Bearing life prediction method based on multi-domain graph migration driven by digital twin model

Abstract

The application relates to a kind of multi-domain graph migration bearing life prediction methods based on digital twin model driving, it includes the following steps, step 1: data acquisition is carried out to bearing equipment, and the digital twin model of bearing is established;Step 2: the whole life data of bearing is generated through digital twin model, and the state monitoring data of degradation stage is extracted;Step 3: the feature of state monitoring data of degradation stage is extracted, and multi-feature graph data is constructed;Step 4: a multi-domain graph migration bearing life prediction model is constructed;Step 5: according to the multi-domain graph migration bearing life prediction model, accurate bearing life prediction value is obtained.The application generates a large amount of data by constructing a digital twin model, presents multi-scale graph data degradation information from multiple angles;Using Graphormer migration model, the global information of feature graph is grasped;Using multi-source domain adaptive counter migration learning, the domain generalization ability of migration and the robustness of model are improved, and bearing residual life prediction is realized.

Description

Technical Field

[0001] This application relates to the field of bearing life prediction technology, and specifically to a bearing life prediction method based on a digital twin model-driven multi-domain graph migration. Background Technology

[0002] Rolling bearings are crucial components widely used in mechanical equipment, and their life prediction is essential for ensuring the normal operation and maintenance of equipment. Currently, existing deep learning models have achieved great success in predicting the remaining life of bearings, but they also face several technical challenges. One major problem is uneven data distribution. Due to differences in equipment operating environments, vibration data of the same type of rolling bearing exhibits different characteristics and distributions under different operating conditions, necessitating modeling for different data distributions during prediction. Another issue is insufficient data samples. Because rolling bearings have a long service life in actual operation, long-term data collection and monitoring are required to obtain sufficient data samples. Furthermore, the different operating conditions of each piece of equipment result in significant differences in the acquired data samples, further exacerbating the problem of insufficient data samples. To address these issues, new technologies and methods need to be developed to improve the accuracy and reliability of rolling bearing life prediction.

[0003] Currently, bearing life prediction methods and systems based on digital twin technology have become a research hotspot, mainly focusing on improving and optimizing existing bearing life prediction methods and systems based on signal processing and data-driven approaches. Signal processing-based methods require manual analysis of signal features, and different processing methods are needed for signals under different operating conditions, making it difficult to meet the needs of online diagnostic applications. Data-driven methods, on the other hand, require extensive pre-training with large amounts of data, making model transfer difficult, and the internal mechanisms of the models remain unclear.

[0004] Therefore, the graph-transfer bearing life prediction method and system driven by digital twin models fully combines the advantages of digital twin technology and machine learning technology, enabling online prediction of bearing life and possessing universality and transferability. This method and system monitor and evaluate the operating status of bearings in real time, predict bearing life in advance, provide a basis for equipment maintenance and replacement, and ensure the safe and efficient operation of rotating machinery. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention generates a large amount of data by constructing a digital twin model, presenting multi-scale graph data degradation information from multiple perspectives; it uses the Graphormer transfer model to grasp the global information of the feature map; and it uses multi-source domain adaptive adversarial transfer learning to improve the domain generalization ability of the transfer and the robustness of the model, thereby achieving bearing remaining life prediction.

[0006] To achieve the above objectives, the solution adopted by this invention is to provide a multi-domain graph migration bearing life prediction method driven by a digital twin model, which includes the following steps:

[0007] Step 1: Collect data from the bearing equipment and establish a digital twin model of the bearing;

[0008] Collect bearing operating data, including vibration, temperature, pressure, and speed; collect bearing parameters, including material properties, structural parameters, and operating condition parameters; perform dynamic stress analysis on the bearing during operation and record the stress state.

[0009] Based on the operational data, parameters, and stress states obtained above, a physical model and a data model are established; the physical model and the data model are then integrated to establish a digital twin model of the bearing, predicting and analyzing the bearing's behavior and health status.

[0010] Step 2: Generate bearing life-cycle data through a digital twin model and extract condition monitoring data during the degradation stage;

[0011] The digital twin model constructed in step 1 generates bearing life-cycle data under different operating conditions. The generated bearing life-cycle data undergoes noise reduction and normalization. The bearing life-cycle data includes a health stage and a degradation stage, which are detected using the root mean square index (RMS). When the RMS meets certain conditions, condition monitoring data for the degradation stage can be obtained. The determination conditions are as follows:

[0012] rms>μ+w*σ;

[0013] In the formula: rms represents the root mean square index; μ represents the first parameter obtained from the vibration data in the initial health stage; σ represents the second parameter obtained from the vibration data in the initial health stage; w represents a random weight in the interval [3,5].

[0014] When the above inequality holds, it serves as the dividing line between the healthy stage and the degenerative stage;

[0015] Step 3: Extract features from the state monitoring data of the degradation stage and construct multi-feature map data;

[0016] Step 31: Extract features from the state monitoring data of the degradation stage obtained in Step 2; first, extract time-domain features from the state monitoring data; then, extract frequency-domain features from the state monitoring data; finally, perform wavelet packet decomposition on the transformed state monitoring data. The extracted features can comprehensively reflect the degradation information of the degradation data.

[0017] Step 32: Construct multi-feature map data; based on the similarity between nodes due to multiple features, determine the similarity distance between each node, and use cosine similarity as the metric. The specific formula is as follows:

[0018]

[0019] In the formula: sim represents cosine similarity; A represents the first vector between nodes; B represents the second vector between nodes; A i B represents the first vector between the i-th nodes; i This represents the second vector between the i-th nodes; i represents the node number; n represents the total number of nodes;

[0020] The similarity distance is used to determine whether there is an edge connecting the two nodes. If the distance is greater than 0, there is an edge connecting the two nodes; otherwise, there is no edge connecting the two nodes.

[0021] Step 4: Construct a multi-domain graph migration bearing life prediction model;

[0022] Step 41: Determine the input data for the multi-domain graph migration bearing life prediction model;

[0023] Step 42: Construct the Graphormer feature extractor to extract the input data features of the multi-domain graph migration bearing life prediction model;

[0024] Step 43: Select the optimal source domain training data, and perform weighted processing on the distance difference between multiple domains to distinguish the source domain data with the smallest distribution difference from the target domain training data, and determine the optimal source domain;

[0025] Step 44: Perform domain adversarial operations on the selected labeled source domain training data and unlabeled target domain training data; feed the target domain training data from the test set into the multi-domain graph migration bearing life prediction model for prediction, and determine whether the cross-operating condition and cross-fault migration prediction function has been achieved by calculating RMSE and SF. The specific calculation formula is as follows:

[0026]

[0027] In the formula: d i The value of RMSE represents the difference between the predicted and actual values ​​of the multi-domain graph migration bearing life prediction model between the i-th nodes; SF represents the model score index; N represents the number of graph data samples; RUL represents the difference between the predicted and actual values. pre This represents the predicted value of the multi-domain graph migration bearing life prediction model; RUL act This represents the true value of the training data in the target domain;

[0028] Step 5: Transfer the bearing life prediction model based on the multi-domain graph to obtain accurate bearing life prediction values;

[0029] Collect bearing operation data from the actual production process, obtain multi-feature map data based on steps 2 and 3, and input it into the multi-domain graph transfer bearing life prediction model trained in step 4 to accurately predict bearing life.

[0030] Preferably, the temporal feature extraction of the status monitoring data in step 31 specifically includes:

[0031] The root mean square index (rms) of the condition monitoring data is as follows:

[0032]

[0033] In the formula: x i This represents the vibration signal of the i-th node;

[0034] The root square amplitude CL of the state monitoring data is as follows:

[0035]

[0036] In the formula: CL represents the root magnitude; max represents the maxima function; x represents the data generated by the digital twin;

[0037] The kurtosis (KU) of the state monitoring data is as follows:

[0038]

[0039] In the formula: KU represents the kurtosis of the condition monitoring data; This represents the mean of the status monitoring data.

[0040] Preferably, in step 31, frequency domain feature extraction is performed on the state monitoring data, specifically as follows:

[0041] First, perform a Fourier transform on the status monitoring data, as follows:

[0042]

[0043] In the formula: S(k) represents the frequency domain of the signal; x(kΔtz) represents the sampled value of the state monitoring data; N represents the number of sampling points; Δtz represents the sampling interval; k represents the index of the time-frequency discrete value; γ represents the complex number sign;

[0044] Then, the centroid frequency (MSF) and mean square frequency (FC) are obtained using the transformed state monitoring data, as shown in the following formulas:

[0045]

[0046] In the formula: MSF represents the centroid frequency of the condition monitoring data; FC represents the mean square frequency of the condition monitoring data; k represents the number of spectral lines; f k (k) represents the frequency value of the kth spectral line.

[0047] Preferably, in step 31, wavelet packet decomposition is performed on the transformed state monitoring data, specifically as follows:

[0048] Time-frequency domain feature extraction obtains component data of different levels and directions of the signal through 3dB wavelet packet decomposition. The original signal before decomposition is as follows:

[0049]

[0050] In the formula: S 0,0 S represents the original signal before decomposition. 1,0 and S 1,1 S represents the first and second wavelet basis coefficient decomposition signals of the first layer, respectively; 2,0 S 2,1 S 2,2 and S 2,3 These represent the first, second, third, and fourth wavelet basis coefficient decomposition signals of the second layer, respectively; S 3,0 S 3,1 S 3,2 S 3,3 S 3,4 S 3,5 S 3,6 and S 3,7 These represent the first, second, third, fourth, sixth, seventh, and eighth wavelet basis coefficient decomposition signals of the third layer, respectively.

[0051] S e,f This represents the decomposed signal corresponding to the f-th node at the e-th scale. The specific decomposition formula is as follows:

[0052]

[0053] In the formula: g(k-2n) represents the wavelet basis function; S e,f (k) represents the decomposition signal of the f-th wavelet basis coefficient at position k in the e-th scale; S e+1,2f (n) represents the decomposition signal of the 2f-th wavelet basis coefficient at position k in the (e+1)-th scale.

[0054] Preferably, the input data for determining the multi-domain graph migration bearing life prediction model in step 41 is as follows:

[0055] The input data is multi-feature map data, as detailed below:

[0056] G = (V, F);

[0057] In the formula: G represents the multi-feature graph data; V represents the set of vertices in the multi-feature constructed graph; E represents the set of edges in the multi-feature constructed graph;

[0058] After preprocessing and constructing multi-scale graphs based on similarity, the data is divided according to the requirements of transfer learning, specifically based on operating conditions and fault types.

[0059] Preferably, in step 42, a Graphormer feature extractor is constructed to extract the input data features of the multi-domain graph migration bearing life prediction model, specifically as follows:

[0060] Step 421: Graphormer can calculate the importance of nodes in a multi-scale graph using the PageRank algorithm and add the node importance to the node features, as follows:

[0061]

[0062] In the formula: P(u) represents the influence of node u; v represents the neighboring nodes of node u; P(v) is the influence of node v; L(v) represents the number of outgoing chains of node v;

[0063] Step 422: Supplement the spatial topological information of the multi-scale feature map, and introduce topological information encoding s∈R when Graphormer calculates similarity. N*N The fusion of points i and j is as follows:

[0064]

[0065] In the formula: A ij h represents the similarity between node i and node j; i and h j W represents the features of the i-th and j-th nodes, respectively; Q W represents the first learned weight parameter; K This represents the second learning weight parameter; d represents the topological information of nodes i and j; d represents vector h. j W k dimensionality;

[0066] Step 423: Integrate the edge information from the multi-scale feature map, and set an edge encoding c∈R. N*N The matrix is ​​as follows:

[0067]

[0068] In the formula: c ij Encoding of edge information; The characteristics representing the nonlinear transformation of the shortest path edge; Indicates the learning weights; T indicates the transpose operation.

[0069] Step 424: By fusing the spatial topological information of points and edge information in the multi-scale feature map, the extracted features are more complete. Furthermore, by utilizing the attention mechanism in Graphermer, global information in the graph can be extracted effectively, thus improving the shortcomings of traditional graph convolutional neural networks and transformers.

[0070] Preferably, in step 43, the optimal target domain training data is selected based on multiple domain difference distances, and the domain difference distances are weighted to determine the optimal source domain. Specifically:

[0071] To measure the feature differences between different source and target domains, a similarity measure D is used. cos Maximum difference measure MMD and related alignment measure D CORAL To implement adaptive weighted metric and determine the optimal source domain, the following steps are taken:

[0072] The similarity measure D cos The method to obtain it is as follows:

[0073]

[0074] In the formula: D cos μ represents the cosine similarity measure. s μ represents the source domain eigenvector; t ||μ represents the feature vector of the target domain; s ||2 represents the 2-norm of the source domain vector; ||μ t ||2 represents the 2-norm of the target domain vector;

[0075] The method for obtaining the maximum difference measure (MMD) is as follows:

[0076]

[0077] Where: MMD 2 (X,Y) represents the square of the maximum difference measure between the source and target domains; φ(X i ) represents mapping domain X to the feature space; n1 represents the number of samples in domain X; φ(Y) represents the number of samples in domain X. j ) represents the mapping of domain Y to the feature space; n2 represents the number of samples in domain Y; X represents the source domain of the maximum difference measure; Y represents the target domain of the maximum difference measure;

[0078] The relevant alignment metric D CORAL The method to obtain it is as follows:

[0079]

[0080] In the formula: D CORAL (X s ,X t ) represents the relevant alignment metric; C s C represents the source domain covariance matrix; t Represents the source domain covariance matrix; The Frobenius norm of the matrix is ​​represented by ds; the feature dimension is represented by X. s Represents the source domain of the relevant alignment metric; Represents the target domain of the relevant alignment metric;

[0081] The adaptive weighting parameter W is used in bearing life prediction via multi-domain graph migration. i Optimize to obtain the metric between the source and target domains, as follows:

[0082] D m (X s ,X t )=∑ a W a D a (X s ,X t );

[0083] In the formula: D m (X s ,X t ) represents the final domain distribution difference; W a D represents the adaptive weight parameter for number a; a This represents the distance metric for index 'a'.

[0084] By measuring the distance between each source domain and the target domain, the source domain with the smallest difference is determined as the optimal source domain.

[0085] Preferably, in step 44, the selected labeled source domain training data and unlabeled target domain training data will undergo domain adversarial operations, specifically:

[0086] Step 441: Input the labeled source domain data into the multi-domain graph migration bearing life prediction model for prediction, and input the predicted values ​​and the true values ​​into the MSELoss function to measure the difference between them, thus obtaining the source domain loss. pre value;

[0087] Step 442: Treat the source domain data label as 0 and feed it into the domain discriminator for discrimination. The label and predicted label are then fed into the NLLLoss function for measurement to obtain the source domain loss. s value;

[0088] Step 443: Treat the target domain label as 1 and feed it into the domain discriminator for discrimination. Similarly, feed the label and the predicted label into the NLLLoss function for measurement to obtain the target domain loss.t value;

[0089] Step 444: Finally, add the loss from multi-source domain adaptation and these three loss values ​​with certain weights to get the loss of the training model and update the model parameters.

[0090] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0091] (1) This invention constructs a physical model, a data model, and integrates a multi-dimensional model to construct a digital twin model. By constructing a physical model, it is possible to gain a deep understanding of the operating principles and interaction relationships of physical systems. Through the data feedback mechanism of the digital twin model, the physical model is continuously optimized and improved.

[0092] (2) This invention constructs a multidimensional digital twin model and uses the model to generate a large amount of simulation data, which not only solves the problem of data shortage in the field of deep learning for remaining life prediction, but also enables the digital twin to monitor and predict with the actual system in real time, thereby improving the accuracy of the prediction model.

[0093] (3) This invention uses multiple features to construct a multi-scale topology map and embeds degradation information from multiple perspectives to help the prediction model extract degradation features more quickly and accurately.

[0094] (4) This invention adopts a domain-adaptive graph transfer model, which obtains global information in the topological graph through Graphormer, thus solving the drawbacks of traditional graph convolutional neural networks and transformers; and this model also greatly improves the cross-domain prediction accuracy.

[0095] (5) The present invention adopts multi-source domain adaptation, selects the most suitable source domain, and then conducts domain adversarial between the source domain and the target domain to achieve domain transfer; thereby improving the domain generalization ability of the transfer and the robustness of the model. Attached Figure Description

[0096] Figure 1 This is a control block diagram of the multi-domain graph migration bearing life prediction method driven by a digital twin model, as described in an embodiment of the present invention.

[0097] Figure 2 This is a flowchart of a multi-domain graph migration bearing life prediction method driven by a digital twin model, as described in an embodiment of the present invention.

[0098] Figure 3 The diagram below is a flowchart of the method for predicting the remaining life of a migrating bearing, as described in an embodiment of the present invention.

[0099] Figure 4 This is a diagram of the original vibration signal of the bearing in an embodiment of the present invention;

[0100] Figure 5This is a vibration signal diagram processed by RMS according to an embodiment of the present invention;

[0101] Figure 6 This is a flowchart of the Graphormer model according to an embodiment of the present invention. Detailed Implementation

[0102] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0103] This invention constructs a digital twin model by building a physical model, a data model, and fusing a multi-dimensional model. This allows for a deep understanding of the operating principles and interactions of the physical system. Furthermore, the data feedback mechanism of the digital twin model continuously optimizes and improves the physical model. The digital twin model generates a large amount of simulation data, solving the data shortage problem in the field of remaining lifetime prediction for deep learning, while simultaneously improving the accuracy of the prediction model. This invention uses Grapher to obtain global information in the topological graph, greatly improving cross-domain prediction accuracy. It employs multi-source domain adaptation, selecting a suitable source domain and then conducting domain adversarial exercises between the source and target domains to achieve domain transfer. This improves the domain generalization ability of the transfer and the robustness of the model. Figure 1 The diagram shown is a control block diagram of the bearing life prediction method based on a digital twin model driven by a multi-domain graph migration method according to an embodiment of the present invention.

[0104] This invention provides a method for predicting bearing life based on a digital twin model-driven multi-domain graph migration model, such as... Figure 2 The diagram shows a flowchart of a multi-domain graph migration bearing life prediction method driven by a digital twin model, according to an embodiment of the present invention. To demonstrate the applicability of the present invention, it is applied to an example, specifically including the following steps:

[0105] S1: Collect data from the bearing equipment and establish a digital twin model of the bearing;

[0106] Collect bearing operating data, including vibration, temperature, pressure, and speed; collect bearing parameters, including material properties, structural parameters, and operating condition parameters; perform dynamic stress analysis on the bearing during operation and record the stress state.

[0107] Based on the obtained operating data, parameters, and stress state, a physical model and a data model are established; the physical model and the data model are then integrated to establish a digital twin model of the bearing, predicting and analyzing the bearing's behavior and health status.

[0108] S2: Generate bearing life data through digital twin model and extract condition monitoring data during degradation stage;

[0109] The digital twin model constructed in S1 generates bearing life data under different operating conditions. Noise reduction and normalization are then applied to this data. The bearing life data is divided into two stages: the health stage and the degradation stage. Condition monitoring data typically exhibits relatively weak degradation characteristics in the early degradation stage. Because condition monitoring data from the early degradation stage can lead to fluctuations and errors in bearing life prediction, it is difficult to effectively utilize the bearing life prediction model. Therefore, the root mean square index (RMS) is used for detection. When the RMS meets certain conditions, condition monitoring data for the degradation stage can be obtained. The criteria are as follows:

[0110] rms>μ+w*σ;

[0111] In the formula: rms represents the root mean square index; μ represents the first parameter obtained from the vibration data of the initial health stage; σ represents the second parameter obtained from the vibration data of the initial health stage; w represents a random weight in the interval [3,5].

[0112] When the above inequality holds, it serves as the dividing line between the healthy stage and the degenerative stage, such as... Figure 4 The image shown is a diagram of the original vibration signal of the bearing in an embodiment of the present invention.

[0113] S3: Extract features from the state monitoring data of the degradation stage and construct multi-feature map data;

[0114] S31: Extract features from the state monitoring data of the degradation stage obtained in S2; such as... Figure 5 The image shown is a vibration signal diagram processed by RMS according to an embodiment of the present invention. First, time-domain features are extracted from the condition monitoring data. The root mean square index (rms) of the condition monitoring data is as follows:

[0115]

[0116] In the formula: x i This represents the vibration signal of the i-th node.

[0117] The root square amplitude CL of the condition monitoring data is as follows:

[0118]

[0119] In the formula: CL represents the square root magnitude; max represents the maximum value function; x represents the data generated by the digital twin.

[0120] The kurtosis (KU) of the condition monitoring data is as follows:

[0121]

[0122] In the formula: KU represents the kurtosis of the condition monitoring data; This represents the mean of the status monitoring data.

[0123] Next, frequency domain features are extracted from the condition monitoring data. First, the condition monitoring data is subjected to Fourier transform, as follows:

[0124]

[0125] In the formula: S(k) represents the frequency domain of the signal; x(kΔtz) represents the sampled value of the state monitoring data; N represents the number of sampling points; Δtz represents the sampling interval; k represents the index of the time-frequency discrete value; γ represents the complex number symbol.

[0126] Then, the centroid frequency (MSF) and mean square frequency (FC) are obtained using the transformed state monitoring data, as shown in the following formulas:

[0127]

[0128] In the formula: MSF represents the centroid frequency of the condition monitoring data; FC represents the mean square frequency of the condition monitoring data; k represents the number of spectral lines; f k (k) represents the frequency value of the kth spectral line.

[0129] Finally, wavelet packet decomposition is performed on the transformed state monitoring data, specifically:

[0130] Time-frequency domain feature extraction obtains component data of different levels and directions of the signal through 3dB wavelet packet decomposition. The original signal before decomposition is as follows:

[0131]

[0132] In the formula: S 0,0 S represents the original signal before decomposition. 1,0 and S 1,1 S represents the first and second wavelet basis coefficient decomposition signals of the first layer, respectively; 2,0 S 2,1 S 2,2 and S 2,3 These represent the first, second, third, and fourth wavelet basis coefficient decomposition signals of the second layer, respectively; S 3,0 S 3,1 S 3,2 S 3,3 S 3,4 S 3,5 S 3,6 and S 3,7 These represent the first, second, third, fourth, sixth, seventh, and eighth wavelet basis coefficient decomposition signals of the third layer, respectively.

[0133] S e,fThis represents the decomposed signal corresponding to the f-th node at the e-th scale. The specific decomposition formula is as follows:

[0134]

[0135] In the formula: g(k-2n) represents the wavelet basis function; S e,f (k) represents the decomposition signal of the f-th wavelet basis coefficient at position k in the e-th scale; S e+1,2f (n) represents the decomposition signal of the 2f-th wavelet basis coefficient at position k in the (e+1)-th scale.

[0136] The extracted features can comprehensively reflect the degradation information of the degraded data.

[0137] S32: Construct multi-feature map data; based on the similarity between nodes due to various features, determine the similarity distance between nodes and use cosine similarity as the metric, with the specific formula as follows:

[0138]

[0139] In the formula: sim represents cosine similarity; A represents the first vector between nodes; B represents the second vector between nodes; A i B represents the first vector between the i-th nodes; i Let represent the second vector between the i-th nodes; i represents the node number; n represents the total number of nodes.

[0140] The similarity distance is used to determine if two nodes are connected by an edge; if the distance is greater than 0, there is no connection. After preprocessing and constructing a multi-scale graph based on similarity, the data needs to be partitioned according to transfer learning requirements. This partitioning is based on whether the data falls under different operating conditions or different fault types. For example, operating condition one has a frequency of 19.05 Hz, and operating condition two has a frequency of 29.05 Hz. Data from different operating conditions is set as the source domain, while data from other operating conditions is set as the target domain. Another example is that some bearings only experience outer and inner ring faults, while others experience faults in the outer ring, inner ring, cage, and rolling elements. Similarly, different fault types are partitioned into source and target domains. This enables cross-operating condition and cross-fault type transfer.

[0141] S4: Construct a multi-domain graph migration bearing life prediction model; such as... Figure 3 The diagram shown is a flowchart of the method for predicting the remaining life of a migrating bearing according to an embodiment of the present invention, used to illustrate the life prediction process.

[0142] S41: Determine the input data for the multi-domain graph migration bearing life prediction model, specifically:

[0143] The input data consists of multiple feature maps, as detailed below:

[0144] G = (V, E);

[0145] In the formula: G represents the multi-feature graph data; V represents the set of vertices in the multi-feature construction graph; E represents the set of edges in the multi-feature construction graph.

[0146] After preprocessing and constructing multi-scale graphs based on similarity, the data is divided according to the requirements of transfer learning, specifically based on operating conditions and fault types.

[0147] S42: Construct a Graphormer feature extractor to extract features from the input data of the multi-domain graph migration bearing life prediction model, such as... Figure 6 The flowchart shown is of the Graphormer model according to an embodiment of the present invention, specifically as follows:

[0148] S421: Graphormer can calculate the importance of nodes in a multi-scale graph using the PageRank algorithm and incorporate the node importance into the node features, as follows:

[0149]

[0150] In the formula: P(u) represents the influence of node u; v represents the neighboring nodes of node u; P(v) is the influence of node v; L(v) represents the number of outgoing chains of node v.

[0151] S422: Supplement the spatial topological information of the multi-scale feature map by introducing topological information encoding s∈R when Graphormer calculates similarity. N*N The fusion of points i and j is as follows:

[0152]

[0153] In the formula: A ij h represents the similarity between node i and node j; i and h j W represents the features of the i-th and j-th nodes, respectively; Q W represents the first learned weight parameter; K This represents the second learning weight parameter; d represents the topological information of nodes i and j; d represents vector h. j W k The dimension of.

[0154] S423: Incorporate edge information from the multi-scale feature map and set an edge encoding c∈R. N*N The matrix is ​​as follows:

[0155]

[0156] In the formula: cij Encoding of edge information; The characteristics representing the nonlinear transformation of the shortest path edge; represents the learning weights; T represents the transpose operation.

[0157] S424: By fusing the spatial topological information of points and edge information in multi-scale feature maps, the extracted features are more comprehensive. Furthermore, by utilizing the attention mechanism in Grapher, global information in the graph can be extracted effectively, thus improving upon the shortcomings of traditional graph convolutional neural networks and transformers.

[0158] S43: Select the optimal source domain training data, and perform weighted processing on the distance differences between multiple domains to distinguish the source domain data with the smallest distribution difference from the target domain training data, thereby determining the optimal source domain. Specifically:

[0159] To measure the feature differences between different source and target domains, a similarity measure D is used. cos Maximum difference measure MMD and related alignment measure D CORAL To implement adaptive weighted metric and determine the optimal source domain, the following steps are taken:

[0160] Similarity metric D cos The method to obtain it is as follows:

[0161]

[0162] In the formula: D cos μ represents the cosine similarity measure. s μ represents the source domain eigenvector; t ||μ represents the feature vector of the target domain; s ||2 represents the 2-norm of the source domain vector; ||μ t ||2 represents the 2-norm of the target domain vector.

[0163] The method for obtaining the Maximum Difference Measure (MMD) is as follows:

[0164]

[0165] Where: MMD 2 (X,Y) represents the square of the maximum difference measure between the source and target domains; φ(X i ) represents mapping domain X to the feature space; n1 represents the number of samples in domain X; φ(Y) represents the number of samples in domain X. j ) represents the mapping of domain Y to the feature space; n2 represents the number of samples in domain Y; X represents the source domain of the maximum difference measure; Y represents the target domain of the maximum difference measure.

[0166] Relevant alignment metric D CORAL The method to obtain it is as follows:

[0167]

[0168] In the formula: D CORAL (X s ,X t ) represents the relevant alignment metric; C s C represents the source domain covariance matrix; t Represents the source domain covariance matrix; The Frobenius norm of the matrix is ​​represented by ds; the feature dimension is represented by X. s Represents the source domain of the relevant alignment metric; represents the target domain of the relevant alignment metric.

[0169] The adaptive weighting parameter W is used in bearing life prediction via multi-domain graph migration. i Optimize to obtain the metric between the source and target domains, as follows:

[0170] D m (X s ,X t )=∑ a W a D a (X s ,X t );

[0171] In the formula: D m (X s ,X t ) represents the final domain distribution difference; W a D represents the adaptive weight parameter for number a; a This represents the distance metric for number 'a'.

[0172] By measuring the distance between each source domain and the target domain, the source domain with the smallest difference is determined as the optimal source domain, which is used to distinguish the source domain data with the smallest distribution difference from the target domain training data.

[0173] S44: The selected labeled source domain training data and unlabeled target domain training data will be subjected to domain adversarial operations, specifically:

[0174] S441: The labeled source domain data is fed into the multi-domain graph migration bearing life prediction model for prediction, and the predicted and true values ​​are fed into the MSELoss function to measure the difference between them, thus obtaining the source domain loss. pre value.

[0175] S442: Treat the source domain data label as 0 and feed it into the domain discriminator for discrimination. The label and predicted label are then fed into the NLLLoss function for measurement to obtain the source domain loss. s value.

[0176] S443: Treat the target domain label as 1 and feed it into the domain discriminator for discrimination. Similarly, feed the label and the predicted label into the NLLLoss function for measurement to obtain the target domain loss. t value.

[0177] S444: Finally, the loss from multi-source domain adaptation and these three loss values ​​are added together with certain weights to obtain the loss of the training model and the model parameters are updated.

[0178] The target domain training data in the test set is fed into the multi-domain graph migration bearing life prediction model for prediction. The RMSE and SF are calculated to determine whether the cross-operating condition and cross-fault migration prediction function has been achieved. The specific calculation formula is as follows:

[0179]

[0180] In the formula: d i The value of RMSE represents the difference between the predicted and actual values ​​of the multi-domain graph migration bearing life prediction model between the i-th nodes; SF represents the model score index; N represents the number of graph data samples; RUL represents the difference between the predicted and actual values. pre This represents the predicted value of the multi-domain graph migration bearing life prediction model; RUL act This represents the true value of the training data in the target domain.

[0181] S5: Obtain accurate bearing life prediction values ​​by transferring the bearing life prediction model based on the multi-domain graph;

[0182] The bearing operation data in the actual production process is collected, and multi-feature map data is obtained based on S2 and S3. This data is then input into the multi-domain graph transfer bearing life prediction model trained in S4 to accurately predict the bearing life.

[0183] In summary, the prediction results of the bearing life prediction method based on the digital twin model driven by multi-domain graph migration in this case demonstrate its excellent performance.

[0184] (1) The embodiments of the present invention construct a digital twin model by building a physical model, a data model and integrating a multi-dimensional model. By constructing a physical model, the operating principle and interaction relationship of the physical system can be deeply understood. The physical model is continuously optimized and improved through the data feedback mechanism of the digital twin model. A large amount of simulation data is generated by using the constructed multi-dimensional digital twin model, which not only solves the problem of data shortage in the field of deep learning in the field of remaining lifetime prediction, but also enables the digital twin to monitor and predict the actual system in real time, thereby improving the accuracy of the prediction model.

[0185] (2) The embodiments of the present invention use multiple features to construct a multi-scale topology graph and embed degradation information from multiple perspectives to help the prediction model extract degradation features more quickly and accurately; adopting a domain adaptation graph transfer model, the global information in the topology graph is obtained through Graphermer, which solves the drawbacks of traditional graph convolutional neural networks and transformers; and this model also greatly improves the cross-domain prediction accuracy; adopting multi-source domain adaptation, the most suitable source domain is selected, and then the source domain and the target domain are subjected to domain adversarial, thereby realizing domain transfer and improving the domain generalization ability of the transfer and the robustness of the model.

[0186] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A bearing life prediction method based on a digital twin model-driven multi-domain graph migration model, characterized in that, It includes the following steps: Step 1: Collect data from the bearing equipment and establish a digital twin model of the bearing; Step 2: Generate bearing life-cycle data through a digital twin model and extract condition monitoring data during the degradation stage; Step 3: Extract features from the state monitoring data of the degradation stage and construct multi-feature map data; Step 4: Construct a multi-domain graph migration bearing life prediction model; Step 41: Determine the input data for the multi-domain graph migration bearing life prediction model; Step 42: Construct the Graphormer feature extractor to extract the input data features of the multi-domain graph migration bearing life prediction model; Step 43: Select the optimal source domain training data based on data features. Weight the distance differences between multiple domains to determine the optimal source domain, which is used to distinguish the source domain data with the smallest distribution difference from the target domain training data. Specifically: to measure the feature differences between different source domains and the target domain, a similarity metric is used. Maximum difference measure MMD and related alignment measure Implement adaptive weighted metric and determine the optimal source domain. Then, use a multi-domain graph transfer model to adjust the adaptive weight parameters. Optimize to obtain the metric between the source and target domains, as follows: ； In the formula: This indicates the final difference in domain distribution; Indicates the source domain of the relevant alignment metric; Indicates the target domain of the relevant alignment metric; Indicates number Adaptive weight parameters; Indicates number The distance metric; The distance between each source domain and the target domain is measured, and the source domain with the smallest difference is determined as the optimal source domain. Step 44: Perform domain adversarial operations on the selected labeled source domain training data and unlabeled target domain training data; feed the target domain training data in the test set into the multi-domain graph migration bearing life prediction model for prediction, and determine whether the cross-operating condition and cross-fault migration prediction function has been achieved by calculating RMSE and SF. The specific calculation formula is as follows: ； In the formula: This represents the difference between the predicted value and the actual value of the bearing life prediction model in the multi-domain graph migration between the i-th nodes; This represents the root mean square error between the predicted and actual values. This represents the model score metric; N represents the number of graph data samples. This represents the predicted value of the multi-domain graph migration bearing life prediction model; This represents the true value of the training data in the target domain; Step 5: Transfer the bearing life prediction model based on the multi-domain graph to obtain accurate bearing life prediction values; Collect bearing operation data from the actual production process, obtain multi-feature map data based on steps 2 and 3, and input it into the multi-domain graph transfer bearing life prediction model trained in step 4 to predict the bearing life.

2. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 1, characterized in that, Step 3 includes: Step 31: Extract features from the state monitoring data of the degradation stage obtained in Step 2; first, extract time-domain features from the state monitoring data; then, extract frequency-domain features from the state monitoring data; finally, perform wavelet packet decomposition on the transformed state monitoring data. The extracted features can comprehensively reflect the degradation information of the degradation data. Step 32: Construct multi-feature map data; based on the similarity between nodes due to multiple features, determine the similarity distance between each node, and use cosine similarity as the metric. The specific formula is as follows: ； In the formula: Indicates cosine similarity; This represents the first vector between nodes; This represents the second vector between nodes; This represents the first vector between the i-th nodes; This represents the second vector between the i-th nodes; i represents the node number; n represents the total number of nodes; The similarity distance is used to determine whether there is an edge connecting the two nodes. If the distance is greater than 0, there is an edge connecting the two nodes; otherwise, there is no edge connecting the two nodes.

3. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 2, characterized in that, The temporal feature extraction of the status monitoring data in step 31 is specifically as follows: Calculate the root mean square index (rms) of the status monitoring data; The root square amplitude CL of the state monitoring data is as follows: ； In the formula: Indicates the root magnitude; This represents a function that takes the maximum value. This refers to data generated by a digital twin; This represents the vibration signal of the i-th node; The kurtosis (KU) of the state monitoring data is as follows: ； In the formula: Indicates the kurtosis of the status monitoring data; This represents the mean of the status monitoring data.

4. The bearing life prediction method based on a digital twin model driven by multi-domain graph migration as described in claim 2, characterized in that, Step 31 involves extracting frequency domain features from the status monitoring data, specifically as follows: First, perform a Fourier transform on the status monitoring data, as follows: ； In the formula: Represents the frequency domain of the signal; This represents the sampled value of the status monitoring data; M represents the number of sampling points; Indicates the sampling interval; Indicates the sequence number of the discrete time-frequency value; Represents the complex number symbol; The center of gravity frequency is then obtained using the transformed state monitoring data. and mean square frequency .

5. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 2, characterized in that, In step 31, the transformed state monitoring data undergoes wavelet packet decomposition, specifically as follows: Time-frequency domain feature extraction obtains component data of different levels and directions of the signal through 3dB wavelet packet decomposition. The original signal before decomposition is as follows: ； In the formula: This represents the original signal before decomposition; and These represent the first and second wavelet basis coefficient decomposition signals of the first layer, respectively; , , and These represent the first, second, third, and fourth wavelet basis coefficient decomposition signals of the second layer, respectively; , , , , , , and These represent the first, second, third, fourth, fifth, sixth, and seventh wavelet basis coefficient decomposition signals of the third layer, respectively. This represents the decomposed signal corresponding to the f-th node at the e-th scale.

6. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 1, characterized in that, In step 41, the input data for determining the multi-domain graph migration bearing life prediction model is specifically as follows: The input data is multi-feature map data, as detailed below: ； In the formula: Represents multi-feature map data; Represents the set of vertices in a multi-feature constructed graph; Represents the set of edges in a multi-feature constructed graph; After preprocessing and constructing multi-scale graphs based on similarity, the data is divided according to the requirements of transfer learning, specifically based on operating conditions and fault types.

7. The bearing life prediction method based on a digital twin model driven by multi-domain graph migration as described in claim 6, characterized in that, In step 42, a Graphormer feature extractor is constructed to extract the input data features of the multi-domain graph migration bearing life prediction model, specifically as follows: Step 421: Graphormer can calculate the importance of nodes in a multi-scale graph using the PageRank algorithm and add the node importance to the node features, as follows: ； In the formula: Represents a node Its influence; Represents a node The adjacent nodes; It is a node Its influence; Representative node The number of outgoing chains; Step 422: Supplement the spatial topological information of the multi-scale feature maps and introduce topological information encoding when Graphormer calculates similarity. For point and points The fusion is as follows: ； In the formula: Represents a node and nodes Similarity; and Represent the features of the i-th and j-th nodes respectively; This represents the first learned weight parameter; This represents the second learning weight parameter; Represents a node and nodes Topological information; d represents a vector dimensionality; Step 423: Integrate the edge information from the multi-scale feature map and set an edge encoding. The matrix is ​​as follows: ； In the formula: Encoding of edge information; The characteristics representing the nonlinear transformation of the shortest path edge; Indicates learning weights; T indicates that a transpose operation is performed; Step 424: By fusing the spatial topological information of points and edge information in the multi-scale feature map, the extracted features are more complete. Furthermore, the attention mechanism in Graphermer can effectively extract global information in the graph, thus improving the shortcomings of traditional graph convolutional neural networks and transformers.

8. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 1, characterized in that, The similarity measurement in step 43 The method to obtain it is as follows: ； In the formula: Represents the cosine similarity measure; Represents the source domain eigenvectors; Represents the feature vector of the target domain; The second norm of a vector in the source domain is denoted by . The second norm of the target domain vector; The method for obtaining the maximum difference measure (MMD) is as follows: ； In the formula: This represents the square of the maximum difference measure between the source and target domains; Indicates the domain Mapped to feature space; Representation domain Sample size; Representation domain Mapped to feature space; Representation domain Number of samples; X represents the source domain of the maximum difference measure; Y represents the target domain of the maximum difference measure; The relevant alignment metric The method to obtain it is as follows: ； In the formula: Indicates the relevant alignment metric; Represents the source domain covariance matrix; Represents the target domain covariance matrix; Denotes the Frobenius norm of a matrix; Indicates the feature dimension; Indicates the source domain of the relevant alignment metric; This indicates the target domain for the relevant alignment metric.

9. The bearing life prediction method based on digital twin model-driven multi-domain graph migration according to claim 1, characterized in that, In step 44, the selected labeled source domain training data and unlabeled target domain training data will undergo domain adversarial operations, specifically: Step 441: Input the labeled source domain data into the multi-domain graph migration bearing life prediction model for prediction, and input the predicted values ​​and the true values ​​into the MSELoss function to measure the difference between them, thus obtaining the source domain data. value; Step 442: Treat the source domain data label as 0 and send it to the domain discriminator for discrimination. The original label and predicted label are then fed into the NLLLoss function for measurement to obtain the source domain's... value; Step 443: Treat the target domain label as 1 and feed it into the domain discriminator for discrimination. Similarly, feed the label and the predicted label into the NLLLoss function for measurement to obtain the target domain's... value; Step 444: Finally, add the loss from multi-source domain adaptation and these three loss values ​​with certain weights to get the loss of the training model and update the model parameters.

Images