Training method of prediction model, prediction method of health state, device and medium

Abstract

This application provides a training method for a prediction model, a method for predicting health status, an apparatus, and a medium, relating to the field of artificial intelligence technology. The method includes: acquiring time-series monitoring signal data of a workpiece at different operating stages; independently processing the time-series monitoring signal data of different operating stages using different data processing branches of the prediction model to obtain target features for different operating stages; using a fusion prediction layer of the prediction model to perform fusion prediction processing on the target features of different operating stages to obtain predicted health status information of the workpiece; and training the prediction model based on the predicted health status information and actual health status information to obtain the trained prediction model. This application utilizes different data processing branches of the prediction model to independently process the time-series monitoring signal data of different operating stages, enabling targeted identification of the evolution patterns of monitoring signals at each operating stage, thereby improving the prediction accuracy of the workpiece's health status information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for training a prediction model, a method for predicting health status, an apparatus, and a medium. Background Technology

[0002] As the service life and operating load of workpieces continue to increase, their health status deteriorates, easily leading to operational failures. Therefore, it is urgent to predict the health status of workpieces. Taking key industrial workpieces such as rolling bearings, gears, and spindles as examples, their failure mechanisms are diverse and complex, and the degradation process exhibits nonlinear and non-stationary evolution characteristics. The degradation characteristics of monitoring signals such as vibration and temperature evolve dynamically over time, and the characteristic evolution patterns of each operating stage throughout the entire lifespan differ significantly.

[0003] In related prediction technologies, unified modeling based on monitoring signals is usually difficult to accurately capture the degradation patterns at different operating stages, which ultimately leads to insufficient prediction accuracy of the prediction model. Summary of the Invention

[0004] The purpose of this application is to provide a method for training a prediction model, a method for predicting health status, an apparatus, a device, and a medium to solve the above-mentioned technical problems.

[0005] Firstly, a method for training a prediction model is provided, including:

[0006] Acquire timing data of monitoring signals from the workpiece at different stages of operation;

[0007] By utilizing different data processing branches of the prediction model, the time series data of the monitoring signals in the different operating stages are processed independently to obtain the target features of the different operating stages;

[0008] By using the fusion prediction layer of the prediction model, the target features of different operating stages are fused and predicted to obtain the health status prediction information of the workpiece.

[0009] The prediction model is trained based on the predicted health status information and the actual health status information to obtain the trained prediction model.

[0010] In some embodiments, the monitoring signal timing data includes multiple time windows, and each time window includes multiple monitoring signals;

[0011] The step of independently processing the time-series monitoring signal data of different operating stages to obtain the target features of different operating stages includes:

[0012] Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed.

[0013] The target features are determined based on the structure of each first graph in the aforementioned operational phase.

[0014] In some embodiments, determining the target feature based on each first graph structure of the operational phase includes:

[0015] Graph convolution processing is performed on each first graph structure of the running phase to obtain first time window features that correspond one-to-one with the time windows;

[0016] Based on the similarity of adjacent first time window features during the operation phase, a second adjacency matrix and a second graph structure corresponding to the second adjacency matrix are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0;

[0017] The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows;

[0018] The target features are determined based on the second time window features of the aforementioned operational phase.

[0019] In some embodiments, determining the target feature based on the second time window features of the running phase includes:

[0020] Based on the similarity between the features of each second time window in the operation phase, the third adjacency matrix and the third graph structure corresponding to the third adjacency matrix are determined.

[0021] The target features are obtained by performing graph convolution processing on the third graph structure.

[0022] In some embodiments, the step of fusing and predicting the target features of the different operating stages to obtain the health status prediction information of the workpiece includes:

[0023] At least a portion of the target time window features of the first running stage are fused into the target time window features of the second running stage to obtain the fused features of the second running stage; wherein, the target time window features include the first time window features or the second time window features, and the first running stage and the second running stage are adjacent;

[0024] Based on the fusion features and target features of the second running stage, the attention coefficient corresponding to the fusion features is calculated;

[0025] Based on the attention coefficient, the fusion feature and the target feature are weighted and fused to obtain the weighted fusion feature corresponding to the second running stage;

[0026] The weighted fusion features of each of the aforementioned operational stages are spliced ​​together and predicted to obtain the health status prediction information.

[0027] In some embodiments, fusing at least a portion of the target time window features of the first running phase into the target time window features of the second running phase to obtain the fused features of the second running phase includes:

[0028] The similarity between the target time window features of the second running stage and the target time window features of the first running stage is determined to obtain multiple similarity sequences, and the similarity sequences correspond one-to-one with the target time window features of the second running stage.

[0029] From each of the similarity sequences, at least one similarity that satisfies a preset high similarity condition is selected to obtain multiple similarity subsequences;

[0030] Based on the multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage;

[0031] Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

[0032] In some embodiments, the operating phase includes a normal operating phase, a slow degradation phase, and an accelerated degradation phase;

[0033] The step of weightedly fusing the fused features and the target features based on the attention coefficient to obtain the weighted fused features corresponding to the second running stage includes:

[0034] When the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature, and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage.

[0035] Wherein, the first fusion feature corresponds to the first attention coefficient, the second fusion feature corresponds to the second attention coefficient, the first fusion feature is obtained by fusing time-series data of monitoring signals during normal operation, and the second fusion feature is obtained by fusing time-series data of monitoring signals during accelerated degradation.

[0036] Secondly, a method for predicting health status is also provided, the method comprising:

[0037] Acquire timing data of the target monitoring signal of the workpiece;

[0038] The target monitoring signal time series data is copied multiple times and input into multiple data processing branches of the trained prediction model in parallel to obtain the health status information of the workpiece; wherein, the trained prediction model is trained by the method described in the first aspect.

[0039] Thirdly, a training device for a prediction model is also provided, comprising:

[0040] The acquisition module is used to acquire the timing data of monitoring signals of the workpiece at different operating stages;

[0041] The data processing module is used to independently process the time series data of the monitoring signals at different operating stages using different data processing branches of the prediction model, so as to obtain the target features of the different operating stages.

[0042] The fusion prediction module is used to perform fusion prediction processing on the target features of different operating stages using the fusion prediction layer of the prediction model to obtain the health status prediction information of the workpiece.

[0043] The training module is used to train the prediction model based on the health status prediction information and the actual health status information to obtain the trained prediction model.

[0044] In some embodiments, the monitoring signal time-series data includes multiple time windows, and each time window includes multiple monitoring signals; the data processing module is specifically used for:

[0045] Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed.

[0046] The target features are determined based on the structure of each first graph in the aforementioned operational phase.

[0047] In some embodiments, the data processing module is further configured to:

[0048] Graph convolution processing is performed on each first graph structure of the running phase to obtain first time window features that correspond one-to-one with the time windows;

[0049] Based on the similarity of adjacent first time window features during the operation phase, a second adjacency matrix and a second graph structure corresponding to the second adjacency matrix are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0;

[0050] The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows;

[0051] The target features are determined based on the second time window features of the aforementioned operational phase.

[0052] In some embodiments, the data processing module is further configured to:

[0053] Based on the similarity between the features of each second time window in the operation phase, the third adjacency matrix and the third graph structure corresponding to the third adjacency matrix are determined.

[0054] The target features are obtained by performing graph convolution processing on the third graph structure.

[0055] In some embodiments, the fusion prediction module is specifically used for:

[0056] At least a portion of the target time window features of the first running stage are fused into the target time window features of the second running stage to obtain the fused features of the second running stage; wherein, the target time window features include the first time window features or the second time window features, and the first running stage and the second running stage are adjacent;

[0057] Based on the fusion features and target features of the second running stage, the attention coefficient corresponding to the fusion features is calculated;

[0058] Based on the attention coefficient, the fusion feature and the target feature are weighted and fused to obtain the weighted fusion feature corresponding to the second running stage;

[0059] The weighted fusion features of each of the aforementioned operational stages are spliced ​​together and predicted to obtain the health status prediction information.

[0060] In some embodiments, the fusion prediction module is further configured to:

[0061] The similarity between the target time window features of the second running stage and the target time window features of the first running stage is determined to obtain multiple similarity sequences, and the similarity sequences correspond one-to-one with the target time window features of the second running stage.

[0062] From each of the similarity sequences, at least one similarity that satisfies a preset high similarity condition is selected to obtain multiple similarity subsequences;

[0063] Based on the multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage;

[0064] Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

[0065] In some embodiments, the operation phase includes a normal operation phase, a slow degradation phase, and an accelerated degradation phase; the fusion prediction module is specifically used for:

[0066] When the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature, and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage.

[0067] Wherein, the first fusion feature corresponds to the first attention coefficient, the second fusion feature corresponds to the second attention coefficient, the first fusion feature is obtained by fusing time-series data of monitoring signals during normal operation, and the second fusion feature is obtained by fusing time-series data of monitoring signals during accelerated degradation.

[0068] Fourthly, a health status prediction device is also provided, the device comprising:

[0069] The acquisition module is used to acquire the timing data of the target monitoring signal of the workpiece;

[0070] The input module is used to copy the time-series data of the target monitoring signal multiple times and input them in parallel into multiple data processing branches of the trained prediction model to obtain the health status information of the workpiece; wherein, the trained prediction model is trained by the method described in the first aspect.

[0071] Fifthly, an electronic device is also provided, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the methods described above.

[0072] In a sixth aspect, a computer-readable storage medium is also provided, the computer-readable storage medium storing a computer program that, when executed by at least one processor, implements the method as described above.

[0073] In this embodiment of the application, different data processing branches of the prediction model are used to process the time series data of monitoring signals at different operating stages independently. This enables separate modeling of the monitoring signals at each operating stage, allowing for targeted identification of the evolution patterns of the monitoring signals at each operating stage and fully exploring the differences in monitoring signals at different operating stages. This, in turn, helps to improve the prediction accuracy of the health status information of the workpiece. Attached Figure Description

[0074] Figure 1 A flowchart illustrating the training method for the prediction model provided in this application embodiment;

[0075] Figure 2This application provides a health status degradation trend diagram for its embodiments.

[0076] Figure 3 A schematic diagram of a fully connected temporal edge mechanism across time windows provided in an embodiment of this application;

[0077] Figure 4 A schematic diagram of the attention-guided multi-level aggregation mechanism provided in the embodiments of this application;

[0078] Figure 5 A flowchart illustrating the health status prediction method provided in this application embodiment;

[0079] Figure 6 A comparison chart of model loss provided for embodiments of this application;

[0080] Figure 7 The ablation experiment results are shown in the embodiments of this application.

[0081] Figure 8 This is a structural diagram of the health status monitoring system for a dual-arm collaborative handling robot provided in an embodiment of this application;

[0082] Figure 9 A structural diagram of the training device for the prediction model provided in the embodiments of this application;

[0083] Figure 10 A structural diagram of the health status prediction device provided in the embodiments of this application;

[0084] Figure 11 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0086] The terms "first" and "second" in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising" and any variations thereof are intended to cover non-exclusive protection. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. The term "multiple" in this application can mean at least two, for example, two, three, or more, and the embodiments of this application do not impose limitations.

[0087] This application provides a method for training a prediction model, a method for predicting health status, an apparatus, a device, and a medium for improving the accuracy of predicting the health status of a workpiece.

[0088] The embodiments of this application can be applied to various workpieces that exhibit different degradation patterns at different operating stages, such as rolling bearings, gears, and spindles.

[0089] Taking rolling bearings as an example, as a core functional component in motor systems, rolling bearings play a crucial role in various industrial robots, especially in high-precision and high-reliability applications such as dual-arm collaborative handling robots. Motors convert electrical energy into mechanical energy to drive the movement of robot joints and actuators. Rolling bearings, as key supporting components in this mechanical energy transmission chain, not only bear the weight of the motor rotor, ensuring stable rotor rotation, but also effectively reduce friction between the rotor and other components, suppressing vibration and noise during operation, thereby improving the smoothness of robot movement, positioning accuracy, and system reliability. It can be said that the health of rolling bearings directly determines the operational performance and safety level of the robot system. During the operation of dual-arm collaborative handling robots, rolling bearings are subjected to alternating loads, high-frequency start-stop cycles, and complex working conditions for extended periods, making them highly susceptible to performance degradation and failure. As an indispensable and vulnerable critical component in mechanical equipment, a failure in a rolling bearing often triggers a series of chain reactions, ranging from reduced robot operating accuracy and task efficiency to equipment damage, production stoppages, and even safety accidents. Therefore, the health of rolling bearings has become an important indicator for measuring the reliability and service life of robot systems. Due to the complex structure of robot systems and the significant coupling relationships between components, coupled with the variable operating environment and conditions, the failure mechanisms of rolling bearings exhibit significant diversity and complexity. Common failures include increased friction and abnormal temperature rise caused by poor lubrication, and material fatigue, raceway wear, bearing eccentricity, and structural damage resulting from long-term high-load operation. These factors work together to cause the degradation process of rolling bearings to exhibit nonlinear and non-stationary evolutionary characteristics, posing a significant challenge to health status prediction. As industrial robots continue to operate, the health status of rolling bearings gradually deteriorates, and the degradation characteristics contained in their vibration and other state signals also change over time, exhibiting degradation patterns at different operational stages. Based on this, the training method of the prediction model, the health status prediction method, the apparatus, equipment, and medium of the embodiments of this application can be applied to the health status prediction of rolling bearings to improve the accuracy of rolling bearing health status prediction.

[0090] The training method of the prediction model in the embodiments of this application will be described in detail below.

[0091] See Figure 1 , Figure 1 This is a flowchart of a training method for a prediction model provided in this application, such as... Figure 1 As shown, the method includes:

[0092] Step 101: Obtain the timing data of monitoring signals of the workpiece at different operating stages.

[0093] The monitoring signal can be a vibration signal or a temperature signal, etc., selected according to the characteristics of the workpiece.

[0094] Different operational stages correspond to the entire life cycle of a workpiece, and the stages can be divided based on the degradation patterns within the entire life cycle. Specifically, for example... Figure 2 As shown, the degradation pattern is generally divided into three stages: normal operation, slow degradation, and accelerated degradation. During normal operation, the workpiece monitoring signal characteristics show no significant fluctuations, exhibiting a stable and constant trend without obvious deterioration. During slow degradation, the monitoring signal characteristics gradually deviate from the baseline state, showing a slow shift and continuous small changes. During accelerated degradation, the rate of degradation of the monitoring signal characteristics accelerates significantly, showing a rapid deterioration and a sharp increase in the degree of deviation.

[0095] In this embodiment of the application, the full monitoring signal time series of the entire life cycle of the workpiece can be obtained, and then the full monitoring signal time series can be divided to obtain monitoring signal time series data of different operating stages.

[0096] For example, the time-domain characteristics, frequency-domain characteristics, or a combination thereof of the monitoring signal can be combined, and then, based on these characteristics, the monitoring signal can be divided into a process of change from normal operation, slow degradation, to accelerated degradation. Specifically, monitoring signals with relatively stable characteristics correspond to the normal operation stage; monitoring signals with persistent shifts or slow changes in characteristics correspond to the slow degradation stage; and monitoring signals with significantly rapid deterioration in characteristics correspond to the accelerated degradation stage. This division of operation stages can be achieved through methods such as empirical thresholds, degradation inflection point detection, statistical analysis, or offline annotation.

[0097] Step 102: Using different data processing branches of the prediction model, the time series data of monitoring signals at different operating stages are processed independently to obtain the target features of different operating stages.

[0098] Each data processing branch is used to process the time-series data of the monitoring signals for the corresponding operational phase. When the operational phase includes a normal operation phase, a slow degradation phase, and an accelerated degradation phase, the prediction model described above has three parallel data processing branches. The network structures of the different data processing branches can be the same or different.

[0099] By processing the time-series data of monitoring signals at different operating stages independently through different data processing branches, the evolution patterns of monitoring signals at different operating stages can be learned during the training phase. This avoids mutual interference between monitoring signals at different operating stages, thereby extracting the unique signal features of each operating stage more accurately and providing reliable feature support for subsequent workpiece health status prediction.

[0100] Step 103: Using the fusion prediction layer of the prediction model, the target features of different operating stages are fused and predicted to obtain the health status prediction information of the workpiece.

[0101] Among them, health status prediction information includes whether one is healthy, the degree of health, or remaining life expectancy.

[0102] By fusing the prediction layer, the target features of different operating stages can be fused first and then predicted. In this way, the target features of each operating stage can be comprehensively utilized, taking into account the stable monitoring signal features of the normal operation stage, the slow offset monitoring signal features of the slow degradation stage, and the rapid deterioration monitoring signal features of the accelerated degradation stage. This can more comprehensively reflect the overall health status of the workpiece and improve the rationality and reliability of the health status prediction information.

[0103] Step 104: Train the prediction model based on the predicted health status information and the actual health status information to obtain the trained prediction model.

[0104] As an example, common loss functions such as mean squared error loss function or cross-entropy loss function can be used, combined with algorithms such as gradient descent, to train the prediction model. By continuously adjusting the model parameters, the error between the predicted health status information and the actual health status information can be reduced, and the above-mentioned trained prediction model can be obtained.

[0105] In this embodiment of the application, different data processing branches of the prediction model are used to process the time series data of monitoring signals at different operating stages independently. This enables separate modeling of the monitoring signals at each operating stage, allowing for targeted identification of the evolution patterns of the monitoring signals at each operating stage and fully exploring the differences in monitoring signals at different operating stages. This, in turn, helps to improve the prediction accuracy of the health status information of the workpiece.

[0106] In some embodiments, the monitoring signal time-series data includes multiple time windows, each time window including multiple monitoring signals; step 102 above includes:

[0107] Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed.

[0108] Based on the structure of each first graph in the operational phase, the target features are determined.

[0109] In this embodiment, the complete monitoring signal time series can be divided to obtain monitoring signal time series for different operating stages. Then, using a preset time window size T and a step size (the step size is less than or equal to T), the monitoring signal time series for different operating stages are slid to obtain multiple time windows corresponding to different operating stages. The number of time windows corresponding to different operating stages may be the same or different.

[0110] The aforementioned similarity can be calculated based on Euclidean distance, Manhattan distance, or cosine similarity, where the similarity is the connection weight between adjacent monitoring signals. As an example, the Euclidean distance between each monitoring signal within a time window can be calculated using the Euclidean distance formula, thus obtaining an Euclidean distance matrix for each time window. This is represented as follows: where i represents the i-th running stage, t=1,2,3,…,n represents the t-th time window, and then the Euclidean distance matrix is ​​taken. The reciprocal of the first adjacency matrix of the time window graph structure is used as the first adjacency matrix. .

[0111] The first graph structure consists of a first adjacency matrix and corresponding monitoring signals. The monitoring signals correspond to nodes in the first graph structure, and the first adjacency matrix is ​​used to characterize the connection relationships and connection weights between nodes.

[0112] The first diagram structure corresponds one-to-one with the time window, and there are usually multiple first diagram structures for one running phase.

[0113] In this embodiment of the application, by constructing a first adjacency matrix and a corresponding first graph structure, the monitoring signals within the time window can be represented in the form of a graph structure, clearly presenting the correlation between each monitoring signal. Based on the first graph structure, the features of the monitoring signals in each operating stage can be extracted more accurately, so that the obtained target features can accurately reflect the evolution law of the monitoring signals.

[0114] In some embodiments, the target features are determined based on the structure of each first graph in the operational phase, including:

[0115] Graph convolution processing is performed on each first graph structure during the running phase to obtain the first time window features that correspond one-to-one with the time window;

[0116] Based on the similarity of adjacent first time window features during the operation phase, the second adjacency matrix and the corresponding second graph structure are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0;

[0117] The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows.

[0118] Target features are determined based on the characteristics of the second time window during the operational phase.

[0119] In this embodiment, after the data processing branch constructs the first graph structure, it can first use multiple graph convolutional networks (GCNs) to perform graph convolution processing on each first graph structure in the time series data of the monitoring signal to obtain the first time window features corresponding to each time window.

[0120] like Figure 3 As shown, Figure 3 An example is provided, illustrating a data processing branch that processes data to obtain target features. Figure 3 middle, Let represent the time window, where i represents the i-th running stage, and t=1,2,3,…,n represents the n-th time window. After constructing the first graph structure and performing graph convolution in the data processing branch, the corresponding first time window features are obtained. Among them, graph convolution processing is performed to obtain... The mathematical expression can be:

[0121] ;

[0122] Where i represents the i-th running stage, t=1,2,3,…,n, For each, there is a corresponding weight matrix. For the first adjacency matrix Perform symmetric normalization. For activation function, The first time window feature represents the output of the t-th time window of the i-th running stage after graph convolution.

[0123] The first time window features mentioned above include features of multiple nodes, with each node corresponding to a monitoring signal.

[0124] After obtaining the first time window features, in order to model the evolution relationship between adjacent time windows, this application embodiment introduces a cross-time window connection time edge mechanism. Based on the similarity of adjacent first time window features in the monitoring signal time series data, the second adjacency matrix and the second graph structure corresponding to the second adjacency matrix are determined. The (t-1)th first time window feature in the second adjacency matrix points to the tth first time window feature.

[0125] In some embodiments, the second adjacency matrix can be used to partially connect the (t-1)th first time window feature and the tth first time window feature. For example, connections can be established only between features whose similarity meets a preset condition.

[0126] In some embodiments, in order to more completely model the evolutionary relationship between adjacent time windows, the second adjacency matrix can be used to perform a full connection on the (t-1)th first time window feature and the tth first time window feature.

[0127] The following explains how to construct the second adjacency matrix for full connectivity.

[0128] In each running phase, for the (t-1)th first time window feature and the tth first time window feature, the full connection operation between all nodes uses... It indicates. Among them, This represents the j-th node in the first time window feature of the i-th running phase (t-1). This represents the j-th node in the t-th first time window feature of the i-th running stage.

[0129] To perform a full connection between the first time window features at time t-1 and the first time window features at time t, it is necessary to calculate the similarity between the features of the nodes and use the similarity as the connection weight.

[0130] As an example, the above similarity can be calculated based on Euclidean distance, as shown in the following expression:

[0131] , Represents the j-th node pair The Euclidean distance between the features.

[0132] After obtaining the Euclidean distance, the similarity can be obtained by taking the reciprocal of the Euclidean distance, as described in the above embodiment. Alternatively, it can be obtained as follows: Figure 3 As shown For similarity, where normalization represents normalization. The similarity is represented by the similarity score. The similarity score serves as the connection weight between the node representing the (t-1)th first-time-window feature and the node corresponding to the tth first-time-window feature.

[0133] In this case, for the last time window of each monitoring signal time series data, there may be a situation where the time window size is smaller than the preset time window size. This would result in the number of nodes in the last first time window feature being less than that in the previous first time window feature, leading to some nodes not corresponding. In this case, the connection weight of the uncorresponding nodes can be set to 0. For example, if no node in the last first time window feature corresponds to the j-th node of the previous first time window feature, the connection weight of the j-th node can be set to 0, indicating that there is no edge connection between the uncorresponding nodes.

[0134] After obtaining the second adjacency matrix composed of the above connection weights (which can be used) After (representation), a second graph structure consisting of the second adjacency matrix and the features of the first time window can be obtained.

[0135] like Figure 3 As shown, Figure 3 An exemplary schematic diagram of the second graph structure is provided. In the second graph structure, the node of the (t-1)th first time window feature is unidirectionally connected to the tth first time window feature, and the feature of the node is the feature of the corresponding monitoring signal in the first time window feature.

[0136] The second graph structure described above can be regarded as a temporally consistent graph. The second graph structure provides a structural prior in the temporal dimension in graph convolutional neural networks, which enables graph convolution to naturally take into account temporal evolution information during propagation, thereby enhancing the sensitivity and robustness of the prediction model to changes in degradation patterns.

[0137] Here is an example of the mathematical expression for graph convolution on the second graph structure:

[0138] ;

[0139] Where i represents the i-th running stage, t=1,2,3,…,n, For each, there is a corresponding weight matrix. This indicates that the second adjacency matrix is ​​being processed. Perform symmetric normalization. For activation function, The feature representing the t-th first time window feature in the i-th running stage, after being fully connected to all nodes and edges of the (t-1)-th first time window feature, is output by graph convolution. Graph convolution can adaptively pass the (t-1)-th first time window feature to the t-th time window feature. To control the proportion of features transferred from the previous time window, it can be set to 0.2.

[0140] The graph convolution mentioned above can be a graph convolution based on residual connections, used to convert the output of the previous graph convolution. The result is passed to the current graph convolution and residual fusion is performed with the current graph convolution result to avoid gradient vanishing or feature degradation problems during deep graph convolution, thereby extracting the monitoring signal features in the first graph structure more stably and ensuring the extraction quality and stability of target features.

[0141] In this implementation, by constructing a second graph structure and performing graph convolution on the second graph structure, historical time window features are fused into the current first time window features. This effectively utilizes the contextual information of the historical window, weakens the dependence on specific time steps, improves the model's fault tolerance and generalization ability, and enhances the current time window's ability to model degradation trends.

[0142] In some embodiments, to further explore potential correlations between discontinuous time periods, target features are determined based on the characteristics of a second time window during the operational phase, including:

[0143] Based on the similarity between the features of each second time window during the operation phase, the third adjacency matrix and the corresponding third graph structure are determined.

[0144] The target features are obtained by performing graph convolution on the third graph structure.

[0145] The structure of the third diagram above corresponds one-to-one with the operation phase, and is used to characterize the relationship between the features of each second time window in the operation phase.

[0146] As an example, the similarity calculation method described in the above embodiment can be referred to, or the similarity between features in each second time window can be obtained by calculating the cosine similarity, thereby generating a similarity matrix. This similarity matrix is ​​the third adjacency matrix, which can be used... This indicates that i represents the i-th operational stage.

[0147] After obtaining the third adjacency matrix, the corresponding third graph structure is generated, which can be represented as follows: Graph convolutional networks can be used to perform graph convolution processing on each third graph structure in each running stage to achieve feature aggregation and obtain the target features. The mathematical expression for the graph convolution is as follows:

[0148] ;

[0149] Where i represents the i-th running stage, Represents the third adjacency matrix Perform symmetric normalization. This is the corresponding weight matrix. For activation function, The feature output after graph convolution of the third graph structure representing the i-th running stage is the target feature.

[0150] In this implementation, by constructing the aforementioned third graph structure, convolution can be performed in the space of the time window, enhancing local perception capabilities and cross-window relationship modeling, and effectively mining the spatial feature relationships between time windows in each running stage.

[0151] After obtaining the target features mentioned above, the fusion prediction layer can obtain health status prediction information by using an attention-guided multi-level feature aggregation mechanism.

[0152] The following section provides a detailed explanation of the fusion prediction process performed by the fusion prediction layer.

[0153] In some embodiments, target features from different operating stages are fused and predicted to obtain workpiece health status prediction information, including:

[0154] At least a portion of the target time window features of the first operation phase are fused into the target time window features of the second operation phase to obtain the fused features of the second operation phase; wherein, the target time window features include the first time window features or the second time window features, and the first operation phase and the second operation phase are adjacent;

[0155] Based on the fusion features and target features of the second running stage, the attention coefficients corresponding to the fusion features are calculated;

[0156] The fusion features and target features are weighted and fused based on the attention coefficient to obtain the weighted fusion features corresponding to the second running stage.

[0157] The weighted fusion features of each operational stage are spliced ​​and predicted to obtain health status prediction information.

[0158] In this embodiment, the first running stage and the second running stage are adjacent. If the second running stage is an intermediate running stage, there are two adjacent first running stages. If the second running stage is an initial running stage or an end running stage, there is one adjacent first running stage.

[0159] If there are two adjacent first operating phases in the second operating phase, then at least a portion of the target time window features of the two first operating phases need to be fused into the second operating phase.

[0160] The target time window features include either the first time window features or the second time window features mentioned above. The second time window features integrate historical time window features, and have a stronger ability to represent temporal evolution and aggregate global information. Therefore, the second time window features can be selected as the target time window features.

[0161] Based on the fusion feature and the target feature, an attention mechanism can be used to calculate the attention coefficient of the fusion feature. The fusion feature and the target feature are then weighted and fused based on the attention coefficient to obtain the weighted fusion feature. Here, the attention coefficient represents the weight of the fusion feature, and the target feature has a weight of 1.

[0162] After calculating the weighted fusion features of each operational stage, they can be concatenated and then input into the fully connected layer for prediction to obtain the aforementioned health status prediction information.

[0163] In this implementation, by splicing and predicting through feature fusion of target time windows across operating stages, attention mechanism weighted fusion, and multi-stage weighted fusion features, a collaborative modeling mechanism can be established between each operating stage. This allows for better capture of complex behavioral patterns such as anomalies, mutations, and non-periodicities during workpiece operation, thereby improving the accuracy and reliability of health status prediction information.

[0164] In some embodiments, at least a portion of the target time window features of the first running phase are fused to the target time window features of the second running phase to obtain the fused features of the second running phase, including:

[0165] The similarity between the target time window features of the second operation phase and the target time window features of the first operation phase is determined, resulting in multiple similarity sequences. Each similarity sequence corresponds one-to-one with the target time window features of the second operation phase.

[0166] From each similarity sequence, at least one similarity that meets the preset high similarity condition is selected to obtain multiple similarity subsequences;

[0167] Based on multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage.

[0168] Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

[0169] In this implementation, a preset high similarity condition is met, such as being greater than a preset similarity threshold, or being ranked in the top k of the similarity sequence. The similarity sequence is sorted from high to low similarity.

[0170] From each similarity sequence, at least one similarity that meets the preset high similarity condition is selected to obtain multiple similarity subsequences. A similarity subsequence refers to a partial set of similarities that meets the preset high similarity condition and is selected from the corresponding similarity sequence for a single target time window feature in the second running stage. It is used to characterize the relationship between the target time window feature and the most relevant target time window feature in the first running stage.

[0171] Based on multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed. That is, for each target time window feature in the second running stage, the connection relationship corresponding to its similarity subsequence is retained, the connection weight is the similarity, and the weights of the other connections are all set to 0.

[0172] In this implementation, through the above steps, representative target time window features in the first operation phase can be efficiently fused into target time window features in the second operation phase, realizing feature association and information aggregation across operation phases, providing more representative fusion features for subsequent weighted fusion and health status prediction processing; at the same time, by pre-setting high similarity conditions to filter key association information, representative fusion features can be highlighted, effectively avoiding interference from redundant features, and improving the pertinence and effectiveness of feature fusion.

[0173] In some embodiments, the operation phase includes a normal operation phase, a slow degradation phase, and an accelerated degradation phase;

[0174] The fusion features and target features are weighted and fused based on the attention coefficient to obtain the weighted fusion features corresponding to the second running stage, including:

[0175] In the case that the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage;

[0176] The first fusion feature corresponds to the first attention coefficient, and the second fusion feature corresponds to the second attention coefficient. The first fusion feature is obtained by fusing the time-series data of the monitoring signals during the normal operation phase, and the second fusion feature is obtained by fusing the time-series data of the monitoring signals during the accelerated degradation phase.

[0177] In this embodiment, when the second operating phase is the intermediate slow degradation phase, its adjacent operating phases include the normal operating phase and the accelerated degradation phase. Therefore, it is necessary to calculate the first fusion feature and the first attention coefficient of the normal operating phase fused to the slow degradation phase, as well as the second fusion feature and the second attention coefficient of the accelerated degradation phase fused to the slow degradation phase.

[0178] When the operation phase includes the three operation phases mentioned above, the following is an example of the mathematical expression for the prediction after concatenating the weighted fusion features:

[0179] ;

[0180] in, This represents the final health status prediction information obtained, such as remaining life expectancy; This indicates the target characteristics of the normal operation phase. This represents the fused features of the normal operation phase after incorporating the target time window features of the slow degradation phase. This corresponds to the attention coefficient; The target characteristics represent the slow degradation stage. This represents the second fused feature of the slow degradation stage, which incorporates the target time window features of the accelerated degradation stage. For the corresponding attention system, This represents the first fused feature of the slow degradation phase, after incorporating the target time window features of the normal operation phase. This corresponds to the attention coefficient; This indicates the target characteristics of the accelerated degradation stage. This represents the fused features of the accelerated degradation stage after incorporating the target time window features of the slow degradation stage. This represents the attention coefficient.

[0181] In this embodiment, when the second operating stage is the intermediate slow degradation stage, the target time window features of the normal operating stage and the accelerated degradation stage are integrated to achieve feature information transfer and collaborative modeling across degradation stages. This not only retains the baseline feature information of the normal operating stage, but also introduces the evolution trend features of the accelerated degradation stage, providing more comprehensive inter-stage correlation information for subsequent weighted fusion and health status prediction, and improving the ability to characterize and predict the degradation behavior of the workpiece throughout its entire life cycle.

[0182] To better understand the technical solutions of the embodiments of this application, the following is combined with... Figure 4 The determination of fusion features and attention coefficients across the above-mentioned operational phases are illustrated by example.

[0183] See Figure 4 In this example, the first level represents the normal operation stage, the second level represents the slow degradation stage, and the third level represents the accelerated degradation stage.

[0184] First, construct a forward edge connection, that is, a one-way connection from the first level to the second level. One-way connection from the second level to the third level ,in This represents all time window nodes in the first level. This represents all time window nodes in the second level. This represents all time window nodes in the third level.

[0185] In determining At that time, a Top-K mechanism is used to select the top K similar time window nodes in the first level for each time window node in the second level time window and retain the edges (similarity is determined based on the target time window features of the time window nodes, and the connection weight of the edges is equal to the similarity), and set the rest to 0, so as to obtain the connection of representative nodes, thereby obtaining the positive node connection matrix between the first and second levels. .exist In this process, a Top-K mechanism is used to select the edges of the top K similar time window nodes in the second level for each time window node in the third level, and set the rest to 0, thus obtaining the connections of representative nodes and obtaining a positive node connection matrix between the second and third levels. Then, a graph convolutional network is used to aggregate features from the forward node connections between the first and second layers, and between the second and third layers, to obtain the corresponding fused features. The calculation process is as follows:

[0186] ;

[0187] ;

[0188] in, The features represented by the output after convolution of the first and second layer forward connection graphs are the fused features of the target time window after the normal operation phase and the slow degradation phase. The features represented by the output after convolution of the forward connection graphs of the second and third layers indicate the fused features of the accelerated degradation stage after incorporating the target time window features of the slow degradation stage. and Each is a pair of adjacency matrices and Perform symmetric normalization. For activation function, Represents the target time window features of both the first and second layers. Represents all target time window features of the second and third layers. and This is the corresponding weight matrix.

[0189] Secondly, construct reverse edge connections, that is, one-way connections from the second level to the first level. A one-way connection from the third level to the second level. .exist The Top-K mechanism is used to select the top K similarity time window nodes in the second level for each node in the first-level time window and retain the edges (similarity is determined based on the target time window features of the time window nodes, and the connection weight of the edges is equal to the similarity). The rest are set to 0 to obtain the connections of representative nodes, thus obtaining the reverse node connection matrix between the first and second levels. .exist Similarly, the Top-K mechanism is used to select the edges of the top K similarity time window nodes in the third level for each node in the second-level time window, and set the rest to 0, to obtain the connection of representative nodes, thus obtaining the reverse node connection matrix between the third level and the second level. Then, a graph convolutional network is used to aggregate features from the inverse node connections between the first and second layers, and between the second and third layers, to obtain the corresponding fused features. The calculation process is as follows:

[0190] ;

[0191] ;

[0192] in The features represented by the output after convolution of the inverse connection graphs of the first and second layers are the fused features of the normal operation phase after incorporating the target time window features of the slow degradation phase. The features represented by the output after convolution of the inverse connection graphs of the second and third layers are the fused features of the slow degradation stage after incorporating the target time window features of the accelerated degradation stage. and For the adjacency matrix and Perform symmetric normalization. For activation function, Represents the target time window features of both the first and second layers. Represents all target time window features of the second and third layers. and This is the corresponding weight matrix.

[0193] After obtaining the fused features, an attention coefficient is calculated through a fusion attention-guided aggregation mechanism to highlight representative multi-level feature information. Specifically:

[0194] First, the target features are output after the third graph structure of the first level is processed by graph convolution. Features fused with representative nodes from the second level After concatenation at the node level, learnable parameters are used... Activation function and The function obtains Corresponding attention coefficient The calculation process is as follows:

[0195] ;

[0196] in This represents a splicing operation. For learnable parameters, For weight parameters, for Activation function.

[0197] Then, the target features are output after the third graph structure of the second level is processed by graph convolution. Features were fused with representative nodes from the third level. Features fused with representative nodes from the first level After concatenation at the node level, learnable parameters are used... Activation function and The function obtains the attention coefficients corresponding to the second-level multi-feature fusion. as well as The calculation process is as follows:

[0198] ;

[0199] .

[0200] Then, the target features are output after the third-level third graph structure is processed by graph convolution. Features fused with representative nodes from the second level After concatenation at the node level, learnable parameters are used... Activation function and The function obtains the attention coefficients corresponding to the third-level multi-feature fusion. The calculation process is as follows:

[0201] .

[0202] See Figure 4 After obtaining the above fusion features and attention coefficients, the above formula is used:

[0203] This will provide health status prediction information.

[0204] See Figure 5 , Figure 5 This is a flowchart of a health status prediction method provided in an embodiment of this application. The method includes:

[0205] Step 501: Obtain the timing data of the target monitoring signal of the workpiece;

[0206] Step 502: Copy the target monitoring signal time series data multiple times and input them in parallel into multiple data processing branches of the trained prediction model to obtain the health status information of the workpiece; wherein, the trained prediction model is obtained through the above embodiment.

[0207] To verify the performance of the prediction model trained in the embodiments of this application, taking the remaining lifespan as the health status prediction information as an example, a test set loss function convergence comparison experiment and a key module ablation verification experiment were carried out respectively.

[0208] from Figure 6As can be seen, the proposed prediction model trained based on multi-level degradation modeling and cross-time window graph convolution (the "Ours" curve in the figure) outperforms the comparative methods in terms of convergence speed and stability of test set loss. The comparative methods include those utilizing Convolutional Neural Network-Long Short Term Memory (CNN-LSTM), Graph Convolutional Network (GCN), Markov Random Field-Graph Convolutional Network (MRF-GCN), or Graph Attention Network (GAT). In the early stages of training, the test loss of each model shows a rapid decreasing trend. However, the traditional comparative methods do not distinguish the feature differences of the workpiece at different running stages and lack explicit modeling of the evolutionary relationship between time windows. Their loss curves exhibit significant oscillations during the decrease, making it difficult to maintain stable convergence. Specifically, the GCN, MRF-GCN, and GAT models still show varying degrees of fluctuation in the mid-to-late training stages. In contrast, this application divides the entire lifecycle of the workpiece into three stages: normal operation, slow degradation, and accelerated degradation. It also introduces a fully connected temporal edge mechanism across time windows at each level, which can model the continuous evolution of workpiece degradation characteristics over time. This results in faster model convergence and a stable maintenance at a low loss level.

[0209] To further verify the contribution of each key module in the above prediction model to the prediction performance of the remaining useful life (RUL) of motor bearings, this application uses the mean absolute error (MAE) as the evaluation index. The smaller the MAE value, the more accurate the prediction result. Three ablation experiments were designed: removing the cross-time window fully connected temporal edge mechanism, removing the multi-level degradation decomposition structure, and removing the attention-guided multi-level aggregation mechanism in the prediction model, respectively, while keeping the other network structures and training strategies consistent.

[0210] Ablation comparison results as follows Figure 7 As shown. From Figure 7It can be seen that the complete model (Ours) has the lowest Mean Absolute Error (MAE) and the smallest prediction error, indicating that multi-module collaborative modeling can improve lifetime prediction accuracy. When the cross-time edge mechanism (without cross-time edges) is removed, the MAE increases significantly, indicating that this mechanism can explicitly model the continuous evolution relationship between adjacent time windows within the same degradation stage, allowing the contextual information of historical windows to be adaptively transferred to the current window, improving the ability to characterize degradation trends. Without this structure, the model's modeling of the temporal consistency of the degradation process is insufficient, leading to an increase in prediction error. When the hierarchical decomposition structure (without hierarchical decomposition) is removed, the MAE increases the most, indicating significant differences in vibration modes among the three stages: normal operation, slow degradation, and accelerated degradation. Without stage-specific modeling, the mixing of cross-stage features can cause ambiguity, especially in the slow degradation stage where it is difficult to detect fault signs in advance, thus reducing lifetime prediction accuracy. Even after removing the attention-guided multi-level feature aggregation mechanism (w / o Attention Aggregation), the MAE still increased, indicating that attention fusion can highlight representative multi-level features and suppress redundant and noisy information in the information interaction between levels, making the fused feature representation more suitable for Remaining Useful Life (RUL) regression prediction.

[0211] Considering the crucial role of key components in various industrial robots, especially in high-precision and high-reliability applications such as dual-arm collaborative handling robots, therefore, such as Figure 8 As shown in the figure, this application embodiment also provides a health status monitoring system for a dual-arm collaborative handling robot based on multi-level degradation modeling (hereinafter referred to as the system).

[0212] The aforementioned system integrates functions such as online vibration signal acquisition, hierarchical modeling of degradation stages, cross-time window graph learning, and attention-guided multi-level feature fusion, forming an intelligent monitoring platform for health assessment and lifespan prediction of critical components. The system collects monitoring signals (e.g., vibration signals) from key robot components (e.g., rolling bearings) in real time using multiple sensors and divides the equipment operation process into three stages: normal operation, slow degradation, and accelerated degradation, based on the health status decay trend. Building upon this, a multi-level decomposed graph convolutional structure is used to hierarchically model the vibration characteristics of different degradation stages. Furthermore, a fully connected temporal edge mechanism across time windows dynamically characterizes the evolution of degradation characteristics over time.

[0213] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0214] Based on the same inventive concept, please refer to Figure 9 As shown, this embodiment provides a training device for a prediction model, including:

[0215] The acquisition module 901 is used to acquire the timing data of monitoring signals of the workpiece at different operating stages;

[0216] Data processing module 902 is used to independently process the time series data of monitoring signals in different operating stages using different data processing branches of the prediction model, so as to obtain the target features of the different operating stages.

[0217] The fusion prediction module 903 is used to perform fusion prediction processing on the target features of different operating stages using the fusion prediction layer of the prediction model to obtain the health status prediction information of the workpiece.

[0218] The training module 904 is used to train the prediction model based on the health status prediction information and the actual health status information to obtain the trained prediction model.

[0219] In some embodiments, the monitoring signal time-series data includes multiple time windows, and each time window includes multiple monitoring signals; the data processing module is specifically used for:

[0220] Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed.

[0221] The target features are determined based on the structure of each first graph in the aforementioned operational phase.

[0222] In some embodiments, the data processing module 902 is further configured to:

[0223] Graph convolution processing is performed on each first graph structure of the running phase to obtain first time window features that correspond one-to-one with the time windows;

[0224] Based on the similarity of adjacent first time window features during the operation phase, a second adjacency matrix and a second graph structure corresponding to the second adjacency matrix are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0;

[0225] The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows;

[0226] The target features are determined based on the second time window features of the aforementioned operational phase.

[0227] In some embodiments, the data processing module 902 is further configured to:

[0228] Based on the similarity between the features of each second time window in the operation phase, the third adjacency matrix and the third graph structure corresponding to the third adjacency matrix are determined.

[0229] The target features are obtained by performing graph convolution processing on the third graph structure.

[0230] In some embodiments, the fusion prediction module 903 is specifically used for:

[0231] At least a portion of the target time window features of the first running stage are fused into the target time window features of the second running stage to obtain the fused features of the second running stage; wherein, the target time window features include the first time window features or the second time window features, and the first running stage and the second running stage are adjacent;

[0232] Based on the fusion features and target features of the second running stage, the attention coefficient corresponding to the fusion features is calculated;

[0233] Based on the attention coefficient, the fusion feature and the target feature are weighted and fused to obtain the weighted fusion feature corresponding to the second running stage;

[0234] The weighted fusion features of each of the aforementioned operational stages are spliced ​​and predicted to obtain the health status prediction information.

[0235] In some embodiments, the fusion prediction module 903 is further configured to:

[0236] The similarity between the target time window features of the second running stage and the target time window features of the first running stage is determined to obtain multiple similarity sequences, and the similarity sequences correspond one-to-one with the target time window features of the second running stage.

[0237] From each of the similarity sequences, at least one similarity that satisfies a preset high similarity condition is selected to obtain multiple similarity subsequences;

[0238] Based on the multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage;

[0239] Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

[0240] In some embodiments, the operation phase includes a normal operation phase, a slow degradation phase, and an accelerated degradation phase; the fusion prediction module is specifically used for:

[0241] When the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature, and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage.

[0242] Wherein, the first fusion feature corresponds to the first attention coefficient, the second fusion feature corresponds to the second attention coefficient, the first fusion feature is obtained by fusing time-series data of monitoring signals during normal operation, and the second fusion feature is obtained by fusing time-series data of monitoring signals during accelerated degradation.

[0243] It should be understood that, for the sake of brevity, some of the content described in the previous embodiments will not be repeated in this embodiment.

[0244] Based on the same inventive concept, please refer to Figure 10 As shown, this embodiment provides a health status prediction device, including:

[0245] The acquisition module 1001 is used to acquire the timing data of the target monitoring signal of the workpiece;

[0246] The input module 1002 is used to copy the time-series data of the target monitoring signal multiple times and input them in parallel into multiple data processing branches of the trained prediction model to obtain the health status information of the workpiece; wherein, the trained prediction model is trained by the training method.

[0247] It should be understood that, for the sake of brevity, some of the content described in the previous embodiments will not be repeated in this embodiment.

[0248] Based on the same inventive concept, embodiments of this application provide a computer device, which may be a server, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for training a predictive model.

[0249] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0250] Based on the same inventive concept, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0251] Acquire timing data of monitoring signals from the workpiece at different stages of operation;

[0252] By utilizing different data processing branches of the prediction model, the time series data of the monitoring signals in the different operating stages are processed independently to obtain the target features of the different operating stages;

[0253] By using the fusion prediction layer of the prediction model, the target features of different operating stages are fused and predicted to obtain the health status prediction information of the workpiece.

[0254] The prediction model is trained based on the predicted health status information and the actual health status information to obtain the trained prediction model.

[0255] In some embodiments, the monitoring signal timing data includes multiple time windows, each time window including multiple monitoring signals, and the processor executes the following steps when executing the computer program:

[0256] Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed.

[0257] The target features are determined based on the structure of each first graph in the aforementioned operational phase.

[0258] In some embodiments, when a processor executes a computer program, it performs the following steps:

[0259] Graph convolution processing is performed on each first graph structure of the running phase to obtain first time window features that correspond one-to-one with the time windows;

[0260] Based on the similarity of adjacent first time window features during the operation phase, a second adjacency matrix and a second graph structure corresponding to the second adjacency matrix are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0;

[0261] The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows;

[0262] The target features are determined based on the second time window features of the aforementioned operational phase.

[0263] In some embodiments, when a processor executes a computer program, it performs the following steps:

[0264] Based on the similarity between the features of each second time window in the operation phase, the third adjacency matrix and the third graph structure corresponding to the third adjacency matrix are determined.

[0265] The target features are obtained by performing graph convolution processing on the third graph structure.

[0266] In some embodiments, when a processor executes a computer program, it performs the following steps:

[0267] At least a portion of the target time window features of the first running stage are fused into the target time window features of the second running stage to obtain the fused features of the second running stage; wherein, the target time window features include the first time window features or the second time window features, and the first running stage and the second running stage are adjacent;

[0268] Based on the fusion features and target features of the second running stage, the attention coefficient corresponding to the fusion features is calculated;

[0269] Based on the attention coefficient, the fusion feature and the target feature are weighted and fused to obtain the weighted fusion feature corresponding to the second running stage;

[0270] The weighted fusion features of each of the aforementioned operational stages are spliced ​​and predicted to obtain the health status prediction information.

[0271] In some embodiments, when a processor executes a computer program, it performs the following steps:

[0272] The similarity between the target time window features of the second running stage and the target time window features of the first running stage is determined to obtain multiple similarity sequences, and the similarity sequences correspond one-to-one with the target time window features of the second running stage.

[0273] From each of the similarity sequences, at least one similarity that satisfies a preset high similarity condition is selected to obtain multiple similarity subsequences;

[0274] Based on the multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage;

[0275] Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

[0276] In some embodiments, the operating phase includes a normal operating phase, a slow degradation phase, and an accelerated degradation phase; the processor executes the following steps when executing the computer program:

[0277] When the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature, and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage.

[0278] Wherein, the first fusion feature corresponds to the first attention coefficient, the second fusion feature corresponds to the second attention coefficient, the first fusion feature is obtained by fusing time-series data of monitoring signals during normal operation, and the second fusion feature is obtained by fusing time-series data of monitoring signals during accelerated degradation.

[0279] Based on the same inventive concept, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0280] Acquire timing data of the target monitoring signal of the workpiece;

[0281] The target monitoring signal time series data is copied multiple times and input into multiple data processing branches of the trained prediction model in parallel to obtain the health status information of the workpiece; wherein, the trained prediction model is trained by the training method.

[0282] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0283] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0284] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for training a prediction model, characterized in that, The method includes: Acquire timing data of monitoring signals from the workpiece at different stages of operation; By utilizing different data processing branches of the prediction model, the time series data of the monitoring signals in the different operating stages are processed independently to obtain the target features of the different operating stages; By using the fusion prediction layer of the prediction model, the target features of different operating stages are fused and predicted to obtain the health status prediction information of the workpiece. The prediction model is trained based on the predicted health status information and the actual health status information to obtain the trained prediction model.

2. The training method for the prediction model according to claim 1, characterized in that, The monitoring signal timing data includes multiple time windows, and each time window includes multiple monitoring signals; The step of independently processing the time-series monitoring signal data of different operating stages to obtain the target features of different operating stages includes: Based on the similarity between the monitoring signals within each time window, a first adjacency matrix and a first graph structure corresponding to the first adjacency matrix are constructed. The target features are determined based on the structure of each first graph in the aforementioned operational phase.

3. The training method for the prediction model according to claim 2, characterized in that, The determination of the target features based on the structure of each first graph in the aforementioned operational phase includes: Graph convolution processing is performed on each first graph structure of the running phase to obtain first time window features that correspond one-to-one with the time windows; Based on the similarity of adjacent first time window features during the operation phase, a second adjacency matrix and a second graph structure corresponding to the second adjacency matrix are determined; wherein, the second adjacency matrix is ​​used to represent that the (t-1)th first time window feature points to the tth first time window feature, where t is an integer greater than 0; The second graph structure is subjected to graph convolution processing to obtain the second time window features that correspond one-to-one with the time windows; The target features are determined based on the second time window features of the aforementioned operational phase.

4. The training method for the prediction model according to claim 3, characterized in that, Determining the target feature based on the second time window features of the operational phase includes: Based on the similarity between the features of each second time window in the operation phase, the third adjacency matrix and the third graph structure corresponding to the third adjacency matrix are determined. The target features are obtained by performing graph convolution processing on the third graph structure.

5. The training method for the prediction model according to claim 3 or 4, characterized in that, The process of fusing and predicting the target features of different operating stages to obtain the health status prediction information of the workpiece includes: At least a portion of the target time window features of the first running stage are fused into the target time window features of the second running stage to obtain the fused features of the second running stage; wherein, the target time window features include the first time window features or the second time window features, and the first running stage and the second running stage are adjacent; Based on the fusion features and target features of the second running stage, the attention coefficient corresponding to the fusion features is calculated; Based on the attention coefficient, the fusion feature and the target feature are weighted and fused to obtain the weighted fusion feature corresponding to the second running stage; The weighted fusion features of each of the aforementioned operational stages are spliced ​​and predicted to obtain the health status prediction information.

6. The training method for the prediction model according to claim 5, characterized in that, The step of fusing at least a portion of the target time window features of the first running stage into the target time window features of the second running stage to obtain the fused features of the second running stage includes: The similarity between the target time window features of the second running stage and the target time window features of the first running stage is determined to obtain multiple similarity sequences, and the similarity sequences correspond one-to-one with the target time window features of the second running stage. From each of the similarity sequences, at least one similarity that satisfies a preset high similarity condition is selected to obtain multiple similarity subsequences; Based on the multiple similarity subsequences, a fourth adjacency matrix corresponding to the second running stage and a fourth graph structure corresponding to the fourth adjacency matrix are constructed; wherein, the target time window feature of the first running stage in the fourth adjacency matrix points to the target time window feature corresponding to the second running stage; Graph convolution is performed on the fourth graph structure to obtain the fusion features of the second running stage.

7. The training method for the prediction model according to claim 5, characterized in that, The operational phases include a normal operation phase, a slow degradation phase, and an accelerated degradation phase; The step of weightedly fusing the fused features and the target features based on the attention coefficient to obtain the weighted fused features corresponding to the second running stage includes: When the second running stage is the slow degradation stage, the first fusion feature, the second fusion feature, and the target feature are weighted and fused based on the first attention coefficient and the second attention coefficient to obtain the weighted fusion feature corresponding to the slow degradation stage. Wherein, the first fusion feature corresponds to the first attention coefficient, the second fusion feature corresponds to the second attention coefficient, the first fusion feature is obtained by fusing time-series data of monitoring signals during normal operation, and the second fusion feature is obtained by fusing time-series data of monitoring signals during accelerated degradation.

8. A method for predicting health status, characterized in that, The method includes: Acquire timing data of the target monitoring signal of the workpiece; The target monitoring signal time series data is copied multiple times and input into multiple data processing branches of the trained prediction model in parallel to obtain the health status information of the workpiece; wherein the trained prediction model is trained by the method of any one of claims 1-7.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by at least one processor, implements the method as described in any one of claims 1-8.

Images