Equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning

Abstract

The present application discloses a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning, and relates to the technical field of equipment performance monitoring. In the method, based on the component service status identification model in the digital twin system, the component service status of the key component is identified according to the multi-source sensor data of the processing process of the key component, and the identification ability of the component service status under the imbalance of samples is balanced through the cyclically consistent generation, discrimination and classification process, thereby overcoming the problems of unbalanced number of samples of different service status and insufficient service status identification accuracy; then, based on the component performance degradation monitoring model under the component service status, according to the multi-source sensor data of the key component, the spatiotemporal coupling characteristics of the dynamic changes of the multi-source sensor data can be fully captured to obtain the component performance index degradation value of the key component. The present application combines status identification and degradation monitoring to comprehensively reflect the in-depth research on the performance degradation process of key components of complex equipment.

Description

Technical Field

[0001] The present application relates to the technical field of equipment performance monitoring, and in particular to a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning. Background Art

[0002] In the modern industrial field, high-end complex equipment such as nuclear power plant equipment, wind turbines, and CNC machine tools are the core driving force for development. The efficient operation of these equipment depends largely on the performance of key components, such as the primary circuit pipelines of nuclear power plants, the gearboxes of wind turbines, and the tools of CNC machine tools. However, with the increase of service life, these key components will inevitably suffer from performance degradation. Since these components usually operate in a complex and constantly changing working environment, performance degradation often manifests itself as hidden and nonlinear changes, which significantly increases the difficulty of equipment monitoring. Therefore, how to efficiently and accurately monitor the performance degradation process of these key equipment components is of vital importance to improving the level of manufacturing and establishing a leading position in the field of science and technology.

[0003] At present, equipment component monitoring technology is mainly divided into direct monitoring and indirect monitoring. The direct monitoring method directly measures the key geometric parameters of the components through equipment such as microscopes and charge-coupled device cameras. Although this type of method has high monitoring accuracy, it has low measurement efficiency and is therefore limited in practical applications. In contrast, researchers at home and abroad prefer to adopt indirect monitoring methods based on multi-source sensor data, that is, to use machine learning, deep learning and other technologies to establish the association between multi-source sensor data and the degree or process of component degradation. This method has achieved certain results, but it still has shortcomings in practical engineering applications: the number of samples of different performance degradation stages obtained from historical actual processing is usually unbalanced, and samples of the stage of rapid performance degradation are often missing, so the accuracy of service status identification is difficult to guarantee; at the same time, the existing performance degradation prediction methods often rely on specific time, space or spatiotemporal sequence methods to extract sensor features, which makes it difficult to fully capture the spatiotemporal coupling characteristics of dynamic changes in multi-source sensor data; in addition, existing performance monitoring research focuses on the identification of service status or the prediction of degradation process, and lacks in-depth research on combining the two to fully reflect the performance degradation process of components. Summary of the invention

[0004] The purpose of this application is to provide a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning, which can achieve high-precision monitoring of the performance degradation process of key components of complex equipment.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning, comprising the following steps:

[0007] Collect multi-source sensor data of the processing of key components of complex equipment. The multi-source sensor data of the processing process refers to the multi-source sensor data of key components during a processing process.

[0008] Based on the component service status identification model in the digital twin system, the component service status of key components is identified according to the multi-source sensor data of the key component processing process. The component service status identification model is a classification enhancement network with cycle consistency. It balances the identification ability of component service status under sample imbalance through cycle consistent generation, discrimination and classification process; the digital twin system is a virtual real-time simulation system with the same characteristics as complex equipment, which is used to monitor the performance indicators of each component in real time.

[0009] Based on the component performance degradation monitoring model under the service state of the component, the component performance index degradation value of the key component is obtained according to the multi-source sensor data of the processing process of the key component, so as to realize the performance degradation monitoring of the key components of complex equipment; the component performance degradation monitoring model includes a time series feature modeling branch and a space feature modeling branch; the space feature modeling branch is used to extract latent spatial features according to the graph structured sensor data corresponding to the multi-source sensor data of the processing process; the time series feature modeling branch is used to extract long-term time series features from the multi-source sensor data of the processing process, and after fusing the long-term time series features and the latent space features, the component performance index degradation value of the key component is obtained.

[0010] Optionally, the component service status identification model is obtained by the following steps:

[0011] A component service status training sample set is obtained; the component service status training sample set includes a plurality of component service status training samples, and the component service status training samples include multi-source sensor data of a processing process of a key component and a corresponding component service status.

[0012] A classification enhancement network with cycle consistency is constructed to obtain a component service status identification model; the component service status identification model includes a generator, a discriminator and a classifier; the generator is used to simulate and generate multi-source sensor data of the processing process under different component service states according to the source domain data features of the component service status training sample set that can characterize the component service status; the multi-source sensor data of the processing process simulated and generated by the generator imitates the distribution of the source domain data to constitute the target domain data; the discriminator is used to determine whether the multi-source sensor data of the processing process input into the discriminator is the source domain data or the target domain data; the classifier is used to identify the component service status of the key component according to the input multi-source sensor data of the processing process of the key component.

[0013] The generator and the discriminator are used to compete with each other to balance the distribution of the service status of each component in the component service status training sample set, and the service status of the component is identified through the classifier.

[0014] Optionally, the generator extracts key features in the source domain data through a multi-layer convolution module, and transforms the key features in the source domain data through a deep residual module so that the source domain data can adapt to the distribution of the target domain, and then restores the transformed features to the original space based on the inverse convolution operation to achieve cross-domain data generation from the source domain to the target domain; the discriminator includes a multi-layer convolution module, which efficiently extracts the spatial features of the input data through the multi-layer convolution module and continuously optimizes the discrimination threshold.

[0015] Optionally, obtaining a component service status training sample set specifically includes the following steps:

[0016] From the historical processing database of key components, a number of multi-source sensor data of the processing process and the corresponding historical performance index degradation values ​​are obtained.

[0017] The component service status is divided according to the slope change of the historical performance index degradation value curve, and the component service status corresponding to the multi-source sensor data of each processing process is determined; the multi-source sensor data of the processing process and the corresponding component service status constitute a component service status training sample; and several component service status training samples constitute a component service status training sample set.

[0018] Optionally, after obtaining a plurality of multi-source sensor data of the processing process and corresponding historical performance index degradation values ​​from the historical processing process database of the key component, the method further comprises the following steps:

[0019] For any multi-source sensor data of a processing process, it is determined whether there are missing values ​​in the historical performance indicator degradation values ​​of the multi-source sensor data of the processing process, and a first determination result is obtained.

[0020] When the first judgment result is yes, the historical performance indicator degradation value completion model is used to obtain the completed historical performance indicator degradation value according to the sensor characteristics of the multi-source sensor data of the processing process; the historical performance indicator degradation value completion model is a variational autoencoder model; the variational autoencoder model includes an encoder and a decoder, the encoder of the variational autoencoder model is used to map the input sensor characteristics to the latent space and generate the mean and standard deviation of the latent variables; the decoder of the variational autoencoder model is used to generate the completed historical performance indicator degradation value according to the latent variables in the latent space.

[0021] When the first judgment result is no, no processing is performed.

[0022] Optionally, the variational autoencoder model is improved based on a Bayesian inference mechanism and a conditional modeling mechanism; the Bayesian inference mechanism assigns a prior distribution to each weight of the variational autoencoder model, and updates the prior distribution by inputting and outputting complete training data, and finally obtains a posterior distribution; the conditional modeling mechanism takes the size and processing technology of key components as a column of additional features, and inputs them into the encoder of the variational autoencoder model together with the sensor features.

[0023] Optionally, the sensing features of multi-source sensing data of the machining process include time domain features, frequency domain features and time-frequency domain features; the time domain features include peak value, mean value, standard deviation, root mean square value, skewness, kurtosis and margin, and the time domain features are used to describe the basic form, fluctuation amplitude and trend of the signal; the frequency domain features include centroid frequency, mean square frequency and variance frequency, and the frequency domain features are used to reflect the frequency distribution and energy concentration of the signal; the time-frequency domain features include high and low frequency component coefficients obtained by three-layer wavelet packet decomposition, which are used to reveal the time-frequency characteristics of the signal at different scales.

[0024] Optionally, the complex equipment is a CNC grinder, and the key component is the grinder tool; by arranging an acoustic emitter, a vibration sensor and a cutting force sensor near the grinder tool, and arranging a current sensor in the electrical cabinet of the CNC grinder, multi-source sensor data of the machining process is collected.

[0025] Optionally, in the spatial feature modeling branch, the K-nearest neighbor method is used to construct graph-structured sensor data of multi-source sensor data for each processing process, and then the latent spatial features between the multi-source sensor data are gradually extracted through three-layer graph convolution operations, so as to capture the spatial distribution pattern of the performance degradation process of key components; combining the bidirectional gated recurrent unit with the global attention mechanism, while capturing the temporal dynamic change law in the multi-source sensor data of the processing process, the influence of different latent spatial features on the prediction task is adjusted.

[0026] Optionally, the time series feature modeling branch adopts an encoder-decoder structure to efficiently model the complex patterns of long time series data. The encoder of the time series feature modeling branch is constructed by three layers of repeatedly stacked modules, each layer contains a multi-head probabilistic sparse self-attention mechanism and a convolutional layer for dimensionality reduction, which can capture the long-term dependencies and local characteristics of multi-source sensor data in the machining process through cascaded feature maps; the decoder of the time series feature modeling branch is composed of two layers of modules, which extracts the evolution law of tool wear through a multi-head probabilistic sparse self-attention mechanism, uses a cross-attention mechanism to construct a global correlation between sensor data and performance indicators, and integrates the long-term dependencies and local characteristics of multi-source sensor data.

[0027] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0028] The present application provides a method for monitoring the performance degradation of equipment components based on digital twin spatiotemporal parallel learning. In this method, multi-source sensor data of key components in a processing process is collected, and based on the component service status identification model in the digital twin system, the component service status of the key components is identified according to the multi-source sensor data of the key components. The component service status identification model is a classification enhancement network with cycle consistency, which balances the ability to identify the component service status under sample imbalance through cycle-consistent generation, discrimination and classification processes, thereby overcoming the problem that the number of samples of different service status obtained in the prior art is usually unbalanced, resulting in low service status identification accuracy. Insufficient problems; then based on the component performance degradation monitoring model under the component service status, according to the multi-source sensor data of the key components, the component performance index degradation value of the key components is obtained, and the performance degradation monitoring of the key components of complex equipment is realized. In the component performance degradation monitoring model of the present application, the timing characteristics and spatial characteristics are captured respectively by the timing feature modeling branch and the space feature modeling branch, and the spatiotemporal coupling characteristics of the dynamic changes of multi-source sensor data are fully captured. The structure and parameters of the component performance degradation detection model are adapted to the current component service status, so as to achieve the combination of state identification and degradation monitoring, so as to comprehensively reflect the in-depth study of the performance degradation process of the key components of complex equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 A flowchart of a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning is provided in accordance with an embodiment of the present application.

[0031] Figure 2 A flowchart of a component service status identification model obtained in a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning provided in one embodiment of the present application.

[0032] Figure 3 A schematic diagram of the model structure of a component performance degradation monitoring model in an equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning provided in one embodiment of the present application.

[0033] Figure 4 A flowchart of step B1 in a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning provided in one embodiment of the present application.

[0034] Figure 5 A flowchart for detecting and completing missing values ​​in a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning is provided in accordance with an embodiment of the present application. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0036] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0037] The embodiment of the present application provides a method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning, such as Figure 1 As shown, the following steps are included:

[0038] A1. Collect multi-source sensor data of the processing process of key components of complex equipment. The multi-source sensor data of the processing process is the multi-source sensor data of the key components during a processing process. For example, the complex equipment is a CNC grinder, and the key component is the grinder tool; by arranging an acoustic emitter, a vibration sensor, and a cutting force sensor near the grinder tool, and arranging a current sensor in the electrical cabinet of the CNC grinder, the multi-source sensor data of the processing process is collected. When the key component is a grinder tool, the service status of the component can represent the tool wear status, and the corresponding performance indicator degradation value is the tool wear value.

[0039] A2. Based on the component service status identification model in the digital twin system, the component service status of key components is identified according to the multi-source sensor data of the key component processing process. The component service status identification model is a classification enhancement network with cycle consistency. It balances the identification ability of component service status under sample imbalance through cycle consistent generation, discrimination and classification process; the digital twin system is a virtual real-time simulation system with the same characteristics as complex equipment, which is used to monitor the performance indicators of each component in complex equipment in real time.

[0040] The number of samples of different service states obtained from actual processing is usually unbalanced, and samples of the stage of rapid performance degradation are often missing, making it difficult to ensure the accuracy of service state identification. To this end, this application solves this problem through a classification enhancement network with cycle consistency. Specifically in this embodiment, Figure 2 As shown in Figure 1, the component service status identification model is obtained through the following steps:

[0041] B1. Obtaining a component service status training sample set. The component service status training sample set includes a number of component service status training samples, and the component service status training samples include multi-source sensor data of the processing process of key components and the corresponding component service status.

[0042] B2. Construct a classification enhancement network with cycle consistency to obtain a component service status identification model. The component service status identification model includes a generator, a discriminator and a classifier; the generator is used to simulate and generate multi-source sensor data of the processing process under different component service states based on the source domain data features of the component service status training sample set that can characterize the component service status; the multi-source sensor data of the processing process simulated by the generator imitates the distribution of the source domain data to form the target domain data; the discriminator is used to determine whether the multi-source sensor data of the processing process input into the discriminator is source domain data or target domain data; the classifier is used to identify the component service status of the key component based on the input multi-source sensor data of the processing process of the key component.

[0043] B3. Use the generator and the discriminator to compete with each other, balance the distribution of the service status of each component in the component service status training sample set, and identify the component service status through the classifier.

[0044] Specifically, by allowing the generator and the discriminator to compete with each other, the generator can generate multi-source sensor data of the processing process that is close to the real data, so as to balance the distribution of the service status of each component in the component service status training sample set; then, these samples containing the component service status (including real samples and generated samples) are input into the classifier, and training is performed under the condition of sample balance to achieve accurate identification of the component service status.

[0045] In an exemplary embodiment, the generator extracts key features in the source domain data through a multi-layer convolution module, and transforms the key features in the source domain data through a deep residual module so that the source domain data can adapt to the distribution of the target domain, and then restores the transformed features to the original space based on the inverse convolution operation to achieve cross-domain data generation from the source domain to the target domain; the discriminator includes a multi-layer convolution module, which helps to capture the detailed information of the input data by efficiently extracting the spatial features of the input data, and improves the discriminator's ability to distinguish between true and false data by continuously optimizing the discrimination threshold.

[0046] A3. Based on the component performance degradation monitoring model under the service state of the component, the component performance index degradation value of the key component is obtained according to the multi-source sensor data of the processing process of the key component, so as to realize the performance degradation monitoring of the key components of the complex equipment; the component performance degradation monitoring model includes a time series feature modeling branch and a space feature modeling branch; the space feature modeling branch is used to extract the latent spatial features according to the graph structured sensor data corresponding to the multi-source sensor data of the processing process; the time series feature modeling branch is used to extract the long-term time series features of the multi-source sensor data of the processing process, and after fusing the long-term time series features and the latent space features, the component performance index degradation value of the key component is obtained.

[0047] Existing performance degradation prediction methods usually rely on specific time, space or spatiotemporal serial methods to extract sensor features, which makes it difficult to fully capture the spatiotemporal coupling relationship of dynamic changes in multi-source sensor data. In addition, existing performance monitoring research mostly focuses on the single task of state identification or degradation process prediction, and lacks the exploration of organically combining the two to comprehensively reflect the performance degradation process of components. In this application, on the basis of determining the service status of components, a component performance degradation monitoring model based on spatiotemporal parallel learning is used to accurately predict the degradation of performance indicators of key components. When the component service status identification model of the digital twin system senses the key components, the component performance degradation monitoring model based on spatiotemporal parallel learning is used to accurately predict the degradation of performance indicators of key components. When the service status of a component changes, the component performance degradation monitoring model is dynamically updated to achieve high-precision monitoring of the component performance degradation process; specifically, when the complex equipment is a CNC grinder and the key component is the grinder tool, the tool wear status identified by the component service status identification model of the digital twin system is used as the driving condition. When the tool wear status changes, the structure and parameters of the component performance degradation monitoring model are adaptively adjusted through the dynamic update mechanism, so that the model can respond to these changes in real time, ensuring the adaptability and prediction accuracy of the monitoring model under complex working conditions, and achieving accurate prediction of tool wear.

[0048] Specifically in this embodiment, Figure 3 The schematic diagram of the model structure shown in the figure uses a graph convolutional network as the main body to construct a spatial feature modeling branch. In the spatial feature modeling branch, the K nearest neighbor method is used to construct the graph structured sensor data of the multi-source sensor data of each processing process, and then the latent spatial features between the multi-source sensor data are gradually extracted through three-layer graph convolution operations, thereby capturing the spatial distribution pattern of the performance degradation process of key components; combined with the bidirectional gated recurrent unit and the global attention mechanism, while capturing the temporal dynamic change law in the multi-source sensor data of the processing process, the influence of different latent spatial features on the prediction task is adjusted.

[0049] The temporal feature modeling branch is constructed with the Informer structure as the main body. The temporal feature modeling branch adopts an encoder-decoder structure to efficiently model the complex patterns of long time series data. The encoder of the temporal feature modeling branch is constructed by three layers of repeatedly stacked modules. Each layer contains a multi-head probabilistic sparse self-attention mechanism and a convolutional layer for dimensionality reduction. It can capture the long-term dependency and local features of multi-source sensor data in the machining process through cascaded feature maps, while also reducing the computational complexity; the decoder of the temporal feature modeling branch is composed of two layers of modules. The evolution law of tool wear is extracted through the multi-head probabilistic sparse self-attention mechanism, and the cross-attention mechanism is used to construct the global association between sensor data and performance indicators, and the long-term dependency and local features of multi-source sensor data are integrated.

[0050] In an exemplary embodiment, Figure 4 As shown, B1 obtains a component service status training sample set, which specifically includes the following steps:

[0051] B11. Obtain a number of multi-source sensor data of the processing process and the corresponding historical performance indicator degradation values ​​from the historical processing database of key components.

[0052] Specifically, when the complex equipment is a CNC grinder and the key component is the grinding machine tool, a pneumatic precision sensor is used to monitor the wear of the grinding wheel. After the grinder has processed a workpiece, the grinding wheel and the air gap injection hole are aligned according to the program of the CNC system. When the distance between them reaches the set threshold, the sensor signal change is triggered, and the vertical coordinates of the CNC machine tool grinding wheel are read. By calculating the coordinate difference between two consecutive processing processes, the grinding wheel wear value of each processing can be determined, and the real-time sensor data and tool wear values ​​collected in each processing process are stored in the digital twin system.

[0053] B12. The service state of the components is divided based on the slope change of the historical performance index degradation value curve, and the service state of the components corresponding to the multi-source sensor data of each processing process is determined. The multi-source sensor data of the processing process and the corresponding service state of the components constitute a component service state training sample; a plurality of component service state training samples constitute a component service state training sample set. That is, in this embodiment, the tool wear state can be divided based on the slope change of the historical tool wear value curve.

[0054] Most existing methods assume that the key performance indicators of components in historical samples are accurate and complete. Although a few studies have supplemented the missing performance indicators, they usually rely on the specific distribution of a small number of samples, ignoring the measurement errors and distribution differences under different working conditions, which makes the credibility of the supplemented results insufficient and the prediction model effect difficult to guarantee. Therefore, in the embodiment provided by the present application, after step B11 obtains a number of multi-source sensor data of the processing process and the corresponding historical performance indicator degradation values, the method also includes a process of detecting missing values ​​and supplementing them, such as Figure 5 As shown, the process includes the following steps:

[0055] C1. For any multi-source sensor data of a processing process, determine whether there are missing values ​​in the historical performance indicator degradation values ​​of the multi-source sensor data of the processing process, and obtain a first determination result.

[0056] C2. When the first judgment result is yes, the historical performance indicator degradation value completion model is used to obtain the completed historical performance indicator degradation value according to the sensor characteristics of the multi-source sensor data of the processing process; the historical performance indicator degradation value completion model is a variational autoencoder model; the variational autoencoder model includes an encoder and a decoder, and the encoder of the variational autoencoder model is used to map the input sensor characteristics to the latent space and generate the mean and standard deviation of the latent variables; the decoder of the variational autoencoder model is used to generate the completed historical performance indicator degradation value according to the latent variables in the latent space. The encoder of the variational autoencoder model describes the distribution in the latent space through the mean and standard deviation of the latent variables, and the decoder of the variational autoencoder model generates the mean and variance of the completed historical performance indicator degradation value after receiving the sampled latent variables, and the completed historical performance indicator degradation value can be obtained by sampling.

[0057] C3. When the first judgment result is no, no processing is performed.

[0058] As a further improved embodiment, the variational autoencoder model is improved based on the Bayesian inference mechanism and the conditional modeling mechanism; the Bayesian inference mechanism assigns a prior distribution to each weight of the variational autoencoder model, and updates the prior distribution by inputting and outputting complete training data, and finally obtains a posterior distribution, so as to fully consider the uncertainty of the sample data, avoid overfitting of the model to the existing training samples, and help improve the completion accuracy of the model; the conditional modeling mechanism takes the size and processing technology of the key components as a column of additional features, and inputs them into the encoder of the variational autoencoder model together with the sensor features, so as to avoid the limitation that the variational autoencoder model only relies on a single data distribution.

[0059] Specifically, the sensing features of multi-source sensor data in the machining process include time domain features, frequency domain features and time-frequency domain features; time domain features include peak value, mean value, standard deviation, root mean square value, skewness, kurtosis and margin, and time domain features are used to describe the basic form, fluctuation amplitude and trend of the signal; frequency domain features include centroid frequency, mean square frequency and variance frequency, and frequency domain features are used to reflect the frequency distribution and energy concentration of the signal; time-frequency domain features include high and low frequency component coefficients obtained by three-layer wavelet packet decomposition, which are used to reveal the time-frequency characteristics of the signal at different scales and provide more detailed local features for non-stationary sensor signals.

[0060] The above-mentioned embodiment of the present application provides a method for monitoring the performance degradation of equipment components based on digital twin spatiotemporal parallel learning. Based on the component service status identification model with cycle consistency in the digital twin system, the component service status of the key component is identified according to the multi-source sensor data of the key component. The ability to identify the component service status under sample imbalance is balanced through the cycle-consistent generation, discrimination and classification process, thereby overcoming the problem of insufficient service status identification accuracy caused by the imbalance in the number of samples of different service statuses obtained in the prior art. Then, based on the component performance degradation monitoring model under the component service status, the component performance index degradation value of the key component is obtained according to the multi-source sensor data of the key component, so as to achieve Performance degradation monitoring of key components of complex equipment. In the component performance degradation monitoring model of the present application, the timing characteristics and spatial characteristics are captured respectively through the timing feature modeling branch and the spatial feature modeling branch, fully capturing the spatiotemporal coupling characteristics of the dynamic changes of multi-source sensor data, and accurately and concurrently capturing the key timing patterns, dynamic spatial relationships and significant feature channels, thereby improving the prediction effect of the model; and the structure and parameters of the component performance degradation detection model are adapted to the current service status of the component, so as to achieve the combination of state identification and degradation monitoring, and drive the dynamic update of the prediction model through the real-time identified service status, thereby improving the prediction accuracy and adaptability of the model under variable working conditions, and realizing accurate monitoring of the entire service cycle of equipment components.

[0061] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.

[0062] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for monitoring equipment component performance degradation based on digital twin spatiotemporal parallel learning, characterized in that: include: Collect multi-source sensor data during the processing of key components of complex equipment; The multi-source sensor data of the processing process is the multi-source sensor data of the key component during one processing process; Based on the component service status identification model in the digital twin system, the component service status of the key component is identified according to the multi-source sensor data of the processing process of the key component; the component service status identification model is a classification enhancement network with cycle consistency, which balances the identification ability of the component service status under sample imbalance through cycle-consistent generation, discrimination and classification processes; The digital twin system is a virtual real-time simulation system with the same characteristics as the complex equipment, and is used to monitor the performance indicators of each component in real time; Based on the component performance degradation monitoring model under the service state of the component, according to the multi-source sensor data of the processing process of the key component, the component performance index degradation value of the key component is obtained to realize the performance degradation monitoring of the key components of complex equipment; The component performance degradation monitoring model includes a time series feature modeling branch and a space feature modeling branch; The spatial feature modeling branch is used to extract potential spatial features according to the graph structured sensor data corresponding to the multi-source sensor data of the processing process; The time series feature modeling branch is used to extract long-term time series features from the multi-source sensor data of the processing process, and fuse the long-term time series features with the latent space features to obtain the component performance index degradation value of the key component; In the spatial feature modeling branch, the K-nearest neighbor method is used to construct the graph structured sensor data of the multi-source sensor data of each processing process, and then the latent spatial features between the multi-source sensor data are gradually extracted through a three-layer graph convolution operation, so as to capture the spatial distribution pattern of the performance degradation process of the key components; the bidirectional gated recurrent unit and the global attention mechanism are combined to capture the temporal dynamic change law in the multi-source sensor data of the processing process while adjusting the influence of different latent spatial features on the prediction task; The temporal feature modeling branch adopts an encoder-decoder structure to efficiently model the complex patterns of long time series data. The encoder of the temporal feature modeling branch is constructed by three layers of repeatedly stacked modules, each layer of which contains a multi-head probabilistic sparse self-attention mechanism and a convolutional layer for dimensionality reduction. It can capture the long-term dependency and local characteristics of multi-source sensor data in the processing process through cascaded feature maps, while also reducing the computational complexity; the decoder of the temporal feature modeling branch is composed of two layers of modules, which extracts the evolution law of tool wear through a multi-head probabilistic sparse self-attention mechanism, uses a cross-attention mechanism to construct a global correlation between sensor data and performance indicators, and integrates the long-term dependency and local characteristics of multi-source sensor data.

2. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 1 is characterized in that: The component service status identification model is obtained by the following steps: Acquire a component service status training sample set; the component service status training sample set includes a plurality of component service status training samples, and the component service status training samples include multi-source sensor data of a key component processing process and a corresponding component service status; A classification enhancement network with cycle consistency is constructed to obtain a component service status identification model; the component service status identification model includes a generator, a discriminator and a classifier; the generator is used to simulate and generate multi-source sensor data of the processing process under different component service states according to the source domain data features of the component service status training sample set that can characterize the component service status; the multi-source sensor data of the processing process simulated and generated by the generator imitates the distribution of the source domain data to constitute the target domain data; the discriminator is used to determine whether the multi-source sensor data of the processing process input into the discriminator is the source domain data or the target domain data; the classifier is used to identify the component service status of the key component according to the input multi-source sensor data of the processing process of the key component; The generator and the discriminator are used to compete with each other to balance the distribution of the service status of each component in the component service status training sample set, and the service status of the component is identified by the classifier.

3. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 2 is characterized in that: The generator extracts key features in the source domain data through a multi-layer convolution module, and transforms the key features in the source domain data through a deep residual module so that the source domain data can adapt to the distribution of the target domain, and then restores the transformed features to the original space based on the inverse convolution operation to achieve cross-domain data generation from the source domain to the target domain; the discriminator includes a multi-layer convolution module, which helps to capture the detailed information of the input data by efficiently extracting the spatial features of the input data, and improves the discriminator's ability to distinguish between true and false data by continuously optimizing the discrimination threshold.

4. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 2 is characterized in that: Obtain component service status training sample sets, including: From the historical processing database of key components, obtain a number of multi-source sensor data of the processing process and the corresponding historical performance index degradation values; The component service status is divided according to the slope change of the historical performance index degradation value curve, and the component service status corresponding to the multi-source sensor data of each processing process is determined; the multi-source sensor data of the processing process and the corresponding component service status constitute a component service status training sample; and a plurality of component service status training samples constitute a component service status training sample set.

5. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 4 is characterized in that: After obtaining a plurality of multi-source sensor data of the processing process and corresponding historical performance index degradation values ​​from the historical processing database of the key component, the method further includes: For any multi-source sensor data of a processing process, determining whether there are missing values ​​in the historical performance indicator degradation value of the multi-source sensor data of the processing process, and obtaining a first determination result; When the first judgment result is yes, the completed historical performance indicator degradation value is obtained according to the sensor characteristics of the multi-source sensor data of the processing process through the historical performance indicator degradation value completion model; the historical performance indicator degradation value completion model is a variational autoencoder model; the variational autoencoder model includes an encoder and a decoder, the encoder of the variational autoencoder model is used to map the input sensor characteristics to a latent space to generate a mean and a standard deviation of the latent variables; the decoder of the variational autoencoder model is used to generate the completed historical performance indicator degradation value according to the latent variables in the latent space; When the first judgment result is no, no processing is performed.

6. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 5 is characterized in that: The variational autoencoder model is improved based on a Bayesian inference mechanism and a conditional modeling mechanism; the Bayesian inference mechanism assigns a prior distribution to each weight of the variational autoencoder model, and updates the prior distribution by inputting and outputting complete training data, and finally obtains a posterior distribution to fully consider the uncertainty of the sample data; the conditional modeling mechanism takes the size and processing technology of key components as a column of additional features, and inputs them into the encoder of the variational autoencoder model together with the sensor features, so as to avoid the limitation that the variational autoencoder model only relies on a single data distribution.

7. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 5 is characterized in that: The sensing features of the multi-source sensing data of the processing process include time domain features, frequency domain features and time-frequency domain features; the time domain features include peak value, mean value, standard deviation, root mean square value, skewness, kurtosis and margin, and the time domain features are used to describe the basic form, fluctuation amplitude and trend of the signal; the frequency domain features include centroid frequency, mean square frequency and variance frequency, and the frequency domain features are used to reflect the frequency distribution and energy concentration of the signal; the time-frequency domain features include high and low frequency component coefficients obtained by three-layer wavelet packet decomposition, which are used to reveal the time-frequency characteristics of the signal at different scales and provide more detailed local features for non-stationary sensing signals.

8. The equipment component performance degradation monitoring method based on digital twin spatiotemporal parallel learning according to claim 1 is characterized in that: By arranging acoustic emitters, vibration sensors and cutting force sensors near key components and current sensors in the electrical cabinet of the CNC grinder, multi-source sensor data of the machining process can be collected.

Images