Rotary machine remaining service life prediction method based on comparison decoupling graph neural network

Abstract

The invention relates to a method for predicting the remaining service life of a rotary machine based on a contrast decoupling graph neural network, and the method is characterized in that the method comprises the following steps: S1, collecting a detection signal of a multi-source sensor through a rotary machine test bench, and converting the detection signal into graph data; s2, decoupling the node features into global degradation features and local relation features by using a node feature comparison decoupling module, and constructing corresponding decoupling constraint terms; s3, calculating edges among the nodes by using the local relation features, and updating the global degradation features and the local relation features based on the edges; s4, recoupling the global degradation feature and the local relation feature of each node; s5, calculating the information consistency and the information propagation probability of the nodes, combining the information consistency and the information propagation probability to construct a node importance weight score, sorting the importance degrees of the nodes by using top-rank, and discarding the nodes with low rankings; and S6, predicting the residual service life of the equipment by combining the graph output by each layer of network.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction of the remaining service life of rotating machinery, and relates to a method for predicting the remaining service life of rotating machinery based on a comparative decoupling graph neural network. Background Art

[0002] Prognostics and health management (PHM) is an advanced technical framework that uses real-time monitoring of the operating status of mechanical systems to predict faults, diagnose, and make maintenance decisions. Its primary goals include improving system reliability, reducing lifecycle maintenance costs, and establishing proactive safety assurance mechanisms. As a core component of PHM, remaining useful life (RUL) prediction has become a key research area in reliability engineering. By analyzing multi-source sensor data to quantify the time interval from a system's current state to failure, RUL prediction provides a key basis for condition-based maintenance strategies.

[0003] With the rapid development of computer technology, deep learning-based RUL prediction methods have rapidly developed, thanks to their powerful feature extraction and nonlinear modeling capabilities. For example, Cheng et al. proposed a bearing RUL prediction method based on a dynamic interaction hybrid model, which effectively quantifies the uncertainty of the prediction results. Zhou et al. developed a novel dual-threaded gated recurrent unit model specifically for gear RUL prediction. Zhu et al. proposed an online RUL prediction framework based on a spatial attention temporal convolutional network. Although existing deep learning-based RUL prediction methods have made significant progress in temporal feature extraction and degradation modeling, they typically treat multi-source monitoring data as independent time series or simply stacked, failing to explicitly model spatial dependencies in sensor networks. In complex industrial machinery systems, multi-source monitoring data often exhibit heterogeneity and spatial distribution characteristics, and key spatial correlations, such as topological relationships between sensor nodes, remain underutilized. Traditional convolutional neural networks and recurrent neural networks struggle to effectively represent spatial interactions across nodes, limiting the accuracy and generalization performance of RUL prediction.

[0004] Multi-source sensor monitoring systems are abstracted as graph-structured data, where nodes represent sensors and edges define the functional relationships between sensors. This allows for an explicit mathematical representation of the topological relationships within the sensor network. As a powerful representational tool for graph-structured data modeling, graph neural networks (GNNs), deep learning models specialized for processing graph-structured data, can adaptively aggregate information from multi-hop neighborhood nodes to mine latent spatial correlations. Compared to traditional deep learning methods that rely on single-dimensional time series modeling, GNNs synergistically fuse spatial relationship information from sensor networks with data features, achieving complementary enhancement of multidimensional information and significantly improving the accuracy of RUL prediction under complex operating conditions. For example, Wang et al. proposed a RUL prediction method based on a comprehensive dynamic structured graph neural network. This method not only fully utilizes the health information implicit in condition monitoring data but also considers the structural characteristics of aircraft engines. Liu et al. developed a multi-layer graph fusion network with uncertainty quantification for RUL prediction. Liang et al. proposed an innovative end-to-end framework that uses a deep adaptive Transformer architecture enhanced by a graph attention network to achieve RUL prediction.

[0005] Despite the progress made, existing methods still have key limitations. First, traditional GNN-based RUL prediction methods usually use node features for life prediction and spatial dependency modeling at the same time, resulting in implicit conflicts in feature learning objectives. Specifically, spatial dependency modeling focuses on capturing the local spatial correlation between nodes, while RUL prediction requires node features to accurately characterize the degradation state of the equipment. When the two tasks share the same feature representation, the node features are forced to encode degradation information and topological relationships at the same time, which may lead to feature entanglement (for example, topological noise interferes with the degradation feature expression) under complex working conditions, ultimately impairing the model prediction performance. Secondly, existing methods lack a screening mechanism for non-critical nodes. Industrial monitoring systems usually contain redundant or weakly correlated sensor nodes, which will introduce noise interference in the global representation learning process and dilute key degradation features. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method for predicting the remaining useful life of rotating machinery based on contrastive decoupling graph neural network (CDGNN), which improves the accuracy of predicting the remaining useful life of rotating machinery by comparing and learning decoupling node features and adopting an adaptive mechanism to discard redundant nodes.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network comprises the following steps:

[0009] S1: Collect detection signals of multiple sensors through the rotating mechanical test bench and convert them into graph data;

[0010] S2: Utilize the node feature contrastive decoupling module (NFCDM) to decouple node features into global degenerate features and local relationship features, and construct corresponding decoupling constraints;

[0011] S3: Use local relationship features to calculate the edges between nodes, and update the global degradation features and local relationship features based on the edges;

[0012] S4: Recouple the global degradation features and local relationship features of each node;

[0013] S5: Calculate the information consistency and information propagation probability of the nodes, combine the two to construct the node importance weight score, use top-rank to sort the importance of the nodes, and discard the low-ranked nodes;

[0014] S6: The graph output by each layer of the network is combined to predict the remaining useful life of the equipment.

[0015] Furthermore, the node feature comparison and decoupling module adopts a cross-fusion-comparison decoupling strategy. The process of this module is as follows:

[0016] The feature of node i in layer l Input to two independent decoupling networks and The global degradation characteristics are obtained from and local relation feature r i l , as shown below:

[0017]

[0018] in, For RUL prediction, r i l Used for spatial dependency modeling;

[0019] Two additional constraints are added to strengthen feature separation:

[0020] ①Impose constraints on spatial dependency tasks:

[0021] First, let R l Represents the local relationship feature matrix, D lRepresents the global degradation characteristic matrix, R l Shuffle by row to get Then D l and Fusion to obtain Φ l , Φ l Line 1 of It is obtained from the following formula:

[0022]

[0023] in Represents element-by-element addition operation;

[0024] Then calculate and Φ l The similarity of other rows in , get vector

[0025]

[0026] in, Represents the similarity measurement network, which consists of three layers of fully connected networks. The number of neurons in the last fully connected layer is 1, and sigmoid is used as the activation function;

[0027] Also given:

[0028]

[0029] When the information separation is good, the following formula exists:

[0030]

[0031] The following decoupling constraints are constructed for the spatial dependency modeling task:

[0032]

[0033] ②Impose constraints on the RUL prediction task:

[0034] Given Represents the graph set of the lth layer during training, from which a graph is randomly selected There must be one So that the following formula holds:

[0035]

[0036] where y χ and y γ Respectively and Corresponding real RUL label

[0037] right and Decouple the node features separately and get Two decoupled characteristic matrices and as well as Two decoupled characteristic matrices and

[0038] right and Fusion As shown in the following formula:

[0039]

[0040] The decoupling constraint for the RUL prediction task is:

[0041]

[0042] Furthermore, the local relationship features in step S3 are used to measure the connection relationship between nodes. Based on the local relationship features of the current layer and the edges of the previous layer, the edges of the current layer are obtained by the following formula:

[0043]

[0044] in, Represents the edge between node i and node j in the lth layer; the updated edge will be used to simultaneously update the global degradation feature and the local relationship feature. The calculation process is as follows:

[0045]

[0046] in, represents the decoupled feature update network, which consists of two fully connected layers.

[0047] Furthermore, in step S4, the global degradation feature and the local relationship feature of each node are recoupled, and the process is as follows:

[0048]

[0049] in, Represents a coupled network, which also consists of two fully connected layers.

[0050] Furthermore, in step S5, the information consistency and information propagation probability of the nodes are calculated, and the two are combined to construct a node importance weight score. The importance of the nodes is ranked using top-rank, and low-ranked nodes are discarded. This is achieved by constructing a proposed entropy and propagation probability-guided graph pooling (EPGPool) layer. The implementation process is as follows:

[0051] First, based on Get the initial graph global representation g l , as shown below:

[0052]

[0053] set up and Respectively represent g l and The αth element in measures the information consistency of each node based on information entropy:

[0054]

[0055] The L2 norm is introduced to construct the following indicator to measure the consistency of node information to identify abnormal patterns:

[0056]

[0057] The smaller the value of , the greater the possibility of node abnormality;

[0058] The propagation probability of the node is used to calculate the auxiliary information of the node weight; let A l represents the adjacency matrix of the lth layer, Indicates A l The sum of the elements in row i, Indicates A l The sum of all elements in , the information propagation probability of node i is given by:

[0059]

[0060] The larger the value, the greater the influence of node i on other nodes.

[0061] based on and Get the weight of node i As shown in the following formula:

[0062]

[0063] Use the top-rank operation to set the global degradation features of the nodes that need to be discarded to 0, and obtain the updated global degradation feature matrix D l :

[0064]

[0065] M l =top-rank(Q l ,ω) (22)

[0066]

[0067] Where ω represents the number of nodes that need to be discarded, M l Represents the mask matrix;

[0068] Finally, we get the global representation of the reconstructed graph:

[0069]

[0070] Furthermore, the graph outputted by each layer of the network in step S6 is used to predict the remaining useful life of the equipment, as shown below:

[0071]

[0072] The loss function of the network is as follows:

[0073]

[0074] Among them, the mean square error loss is used to quantify the difference between the actual RUL and the predicted RUL, λ r and λ d is a hyperparameter.

[0075] The beneficial effects of the present invention are as follows: the present invention adopts the proposed node feature contrastive decoupling module (NFCDM) to decouple node features into global degradation features (for life prediction) and local relationship features (for spatial dependency modeling). By constructing task-specific contrast decoupling constraints, the cross-interference of task-irrelevant information is effectively suppressed, thereby enhancing the task relevance and information independence of the decoupled features. In addition, a proposed entropy and propagation probability-guided graph pooling (EPGPool) layer is proposed, which can establish a node importance evaluation index based on node information consistency and information propagation probability, realize adaptive selection of key nodes, and effectively suppress noise interference caused by redundant or weakly correlated nodes in the feature aggregation stage, further improving the network's degradation state characterization capability under complex working conditions.

[0076] (1) The present invention can realize the RUL prediction of rotating machinery.

[0077] (2) The method proposed in this invention has higher accuracy than existing classical and cutting-edge RUL prediction methods.

[0078] (3) The node feature contrast decoupling module proposed in this invention can effectively enhance the task relevance and information independence of the decoupled features, thereby improving the prediction accuracy.

[0079] (4) The graph pooling layer guided by entropy and information propagation probability proposed in this paper can effectively suppress the impact of redundant nodes on prediction accuracy.

[0080] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0082] Figure 1 The flowchart of the method for predicting the remaining useful life of rotating machinery based on comparative decoupling graph neural network is shown in FIG.

[0083] Figure 2 Decouple module structure for node comparison;

[0084] Figure 3 A graph pooling layer structure guided by entropy and information propagation probability;

[0085] Figure 4 Comparison results between actual RUL and predicted RUL on four subsets; (a) is the FD001 subset, (b) is the FD002 subset, (c) is the FD003 subset, and (d) is the FD004 subset;

[0086] Figure 5 The feature visualization results on the FD001 and FD003 datasets, where (a) is the result of Model1 and (b) is the result of CDGNN;

[0087] Figure 6 (a)-(c) in the figure are the prediction results of CDGNN on three tasks;

[0088] Figure 7 (a)-(c) in the figure are the feature visualization results of CDGNN on three tasks. DETAILED DESCRIPTION

[0089] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0090] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0091] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0092] Example 1:

[0093] like Figure 1As shown, the present invention provides a method for predicting the remaining service life of rotating machinery based on a comparative decoupling graph neural network, comprising the following steps:

[0094] S1: Collect detection signals of multi-source sensors through the rotating mechanical test bench and convert them into graph data;

[0095] S2: Node feature comparison and decoupling module is used to decouple node features into global degradation features and local relationship features, and corresponding decoupling constraints are constructed. The node feature comparison and decoupling module adopts a cross-fusion-contrast decoupling strategy to effectively suppress the cross-interference of task-irrelevant information, thereby enhancing the task relevance and information independence of the decoupled features. Its structure is as follows: Figure 2 As shown. The process of this module is as follows: Given Represents the features of node i in the lth layer, which is input into two independent decoupled networks and The global degradation characteristics are obtained from and local relation feature r i l , as shown below:

[0096]

[0097] in, Will be used for RUL prediction, r i l will be used to model spatial dependencies. Ideally, Should not contain information for modeling spatial dependencies, and r i l It does not contain information for RUL prediction. However, it is difficult to achieve such accurate feature separation by simply relying on two independent decoupling networks. Therefore, the present invention adds two additional constraints to enhance the decoupling effect. First, let R l Represents the local relationship feature matrix, D l Represents the global degradation characteristic matrix, R l Shuffle by row to get Then D l and Fusion to obtain Φ l , so Φ l Line 1 of It can be obtained by the following formula:

[0098]

[0099] in It represents the element-by-element addition operation. It can be seen that formula (3) and It was integrated, It also contains the information of node i for RUL prediction and node j for spatial dependency modeling. and Φ l The similarity of other rows in , we get the vector Right now:

[0100]

[0101] in, Represents the similarity measurement network, which consists of three layers of fully connected networks. The number of neurons in the last fully connected layer is 1, and sigmoid is used as the activation function.

[0102] Also given:

[0103]

[0104] The present invention believes that when the information separation is good, the following formula exists:

[0105]

[0106] Therefore, the following decoupling constraints can be constructed for the spatial dependency modeling task:

[0107]

[0108] The above process imposes constraints on the spatial dependency task. Next, constraints will be imposed on the RUL prediction task. Represents the graph set of the lth layer during training, from which a graph is randomly selected There must be one So that the following formula holds:

[0109]

[0110] where y χ and y γ Respectively and The corresponding real RUL label. and Decoupling node features separately, we can get Two decoupled characteristic matrices and as well as Two decoupled characteristic matrices and Next, and Fusion As shown in the following formula:

[0111]

[0112] As you can see, Fusion Information for modeling spatial dependencies and Information for RUL prediction. To this end, we hope that the network is based on The predicted RUL label should be as close as possible to Similarly, we can obtain the decoupling constraint term for the RUL prediction task:

[0113]

[0114] It can be seen that for f sim (·) and Equivalent to adding and r1 l The interference term on f p (·) Equivalent to adding Therefore, the proposed cross-fusion contrast decoupling strategy can not only improve the independence and task-specificity of decoupled features, but also enhance the anti-interference ability and robustness of the network.

[0115] S3: Use local relationship features to calculate the edges between nodes, and update the global degradation features and local relationship features based on the edges; local relationship features will be used to measure the connection relationship between nodes. Based on the local relationship features of the current layer and the edges of the previous layer, the edges of the current layer can be obtained by the following formula:

[0116]

[0117] in, Represents the edge between node i and node j in layer l. The updated edge will be used to update the global degradation feature and local relationship feature at the same time. The calculation process is as follows:

[0118]

[0119] in, represents the decoupled feature update network, which consists of two fully connected layers.

[0120] S4: Recouple the global degradation features and local relationship features of each node; the coupling process is as follows:

[0121]

[0122] in, Represents a coupled network, which also consists of two fully connected layers.

[0123] S5: Calculate the information consistency and information propagation probability of the nodes, combine the two to construct the node importance weight score, use top-rank to sort the importance of the nodes, and discard the nodes with low rankings. Specifically, in this step, the proposed entropy and propagation probability-guided graph pooling (EPGPool) layer is designed to effectively reduce the impact of redundant nodes on prediction accuracy. Its structure is as follows: Figure 3 The specific implementation process is as follows:

[0124] First, based on Get the initial graph global representation g l , as shown below:

[0125]

[0126] set up and Respectively represent g l and The present invention measures the information consistency of each node based on information entropy:

[0127]

[0128] when When the value is low, it indicates that the node may have significant group deviation in a specific feature dimension; When the value is high, although it does not reflect the deviation of local dimensions, there may be a global feature shift in multiple dimensions. In order to fully identify these two completely different abnormal patterns, this paper introduces the L2 norm and constructs the following indicator to measure the consistency of node information:

[0129]

[0130] The smaller the value of , the greater the possibility that the node is abnormal. Subsequently, the propagation probability of the node is used as auxiliary information to calculate the node weight. Let A l represents the adjacency matrix of the lth layer, Indicates A l The sum of the elements in row i, Indicates A l The sum of all elements in . The information propagation probability of node i is given by:

[0131]

[0132] The larger the value of , the greater the influence of node i on other nodes. and The weight of node i can be obtained As shown in the following formula:

[0133]

[0134] Then, the top-rank operation is used to set the global degradation features of the nodes that need to be discarded to 0, and the updated global degradation feature matrix D is obtained. l :

[0135]

[0136] M l =top-rank(Q l ,ω) (22)

[0137]

[0138] Where ω represents the number of nodes that need to be discarded, M l Represents the mask matrix. Finally, we can get the global representation of the reconstructed graph:

[0139]

[0140] S6: The graph output by each layer of the network is combined to predict the remaining useful life of the equipment.

[0141] The global representation of the reconstructed graphs of all layers will be used to obtain the final predicted RUL, as shown below:

[0142]

[0143] The loss function of the network is as follows:

[0144]

[0145] Among them, the mean square error loss is used to quantify the difference between the actual RUL and the predicted RUL, λ r and λ d is a hyperparameter.

[0146] Example 2:

[0147] To verify the effectiveness of the proposed method, this study was conducted on two datasets, including the CMASPSS dataset and the TWM dataset.

[0148] The CMAPSS dataset is a benchmark dataset widely used in the field of RUL prediction research. This dataset is generated by a high-fidelity turbofan engine simulation model and contains four subsets (FD001-FD004). Each subset simulates the monitoring data of 21 sensors under different fault modes and operating conditions. Considering that not all sensor recording data contain degradation-related information, the present invention selects 14 sensors (numbered: 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20 and 21) for RUL prediction. To facilitate subsequent prediction tasks, the maximum-minimum normalization method is used to standardize the sensor monitoring data. In the data preprocessing stage, a sliding time window with a step size of 1 and a window length of T is used to divide the monitoring data into samples, and the RUL value corresponding to the last time step of each sample is used as the label. In particular, for samples with less than T time points, the spline interpolation method is used to complete the data.

[0149] In order to quantitatively describe the performance of the proposed method, the root mean square error (RMSE) and scoring function (SF) are used as performance evaluation indicators, and the definitions of the two are as follows:

[0150]

[0151] To verify the performance of the proposed method, methods from nine representative papers were selected for comparative analysis. The comparative methods cover three typical architectures: BiGRU-TSAM and DA-LSTM based on recurrent neural networks (RNNs), STRUL based on convolutional neural networks (CNNs), and GAT-DAT, MST-GAT, HAGCN, ConvGAT, MGCAL-UQ, and CN-PR based on GNNs. By comparing with RNN and CNN methods, the significant advantages of GNN in the field of RUL prediction were verified; and the comparison with other GNN methods highlighted the advanced nature of the method proposed in this invention.

[0152] The experiment implemented the CDGNN model using the PyTorch machine learning framework. The Adam optimizer was used for parameter optimization (with a learning rate of 0.001), a batch size of 100, and 200 training epochs. Table 1 shows the RMSE and SF metrics of each method on the CMAPSS dataset. As can be seen, the proposed method significantly outperforms the other compared methods in the overall performance, fully demonstrating its superiority in the RUL prediction task. Figure 4(a)-(d) show the comparison curves of the predicted RUL and the actual RUL of CDGNN on the four CMAPSS sub-datasets. It can be seen that the predicted results are highly consistent with the actual values, which is mutually confirmed with the quantitative results in Table 1, fully demonstrating the feasibility of this method.

[0153] Table 1

[0154]

[0155] In addition, a series of ablation experiments were designed to evaluate the impact of each component in the CDGNN model on the overall performance. Specifically, four model variants were designed: (1) Model 1 removes both the CFCDM module and the EPGPool layer; (2) Model 2 removes the EPGPool layer and uses two independent neural networks to achieve node feature decoupling; (3) Model 3 retains only the CFCDM module; (4) Model 4 retains only the EPGPool layer. The other configurations of all models remain consistent to ensure fair comparison. The ablation experiment results are shown in Tables 2 and 3. The experimental results show that CDGNN outperforms the four ablation variants on all four sub-datasets. It is worth noting that Model 3 performs better than Model 2, which shows that CFCDM effectively enhances the feature decoupling effect by introducing two decoupling constraints. The comparison between Model 1 and Model 4 further confirms that the EPGPool layer improves the model RUL prediction performance by discarding redundant nodes. In summary, each component in the CDGNN model plays an irreplaceable role in improving the overall performance. To further demonstrate the superiority of the proposed CDGNN, Figure 5 Figures (a) and (b) show the feature visualization results of Model 1 and CDGNN on the FD001 and FD003 datasets. It can be observed that due to its node feature decoupling capability, CDGNN achieves better task-level feature optimization, resulting in a more significant degradation trend in the extracted features compared to Model 1, which does not have feature decoupling capabilities.

[0156] Table 2

[0157]

[0158] Table 3

[0159]

[0160] The Tool Wear Monitoring (TWM) dataset originated from the 2010 Tool Health Prediction Competition held by the PHM Association. This dataset records the full lifecycle signals of ball-end milling cutters on high-speed CNC machine tools using a variety of sensors, including cutting forces, vibration signals, and RMS values of acoustic emission signals in the X, Y, and Z directions. The dataset contains wear data from six repeated experiments and is therefore composed of six sub-datasets: C1, C2, C3, C4, C5, and C6. The labeled sub-datasets C1, C4, and C6 were selected for the experiment. The RMS values of the seven sensor measurement data were selected as the original features. A graph was constructed by extracting 30 RMS values from each sample, and the corresponding remaining cutting cycle was used as the RUL label.

[0161] This paper uses a three-fold cross-validation strategy on the TWM dataset to construct three prediction tasks. Specifically, two subsets are selected as training sets each time, and the remaining subset is used as the test set, as shown in Table 4. To verify the superiority of the proposed CDGNN, five comparison methods (STRUL, TCN-SA, HAGCN, MST-GAT, and GAT-DAT) are selected for comparative experiments, and the corresponding results are shown in Table 5. The experimental results show that CDGNN outperforms the other comparison methods in all three tasks. Figure 6 (a)-(c) in the figure show the prediction results of CDGNN on three tasks. It can be seen that the predicted RUL is highly consistent with the actual RUL. Figure 7 Figures (a)-(c) show the feature visualization results of CDGNN on three tasks. It can be observed that the features extracted by the model show a clear degradation trend. This proves that the proposed CDGNN has excellent performance in the remaining useful life prediction task.

[0162] Table 4

[0163]

[0164] Table 5

[0165]

[0166] In the above embodiments, references to "this embodiment" in the specification indicate that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple occurrences of "this embodiment" do not necessarily refer to the same embodiment.

[0167] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the present invention are intended to encompass all such alternatives, modifications, and variations that fall within the broad scope of the appended claims.

[0168] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, any one of the methods in this embodiment is implemented.

[0169] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0170] The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes any one of the methods in this embodiment.

[0171] Regarding the computer-readable storage medium in this embodiment, those skilled in the art will appreciate that all or part of the steps in the aforementioned method embodiments can be implemented using hardware associated with the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps in the aforementioned method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0172] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication with each other. The memory is used to store computer programs, the communication interface is used for communication, and the processor and the transceiver are used to run computer programs so that the electronic terminal executes the various steps of the above method.

[0173] In this embodiment, the memory may include a random access memory (RAM), and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0174] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0175] The present invention can be used in a wide variety of general-purpose or special-purpose computing system environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above.

[0176] The present invention may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network, characterized by: The following steps are involved: S1: Collect detection signals of multiple sensors through the rotating mechanical test bench and convert them into graph data; S2: Utilize the node feature contrast decoupling module NFCDM to decouple node features into global degradation features and local relationship features, and construct corresponding decoupling constraints; S3: Use local relationship features to calculate the edges between nodes, and update the global degradation features and local relationship features based on the edges; S4: Recouple the global degradation features and local relationship features of each node; S5: Calculate the information consistency and information propagation probability of the nodes, combine the two to construct the node importance weight score, use top-rank to sort the importance of the nodes, and discard the low-ranked nodes; S6: The graph output by each layer of the network is combined to predict the remaining useful life of the equipment.

2. The method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network according to claim 1, characterized in that: The node feature comparison and decoupling module adopts a cross-fusion-comparison decoupling strategy. The process of this module is as follows: The feature of node i in layer l Input to two independent decoupling networks and The global degradation characteristics are obtained from and local relationship features As shown in the following formula: in, For RUL prediction, r i l Used for spatial dependency modeling; Two additional constraints are added to strengthen feature separation: ①Impose constraints on spatial dependency tasks: First, let R l Represents the local relationship feature matrix, D l Represents the global degradation characteristic matrix, R l Shuffle by row to get Then D l and Fusion to obtain Φ l , Φ l Line 1 of It is obtained from the following formula: in Represents element-by-element addition operation; Then calculate and Φ l The similarity of other rows in , we get the vector in, Represents the similarity measurement network, which consists of three layers of fully connected networks. The number of neurons in the last fully connected layer is 1, and sigmoid is used as the activation function; Also given: When the information separation is good, the following formula exists: The following decoupling constraints are constructed for the spatial dependency modeling task: ②Impose constraints on the RUL prediction task: Given Represents the graph set of the lth layer during training, from which a graph is randomly selected There must be one So that the following formula holds: where y χ and y γ Respectively and Corresponding real RUL label right and Decouple the node features separately and get Two decoupled characteristic matrices and as well as Two decoupled characteristic matrices and right and Fusion As shown in the following formula: The decoupling constraint for the RUL prediction task is:

3. The method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network according to claim 1, characterized in that: The local relationship features in step S3 are used to measure the connection relationship between nodes. Based on the local relationship features of the current layer and the edges of the previous layer, the edges of the current layer are obtained by the following formula: in, Represents the edge between node i and node j in the lth layer; the updated edge will be used to simultaneously update the global degradation feature and the local relationship feature. The calculation process is as follows: in, represents the decoupled feature update network, which consists of two fully connected layers.

4. The method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network according to claim 1, characterized in that: In step S4, the global degradation features and local relationship features of each node are recoupled. The process is as follows: in, Represents a coupled network, which also consists of two fully connected layers.

5. The method for predicting the remaining useful life of rotating machinery based on comparative decoupling graph neural network according to claim 1, characterized in that: In step S5, the information consistency and information propagation probability of the nodes are calculated, and the two are combined to construct the node importance weight score. The importance of the nodes is sorted using top-rank, and low-ranked nodes are discarded. This is achieved by constructing a graph pooling EPGPool layer guided by entropy and information propagation probability. The implementation process is as follows: First, based on Get the initial graph global representation g l , as shown below: set up and Respectively represent g l and The αth element in measures the information consistency of each node based on information entropy: The L2 norm is introduced to construct the following indicator to measure the consistency of node information to identify abnormal patterns: The smaller the value of , the greater the possibility of node abnormality; The propagation probability of the node is used to calculate the auxiliary information of the node weight; let A l represents the adjacency matrix of the lth layer, Indicates A l The sum of the elements in row i, Indicates A l The sum of all elements in , the information propagation probability of node i is given by: The larger the value, the greater the influence of node i on other nodes. based on and Get the weight of node i As shown in the following formula: Use the top-rank operation to set the global degradation features of the nodes that need to be discarded to 0, and obtain the updated global degradation feature matrix D l : M l =top-rank(Q l ,ω) (22) Where ω represents the number of nodes that need to be discarded, M l Represents the mask matrix; Finally, we get the global representation of the reconstructed graph:

6. The method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network according to claim 1, characterized in that: The graph output from each layer of the network described in step S6 represents the remaining useful life of the predicted equipment, as shown below: The loss function of the network is as follows: Among them, the mean square error loss is used to quantify the difference between the actual RUL and the predicted RUL, λ r and λ d is a hyperparameter.

7. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network as described in any one of claims 1 to 6 when executing the computer program.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network as described in any one of claims 1 to 6 is implemented.

9. A computer program product, characterized in that: It includes a computer program, which, when executed by a processor, implements the method for predicting the remaining useful life of rotating machinery based on a comparative decoupling graph neural network as described in any one of claims 1 to 6.

Images