S3: Introduce the output RUL into the physical information loss function. If the condition is satisfied, perform forward propagation; otherwise, perform backward optimization.

[0049] Step S4: Finally, the RUL prediction value is output through global average pooling and dimensionality reduction of the fully connected layer.

[0050] In step S1, to obtain the vibration signal of the turntable bearing, an accelerated life test of the turntable bearing is conducted. This mainly includes mechanical, hydraulic, and measurement and control components. The test bench is driven by a hydraulic motor, which drives the turntable bearing meshing with it to rotate via a pinion gear. The test bench is as follows... Figure 2As shown. The experimental object was a QNA-730-22 type single-row ball slewing bearing. In the first group of tests, the axial force of 96kN and the overturning force of 240kNm were selected as the ultimate loads based on the specifications and load-bearing capacity of the slewing bearing. The experiment lasted for 11 days. All components of the slewing bearing showed performance degradation until they seized. At this point, the slewing bearing was in a state of wear failure. Data sets of its vibration acceleration, temperature and torque were obtained.

[0051] First, modal synchronization needs to be performed. The acquired signals are sampled, and 1000 data points are uniformly extracted from the temperature and torque data. Then, the vibration acceleration dataset is collected every 100 points. Figure 3 The collected vibration signals underwent preliminary noise reduction to remove outliers, specifically those exceeding six standard deviations. After denoising the vibration data, 2,051,282 vibration data points remained. 2,048 of these data points were grouped into a single sample, resulting in 1,000 samples. Additionally, 3,000 temperature and torque data points were sampled into another 1,000 samples, synchronizing the multimodal data (1,000 groups) and compressing the prediction time. Ten feature parameters were then extracted from the 1,000 vibration samples in the time-frequency domain: mean, standard deviation, maximum, minimum, root mean square, peak value, skewness, kurtosis, frequency domain mean, and frequency domain maximum. Finally, the vibration, temperature, and torque data were normalized using minmaxscaler, and zero-padding was used to ensure consistent feature dimensions across different modes.

[0052] In step S2, all processed vibration acceleration, temperature, and torque datasets are introduced into a graph neural network as three nodes in a fully connected manner, allowing for full interaction between multiple sensors. The edges of the graph represent the mutual influence relationships between different physical quantities, and this is the network structure. Figure 4 As shown, this demonstrates multimodal fusion. The input is a concatenation of each node's features and physical information. After three layers of GCNConv extract spatial relationships, physical information is concatenated at each layer, and high-order interaction features between nodes are extracted layer by layer. Then, multi-head self-attention learns weights for different sensor information on the node features of each graph. The features of the three nodes (sensors) are concatenated and processed through a fully connected layer, using dropout to prevent overfitting, and aggregated into global features. The forward propagation process first aggregates neighbor node information layer by layer through graph convolution, then reorganizes the graph data into the format required for attention, the attention mechanism calculates weights, then combines multi-node features into a unified representation, and finally reduces dimensionality through global average pooling and a fully connected layer to output the RUL prediction value. Its lifetime prediction process is as follows: Figure 5 As shown.

[0053] In step S3, the physical information loss function is used in the loss function. Because RUL has a high correlation with temperature, the first interval monotonicity loss function is adopted, which represents a portion of RUL decreasing as temperature increases. Specific parameters include topk, margin, and the physical loss coefficient λ. The formula for the total loss is... In the formula, The loss function of a neural network; This represents the error between the predicted value and the actual value. for The weights; For physical constraints; for The weights are denoted by N, which represents the total number of physical constraint terms and serves to balance the various loss terms. The physical information loss function module is as follows: Figure 6 As shown. Another trend slope constraint function (slope loss) is chosen as the physical information loss function for comparison. The MSE, RMSE, and MAE of the predicted RUL values ​​and the actual RUL values ​​are used to reflect the prediction effect; the smaller the index, the higher the prediction accuracy. The MSE values ​​in the bar chart have been magnified 10 times for comparison. The prediction results for different parameters topk and margin are analyzed as follows. Figure 7 As shown. Figure 8 The diagrams compare the prediction performance of different parameters. The dashed line represents the ideal prediction index for the sample, while the point cloud represents the model's prediction performance. A higher degree of overlap between the two indicates a better prediction performance. The comparison shows significant differences in remaining lifetime prediction performance between models with different margin and topk settings. In terms of prediction accuracy, stability, and efficiency, Scheme 2 demonstrates better prediction performance and outperforms the slope loss function. Through systematic parameter comparison experiments, the optimal parameter combination can be identified, thereby significantly improving prediction accuracy.

[0054] In step S4, the dataset is divided into training, validation, and test sets according to 60%, 25%, and 15% respectively. The training process uses DataLoader for batch processing to improve efficiency. Validation is performed in each epoch, and the partial derivative of the loss function is calculated to backpropagate the error, causing the loss function to continuously decrease. Training ends when the loss function reaches the convergence condition, overfitting is monitored, and finally, the RUL is output. Figure 9 The diagram shows a detailed analysis of the prediction results, including the training loss and prediction error distribution of the method proposed in this invention in this dataset prediction task. The prediction results accurately output the turntable bearing RUL.

[0055] Then, the performance of this invention in predicting the lifespan of turntable bearings is compared with four other methods: Physics-Attention GNNs, Attention GNNs, Physics GNNs, and Physics-Attention LSTM. The results are as follows: Figure 10 As shown in the detailed comparison chart of the prediction performance of each model, see below. Figure 11 As shown, the Physics-Attention GNNs model significantly outperforms other comparative models in predicting remaining lifetime. Through systematic ablation experiments, Physics-Attention GNNs demonstrates a significant advantage in integrating prior physical knowledge with attention mechanisms, enabling it to more accurately capture key features during bearing degradation. In contrast, single Attention GNNs or Physics GNNs exhibit significant deficiencies in overall prediction accuracy and stability. While the Physics-Attention LSTM model incorporates physical constraints, its LSTM architecture limits its processing of structured data, neglecting potential information. These comparative models could potentially improve their prediction performance through in-depth parameter optimization and architecture adjustments, but this requires substantial expert time and computational resources. In contrast, Physics-Attention GNNs, by cleverly integrating physical constraints and attention mechanisms, maintains high prediction accuracy while exhibiting better generalization ability and robustness, making it more competitive in terms of prediction accuracy, stability, and practicality. There are five ways to connect graph nodes, such as... Figure 12 As shown, from left to right, the graphs represent fully connected, chained, star-shaped, and ring-shaped connections. In a fully connected graph, each node is interconnected with the other two nodes, allowing direct information transfer between any two nodes. In a chained graph, nodes are connected sequentially end-to-end, forming a "chain," but the first and last nodes are not directly connected. In a star-shaped graph, the vibration node is the center, and temperature and torque nodes are only connected to the vibration node, not directly to each other. In a ring-shaped graph, each node is connected only to the next node, with the last node connecting back to the first node, forming a closed loop. Ring-shaped connections can be clockwise or counterclockwise, which in this case represents the sequential connection of temperature and torque. Comparing these graph construction methods yields the following prediction results: Figure 13 As shown in the diagram, a detailed comparison of the prediction results for different connection methods is presented. Figure 14As shown in the figure. The comparison reveals that the fully connected approach significantly outperforms other connection methods in remaining lifetime prediction. System comparison experiments show that the fully connected approach performs best in bearing remaining lifetime prediction. Its global information interaction mechanism ensures that each node can fully acquire all feature information in the network, effectively avoiding the information decay of chain connections, the edge interaction limitations of star connections, and the propagation delay problems of ring connections. The fully connected approach not only significantly improves prediction accuracy but also demonstrates significant advantages in model stability and generalization ability, providing the optimal network topology solution for accurately capturing complex bearing degradation features. Finally, to verify the rationality of the dataset partitioning, the prediction results of datasets with different proportions were merged and compared. The comparison results are shown in the figure. Figure 15 As shown in the figure, a detailed comparison of the prediction performance of datasets with different proportions is presented. Figure 16 As shown, different dataset partitioning ratios result in significant differences in model training and prediction performance. This embodiment employs a data partitioning strategy of 60%, 25%, and 15%, which yields the best prediction performance. Through comparative experiments on dataset partitioning ratios, the optimal data partitioning strategy can be identified, largely avoiding overfitting or underfitting problems caused by improper data partitioning, thereby significantly improving the model's generalization ability and prediction stability.

[0056] The embodiments disclosed in this invention are merely preferred implementations for illustrating the technical solutions and are not intended to limit the scope of patent protection. Those skilled in the art should understand that any technical adjustments, equivalent substitutions of embodiments, or structural modifications and process optimizations made based on the description and drawings, without departing from the core principles and spirit of this invention, for specific application scenarios, all fall within the scope of protection of this invention. Whether directly implemented or indirectly applied to other fields through technical connections, any implementation that conforms to the technical concept of this invention and falls within the scope defined by the claims is protected by patent law.

Claims

1. A method for predicting the remaining life of a turntable bearing based on the synergy of graph neural network and physical information neural network, characterized in that, Includes the following steps: Step S1: Obtain the vibration acceleration, temperature, and torque dataset of the large turntable bearing; extract the sample size synchronized with the torque and temperature modes from the vibration acceleration dataset; and perform data processing on the sample size. Step S2: After processing the data, input it into the Physics-Attention GNNs model for training through a fully connected method. The model includes: a three-layer GCNConv to extract spatial relationships and a multi-head attention mechanism for the node features of each graph. Step S3: Introduce the output RUL into the physical information loss function. If the condition is satisfied, perform forward propagation; otherwise, perform backward optimization. Step S4: Finally, the RUL prediction value is output through global average pooling and dimensionality reduction of the fully connected layer.

2. The method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network as described in claim 1, characterized in that, Step S1 specifically includes: Step S1.1: Obtain vibration acceleration, temperature and torque datasets during the operation of the turntable bearing using different sensors, and perform wavelet decomposition and threshold noise reduction on the original vibration dataset; Step S1.2: Ten-dimensional statistical and frequency domain features were extracted from the denoised signal, namely the time-series features of mean, standard deviation, maximum value, minimum value, root mean square, peak value, skewness, kurtosis, frequency domain mean, and frequency domain maximum value. Step S1.3: Normalize vibration, temperature and torque using minmaxscaler, and use zero padding to keep the feature dimensions of different modes consistent so that multiple sensors can interact fully.

3. The method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network according to claim 1, characterized in that, Step S2 specifically includes: Step S2.1: By introducing vibration acceleration, temperature and torque as three nodes into the graph neural network and using fully connected processing, feature vectors are formed for each node, and edges represent the physical connection relationships between modes, thus demonstrating multimodal fusion; Step S2.2: The task involves three graph convolution processes concatenating physical information. After three layers of GCNConv extracting spatial relationships, higher-order interaction features between nodes are extracted layer by layer. Physical information is concatenated at each layer, strengthening the spatiotemporal correlation. The computational function can be simplified as follows: In the formula, It is a node A set of randomly sampled neighbor nodes, (·) represents an aggregation operation. This represents the physical information splicing operation. It is a vector dot product. It is the network layer parameter matrix. (·) is the sigmoid activation function.

4. The method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network according to claim 1, characterized in that, Step S3 specifically includes: Step S3.1: The first type of interval monotonicity loss function is used as the physical information loss function, with specific parameters including topk, margin, and physical loss coefficient λ. After parameter tuning, a comparison is performed. In the formula, The loss function of a neural network; This represents the error between the predicted value and the actual value. for The weights; For physical constraints; for The weights; N is the total number of physical constraint terms; Step S3.2: Perform forward propagation of information during data training. The forward propagation process is to first aggregate the physical information of neighboring nodes layer by layer through graph convolution, then calculate the weights through the attention mechanism, and finally propagate the error backward by taking the partial derivative of the loss function.

5. The method for predicting the remaining life of a turntable bearing based on the collaboration of graph neural network and physical information neural network according to claim 1, characterized in that, Step S4 specifically includes: Step S4.1: The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) between the predicted and actual RUL values ​​reflect the prediction performance. , , In the formula, n is the total number of samples; For the first The true value of each sample; For the first Predicted values ​​for each sample; Step S4.2: Dimensionality reduction is achieved through global average pooling and fully connected layers. Partial derivatives of the loss function are then backpropagated to continuously decrease the loss function. Training ends when the loss function reaches convergence. The final output is the RUL prediction value.

Images