A digital twin simulation method and system for industrial simulation platform

Abstract

The present invention provides a digital twin simulation method and system for an industrial simulation platform, the method comprising: collecting the operation data of the industrial training production line through a sensor array to obtain a real-time mapping data stream; constructing a digital twin model according to the real-time mapping data stream, using a semantic ontology library and a graph neural network to perform correlation analysis on the multimodal data of the physical space and the digital twin model, and obtaining a twin knowledge graph of the physical-digital dual space; according to the twin knowledge graph and a preset expert knowledge base, using a reinforcement learning method to perform collaborative learning on the operation conditions of the physical-digital dual space to obtain an elastic fault-tolerant boundary; according to the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the real-time action of the operator in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output. The present invention overcomes the limitations of the existing fixed standard action model and provides personalized guidance for operators with different proficiency levels.

Description

Technical Field

[0001] The present invention relates to the field of simulation technology, and in particular to a digital twin simulation method and system for an industrial simulation platform. Background Art

[0002] At present, the simulation system of industrial training production line has been widely used in enterprise training and teaching. Traditional simulation systems mainly use fixed standard action models for action evaluation and guidance. In this type of system, the system will pre-set a set of standard operation action sequences and corresponding evaluation criteria, and all operators need to perform operation tasks according to this fixed standard. However, in the actual industrial production environment, due to the large differences in workers' proficiency and the dynamic changes in the processing characteristics and operating environment of different workpieces, the use of a unified and fixed standard action model for evaluation often leads to the problem of too strict or too loose evaluation standards. For example, for novice operators, too strict standards will dampen their enthusiasm for learning; while for skilled workers, too loose standards will not promote their skill improvement. This lack of flexibility in the evaluation mechanism makes it impossible for the system to provide personalized guidance based on the characteristics of different operators, which ultimately affects the training effect. Summary of the invention

[0003] The main purpose of the present invention is to solve the technical problem that in the existing industrial training production line simulation system, due to the use of fixed standard action models for evaluation, it is impossible to provide personalized guidance according to the operator's proficiency, workpiece characteristics and dynamic changes in the operating environment; A first aspect of the present invention provides a digital twin simulation method for an industrial simulation platform, the digital twin simulation method for the industrial simulation platform comprising: Through the multi-type sensor array, the operation data of different levels of the industrial training production line are collected in layers and partitions to obtain real-time mapping data streams; A digital twin model is constructed according to the real-time mapping data stream, and a semantic ontology library and a graph neural network are used to perform correlation analysis on multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; According to the twin knowledge graph and the preset expert knowledge base, a reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space; According to the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the operator's real-time action in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output in the physical-digital dual space according to different deviation levels.

[0004] Optionally, in a first implementation of the first aspect of the present invention, the digital twin model is constructed according to the real-time mapping data stream, and the multimodal data of the physical space and the digital twin model are associated and analyzed using a semantic ontology library and a graph neural network to obtain a twin knowledge graph of the physical-digital dual space, including: Performing spatiotemporal feature extraction and data cleaning on the real-time mapping data stream to obtain a normalized data set of the physical space; Constructing a geometric topological structure and a dynamic attribute mapping relationship of a digital twin model according to the normalized data set to obtain a digital twin model of a digital space; Constructing node features of a graph neural network based on the digital twin model and the semantic ontology library, and establishing edge connections according to the mapping relationship and operation process of the physical-digital space to obtain an association graph network of the physical-digital dual space; Graph attention calculation and semantic clustering are performed on the association graph network to obtain a twin knowledge graph that represents the interactive relationship between the physical and digital dual spaces.

[0005] Optionally, in a second implementation of the first aspect of the present invention, performing graph attention calculation and semantic clustering on the association graph network to obtain a twin knowledge graph representing the interactive relationship between physical and digital dual spaces includes: Performing multi-head attention calculation on the nodes in the association graph network to obtain an attention weight matrix between the nodes; Perform message passing and feature aggregation on the association graph network according to the attention weight matrix to obtain updated node features; Perform hierarchical semantic clustering on the updated node features, and build a subgraph structure based on the clustering results to obtain a hierarchical knowledge subgraph set; A mapping rule between physical and digital space is established based on the hierarchical knowledge subgraph set, and the knowledge subgraphs at each layer are integrated to obtain a twin knowledge graph.

[0006] Optionally, in a third implementation of the first aspect of the present invention, the operation conditions of the physical-digital dual space are collaboratively learned using a reinforcement learning method according to the twin knowledge graph and a preset expert knowledge base to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space, including: Constructing a state space and an action space of reinforcement learning according to the operation sequence in the twin knowledge graph and the operation experience in the expert knowledge base to obtain an initial learning environment; Performing Monte Carlo sampling on the initial learning environment to simulate state transition sequences under different operating conditions to obtain a state-action trajectory set; Calculating the reward value of each operation state according to the state-action trajectory set, and updating the state-action value function based on the temporal difference algorithm to obtain the optimal operation strategy; The operating error range is dynamically divided according to the optimal operating strategy to obtain a dynamic and adaptive elastic fault-tolerant boundary of the physical-digital dual space.

[0007] Optionally, in a fourth implementation of the first aspect of the present invention, dynamically dividing the operation error range according to the optimal operation strategy to obtain a dynamic adaptive elastic fault-tolerant boundary of the physical-digital dual space includes: Performing cluster analysis on the state-action pairs in the optimal operation strategy to obtain feature clusters of different operation stages; Calculate the standard deviation and confidence interval of each operation stage according to the feature cluster to obtain an initial fault tolerance boundary; Dynamically calculating the adjustment factor of the initial fault tolerance boundary, setting the boundary scaling factor based on the operator proficiency and the workpiece characteristics, and obtaining an adaptive boundary parameter; The initial fault-tolerant boundary is dynamically updated according to the adaptive boundary parameter to obtain a flexible fault-tolerant boundary of the physical-digital dual space.

[0008] Optionally, in a fifth implementation of the first aspect of the present invention, the step of performing a graded judgment on the deviation between the real-time action of the operator in the physical space and the standard action in the digital space according to the twin knowledge graph and the elastic fault-tolerant boundary, and outputting corresponding feedback and warning information in the physical-digital dual space according to different deviation levels includes: Acquire a standard action model according to the twin knowledge graph, and perform feature extraction on the real-time action data of the operator in the physical space to obtain an action feature vector; Calculating the similarity between the action feature vector and the standard action model, and classifying the deviation levels based on the elastic fault tolerance boundary to obtain a graded deviation result; Selecting a corresponding feedback mode according to the graded deviation result, and triggering tactile feedback in the physical space or displaying guidance information in the digital space to obtain a real-time feedback signal; The operator's action trend is sequentially predicted, and the risk probability is calculated based on historical deviation data. When the risk probability exceeds the preset threshold, a warning signal is output.

[0009] Optionally, in a sixth implementation of the first aspect of the present invention, the similarity calculation is performed on the action feature vector and the standard action model, and the deviation level is divided based on the elastic fault tolerance boundary to obtain the graded deviation result, which includes: Performing dynamic time warping on the action feature vector and the standard action model to obtain an aligned feature sequence; Calculate the Euclidean distance and cosine similarity according to the aligned feature sequences to obtain a multi-dimensional similarity index; Performing weighted fusion on the multi-dimensional similarity indexes, and setting a grading threshold based on the elastic fault tolerance boundary to obtain a preliminary deviation grade; Bayesian correction is performed based on the preliminary deviation level and historical deviation data to obtain a graded deviation result.

[0010] A second aspect of the present invention provides a digital twin simulation system of an industrial simulation platform, the digital twin simulation system of the industrial simulation platform comprising: The data acquisition module is used to collect the operation data of different levels of the industrial training production line in layers and partitions through the multi-type sensor arrays to obtain real-time mapping data streams; A model building module is used to build a digital twin model according to the real-time mapping data stream, and use a semantic ontology library and a graph neural network to perform correlation analysis on the multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; A boundary generation module is used to collaboratively learn the operating conditions of the physical-digital dual space using a reinforcement learning method according to the twin knowledge graph and a preset expert knowledge base, so as to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space; The feedback warning module is used to grade the deviation between the operator's real-time actions in the physical space and the standard actions in the digital space according to the twin knowledge graph and the elastic fault-tolerant boundary, and output corresponding feedback and warning information in the physical-digital dual space according to different deviation levels.

[0011] The digital twin simulation method and system of the above industrial simulation platform collects the operation data of the industrial training production line through the sensor array to obtain a real-time mapping data stream; constructs a digital twin model based on the real-time mapping data stream, uses the semantic ontology library and the graph neural network to perform correlation analysis on the multimodal data of the physical space and the digital twin model, and obtains the twin knowledge graph of the physical-digital dual space; according to the twin knowledge graph and the preset expert knowledge base, the reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain an elastic fault-tolerant boundary; according to the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the real-time action of the operator in the physical space and the standard action in the digital space is graded and judged, and the corresponding feedback and warning information are output. The present invention overcomes the limitations of the existing fixed standard action model and provides personalized guidance for operators with different proficiency levels.

[0012] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0013] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of a first embodiment of a digital twin simulation method for an industrial simulation platform in an embodiment of the present invention; Figure 2 It is a schematic diagram of an embodiment of a digital twin simulation system of an industrial simulation platform in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0017] To facilitate understanding of this embodiment, a digital twin simulation method for an industrial simulation platform disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps: 101. Through the multi-type sensor array, the operation data of different levels of the industrial training production line are collected in layers and partitions to obtain real-time mapping data streams; In one embodiment of the present invention, the arrangement of multi-type sensor arrays covers the operation links of operators in the production line, the operation parts of the equipment, and the environmental interaction areas. First, high-precision accelerometers and gyroscopes are installed near the mechanical processing unit of the production line to sense the motion data of the rotating parts and the workpiece handling process, so as to record the operation speed and posture changes. At this time, it is necessary to set a fixed frequency sampling strategy inside the accelerometer to continuously obtain a stable motion sequence. The data of the accelerometer and the gyroscope will be transmitted to the data aggregation unit through the local area network. The data aggregation unit will uniformly manage the timestamps to ensure that the timing of each sensor is consistent. Then, an infrared depth camera and a visual recognition module are configured at the workstation where the operator is located to monitor limb movements, standing distances, and operating postures. The infrared depth camera needs to calibrate the viewing angle and ranging reference during installation, and obtain accurate depth imaging parameters through the calibration plate. After the calibration is completed, depth information that can be used for three-dimensional model reconstruction will be generated, and the visual recognition module will extract bone points based on the synchronized video frames. The depth matrix and bone coordinates output by the infrared depth camera will be synchronously transmitted to the data aggregation unit and aligned with the timing tags of the accelerometer and gyroscope. Industrial environmental sensors are deployed in the temperature, humidity, smoke and light distribution areas of the workshop. The sampling frequency of these sensors is different from that of mechanical sensors and visual sensors, and independent data caching strategies need to be set to avoid interference with high-speed data streams. Temperature and humidity sensors can continuously detect air parameters within a certain range and send these parameters to the data aggregation unit. Smoke and light sensors transmit real-time values ​​to the status monitoring node. When the smoke concentration or light brightness exceeds the preset threshold, the data marking program is triggered to form an additional timestamp event record. Radio frequency identification (RFID) or proximity sensing modules are also installed on operating tools and workpieces to identify the flow path of specific parts. The relevant data helps to distinguish different material numbers during operation to avoid confusion. The industrial control system itself provides interfaces between PLC (programmable logic controller) and SCADA (data acquisition and monitoring system), which include industrial real-time variables such as equipment speed, spindle load, current and voltage parameters. The data aggregation unit continuously monitors the network output ports of PLC and SCADA, and incorporates these industrial real-time variables into a unified data structure according to the same time base. In this way, sensor data at all levels can be integrated into a real-time mapping data stream containing operator actions, equipment status, environmental parameters and workpiece identification in the same timeline. This data stream will be partitioned and managed by category in the structured storage module. For example, motion capture data is stored in the skeletal motion database, environmental sensing data is stored in the environmental monitoring database, and industrial equipment status parameters are stored in the production process database. Whenever new sensor data arrives at the data aggregation unit, the real-time mapping data stream will add a comprehensive record with a timestamp. The record content contains multiple subfields and matches the corresponding index identifier.In this way, the operational data of the industrial training production line at different levels are reasonably split and integrated, and the results of various sensors can obtain complete action and process mapping within the same time frame, thereby forming a real-time mapping data stream that can be used by the digital twin model.

[0018] 102. Build a digital twin model based on real-time mapping data streams, and use semantic ontology library and graph neural network to perform correlation analysis on multimodal data of physical space and digital twin model to obtain twin knowledge graph of physical-digital dual space; In one embodiment of the present invention, the digital twin model is constructed according to the real-time mapping data stream, and the multimodal data of the physical space and the digital twin model are associated with each other using a semantic ontology library and a graph neural network to obtain a twin knowledge graph of the physical-digital dual space, including: performing spatiotemporal feature extraction and data cleaning on the real-time mapping data stream to obtain a normalized data set of the physical space; constructing the geometric topological structure and dynamic attribute mapping relationship of the digital twin model according to the normalized data set to obtain a digital twin model of the digital space; constructing the node features of the graph neural network based on the digital twin model and the semantic ontology library, and establishing edge connections according to the mapping relationship and operation process of the physical-digital space to obtain an association graph network of the physical-digital dual space; performing graph attention calculation and semantic clustering on the association graph network to obtain a twin knowledge graph that characterizes the interaction relationship between the physical and digital dual spaces.

[0019] Specifically, the real-time mapping data stream is subjected to spatiotemporal feature extraction and data cleaning to obtain a normalized data set of the physical space; in this process, the data streams from different sources such as the acceleration sensor, the visual recognition module, and the environmental detection module are aligned according to the timestamp, and the dimensions such as the action characteristics, the device operation status, and the environmental parameters are distinguished according to the data type, and then the continuous sequence of posture changes and the corresponding time tags are extracted from the action information by means of spatiotemporal window segmentation, so as to obtain the basic vector of the behavior pattern. At this time, it is necessary to perform a removal or interpolation strategy for some abnormal or missing values ​​to ensure that the numerical dimension maintains integrity and coherence. For the device operation information, it is necessary to screen those scattered fragments that do not meet the acquisition threshold, and use statistical methods and domain knowledge to correct or discard the values ​​that obviously deviate from the normal range. The sampling frequency of the environmental detection data is different from that of other channels, and it is necessary to merge it into a corresponding structured record under a unified time series benchmark to facilitate the real mapping of multimodal information in the same period. There are also large differences between action data and device data, and it is necessary to pre-define the corresponding field labels and type constraints in the storage structure to avoid data type conflicts in subsequent association analysis. In this way, all the original information is reshaped into a normalized data set containing multi-dimensional vectors, removing dirty data and having a consistent time series benchmark, and retaining index references to different data sources at the record level, making it easier to establish a corresponding relationship between the physical and digital spaces. Specifically, the geometric topological structure and dynamic attribute mapping relationship of the digital twin model are constructed according to the normalized data set to obtain the digital twin model of the digital space; this operation first builds the corresponding geometric topological skeleton according to the layout information of the physical equipment and process flow on the production line, and uses graphics or three-dimensional grids to express the spatial distribution of processing units, transmission paths and workstations. Then, combined with the dynamic attributes extracted from the normalized data set, real-time values ​​such as operating speed, load parameters or motion information are added to each key node, so that these nodes not only have coordinate characteristics, but also can reflect the operating status of the equipment in the physical space. Motion capture and environmental sensing data will also be mapped to the corresponding elements in the digital model to form the motion trajectory of the robot or operator in the three-dimensional scene. The construction process of the topological structure needs to set the edge relationship according to the connection method of the actual equipment, for example, the front and rear stations on the same conveyor belt are defined as upstream and downstream nodes. Once the model is completed, the spatial connection status between the processing units will be presented at the visualization level, and it will have continuously updated attribute fields to record real-time changing numerical information. The attributes between nodes and edges will maintain time synchronization. Once the arrival of a new round of data is detected, the status of the corresponding nodes and the flow information of the connecting edges will be refreshed. Then all the geometric and attribute information together constitute a complete twin model in the digital space. Specifically, the node features of the graph neural network are constructed based on the digital twin model and the semantic ontology library, and the edge connection is established according to the mapping relationship between the physical and digital space and the operation process to obtain the association graph network of the physical and digital dual space; at this stage, in order to correspond the geometric nodes in the digital space with the objects in the actual physical space, it is necessary to use the terms and context information accumulated in the semantic ontology library to annotate the meaning of the attributes of the nodes or edges, and ensure that elements such as action types, process steps and equipment functions can have recognizable semantic labels. When receiving these labels, the graph neural network will construct a high-dimensional feature vector for each node, including the operating indicators, maintenance cycles or process restrictions of the equipment represented by the node, and also encode the behavioral characteristics of the operator or robot arm into a numerical form that can be calculated by the network. At this time, according to the operation process and production line layout, edge connections are created for the nodes in the graph structure, examples of which include the sequential relationship between workstation A and workstation B, or multiple associations between the same equipment and different operation steps. The generation of the association graph network needs to take the operation timing into consideration, combine the action information of multiple time periods into a node sequence, and use the graph structure to maintain the intertwined relationship between timing and space. This network includes both devices and people in the physical space, and also integrates the geometric topology and attribute mapping in the digital space, achieving a synchronous description of the two spaces; Specifically, the association graph network is subjected to graph attention calculation and semantic clustering to obtain a twin knowledge graph that characterizes the interactive relationship between physical and digital dual spaces; this part relies on the multi-head attention mechanism in the graph neural network to combine the feature similarity between nodes with the correlation in the actual operation process, and calculate the attention weights of different nodes in the form of a matrix. The higher the weight value, the more significant the influence of the node on the overall network. The attention mechanism iteratively updates the feature representation of the node so that nodes in the same process link or with similar operating states have similar vector distributions. Then, the semantic clustering step is entered. The system performs hierarchical division based on the embedded representation and label information of the node, and groups several nodes with similar functions or close connection in time sequence into the same subgraph. Processing equipment and corresponding tools can be aggregated into the same semantic cluster, and the skeleton nodes of the operator and the process steps they perform will also be judged as a significant interaction group. After the sub-graph structure is integrated, a knowledge graph will appear that reflects the actual coupling relationship between the physical and digital dual spaces in terms of operational processes and spatial distribution. This graph can display the corresponding semantic meanings and dynamic parameters on nodes and edges, thereby forming a twin knowledge graph for industrial training.

[0020] Furthermore, the graph attention calculation and semantic clustering are performed on the association graph network to obtain a twin knowledge graph that characterizes the interactive relationship between physical and digital dual spaces, including: performing multi-head attention calculation on the nodes in the association graph network to obtain an attention weight matrix between nodes; performing message passing and feature aggregation on the association graph network according to the attention weight matrix to obtain updated node features; performing hierarchical semantic clustering on the updated node features, and constructing a subgraph structure based on the clustering results to obtain a hierarchical knowledge subgraph set; establishing a mapping rule for the physical-digital space according to the hierarchical knowledge subgraph set, and integrating the knowledge subgraphs of each layer to obtain a twin knowledge graph.

[0021] Specifically, multi-head attention calculation is performed on the nodes in the associated graph network to obtain the attention weight matrix between nodes; in this process, the graph neural network will set multiple sets of learnable linear transformation parameters for each node to map different levels of feature channels respectively. Each set of parameters will score the edges in the associated graph network to form an attention energy value, and then perform a normalization operation between multiple neighbors of the same node to obtain the importance coefficient of each neighbor node to the current node. In order to express the calculation process of multi-head attention, an attention scoring function is introduced To indicate the Head attention at node With Node The original energy value between , and then apply Softtmax activation to obtain the normalized weight . An exponential transformation can be cited ; in Representation Node The neighbor set of Usually obtained by a feedforward network or dot product calculation. Multi-head attention allows the same node relationship to be evaluated from different perspectives, making it easier to capture the complex dependencies of nodes in multiple dimensions such as operation processes, semantic concepts, and spatial locations. In order to connect this process with the contextual knowledge of the physical-digital dual space, the attribute type of the node in the semantic ontology library, such as equipment features, process step labels, etc., is read during the attention calculation process, and this information is embedded into the generated 's input vector. In this way, each edge can reflect the difference in device type, operation timing or spatial connection at the level of attention energy value. The attention weight matrix between nodes will be continuously tuned with the iterative update of the network to avoid noise interference in the multi-head mechanism. After multiple rounds of forward propagation and backpropagation, the attention parameters will converge to a numerical distribution that can highlight the key relationships, so that the dependencies between nodes will gradually form a clear primary and secondary distinction in the graph structure. The more multi-head attention heads there are, the more detailed the splitting of feature channels will be, but it is necessary to reasonably allocate computing resources and parameter scale. Once the attention weight matrix between nodes is generated, weakly associated edges can be filtered out by sparsification or thresholding, so that the association graph network tends to highlight highly correlated node pairs, which is convenient for subsequent feature aggregation and clustering analysis.

[0022] According to the attention weight matrix, the associated graph network is message-passed and feature-aggregated to obtain updated node features; during message passing, the graph neural network will weightedly merge the feature vectors of neighboring nodes according to the attention weights, and then superimpose them with the feature vector of the current node itself. To show the form of the aggregation process, the following example can be used: ; in represents the updated node features, is the activation function, is the number of longs, For the The attention weight of the head, is the corresponding learnable mapping matrix, Represents neighbor nodes The multi-head mechanism merges the weighted summation results of different heads and obtains the new features of the current node through nonlinear mapping. This process requires reading the edge relationship in the association graph network and controlling the direction of message transmission according to the timestamp or process step mark to avoid mistakenly mixing the reverse information into the current node. The goal of updating node features is to enable the representation vector of the node to absorb the spatial position, process link and behavior pattern of the neighboring nodes, and then show the difference in the high-dimensional feature space that is consistent with the actual operation situation. If some nodes are far apart in the process flow, and the attention weight matrix shows that the two lack direct connection, then the corresponding message transmission weight value will be low, resulting in a small contribution ratio in the merging stage. This mechanism can highlight the feature integration value of local associations in the graph structure and suppress the noise brought by irrelevant nodes. At this time, the new process will also consider the self-connection weight of the node, so that the node retains the original feature component to a certain extent, avoiding the smoothing of the information due to the reliance on external input. After multiple rounds of iterations, the vector distribution of the node gradually shows an effective encoding of the details of the physical-digital dual space interaction.

[0023] The updated node features are hierarchically semantically clustered, and a subgraph structure is constructed based on the clustering results to obtain a hierarchical knowledge subgraph set. In this stage, when the nodes are divided by the semantic clustering algorithm, the feature vector updated by the attention network is used as a mark in the input vector space, and nodes with similar operation processes or equipment types are clustered into the same cluster to facilitate the identification of actual business meanings such as production units, operation teams or functional modules. Hierarchical semantic clustering will first measure the distance between nodes with a similarity index. If two nodes have a smaller distance on the feature vector, they have a higher probability of being identified as belonging to the same cluster. When implementing the clustering algorithm, it is necessary to combine additional label information, such as equipment classification, process sequence number or job role identification. These labels come from the semantic ontology library and are not the main reference for the spatial position of the nodes, but are summarized according to the similarity in the process or attributes. When the clustering process is executed to a certain level, multiple semantic clusters will appear, and the nodes in each semantic cluster usually present similar functional positioning and behavior patterns. The subgraph structure is then formed through the screening and reconstruction of edge relationships, retaining high-frequency interaction edges within the same cluster, or merging weakly associated edges across clusters into a higher-level graph. This hierarchical subgraph can express the multi-level logic of the actual training production line, such as the hierarchical segmentation of local equipment clusters and the entire line system. The resulting subgraph set contains nodes, edges, and their respective attribute information, and retains the attention features of the nodes after the clustering algorithm is executed, so that the hierarchical subgraph remains traceable at different aggregation scales. Timing factors are also taken into account during hierarchical semantic clustering. When there are obvious differences in node features or operation behaviors in different time periods, the clustering module will treat these differences as separate groups, allowing the subgraph to remain compatible with temporal evolution; According to the hierarchical knowledge subgraph set, the mapping rules of physical-digital space are established, and the knowledge subgraphs of each layer are integrated to obtain the twin knowledge graph. During integration, it is necessary to retrieve the mapping relationships that can be paired in the physical equipment, operators, action sequences and simulation model nodes on each layer of the subgraph, and customize these pairing relationships into a list of items according to the standard description of the semantic ontology library. If the physical layer device is defined as a certain label, such as "CNC lathe", and the corresponding node in the digital model contains the "CNC processing unit" label, the mapping rules between the two can be established through hash matching or vocabulary similarity calculation. The action sequence nodes and the skeleton trajectory nodes in the digital simulation model are also mapped and matched. Once the identifiers of the same time period and the same operation gesture are identified, the edge information representing the synchronization relationship is generated in the knowledge graph. When all the hierarchical subgraphs are processed, they can be merged to form a more complete cross-layer interaction network, which contains information about each layer of equipment groups, operation processes and digital simulation nodes, and can also reflect the multiple intersections of action coordination and process connection. The final twin knowledge graph includes not only nodes and edges, but also dynamic elements such as node attribute vectors and attention weights. In this way, the entire map establishes traceable mapping rules between the physical and digital spaces. Any specific equipment, personnel or action node can be traced back to the real operation scenario, and the correlation between nodes will reflect the core interaction relationship based on attention calculation and clustering results, thereby showing the dynamic evolution logic of the industrial training production line in multiple dimensions.

[0024] 103. Based on the twin knowledge graph and the preset expert knowledge base, the reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain the dynamic adaptive elastic fault-tolerant boundary of the physical-digital dual space; In one embodiment of the present invention, the operation conditions of the physical-digital dual space are collaboratively learned using a reinforcement learning method based on the twin knowledge graph and the preset expert knowledge base to obtain a dynamic and adaptive elastic fault-tolerant boundary of the physical-digital dual space, including: constructing a state space and an action space of reinforcement learning according to the operation sequence in the twin knowledge graph and the operation experience in the expert knowledge base to obtain an initial learning environment; performing Monte Carlo sampling on the initial learning environment to simulate the state transition sequence under different operating conditions to obtain a state-action trajectory set; calculating the reward value of each operating state according to the state-action trajectory set, and updating the state-action value function based on the temporal difference algorithm to obtain the optimal operation strategy; dynamically dividing the operation error range according to the optimal operation strategy to obtain a dynamic and adaptive elastic fault-tolerant boundary of the physical-digital dual space.

[0025] Specifically, the state space and action space of reinforcement learning are constructed according to the operation sequence in the twin knowledge graph and the operation experience in the expert knowledge base to obtain the initial learning environment; in this link, it is necessary to first retrieve all the operation nodes in the twin knowledge graph, and combine the description of the process steps and equipment operation elements in the expert knowledge base to convert this information into a state vector that can be represented discretely or continuously. For example, the action node contains the current load of the equipment, the action sequence and related parameters of the operator in the time dimension, and the expert knowledge base contains the processing parameter threshold, action duration requirements and safety criteria. The system merges this information according to the business logic to obtain the corresponding state description under each time slice. The construction of the action space is also based on the operation experience provided by the expert knowledge base, and the specific adjustment means or correction methods are defined as executable actions. Common examples include adjusting the feed speed, changing the tool, or intervening in human-machine collaboration in a certain process link. In order to establish a rigorous mapping rule between state and action, it is necessary to retrieve different industry standards or safety criteria, and combine the process sequence and physical space layout reflected in the twin knowledge graph to define a reasonable value range for each dimension in the state space. When the numerical state exceeds the equipment safety threshold recorded in the expert knowledge base, the action space will add additional emergency shutdown or alarm options to prevent the system from deviating from legal working conditions during reinforcement learning exploration. The initial learning environment is thus formed, including state descriptions of the physical and digital dual spaces and a list of executable actions derived from expert experience. The environment will set up internal simulation logic when it is implemented. When the reinforcement learning algorithm selects a specific action, the environment will update the state vector based on the graph relationship and expert rules, and present it to the algorithm at the next time step. This ensures that reinforcement learning is iterated under conditions that are as close to the actual operating process as possible; Monte Carlo sampling is performed on the initial learning environment to simulate the state transition sequence under different operating conditions to obtain a set of state-action trajectories; this process requires randomly selecting the initial state in reinforcement learning and selecting the next action from the action space at each time step, and then advancing the simulation results of the physical-digital dual space according to the process sequence or equipment association relationship mapped by the twin knowledge graph. Monte Carlo sampling not only adds a certain degree of randomness to the action selection, but also diversifies the setting of the initial state, so that the sampling sequence covers a variety of working conditions such as high load state, normal load state and low load state. Each sampling will generate one or more state-action transfer trajectories, each node in the trajectory represents the state value and the action performed at a certain moment, and the node immediately following it represents the evolution of the state in the next time step. In order to make the sampling more comprehensive, the process will introduce parameters of different personnel proficiency or equipment operation cycle in the operation sequence, so that the system can experience a wider range of operation scenarios during simulation. After the action is executed, the environment updates the state of the next time step based on the equipment operation information in the twin knowledge graph and the safety threshold in the expert knowledge base, which may generate equipment failure information, alarm signals, or process completion marks. Each trajectory in the trajectory set starts from the starting state, and after multiple action executions, it finally reaches the end condition or triggers a fault state. After sampling, the trajectory set is stored and used in the form of tables or sequence data for subsequent reward calculations and value function updates. The sampling scale can be increased by adjusting the number of Monte Carlo rounds. Each round initializes the state under different random seeds, and introduces discrete event interpolation operations in some rounds, so that uncommon extreme situations in the state space can also be included in the statistics, making the sampling more extensive; The reward value of each operation state is calculated according to the state-action trajectory set, and the state-action value function is updated based on the temporal difference algorithm to obtain the optimal operation strategy; the formulation of the reward value needs to refer to the operation specifications and safety standards in the expert knowledge base, and efficiency indicators or cost considerations will also be incorporated. If the operation in a certain state meets the judgment conditions for precise processing and safe production in the expert experience, a positive reward will be given to the corresponding state-action combination. If it deviates from the optimal parameter range or causes potential accidents, a negative reward will follow. The system will calculate the discounted return for each sampled trajectory set, convert the accumulated gains or losses in the future back to the current moment, and form a value assessment for each state-action pair. Since reinforcement learning introduces a temporal difference algorithm here, such as a mechanism based on Q-learning or SARSA, the value function update method will use an iterative equation to gradually approach the state value or action value to the optimal solution. When there are several consecutive state-action accumulations of positive rewards in the trajectory, the value function of the corresponding strategy will move in the direction of high returns; if the trajectory contains recurring negative rewards, this unfavorable information will also be propagated to the previous state-action pair through temporal differences, so that the strategy tends to avoid such operations in the next decision. To improve search efficiency, temporal differences will update the value estimate immediately after the state transfer, avoiding waiting for the entire sequence to end and then backtracking at once. After multiple rounds of iterations, this process will gradually converge to the optimal operating strategy that can balance safety, efficiency, and compliance in the physical-digital dual space, and present a value function mapping that evaluates each state-action pair at the numerical level, so that the algorithm can quickly find the appropriate action when faced with a new state; According to the optimal operation strategy, the operation error range is dynamically divided into boundaries to obtain a dynamic and adaptive elastic fault-tolerant boundary of the physical-digital dual space. The definition of the operation error range will be combined with the process tolerance in the expert knowledge base and the tolerance for human misoperation, and a certain width of error limit will be reserved for beginners or equipment debugging stage, while the corresponding tolerance will be tightened in the skilled or high-speed processing mode. According to the results of the optimal strategy, reinforcement learning will identify the trade-off between operation efficiency and safety, and quantify the benefits or penalties generated by various operation results in the state value function. The system will extract the strategy distribution on different state space sections according to the value function, identify the critical point that separates the safe zone and the dangerous zone in each operation dimension, and thus set the elastic fault-tolerant boundary. At this time, if an operation deviates from the optimal trajectory but is still within the safety threshold, the strategy will give a lower negative reward, reflecting the room for adjustment. If it exceeds this boundary, it will bring serious negative evaluation and trigger a warning or stop action. After this division, the physical and digital dual spaces will execute tolerance updates synchronously. When the simulation value in the digital space enters the corresponding area, an early warning will be issued, and the actual operation in the physical space will also be monitored and intervened in time. The dynamically adaptive elastic fault tolerance boundary not only includes the limits of coordinates, speed or equipment load at the numerical level, but also involves the verification of action timing and process sequence. In this way, regardless of the proficiency of the operator, there will be a corresponding elastic range to ensure safety, and clear guidance and constraints will be obtained during the learning or improvement stage, forming a more refined and personalized training evaluation mechanism.

[0026] Furthermore, the dynamic boundary division of the operation error range according to the optimal operation strategy to obtain the elastic fault-tolerant boundary of the physical-digital dual space dynamically adaptive includes: performing cluster analysis on the state-action pairs in the optimal operation strategy to obtain feature clusters of different operation stages; calculating the standard deviation and confidence interval of each operation stage according to the feature cluster to obtain the initial fault-tolerant boundary; dynamically calculating the adjustment factor of the initial fault-tolerant boundary, setting the boundary scaling coefficient based on the operator proficiency and the workpiece characteristics to obtain the adaptive boundary parameters; dynamically updating the initial fault-tolerant boundary according to the adaptive boundary parameters to obtain the elastic fault-tolerant boundary of the physical-digital dual space.

[0027] Specifically, a cluster analysis is performed on the state-action pairs in the optimal operation strategy to obtain feature clusters of different operation stages; this step requires representing the state-action pairs that appear in the optimal strategy as multi-dimensional vectors and placing these vectors in the same feature space. In order to reduce redundant dimensions and retain the main information, a dimensionality reduction algorithm can be used to map high-dimensional vectors to a relatively compact low-dimensional subspace. After the feature vector is reduced in dimension, the clustering module groups the samples in the vector space based on similarity or distance metrics, and finally divides the clusters of several operation stages. Each cluster represents a group of state-action pairs that are highly similar in terms of process flow, operation load or personnel action characteristics. If a cluster is defined as , then it contains Equivalent feature vectors can be used on this basis to estimate the center and dispersion of the cluster using kernel methods or prototype-based algorithms. The clustering results of different operation stages will be accompanied by their own sample distribution information. For example, there is a central trend or obvious difference in the two dimensions of machine tool load and feed speed. In this way, the optimal strategy can be globally split into multiple stage modes. Each mode corresponds to a consistent operating style, equipment status or action form. The standard deviation and confidence interval of each operation stage are calculated based on the feature cluster to obtain the initial fault-tolerant boundary. In this process, it is necessary to perform statistical analysis on each cluster obtained by clustering, and construct a multidimensional confidence interval to describe the main distribution area of ​​the vector samples within the cluster. If the cluster The data are considered to approximately follow a multivariate normal distribution, and the mean vector is denoted as , the covariance matrix is ​​denoted as , then the preliminary multidimensional confidence interval can be described in the form of an ellipsoid, which can be described using a constraint ; in yes The chi-square critical value at the significance level is Represents the vector dimension. The ellipsoid depicts the area where most samples are distributed in a statistical sense, and can also be used as a benchmark for the initial tolerance range. If non-normal distribution characteristics are found in some clusters, non-parametric methods can be used to estimate the contour surface corresponding to the density distribution of the vector, and then define the core area where most samples fall. For dimensions that can be quantified individually, such as spindle speed, tool wear or ambient temperature, standard deviation and confidence interval can also be used to give upper and lower bounds of the interval according to traditional one-dimensional statistical formulas. After completing this step, each operation stage has an interval or multidimensional ellipsoid set that can reflect the concentrated distribution characteristics, which is called the initial tolerance boundary of that stage. The initial fault tolerance boundary is dynamically adjusted by calculating the factor, and the boundary scaling factor is set based on the operator proficiency and workpiece characteristics to obtain the adaptive boundary parameters; this process requires introducing a set of weights or exponential factors into the initial fault tolerance boundary to represent the skill levels of different personnel and the accuracy requirements of the parts to be processed. If the scaling factor is defined as ,in Indicates the operator proficiency index, represents the key parameters of the workpiece (hardness, dimensional tolerance, etc.), then the overall adjustment can be written as ; This formula means that in the original covariance matrix A variable magnification is introduced at the inverse matrix position ,like If it is less than 1, it means the boundary is relaxed, and if it is less than 1, it means it is tightened. The larger the value, the wider the acceptable error range will be. The more stringent the workpiece requirements are, the lower the , pushing the fault tolerance range to shrink in a more precise direction. The scaling operation can achieve synchronous scaling on the multi-dimensional ellipsoid, and can also perform additive or multiplicative corrections on specific dimensions, so as to make special adjustments on sensitive dimensions such as tool wear or operation time; the initial fault tolerance boundary is dynamically updated according to the adaptive boundary parameters to obtain a flexible fault tolerance boundary of the physical-digital dual space. When updating, it is necessary to associate the distribution information generated during clustering with the dynamic adjustment factor. Different stages have independent scaling factors, which will also be updated over time or with changes in the operator's skill level. When the system detects that the personnel's operating method or workpiece type has switched, the program will immediately recalibrate the corresponding , and perform new numerical calculations on the ellipsoid or confidence interval to obtain the tolerance upper and lower bounds under the current transient conditions. The digital twin model maps these upper and lower bounds to the simulation view or virtual axis, and the physical space monitors the state vector of the actual operation in real time through sensors. If detected Exceed If the threshold is exceeded, it will determine that the operation has serious deviations and trigger an early warning or automatic intervention. In this way, the fault tolerance strategy of the entire production line is no longer fixed to a single value, but is based on the comprehensive calculation results of cluster analysis, expert experience and real-time status, ultimately achieving adaptive elastic control of different personnel, different workpieces and different time periods, so that the industrial training production line can achieve more refined fault tolerance management and safety assurance.

[0028] 104. Based on the twin knowledge graph and elastic fault-tolerant boundary, the deviation between the operator's real-time action in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output in the physical-digital dual space according to different deviation levels.

[0029] In one embodiment of the present invention, the deviation between the real-time action of the operator in the physical space and the standard action in the digital space is graded and judged according to the twin knowledge graph and the elastic fault-tolerant boundary, and corresponding feedback and warning information is output in the physical-digital dual space according to different deviation levels, including: obtaining a standard action model according to the twin knowledge graph, and extracting features of the real-time action data of the operator in the physical space to obtain an action feature vector; calculating the similarity between the action feature vector and the standard action model, and dividing the deviation level based on the elastic fault-tolerant boundary to obtain a graded deviation result; selecting a corresponding feedback mode according to the graded deviation result, and triggering tactile feedback in the physical space or displaying guidance information in the digital space to obtain a real-time feedback signal; performing sequence prediction on the operator's action trend, and calculating the risk probability based on historical deviation data, and outputting a warning signal when the risk probability exceeds a preset threshold.

[0030] Specifically, a standard action model is obtained according to the twin knowledge graph, and the real-time action data of the operator in the physical space is feature extracted to obtain an action feature vector; in this process, it is necessary to first retrieve the standard action information corresponding to the target operation link from the twin knowledge graph, including the key frame parameters of each joint or posture in the time dimension. The model usually stores a standardized action sequence, the angle range of each joint and the timing annotation. The sensor array in the physical space monitors the operator's limb posture, force direction or operation rhythm, and aligns these raw data by timestamp or frame number. Subsequently, noise reduction and posture recognition are performed on the measured time series data to convert the joint point coordinates or acceleration vectors into more abstract action descriptions. If some joint points are blocked or lost, it is necessary to restore the posture continuity in the data structure by interpolation or compensation so that the feature extraction algorithm keeps the output dimension consistent. After obtaining the basic posture, the operator's action process can be expressed in a variety of ways such as bone vectors, gesture trajectories or speed curves. If the operation involves device control or human-computer interaction, the joystick displacement, button trigger timing and body joint points can be included in the feature dimension. The generated action feature vector will retain the continuous relationship between the time sequence nodes, which is convenient for subsequent matching with the standard action model, while avoiding the loss of details such as turning actions or local fine-tuning. The feature vector formed in this way has a high degree of recognition and can correspond to the elements of the standard action one by one, realizing more targeted deviation measurement; The similarity between the action feature vector and the standard action model is calculated, and the deviation level is divided based on the elastic fault tolerance boundary to obtain the graded deviation result; when the system obtains the action feature vector of the operator, the key frame or continuous posture sequence in the standard action model is called, and the two are compared frame by frame or in multiple dimensions in parallel through the similarity algorithm. Common methods include aligning the feature vector and the model vector on the time axis, then calculating the deviation in both the posture dimension and the duration dimension, and then quantifying it with distance or similarity indicators. If the action contains multiple sub-steps or more complex trajectory deformations, local alignment needs to be set to avoid large phase misalignment caused by global alignment. When the difference values ​​of each time slice are summarized, the system will compare the results with the elastic fault tolerance boundary to determine the interval in which the current deviation is located. If the deviation value is within the minimum interval, the action is regarded as a slight deviation and only needs to be prompted for improvement; if it is close to the middle section, it enters the moderate deviation range, which often means that there are hidden dangers in efficiency or safety; if the deviation accumulates and exceeds the maximum threshold of the elastic fault tolerance boundary, it is identified as a serious deviation and requires immediate intervention, correction or warning. The grading process can respond flexibly according to the defined upper and lower limits of the error. The tolerance range under different operator or workpiece requirements is not fixed, so adaptability can be reflected in the judgment stage. The grading results can be marked as numerical levels or text labels to drive subsequent feedback forms, so that the entire process can have classified control over minor differences or major errors; According to the graded deviation result, the corresponding feedback mode is selected, and tactile feedback is triggered in the physical space or guidance information is displayed in the digital space to obtain a real-time feedback signal; this step will select an appropriate feedback method according to the previous deviation level mark. If the system identifies a mild deviation, a slight vibration reminder will be issued on the vibration wristband worn by the operator, or a simple text prompt will be displayed on the digital twin model interface to guide the operator to pay attention. Moderate deviations usually mean that more detailed correction information is required, such as virtual guidance arrows or voice explanations appearing in the operator's field of view, indicating which joint or process has obvious errors. In severe deviation scenarios, the system will initiate emergency intervention strategies, such as temporarily stopping the operation of the robot arm or production equipment, and popping up a highlighted warning message on the operator's head-mounted display to ensure the safety of the production line. The feedback form of the physical space mainly relies on wearable devices, lights or warning sounds, while the feedback of the digital space will display specific correction suggestions through eye-catching animations or icons in the visual interface. Once a certain deviation level is determined, the feedback module will respond in a very short time and deliver the required information flow to the operator. In this way, operators can not only perceive the gap between their own actions and the standard in a timely manner, but also make adjustments in the shortest possible period, thereby reducing delays or waste in training or actual production; The operator's action trend is predicted in sequence, and the risk probability is calculated based on the historical deviation data. When the risk probability exceeds the preset threshold, an early warning signal is output. This stage introduces time series modeling of action evolution. By analyzing the operator's recent posture records or operation gestures, the general direction of the next behavior is determined. If the sequence trend shows a direction of expanding deviation in the next time slice, the system will use the statistical results in the historical deviation data to infer the probability of safety hazards or efficiency losses. The risk probability calculation is usually based on indicators such as error type, frequency of occurrence, and deviation amplitude in the past period of time, and combined with the current action trend to infer the deviation level in the next stage. If this action has gradually approached the upper limit of the elastic fault tolerance boundary or is accelerating away from the standard, the risk probability will increase significantly. When the probability exceeds a certain strict control threshold, an early warning will be issued to the operator and the management end, indicating that a high-risk operation state is about to be entered or there is an overload risk. The digital space will also display relevant warning information, and the physical production line may link the safety device to trigger an emergency brake or warning light signal, thereby reducing the duration of potential dangers or non-compliant operations. This prediction and risk assessment mechanism can, to a certain extent, detect adverse trends in advance, allowing operators to make corrections before larger errors occur. It can also archive high-risk scenarios through logging for subsequent analysts to trace and improve them.

[0031] Furthermore, the similarity between the action feature vector and the standard action model is calculated, and the deviation level is divided based on the elastic fault-tolerant boundary to obtain a graded deviation result, including: dynamic time warping of the action feature vector and the standard action model to obtain an aligned feature sequence; calculating the Euclidean distance and cosine similarity based on the aligned feature sequence to obtain a multidimensional similarity index; weighted fusion of the multidimensional similarity index, and setting a graded threshold based on the elastic fault-tolerant boundary to obtain a preliminary deviation level; and performing Bayesian correction based on the preliminary deviation level and historical deviation data to obtain a graded deviation result.

[0032] Specifically, the action feature vector and the standard action model are dynamically time-warped to obtain an aligned feature sequence; the core of this link is to find the optimal time mapping relationship between the physically collected action sequence and the standard action model, so as to eliminate the mismatch in execution speed or rhythm. If the physical action sequence is recorded as , the standard model sequence is denoted by , can be done in two-dimensional dynamic programming grid The cumulative cost function is defined above ; in Indicates Frame and The local distance between frames is usually calculated based on a multi-dimensional posture difference metric. The upper left corner and the lower right corner of the grid correspond to the start and end positions of the sequence, respectively. Dynamic programming searches for the minimum cumulative cost path at the global level. If there is a large difference in the sequence duration, the algorithm inserts repeated matches or jump operations at appropriate nodes so that the two sequences can be well aligned in the key action segments. Finally, an optimal matching trajectory and a pair of sequences of equal length are output. and After alignment, each time frame can be directly compared to grasp the source of deviation of local action or posture. This process can be performed in multi-dimensional space. If the operator's action includes multiple factors such as position, speed or angle, local distance can be accumulated dimension by dimension. The alignment sequence will be recorded in the data structure for subsequent similarity calculation or classification judgment. The Euclidean distance and cosine similarity are calculated based on the aligned feature sequence to obtain a multi-dimensional similarity index; in the alignment result, each time kernel index You can find the physical action vector With standard motion vector In order to more comprehensively reflect the degree of deviation, the system calculates similarity in terms of amplitude and direction. The amplitude difference can be measured by Euclidean distance, and the directional consistency can be characterized by cosine similarity. The absolute difference in the amplitude of the action is expressed, while the cosine similarity CS_i focuses on the angle. When globally summarized, the ED_i or The data are combined into several numerical indicators by averaging, weighting or extreme value, and then counted in multiple dimensions (position, speed, rotation angle, etc.). If a peak deviation or continuous deviation occurs in a certain dimension, the system will mark it as a potential risk point, providing a fine-grained basis for subsequent threshold judgment. The multi-dimensional similarity indicator vector thus formed can be summarized as: , where each component corresponds to a different evaluation focus; weighted fusion is performed on all multidimensional similarity indicators, and a grading threshold is set based on the elastic fault tolerance boundary to obtain a preliminary deviation level; when the system obtains multidimensional indicators including Euclidean distance and cosine similarity, different weights need to be set for these indicators to highlight the importance of certain dimensions. The fused score can be recorded as , the specific calculation method can be a linear combination or a more complex nonlinear mapping. The elastic fault tolerance boundary is obtained by previous reinforcement learning and cluster statistics, and stores the acceptable deviation range for different operators or workpiece situations. When the Score exceeds a certain grading threshold, the action is judged as a slight, moderate or even severe deviation, and each level represents a different degree of operational error or safety risk. If the operator's proficiency is low, the threshold will be automatically relaxed to leave a larger tolerance interval for the action; if the workpiece precision requirements are extremely high, the threshold will be tightened so that any minor errors can be quickly captured. Therefore, the preliminary deviation level will be attached to the action instance in the form of an integer or text label, and enter the subsequent correction or intervention module, and Bayesian correction will be performed based on the preliminary deviation level and historical deviation data to obtain a graded deviation result. At this time, it is necessary to include the operator's action performance and deviation statistics in the past period of time to correct occasional fluctuations or noise that occur during a measurement process. If the deviation level is set There are discrete value sets, such as , we can define the prior probability Indicates the frequency of occurrence of this level in historical data, observation likelihood Represents the current multidimensional similarity index observed under a certain level condition The posterior inference process can be updated using the Bayesian formula ; If a certain level rarely appears in history or the operator has a high degree of mastery of similar actions in the past, then the prior will be relatively low, making the system tend to make a milder judgment when facing this observation; if the multidimensional similarity index It has a strong correlation with this level, and its posterior probability will also rise, so that the corrected grading result will be transferred to this level. After this round of Bayesian correction, the probability of false alarm or missed judgment caused by instantaneous fluctuations in measurement can be reduced, and the final deviation level result can be output. If the posterior probability still shows an unclear multi-peak distribution, a secondary identification can be made with the help of upper-level scheduling strategies or more sophisticated expert knowledge. After the system outputs the grading deviation result, it will be recorded in the operation log or knowledge graph database to facilitate the continuous accumulation of statistical references for the next judgment. In this way, the action deviation grading process not only covers the preliminary judgment of alignment, similarity calculation and elastic threshold, but also considers historical data and prior knowledge in the correction stage, and finally obtains a more stable and accurate grading deviation result.

[0033] In this embodiment, the operation data of the industrial training production line is collected through a sensor array to obtain a real-time mapping data stream; a digital twin model is constructed based on the real-time mapping data stream, and the multimodal data of the physical space and the digital twin model are analyzed by using a semantic ontology library and a graph neural network to obtain a twin knowledge graph of the physical-digital dual space; based on the twin knowledge graph and a preset expert knowledge base, a reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain an elastic fault-tolerant boundary; based on the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the real-time action of the operator in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output. The present invention overcomes the limitations of the existing fixed standard action model and provides personalized guidance for operators with different proficiency levels.

[0034] The digital twin simulation method of the industrial simulation platform in the embodiment of the present invention is described above. The digital twin simulation system of the industrial simulation platform in the embodiment of the present invention is described below. Figure 2 , an embodiment of the digital twin simulation system of the industrial simulation platform in the embodiment of the present invention includes: The data acquisition module 201 is used to perform hierarchical and partitioned data acquisition on the operation data of different levels of the industrial training production line through a set multi-type sensor array to obtain a real-time mapping data stream; A model building module 202 is used to build a digital twin model according to the real-time mapping data stream, and use a semantic ontology library and a graph neural network to perform correlation analysis on multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; The boundary generation module 203 is used to collaboratively learn the operating conditions of the physical-digital dual space using a reinforcement learning method according to the twin knowledge graph and a preset expert knowledge base, so as to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space; The feedback warning module 204 is used to grade the deviation between the operator's real-time action in the physical space and the standard action in the digital space according to the twin knowledge graph and the elastic fault-tolerant boundary, and output corresponding feedback and warning information in the physical-digital dual space according to different deviation levels.

[0035] In an embodiment of the present invention, the digital twin simulation system of the industrial simulation platform runs the digital twin simulation method of the above industrial simulation platform, and the digital twin simulation system of the industrial simulation platform collects the operation data of the industrial training production line through a sensor array to obtain a real-time mapping data stream; a digital twin model is constructed according to the real-time mapping data stream, and a semantic ontology library and a graph neural network are used to perform correlation analysis on the multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; according to the twin knowledge graph and the preset expert knowledge base, a reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain an elastic fault-tolerant boundary; according to the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the real-time action of the operator in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output. The present invention overcomes the limitations of the existing fixed standard action model and provides personalized guidance for operators with different proficiency levels.

[0036] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0037] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0038] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital twin simulation method for an industrial simulation platform, characterized in that: The digital twin simulation method of the industrial simulation platform includes: Through the multi-type sensor array, the operation data of different levels of the industrial training production line are collected in layers and partitions to obtain real-time mapping data streams; A digital twin model is constructed according to the real-time mapping data stream, and a semantic ontology library and a graph neural network are used to perform correlation analysis on multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; According to the twin knowledge graph and the preset expert knowledge base, a reinforcement learning method is used to collaboratively learn the operating conditions of the physical-digital dual space to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space; According to the twin knowledge graph and the elastic fault-tolerant boundary, the deviation between the operator's real-time action in the physical space and the standard action in the digital space is graded and judged, and corresponding feedback and warning information are output in the physical-digital dual space according to different deviation levels.

2. The digital twin simulation method of the industrial simulation platform according to claim 1, characterized in that: The digital twin model is constructed according to the real-time mapping data stream, and the multimodal data of the physical space and the digital twin model are analyzed for association using the semantic ontology library and the graph neural network to obtain the twin knowledge graph of the physical-digital dual space, including: Performing spatiotemporal feature extraction and data cleaning on the real-time mapping data stream to obtain a normalized data set of the physical space; Constructing a geometric topological structure and a dynamic attribute mapping relationship of a digital twin model according to the normalized data set to obtain a digital twin model of a digital space; Constructing node features of a graph neural network based on the digital twin model and the semantic ontology library, and establishing edge connections according to the mapping relationship and operation process of the physical-digital space to obtain an association graph network of the physical-digital dual space; Graph attention calculation and semantic clustering are performed on the association graph network to obtain a twin knowledge graph that represents the interactive relationship between the physical and digital dual spaces.

3. The digital twin simulation method of the industrial simulation platform according to claim 2, characterized in that: The graph attention calculation and semantic clustering are performed on the association graph network to obtain a twin knowledge graph representing the interactive relationship between physical and digital dual spaces, including: Performing multi-head attention calculation on the nodes in the association graph network to obtain an attention weight matrix between the nodes; Perform message passing and feature aggregation on the association graph network according to the attention weight matrix to obtain updated node features; Perform hierarchical semantic clustering on the updated node features, and build a subgraph structure based on the clustering results to obtain a hierarchical knowledge subgraph set; A mapping rule between physical and digital space is established based on the hierarchical knowledge subgraph set, and the knowledge subgraphs at each layer are integrated to obtain a twin knowledge graph.

4. The digital twin simulation method of the industrial simulation platform according to claim 1, characterized in that: The method of collaboratively learning the operating conditions of the physical-digital dual space using a reinforcement learning method based on the twin knowledge graph and the preset expert knowledge base to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space includes: Constructing a state space and an action space of reinforcement learning according to the operation sequence in the twin knowledge graph and the operation experience in the expert knowledge base to obtain an initial learning environment; Performing Monte Carlo sampling on the initial learning environment to simulate state transition sequences under different operating conditions to obtain a state-action trajectory set; Calculating the reward value of each operation state according to the state-action trajectory set, and updating the state-action value function based on the temporal difference algorithm to obtain the optimal operation strategy; The operating error range is dynamically divided according to the optimal operating strategy to obtain a dynamic and adaptive elastic fault-tolerant boundary of the physical-digital dual space.

5. The digital twin simulation method of the industrial simulation platform according to claim 4, characterized in that: The step of dynamically dividing the operating error range according to the optimal operating strategy to obtain a dynamic adaptive elastic fault-tolerant boundary of the physical-digital dual space includes: Performing cluster analysis on the state-action pairs in the optimal operation strategy to obtain feature clusters of different operation stages; Calculate the standard deviation and confidence interval of each operation stage according to the feature cluster to obtain an initial fault tolerance boundary; Dynamically calculating the adjustment factor of the initial fault tolerance boundary, setting the boundary scaling factor based on the operator proficiency and the workpiece characteristics, and obtaining an adaptive boundary parameter; The initial fault-tolerant boundary is dynamically updated according to the adaptive boundary parameter to obtain a flexible fault-tolerant boundary of the physical-digital dual space.

6. The digital twin simulation method of the industrial simulation platform according to claim 1, characterized in that: The step of performing graded judgment on the deviation between the operator's real-time action in the physical space and the standard action in the digital space according to the twin knowledge graph and the elastic fault-tolerant boundary, and outputting corresponding feedback and warning information in the physical-digital dual space according to different deviation levels includes: Acquire a standard action model according to the twin knowledge graph, and perform feature extraction on the real-time action data of the operator in the physical space to obtain an action feature vector; Calculating the similarity between the action feature vector and the standard action model, and classifying the deviation levels based on the elastic fault tolerance boundary to obtain a graded deviation result; Selecting a corresponding feedback mode according to the graded deviation result, and triggering tactile feedback in the physical space or displaying guidance information in the digital space to obtain a real-time feedback signal; The operator's action trend is sequentially predicted, and the risk probability is calculated based on historical deviation data. When the risk probability exceeds the preset threshold, a warning signal is output.

7. The digital twin simulation method of the industrial simulation platform according to claim 6, characterized in that: The similarity calculation between the action feature vector and the standard action model and the deviation classification based on the elastic fault tolerance boundary to obtain the graded deviation result include: Performing dynamic time warping on the action feature vector and the standard action model to obtain an aligned feature sequence; Calculate the Euclidean distance and cosine similarity according to the aligned feature sequences to obtain a multi-dimensional similarity index; Performing weighted fusion on the multi-dimensional similarity indexes, and setting a grading threshold based on the elastic fault tolerance boundary to obtain a preliminary deviation grade; Bayesian correction is performed based on the preliminary deviation level and historical deviation data to obtain a graded deviation result.

8. A digital twin simulation system for an industrial simulation platform, characterized in that: The digital twin simulation system of the industrial simulation platform includes: The data acquisition module is used to collect the operation data of different levels of the industrial training production line in layers and partitions through the multi-type sensor arrays to obtain real-time mapping data streams; A model building module is used to build a digital twin model according to the real-time mapping data stream, and use a semantic ontology library and a graph neural network to perform correlation analysis on the multimodal data of the physical space and the digital twin model to obtain a twin knowledge graph of the physical-digital dual space; A boundary generation module is used to collaboratively learn the operating conditions of the physical-digital dual space using a reinforcement learning method according to the twin knowledge graph and a preset expert knowledge base, so as to obtain a dynamically adaptive elastic fault-tolerant boundary of the physical-digital dual space; The feedback warning module is used to grade the deviation between the operator's real-time actions in the physical space and the standard actions in the digital space according to the twin knowledge graph and the elastic fault-tolerant boundary, and output corresponding feedback and warning information in the physical-digital dual space according to different deviation levels.

Images