Smart station multi-source heterogeneous data acquisition and remote collaborative analysis method and system

Abstract

The invention provides an intelligent station multi-source heterogeneous data acquisition and remote collaborative analysis method and system, and relates to the technical field of heterogeneous data, and the method comprises the steps: collecting multi-source heterogeneous data, carrying out the standardization processing, constructing a station real-time adaptive three-dimensional model through a self-calibration laser scanning technology, and carrying out the semantic mapping of the model and a BIM model; constructing a self-evolution feature library based on a digital twinborn model, and carrying out space-time correlation analysis by adopting a graph convolutional neural network; a remote cooperation platform is established to realize real-time cooperation of multiple experts, and a diagnosis process is recorded through a block chain technology. According to the invention, efficient acquisition, analysis and remote collaborative diagnosis of the data of the smart station are realized, and the accuracy and efficiency of fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to heterogeneous data technology, and in particular to a method and system for multi-source heterogeneous data collection and remote collaborative analysis at smart stations. Background Art

[0002] With the rapid development of smart stations, the number and variety of equipment continues to increase, and system complexity continues to rise, placing higher demands on intelligent operation and maintenance and remote collaborative diagnosis. Smart stations contain a large number of sensors, controllers, and other devices, generating massive amounts of multi-source, heterogeneous data. Currently, smart stations primarily use traditional data collection and analysis methods, achieving station monitoring through the establishment of three-dimensional models and data visualization platforms, and relying on expert experience for fault diagnosis and resolution.

[0003] In practical applications, existing technologies have the following deficiencies: First, traditional data acquisition and processing methods are difficult to effectively handle multi-source heterogeneous data in smart stations. The lack of a unified standardized processing mechanism between various types of data makes it difficult to conduct in-depth data fusion analysis, affecting the accuracy of fault diagnosis. Secondly, existing three-dimensional modeling technology lacks real-time adaptive capabilities and cannot promptly reflect the dynamic changes of station equipment, making it difficult to support the real-time update requirements of digital twin models. Thirdly, traditional remote collaborative diagnosis methods have problems such as low collaborative efficiency and difficulty in effectively accumulating and inheriting expert experience. In addition, there is a lack of reliable expert knowledge base support, making it difficult to provide accurate decision-making recommendations for fault diagnosis.

[0004] To address these issues, there is an urgent need to develop a technical solution that can efficiently collect multi-source heterogeneous data from smart stations, update digital twin models in real time, and enable remote collaborative diagnosis. This will improve the operational efficiency and fault diagnosis capabilities of smart stations. This solution should support standardized processing of multi-source heterogeneous data, real-time adaptive 3D modeling, the effective accumulation and application of expert knowledge, and ensure the reliability and traceability of the remote collaborative diagnosis process. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for multi-source heterogeneous data collection and remote collaborative analysis of smart stations, which can solve the problems in the existing technology.

[0006] According to a first aspect of the embodiments of the present invention, Provides multi-source heterogeneous data collection and remote collaborative analysis methods for smart stations, including: Collect multi-source heterogeneous data of smart stations, standardize the multi-source heterogeneous data to generate standard data; use multi-sensor fusion self-calibration laser scanning technology to build a real-time adaptive 3D model of the station, and semantically map the adaptive 3D model with the BIM model to establish an equipment geometry and attribute information library; use improved WebGL technology to build a digital twin visualization platform, map the standard data with the real-time adaptive 3D model of the station in real time to form a digital twin model with dynamic update capabilities; use a distributed database cluster to store the standard data, and establish a real-time data update mechanism based on incremental synchronization with the digital twin model; Based on the digital twin model, a self-evolving feature library containing device parameter anomaly features, system performance degradation features, and safety risk warning features is constructed, and an improved graph convolutional neural network is used to perform spatiotemporal correlation analysis on the standard data to obtain anomaly features. Based on the device geometry and attribute information library, a knowledge graph device group analysis model integrating causal reasoning is established, and the dynamic correlation influence between devices is calculated through a graph neural network to obtain device association data. A transfer learning method is used to combine the anomaly features and the device association data to train an adaptive fault diagnosis model. Based on the diagnosis results of the adaptive fault diagnosis model, a multi-level evaluation system is established to classify the degree of anomaly, and graded alarm information is pushed to remote expert terminals through a distributed message queue. After receiving the graded alarm information, a remote collaboration platform supporting a weak network environment is established, and the digital twin model is called in the remote collaboration platform to realize real-time collaboration among multiple experts; intelligent layered SVG vector graphics technology is used to support real-time labeling by multiple people, and causal reasoning analysis is performed on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model to establish an expert knowledge base; a collaborative decision-making recommendation system based on deep reinforcement learning is integrated with the expert knowledge base to provide experts with diagnostic solution suggestions; the diagnostic process is recorded through blockchain technology to ensure that the operation traces cannot be tampered with; the diagnostic results confirmed by the experts are fed back to the digital twin model to continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

[0007] The digital twin visualization platform is constructed using improved WebGL technology, and the standard data is mapped and integrated with the real-time adaptive three-dimensional model of the station in real time to form a digital twin model with dynamic update capability, including: Improve the vertex shader of the WebGL rendering pipeline to a parallel computing architecture to achieve batch preprocessing capabilities for scene data; introduce a hierarchical caching mechanism in the fragment shading stage and optimize the depth test algorithm; design a dynamic vertex buffer management mechanism to achieve dynamic scheduling of vertex data, and form an optimized basic rendering pipeline based on this dynamic adjustment; build an adaptive shader compilation system based on this optimized basic rendering pipeline, dynamically generate and cache GLSL shader code based on scene complexity, and output a rendering engine that supports real-time dynamic compilation; Utilizing the rendering engine supporting real-time dynamic compilation, the three-dimensional models of the equipment within the station are merged and processed through instanced rendering technology to generate a basic static scene; the basic static scene is spatially divided into an octree to obtain a hierarchical grid structure, and the models in the hierarchical grid structure are removed in real time based on a frustum clipping algorithm to output an optimized scene grid; based on the distance relationship between the camera and each model in the optimized scene grid, a multi-level precision model library and a distance threshold mapping table are established, and a real-time adaptive three-dimensional model of the station is generated by dynamically adjusting the model precision level; Construct a device identifier index table to establish a unique mapping relationship between externally collected standard data and the corresponding equipment in the real-time adaptive three-dimensional model of the station; hierarchically process the mapped standard data into a geometric parameter layer, an operating status layer, and an alarm information layer to form a hierarchical data structure; implement incremental push of the hierarchical data structure based on the WebSocket protocol, and smooth the data update frequency to output a real-time data stream; use a distributed state manager to establish a binding relationship between the real-time data stream and the device status, and store the scene status information in the corresponding state tree according to the hierarchical data structure; Based on the observer mode, a dynamic association is established between the state tree and the real-time data stream to realize automatic status update when the data changes; using the rendering engine that supports real-time dynamic compilation, the material and effect of the equipment in the real-time adaptive three-dimensional model of the station are adjusted in real time according to the latest status information in the state tree, the standard data and the three-dimensional model are mapped and integrated in real time, and the initial digital twin model is output; a two-way data synchronization mechanism is constructed to realize real-time data interaction between the initial digital twin model and the state tree, and the user operation records are synchronously updated to the state tree, continuously triggering the rendering engine to update the screen, and finally generating a complete digital twin model with dynamic interaction and real-time visualization capabilities.

[0008] An improved graph convolutional neural network is used to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features. A knowledge graph device group analysis model integrating causal reasoning is established based on the device geometry and attribute information library. The dynamic correlation influence between devices is calculated through the graph neural network to obtain device association data including: A temporal self-attention module is designed in the convolutional layer, including a multi-head attention mechanism to extract time-scale features and a position encoding unit to maintain temporal information, outputting enhanced temporal features. An adaptive gating mechanism is designed based on the enhanced temporal features, which dynamically adjusts the propagation of information between nodes by updating the gate to retain historical information and resetting the gate to filter redundant information. The gated inter-node information is input into a multi-scale feature fusion module to obtain fused features. A feature pyramid is used to extract hierarchical features, and residual connections are used to retain the original information. Causal constraints and feature compression are applied to the fused features, and graph adversarial training is used to improve robustness, resulting in an improved graph convolutional neural network. The fused features are dynamically filtered using the adaptive gating mechanism to output standardized device spatiotemporal features; the standardized device spatiotemporal features are input into the improved graph convolutional neural network, and the causal perception module is used to causally constrain the standardized device spatiotemporal features to ensure the causal consistency of the features and obtain causal relationships; the features are compressed based on a hierarchical graph pooling strategy to obtain node-level anomaly representations; a variational autoencoder is constructed to encode and reconstruct the node-level anomaly representations, and the device anomaly features are output through optimization using a combined loss function of reconstruction error and KL divergence; Utilizing the device anomaly characteristics to calculate the time delay mutual information between nodes; determining the optimal time delay between nodes based on maximizing the mutual information; fusing the causal relationship with device attribute information to construct an initial knowledge graph; utilizing the causal perception module of the improved graph convolutional neural network to dynamically fuse the device anomaly characteristics with the initial knowledge graph; achieving knowledge compression through the hierarchical graph pooling strategy to obtain a device group causal analysis model; The device anomaly characteristics are input into the device group causal analysis model, and the dynamic coupling strength between nodes is calculated through the improved graph convolutional neural network; device groups are hierarchically clustered based on the dynamic coupling strength; an anomaly propagation network is constructed by combining the device anomaly characteristics and the causal relationship; the anomaly propagation path is analyzed using the temporal self-attention module; the dynamic coupling strength, the optimal delay and the causal relationship are integrated in multiple dimensions to output the final device association data.

[0009] A transfer learning method is used to combine the abnormal features with the device-related data to train an adaptive fault diagnosis model. A multi-level evaluation system is established based on the diagnosis results of the fault diagnosis model to classify the degree of abnormality. The graded alarm information is pushed to the remote expert terminal through a distributed message queue, including: The abnormal features are fused with the device-related data to construct a multimodal feature; historical data with sufficient fault labels is used as the source domain feature, and the new device data to be diagnosed is used as the target domain feature. The multimodal features are adaptively weighted and fused using an attention mechanism to obtain an adaptive fusion feature; The adaptive fusion feature is used to train an adaptive fault diagnosis model. The maximum mean difference between the source domain feature and the target domain feature is calculated and the difference is minimized to reduce the feature distribution difference. A feature discriminator is constructed to distinguish the feature distributions of the source domain feature and the target domain feature to obtain a feature discrimination result. Based on the feature discrimination result, an adversarial learning method is used to train a feature extractor to generate a domain-invariant feature representation. The expert knowledge in the source domain feature is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model. Input the diagnostic results of the adaptive fault diagnosis model into a multi-level evaluation system, perform a device-level quantitative score on individual devices based on fault severity parameters, development rate parameters, and propagation risk parameters, input the device-level quantitative score into a system-level evaluation unit, calculate the importance of the faulty device using weighted network centrality, and evaluate the system impact by combining the device-related data; The degree of impact of the system is input into a distributed message queue for hierarchical caching. The fault mode similarity is calculated for the evaluation result of the hierarchical cache based on time correlation and spatial correlation, and the alarm information with similar fault modes is merged. The secondary alarms in the merged alarm information are filtered using alarm suppression rules to obtain filtered hierarchical alarm information, and the filtered hierarchical alarm information is pushed to the remote expert terminal through a publish-subscribe model.

[0010] The adaptive fault diagnosis model is trained using the fusion features, the feature distribution difference is reduced by calculating the maximum mean difference between the source domain features and the target domain features and minimizing the difference value, a feature discriminator is constructed to distinguish the feature distributions of the source domain features and the target domain features to obtain a feature discrimination result, an adversarial learning method is used to train a feature extractor based on the feature discrimination result to generate a domain-invariant feature representation, and the expert knowledge in the source domain features is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model, including: The fusion feature is used to train an adaptive fault diagnosis model, the source domain feature and the target domain feature are mapped to a reproducing kernel Hilbert space, the empirical mean of the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated to obtain distribution difference data, and a multi-kernel maximum mean difference method is used to perform adaptive kernel function optimization on the distribution difference data to obtain a feature distribution difference value; A dual-branch feature discriminator is constructed to discriminate the feature distributions of the source domain features and the target domain features, the network parameters of the feature discriminator are updated by fixing the feature extractor parameters to obtain a feature discrimination result, the network parameters of the feature discriminator are updated by fixing the feature discriminator parameters based on the feature discrimination result, a gradient penalty term is introduced into the feature discriminator, and a spectral normalization technique is used to control the discrimination ability; The feature extractor constructs an optimization objective function based on classification loss, adversarial loss and maximum mean difference loss, and adopts an adaptive weight balancing strategy to iteratively optimize the optimization objective function to generate a domain-invariant feature representation; The expert knowledge in the source domain features is divided into a multi-level knowledge representation according to the degree of abstraction, and the multi-level knowledge representation is temperature-adaptively adjusted to obtain soft label data. The soft label data and the multi-level knowledge representation are input into a progressive knowledge distillation network, and a curriculum learning strategy is used to transfer the expert knowledge to the domain-invariant feature representation to obtain an adaptive fault diagnosis model.

[0011] Intelligent layered SVG vector graphics technology is used to support real-time annotation by multiple people. Causal reasoning analysis is performed on expert annotation data and the diagnostic results of the adaptive fault diagnosis model to establish an expert knowledge base. A collaborative decision-making recommendation system based on deep reinforcement learning is integrated with the expert knowledge base to provide experts with diagnostic solution suggestions including: Fault data containing time series data, feature data, and parameter data is acquired, and the fault data is converted into SVG graphic elements using intelligent layered SVG vector graphics technology to construct a data visualization base layer. An expert annotation operation layer is constructed on the data visualization base layer to support experts in sequentially selecting fault areas to obtain fault area data, marking fault parameters to obtain fault parameter data, and annotating causal relationships to obtain causal link data. An operation conversion algorithm is used to perform consistency processing on the concurrent causal link data of multiple experts, and the processed data is synchronized to the interactive response layer via WebSocket to form the expert annotation results. Causal reasoning analysis is performed on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model. The expert annotation results and the diagnostic results are aligned in time and space using a dynamic time warping algorithm. The causal relationship between the variables of the aligned data is identified based on a conditional independence test to construct an initial causal network. The initial causal network is modified based on expert prior knowledge and the strength of the causal relationship is quantified to obtain weighted causal knowledge. The weighted causal knowledge is encoded as triples and stored in the blockchain system to form an expert knowledge base. Based on the expert knowledge base, a deep reinforcement learning collaborative decision-making system is constructed. The fault scenario characteristics are integrated with historical diagnostic records to form a state vector. According to the causal relationship in the expert knowledge base, prior probabilities are assigned to diagnostic actions to construct an action space. A recurrent neural network is used to capture the temporal dependencies of the decision sequence to establish a double Q network. Through knowledge distillation, the experience of the expert knowledge base is transferred to the policy network. The input fault scenario is analyzed and diagnostic suggestions are generated and pushed to the experts.

[0012] Performing causal reasoning analysis on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model, using a dynamic time warping algorithm to perform spatiotemporal alignment on the expert annotation results and the diagnostic results, identifying the causal relationship between variables in the aligned data based on a conditional independence test to construct an initial causal network, modifying the initial causal network based on expert prior knowledge and quantifying the strength of the causal relationship to obtain weighted causal knowledge, encoding the weighted causal knowledge into triples and storing them in the blockchain system to form an expert knowledge base, including: Based on the adaptive fault diagnosis model, causal reasoning analysis is performed on the expert annotation results and the diagnosis results to obtain initial variable association data; based on the initial variable association data, a dynamic time warping algorithm is used to construct a cumulative distance matrix, and the optimal alignment path between the expert annotation results and the diagnosis results is calculated according to the cumulative distance matrix to obtain first time-aligned data, and spatial feature matching is performed on the first time-aligned data through a multi-scale feature pyramid to obtain second spatiotemporal alignment data, and the first spatiotemporal alignment data and the second spatiotemporal alignment data are weightedly combined to form a spatiotemporal alignment dataset; Based on the spatiotemporal aligned dataset, conditional mutual information between variable pairs is calculated to obtain a mutual information matrix, a Fisher-Z test is used to perform conditional independence judgment on the mutual information matrix to obtain an independence judgment result, and the causal direction between variables is identified based on the independence judgment result to construct an initial causal network; expert prior knowledge is converted into network constraints, the structure of the initial causal network is optimized according to the network constraints to obtain a modified causal network, and a Bayesian network learning method is used to calculate the weight strength of each causal edge in the modified causal network to obtain weighted causal knowledge; The causal relationship in the weighted causal knowledge is encoded as a triple representation, where each triple contains a head entity, a tail entity and a relationship attribute, and the relationship attribute includes the causal relationship type and weight information; the blockchain system is used to encapsulate the triple representation into a block transaction unit, and the knowledge validity and version information of the block transaction unit are verified through a smart contract. The block transaction unit that passes the verification is written into the blockchain system in a time sequence to form an expert knowledge base.

[0013] According to a second aspect of the embodiments of the present invention, Provides a smart station multi-source heterogeneous data collection and remote collaborative analysis system, including: The first unit is used to collect multi-source heterogeneous data of smart stations, standardize the multi-source heterogeneous data to generate standard data; use multi-sensor fusion self-calibration laser scanning technology to build a real-time adaptive three-dimensional model of the station, and semantically map the adaptive three-dimensional model with the BIM model to establish an equipment geometry and attribute information library; use improved WebGL technology to build a digital twin visualization platform, map and integrate the standard data with the real-time adaptive three-dimensional model of the station in real time to form a digital twin model with dynamic update capabilities; use a distributed database cluster to store the standard data, and establish a real-time data update mechanism based on incremental synchronization with the digital twin model; The second unit is used to construct a self-evolving feature library containing equipment parameter abnormality features, system performance degradation features and safety risk warning features based on the digital twin model, and use an improved graph convolutional neural network to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; based on the equipment geometry and attribute information library, a knowledge graph equipment group analysis model integrating causal reasoning is established, and the dynamic correlation influence between devices is calculated through the graph neural network to obtain equipment association data; a transfer learning method is used to combine the abnormal features and the equipment association data to train an adaptive fault diagnosis model, and a multi-level evaluation system is established according to the diagnosis results of the adaptive fault diagnosis model to classify the degree of abnormality, and the graded alarm information is pushed to the remote expert terminal through a distributed message queue; The third unit is used to build a remote collaboration platform that supports weak network environments after receiving the graded alarm information, and call the digital twin model in the remote collaboration platform to realize real-time collaboration among multiple experts; use intelligent layered SVG vector graphics technology to support real-time labeling by multiple people, perform causal reasoning analysis on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model, and establish an expert knowledge base; integrate a collaborative decision-making recommendation system based on deep reinforcement learning based on the expert knowledge base to provide experts with diagnostic solution suggestions; record the diagnostic process through blockchain technology and ensure that the operation traces cannot be tampered with; feed back the diagnostic results confirmed by the experts to the digital twin model, and continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] The beneficial effects of this application are as follows: Through multi-sensor fusion self-calibration laser scanning technology, a real-time adaptive three-dimensional model of the station is constructed, and semantic mapping is performed with the BIM model. Combined with improved WebGL technology, a digital twin visualization platform is built, which realizes the standardized processing and real-time dynamic update of smart station data, ensures the accuracy and real-time nature of data collection, and provides a reliable data foundation for subsequent analysis.

[0017] A self-evolving feature library is constructed based on the digital twin model, and an improved graph convolutional neural network is used for spatiotemporal correlation analysis. Combined with the equipment group analysis model, accurate identification of abnormal features is achieved. An adaptive fault diagnosis model is trained through transfer learning methods, and a multi-level evaluation system is established. This improves the accuracy and efficiency of fault diagnosis and enables timely warning of abnormal conditions.

[0018] A remote collaboration platform that supports weak network environments is built, and intelligent layered SVG vector graphics technology is used to support real-time annotation by multiple people. A collaborative decision-making recommendation system based on deep reinforcement learning is used to provide experts with diagnostic solution suggestions. Blockchain technology is used to record the diagnostic process, thereby achieving effective accumulation and utilization of expert knowledge, improving the efficiency and reliability of remote collaborative diagnosis, and continuously optimizing the self-evolving feature library and adaptive fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for multi-source heterogeneous data collection and remote collaborative analysis in a smart station according to an embodiment of the present invention; Figure 2 This is a structural diagram of the multi-source heterogeneous data collection and remote collaborative analysis system for smart stations according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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 making creative efforts shall fall within the scope of protection of the present invention.

[0021] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0022] Figure 1 This is a flow chart of a method for multi-source heterogeneous data collection and remote collaborative analysis of a smart station according to an embodiment of the present invention. Figure 1 As shown, the method includes: S11. Collect multi-source heterogeneous data from the smart station and standardize the data to generate standard data; construct a real-time adaptive 3D model of the station using multi-sensor fusion self-calibration laser scanning technology, and semantically map the adaptive 3D model with the BIM model to establish a device geometry and attribute information library; construct a digital twin visualization platform using improved WebGL technology, and integrate the standard data with the real-time adaptive 3D model of the station in real time to form a digital twin model with dynamic update capabilities; use a distributed database cluster to store the standard data and establish a real-time data update mechanism based on incremental synchronization with the digital twin model; S12. Based on the digital twin model, a self-evolving feature library containing device parameter anomaly features, system performance degradation features, and safety risk warning features is constructed. An improved graph convolutional neural network is used to perform spatiotemporal correlation analysis on the standard data to obtain anomaly features. A knowledge graph device cluster analysis model integrating causal reasoning is established based on the device geometry and attribute information library. The dynamic correlation effects between devices are calculated using a graph neural network to obtain device association data. A transfer learning method is used to combine the anomaly features and the device association data to train an adaptive fault diagnosis model. A multi-level evaluation system is established based on the diagnostic results of the adaptive fault diagnosis model to classify the degree of anomaly, and graded alarm information is pushed to remote expert terminals via a distributed message queue. S13. After receiving the graded alarm information, a remote collaboration platform that supports weak network environments is established, and the digital twin model is called in the remote collaboration platform to realize real-time collaboration among multiple experts; intelligent layered SVG vector graphics technology is used to support real-time labeling by multiple people, and causal reasoning analysis is performed on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model to establish an expert knowledge base; a collaborative decision-making recommendation system based on deep reinforcement learning is integrated into the expert knowledge base to provide experts with diagnostic solution suggestions; the diagnostic process is recorded through blockchain technology to ensure that the operation traces cannot be tampered with; the diagnostic results confirmed by the experts are fed back to the digital twin model to continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

[0023] In an optional embodiment, a digital twin visualization platform is constructed using improved WebGL technology, and the standard data is mapped and integrated with the real-time adaptive three-dimensional model of the station in real time to form a digital twin model with dynamic update capability, including: Improve the vertex shader of the WebGL rendering pipeline to a parallel computing architecture to achieve batch preprocessing capabilities for scene data; introduce a hierarchical caching mechanism in the fragment shading stage and optimize the depth test algorithm; design a dynamic vertex buffer management mechanism to achieve dynamic scheduling of vertex data, and form an optimized basic rendering pipeline based on this dynamic adjustment; build an adaptive shader compilation system based on this optimized basic rendering pipeline, dynamically generate and cache GLSL shader code based on scene complexity, and output a rendering engine that supports real-time dynamic compilation; Utilizing the rendering engine supporting real-time dynamic compilation, the three-dimensional models of the equipment within the station are merged and processed through instanced rendering technology to generate a basic static scene; the basic static scene is spatially divided into an octree to obtain a hierarchical grid structure, and the models in the hierarchical grid structure are removed in real time based on a frustum clipping algorithm to output an optimized scene grid; based on the distance relationship between the camera and each model in the optimized scene grid, a multi-level precision model library and a distance threshold mapping table are established, and a real-time adaptive three-dimensional model of the station is generated by dynamically adjusting the model precision level; Construct a device identifier index table to establish a unique mapping relationship between externally collected standard data and the corresponding equipment in the real-time adaptive three-dimensional model of the station; hierarchically process the mapped standard data into a geometric parameter layer, an operating status layer, and an alarm information layer to form a hierarchical data structure; implement incremental push of the hierarchical data structure based on the WebSocket protocol, and smooth the data update frequency to output a real-time data stream; use a distributed state manager to establish a binding relationship between the real-time data stream and the device status, and store the scene status information in the corresponding state tree according to the hierarchical data structure; Based on the observer mode, a dynamic association is established between the state tree and the real-time data stream to realize automatic status update when the data changes; using the rendering engine that supports real-time dynamic compilation, the material and effect of the equipment in the real-time adaptive three-dimensional model of the station are adjusted in real time according to the latest status information in the state tree, the standard data and the three-dimensional model are mapped and integrated in real time, and the initial digital twin model is output; a two-way data synchronization mechanism is constructed to realize real-time data interaction between the initial digital twin model and the state tree, and the user operation records are synchronously updated to the state tree, continuously triggering the rendering engine to update the screen, and finally generating a complete digital twin model with dynamic interaction and real-time visualization capabilities.

[0024] First, the WebGL rendering pipeline is optimized and improved. By designing a vertex shader with a parallel computing architecture, batch preprocessing of scene data is achieved. In the specific implementation, vertex data is grouped according to spatial position relationships, and computing resources are independently allocated to each group of data, while processing multiple groups of vertex transformation operations. A three-level cache mechanism is introduced in the fragment shading stage, including an L1 cache to store the most recently accessed depth values, an L2 cache to store commonly used depth comparison results, and an L3 cache as backup storage. When optimizing the depth test, the L1 cache is checked first. If a hit occurs, the result is used directly. If a miss occurs, the L2 cache is queried, and finally the L3 cache is accessed, significantly improving the efficiency of the depth test.

[0025] When designing the dynamic vertex buffer management mechanism, vertex data is divided into two categories: static data and dynamic data. Static data such as vertex coordinates and normals are stored in a fixed buffer, while dynamic data such as transformation matrices and material parameters are stored in a variable buffer. Using double buffering technology, when dynamic data is updated, the new data is written to a backup buffer. After the current frame is rendered, the buffer is quickly switched to avoid rendering pauses caused by data updates.

[0026] Based on the optimized rendering pipeline, an adaptive shader compilation system is built. Based on scene parameters such as the number of light sources, shadow types, and material properties, corresponding GLSL code snippets are dynamically generated. These snippets are stored in a shader code library, and a mapping relationship between parameters and code snippets is established. When scene parameters change, the system automatically combines the corresponding code snippets to generate a complete shader program.

[0027] When using instanced rendering technology to process station equipment models, devices with similar geometric features are first grouped together, with a copy of the underlying model data stored for each type. An instanced matrix array is used to store the position, rotation, and scale information of different instances. This data is then passed all at once during rendering, significantly reducing the number of draw calls.

[0028] The scene is partitioned into an octree space, starting with the scene bounding box and recursively partitioning the space until a preset minimum grid size or a threshold number of models within the grid is reached. Each grid node records a list of contained models and their bounding box information. During real-time rendering, the octree is traversed starting from the root node, and models outside the field of view are quickly culled by performing an intersection test between the view frustum and the node's bounding box.

[0029] When building a multi-level precision model, four different precision versions are generated for each device model. Precision level 0 retains the original model, level 1 simplifies to 50% of the original vertex count, level 2 simplifies to 25%, and level 3 simplifies to 10%. Distance thresholds are set as follows: level 0 for distances within 50 meters, level 1 for distances between 50 and 100 meters, level 2 for distances between 100 and 200 meters, and level 3 for distances above 200 meters.

[0030] A hierarchical hash table structure is used to construct the device identifier index table. The first-level hash table uses the device type as the key and the value as the secondary hash table for devices of that type. The second-level hash table uses the device number as the key and the value as the corresponding model reference. External data contains device type and number information, allowing for rapid location of target models through two-level hash queries.

[0031] In the layered data structure design, the geometric parameter layer stores static attributes such as device location and dimensions, the operating status layer stores real-time data such as temperature and pressure, and the alarm information layer records abnormal conditions such as faults. Data in each layer is updated and pushed independently, reducing data transmission volume.

[0032] A persistent connection is established over WebSocket, and the server uses an incremental update strategy, pushing only changed data. The client uses a sliding window averaging algorithm to handle data update frequency, with a window size set to 1 second to ensure smooth status updates.

[0033] The distributed state manager uses a tree structure, with the scene state as the root node and the device state as child nodes. Each device node has three child nodes: geometry parameters, operating status, and alarm information. When the state is updated, only the subtree containing the changed data is modified to avoid global updates.

[0034] The solution of this application can: Improvements to the WebGL rendering pipeline and instanced rendering technology significantly enhance the rendering performance of large-scale 3D scenes, enabling a smooth real-time interactive experience. The adoption of a layered data structure and incremental push mechanism reduces data transmission overhead, improves system responsiveness, and ensures the real-time and accuracy of digital twin models. Based on the observer model and bidirectional data synchronization mechanism, precise alignment between digital twin models and actual device states is achieved, supporting real-time monitoring and intelligent analysis of complex scenarios.

[0035] In an optional embodiment, an improved graph convolutional neural network is used to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; a knowledge graph device group analysis model integrating causal reasoning is established based on the device geometry and attribute information library, and the device association data obtained by calculating the dynamic association influence between devices through the graph neural network includes: A temporal self-attention module is designed in the convolutional layer, including a multi-head attention mechanism to extract time-scale features and a position encoding unit to maintain temporal information, outputting enhanced temporal features. An adaptive gating mechanism is designed based on the enhanced temporal features, which dynamically adjusts the propagation of information between nodes by updating the gate to retain historical information and resetting the gate to filter redundant information. The gated inter-node information is input into a multi-scale feature fusion module to obtain fused features. A feature pyramid is used to extract hierarchical features, and residual connections are used to retain the original information. Causal constraints and feature compression are applied to the fused features, and graph adversarial training is used to improve robustness, resulting in an improved graph convolutional neural network. The fused features are dynamically filtered using the adaptive gating mechanism to output standardized device spatiotemporal features; the standardized device spatiotemporal features are input into the improved graph convolutional neural network, and the causal perception module is used to causally constrain the standardized device spatiotemporal features to ensure the causal consistency of the features and obtain causal relationships; the features are compressed based on a hierarchical graph pooling strategy to obtain node-level anomaly representations; a variational autoencoder is constructed to encode and reconstruct the node-level anomaly representations, and the device anomaly features are output through optimization using a combined loss function of reconstruction error and KL divergence; Utilizing the device anomaly characteristics to calculate the time delay mutual information between nodes; determining the optimal time delay between nodes based on maximizing the mutual information; fusing the causal relationship with device attribute information to construct an initial knowledge graph; utilizing the causal perception module of the improved graph convolutional neural network to dynamically fuse the device anomaly characteristics with the initial knowledge graph; achieving knowledge compression through the hierarchical graph pooling strategy to obtain a device group causal analysis model; The device anomaly characteristics are input into the device group causal analysis model, and the dynamic coupling strength between nodes is calculated through the improved graph convolutional neural network; device groups are hierarchically clustered based on the dynamic coupling strength; an anomaly propagation network is constructed by combining the device anomaly characteristics and the causal relationship; the anomaly propagation path is analyzed using the temporal self-attention module; the dynamic coupling strength, the optimal delay and the causal relationship are integrated in multiple dimensions to output the final device association data.

[0036] First, standard data is preprocessed and normalized. Using a sliding time window, continuous time series data is segmented into fixed-length segments, each containing device status information for multiple time steps. For each time step, the data is normalized to its maximum and minimum values, ensuring that the values are distributed between 0 and 1.

[0037] During the temporal feature extraction phase, a multi-head attention mechanism is designed to capture dependencies at different time scales. Specifically, the input sequence undergoes a linear transformation to obtain a query vector, a key vector, and a value vector. The similarity between the query vector and the key vector is calculated to obtain the attention weight. Eight attention heads are used, each with an output dimension of 64. Finally, the outputs of multiple heads are concatenated to generate enhanced temporal features. Positional encoding is also introduced to preserve the order of the sequence, using a sine-cosine positional encoding scheme.

[0038] An adaptive gating mechanism is designed to dynamically regulate information propagation. The update gate uses a sigmoid function to weight the current input features with the historical state, with the weights varying dynamically between 0 and 1. The reset gate also uses a sigmoid function to selectively retain historical information. The parameters of the gating mechanism are optimized through backpropagation to achieve adaptive learning.

[0039] During the feature fusion stage, multi-scale convolution is used to extract local features with different receptive fields. Convolutional layers are set with kernel sizes of 3, 5, and 7, stride of 1, and padding of the same. Residual connections are used to preserve the original feature information, and features at different scales are concatenated to generate fused features.

[0040] The causal constraint module calculates the causal strength between nodes based on an attention mechanism. For each pair of nodes, the causal direction and strength are determined based on their feature similarity and temporal dependencies. Graph adversarial training is used to improve model robustness. The discriminator determines the authenticity of causal relationships, while the generator generates feature representations that conform to causal constraints.

[0041] During the anomaly feature extraction phase, a variational autoencoder is used for feature encoding and reconstruction. The encoder maps input features to a low-dimensional latent space, and the decoder reconstructs the latent representation back into the original feature space. This is optimized by minimizing a combined loss function of reconstruction error and KL divergence. Anomalous samples often have large reconstruction errors.

[0042] The optimal delay between nodes is determined based on the principle of maximizing mutual information. The mutual information between node pairs at different time delays is calculated, and the delay with the maximum mutual information is selected as the optimal delay. Device attribute information is combined with causal relationships to construct an initial knowledge graph, where nodes represent devices and edges represent relationships between devices.

[0043] Finally, the anomaly features are dynamically integrated with the knowledge graph. The dynamic coupling strength between nodes is calculated, and hierarchical clustering is performed based on this coupling strength to generate device groups. The propagation paths of anomalies between device groups are analyzed, and the final device association data is output by combining multi-dimensional information such as dynamic coupling strength, optimal latency, and causal relationships.

[0044] The solution of this application can: By designing a temporal self-attention module and adaptive gating mechanism, we can effectively capture the temporal dependencies of device states, improve the accuracy and timeliness of anomaly detection, and reduce false alarm rates. By combining causal constraints with graph adversarial training, we ensure the causal consistency of feature representation, enhance the model's interpretability and generalization capabilities, and make anomaly detection more reliable. A knowledge graph-based device group analysis model enables dynamic modeling of device relationships, accurately identifying anomaly propagation paths, and providing strong support for equipment fault diagnosis and predictive maintenance.

[0045] In an optional embodiment, a transfer learning method is used to combine the abnormal characteristics with the device-related data to train an adaptive fault diagnosis model, a multi-level evaluation system is established based on the diagnosis results of the fault diagnosis model to classify the degree of abnormality, and the graded alarm information is pushed to the remote expert terminal through a distributed message queue, including: The abnormal features are fused with the device-related data to construct a multimodal feature; historical data with sufficient fault labels is used as the source domain feature, and the new device data to be diagnosed is used as the target domain feature. The multimodal features are adaptively weighted and fused using an attention mechanism to obtain an adaptive fusion feature; The adaptive fusion feature is used to train an adaptive fault diagnosis model. The maximum mean difference between the source domain feature and the target domain feature is calculated and the difference is minimized to reduce the feature distribution difference. A feature discriminator is constructed to distinguish the feature distributions of the source domain feature and the target domain feature to obtain a feature discrimination result. Based on the feature discrimination result, an adversarial learning method is used to train a feature extractor to generate a domain-invariant feature representation. The expert knowledge in the source domain feature is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model. Input the diagnostic results of the adaptive fault diagnosis model into a multi-level evaluation system, perform a device-level quantitative score on individual devices based on fault severity parameters, development rate parameters, and propagation risk parameters, input the device-level quantitative score into a system-level evaluation unit, calculate the importance of the faulty device using weighted network centrality, and evaluate the system impact by combining the device-related data; The degree of impact of the system is input into a distributed message queue for hierarchical caching. The fault mode similarity is calculated for the evaluation result of the hierarchical cache based on time correlation and spatial correlation, and the alarm information with similar fault modes is merged. The secondary alarms in the merged alarm information are filtered using alarm suppression rules to obtain filtered hierarchical alarm information, and the filtered hierarchical alarm information is pushed to the remote expert terminal through a publish-subscribe model.

[0046] First, we acquire abnormal characteristic data and device-related data during equipment operation. Abnormal characteristic data includes multi-source sensor data such as equipment vibration signals, temperature signals, and pressure signals. Device-related data includes information such as equipment topology, operating parameters, and maintenance records.

[0047] Multimodal feature fusion is performed on the collected abnormal features and device-related data. A deep feature extraction network is used to extract feature representations of different modal data. For example, a convolutional neural network is used to extract the time-frequency features of vibration signals, and a recurrent neural network is used to extract time series features such as temperature and pressure. The extracted features are adaptively weighted and fused using an attention mechanism to achieve optimal feature combination.

[0048] Based on feature fusion, a transfer learning approach is used to train an adaptive fault diagnosis model. Historical data with sufficient fault labels is selected as the source domain, and data from new equipment to be diagnosed is used as the target domain. By calculating the distribution differences between source and target domain features, an adversarial learning approach is used to train a feature extractor, generating domain-invariant feature representations. Furthermore, a progressive knowledge distillation strategy is used to transfer expert diagnostic knowledge from the source domain to the feature representation, improving the model's diagnostic capabilities.

[0049] Based on the trained fault diagnosis model, a multi-level evaluation system is established to classify the severity of anomalies. At the device level, individual devices are quantitatively scored based on comprehensive considerations such as fault severity, development rate, and propagation risk. At the system level, the importance of the faulty device within the system is calculated based on the relationships between devices, assessing the overall impact on the system.

[0050] Finally, the evaluation results are entered into a distributed message queue for hierarchical caching. Fault pattern similarity analysis is performed on the cached evaluation results, and alarm information with similar fault characteristics is merged. Secondary alarms are filtered using alarm suppression rules to ensure accurate and effective alarm information is pushed. Using a publish-subscribe model, filtered, hierarchical alarm information is pushed to remote expert terminals, providing timely early warning of faults.

[0051] The solution of this application can: Through multimodal feature fusion and transfer learning methods, adaptive optimization of the fault diagnosis model is achieved, improving the model's diagnostic accuracy for new equipment and new fault types while reducing reliance on labeled data. A multi-level evaluation system, encompassing comprehensive assessments from the device to the system level, accurately quantifies the impact of faults and provides a scientific basis for decision-making in troubleshooting. A distributed message queue-based alarm push mechanism enables intelligent merging and filtering of alarm information, reducing redundant alarms and improving the real-time and accuracy of fault warnings, effectively supporting remote experts in fault diagnosis and resolution.

[0052] In an optional embodiment, the fusion features are used to train an adaptive fault diagnosis model, the maximum mean difference between the source domain features and the target domain features is calculated and the difference is minimized to reduce the feature distribution difference, a feature discriminator is constructed to distinguish the feature distributions of the source domain features and the target domain features to obtain a feature discrimination result, an adversarial learning method is used to train a feature extractor based on the feature discrimination result to generate a domain-invariant feature representation, and the expert knowledge in the source domain features is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model, including: The fusion feature is used to train an adaptive fault diagnosis model, the source domain feature and the target domain feature are mapped to a reproducing kernel Hilbert space, the empirical mean of the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated to obtain distribution difference data, and a multi-kernel maximum mean difference method is used to perform adaptive kernel function optimization on the distribution difference data to obtain a feature distribution difference value; A dual-branch feature discriminator is constructed to discriminate the feature distributions of the source domain features and the target domain features, the network parameters of the feature discriminator are updated by fixing the feature extractor parameters to obtain a feature discrimination result, the network parameters of the feature discriminator are updated by fixing the feature discriminator parameters based on the feature discrimination result, a gradient penalty term is introduced into the feature discriminator, and a spectral normalization technique is used to control the discrimination ability; The feature extractor constructs an optimization objective function based on classification loss, adversarial loss and maximum mean difference loss, and adopts an adaptive weight balancing strategy to iteratively optimize the optimization objective function to generate a domain-invariant feature representation; The expert knowledge in the source domain features is divided into a multi-level knowledge representation according to the degree of abstraction, and the multi-level knowledge representation is temperature-adaptively adjusted to obtain soft label data. The soft label data and the multi-level knowledge representation are input into a progressive knowledge distillation network, and a curriculum learning strategy is used to transfer the expert knowledge to the domain-invariant feature representation to obtain an adaptive fault diagnosis model.

[0053] First, the source domain data and target domain data are obtained and preprocessed to obtain initial features. The source domain features and target domain features are extracted separately through a feature extractor, and the extracted features are fused to obtain fused features.

[0054] When training an adaptive fault diagnosis model based on fused features, the source and target domain features are mapped into a reproducing kernel Hilbert space. The empirical mean of the two domain features is calculated in this space to obtain data representing the distribution difference between the two domains. Using the multi-kernel maximum mean difference method, the distribution difference data is optimized using a Gaussian kernel function to obtain a feature distribution difference value. A smaller difference value indicates closer feature distributions between the two domains.

[0055] Next, a feature discriminator with a two-branch structure is constructed, taking in source and target domain features. During training, the parameters of the feature extractor are first fixed, and only the network parameters of the feature discriminator are updated to obtain the feature discrimination results. Based on the discrimination results, the discriminator parameters are then fixed and updated. A gradient penalty term is introduced in the feature discriminator, and spectral normalization techniques are used to control the discriminative power and prevent overfitting.

[0056] The feature extractor's optimization objective function consists of three parts: classification loss, adversarial loss, and maximum mean difference loss. The classification loss ensures feature discriminability, the adversarial loss generates domain-invariant features, and the maximum mean difference loss reduces inter-domain differences. An adaptive weight balancing strategy dynamically adjusts and iteratively optimizes these three losses, ultimately generating a domain-invariant feature representation.

[0057] During the knowledge transfer phase, expert knowledge from source domain features is divided into multiple levels of abstraction into knowledge representations. These knowledge representations are temperature-adaptively adjusted to generate soft-label data. This soft-label data and the multi-level knowledge representations are fed into a progressive knowledge distillation network. Using a curriculum learning strategy, the expert knowledge is gradually transferred to domain-invariant feature representations, ultimately resulting in an adaptive fault diagnosis model.

[0058] The solution of this application can: By combining the multi-kernel maximum mean difference method with adversarial learning, the distribution difference between source and target domain features can be effectively reduced, improving the domain adaptation performance of the features and enabling the model to generalize better. A dual-branch feature discriminator structure, combined with gradient penalty and spectral normalization techniques, effectively controls the discriminant's discriminative power, avoids vanishing or exploding gradients during training, and improves the model's training stability. A progressive knowledge distillation strategy and curriculum learning method enable hierarchical transfer of expert knowledge, fully leveraging prior knowledge in the source domain to guide fault diagnosis in the target domain, thereby improving the diagnostic accuracy and reliability of the diagnostic model in the target domain.

[0059] In an optional embodiment, intelligent layered SVG vector graphics technology is used to support real-time annotation by multiple people. Causal reasoning analysis is performed on the expert annotation data and the diagnostic results of the adaptive fault diagnosis model to establish an expert knowledge base. A collaborative decision-making recommendation system based on deep reinforcement learning is integrated with the expert knowledge base to provide diagnostic solution recommendations to experts, including: Fault data containing time series data, feature data, and parameter data is acquired, and the fault data is converted into SVG graphic elements using intelligent layered SVG vector graphics technology to construct a data visualization base layer. An expert annotation operation layer is constructed on the data visualization base layer to support experts in sequentially selecting fault areas to obtain fault area data, marking fault parameters to obtain fault parameter data, and annotating causal relationships to obtain causal link data. An operation conversion algorithm is used to perform consistency processing on the concurrent causal link data of multiple experts, and the processed data is synchronized to the interactive response layer via WebSocket to form the expert annotation results. Causal reasoning analysis is performed on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model. The expert annotation results and the diagnostic results are aligned in time and space using a dynamic time warping algorithm. The causal relationship between the variables of the aligned data is identified based on a conditional independence test to construct an initial causal network. The initial causal network is modified based on expert prior knowledge and the strength of the causal relationship is quantified to obtain weighted causal knowledge. The weighted causal knowledge is encoded as triples and stored in the blockchain system to form an expert knowledge base. Based on the expert knowledge base, a deep reinforcement learning collaborative decision-making system is constructed. The fault scenario characteristics are integrated with historical diagnostic records to form a state vector. According to the causal relationship in the expert knowledge base, prior probabilities are assigned to diagnostic actions to construct an action space. A recurrent neural network is used to capture the temporal dependencies of the decision sequence to establish a double Q network. Through knowledge distillation, the experience of the expert knowledge base is transferred to the policy network. The input fault scenario is analyzed and diagnostic suggestions are generated and pushed to the experts.

[0060] First, the fault data is visualized. The system receives time-series data on the equipment's operating status (e.g., temperature, pressure, vibration, and other parameters changing over time), feature data (e.g., spectral and statistical characteristics), and equipment parameter data (e.g., equipment model and operating time). Using SVG technology, this data is converted into visual graphics: time-series data is converted into a line graph, feature data is converted into a scatter plot, and parameter data is displayed as text. SVG graphics employ a layered structure: a bottom layer for data visualization, a middle layer for expert annotation, and a top layer for interactive response.

[0061] In the expert annotation operation layer, multiple experts are supported to annotate online at the same time. Experts can use the rectangular selection tool to select the fault area, and the system records the coordinate information of the selection box and the corresponding time period. Experts can add parameter tags such as fault type and severity to the selected area. It also supports experts to draw causal relationship arrows to connect the associations between different parameters. When multiple experts annotate at the same time, the system uses a timestamp-based operation conversion algorithm to handle concurrency conflicts. In specific implementation, a unique timestamp is assigned to each annotation operation, and the operations are applied in sequence according to the timestamp order. For the annotation of overlapping areas, a voting mechanism is used to determine the final result. The processed annotation data is synchronized to the interactive response layer of all online experts in real time via WebSocket.

[0062] The system compares and analyzes the expert annotation results with the diagnostic results of the adaptive fault diagnosis model. First, the dynamic time warping algorithm is used to align the time series of the two sets of data. In the aligned data, the causal relationship between variables is identified through the conditional independence test. For example, it analyzes whether the abnormality of a certain parameter leads to the abnormality of other parameters and establishes a causal link between the parameters. The causal network is modified based on the prior knowledge provided by the experts (such as equipment structure, physical principles, etc.). A weight value is assigned to each causal relationship, and the weight value is calculated based on the data support and expert confidence. The weighted causal knowledge is encoded as a subject-relationship-object triple and stored in the blockchain system to form a knowledge base.

[0063] A collaborative decision-making recommendation system is built based on an expert knowledge base. The system combines the feature vector of the current fault scenario with the handling methods of similar cases in historical diagnostic records to form a state vector. The prior probability of each possible diagnostic action is calculated based on the causal relationships in the knowledge base. A recurrent neural network is used to establish a double-Q network structure, which uses historical decision sequences as input to predict the value of different diagnostic actions. Through knowledge distillation, the experience in the expert knowledge base is converted into parameters for the policy network. Finally, the system generates diagnostic recommendations based on the current state, including fault cause analysis and recommended treatment measures, and sends them to experts for decision-making reference.

[0064] The solution of this application can: The system enables efficient collection and sharing of expert knowledge. It supports real-time annotation by multiple people through intelligent layered SVG technology, ensuring consistency in annotation data and improving knowledge acquisition efficiency. The system establishes a complete causal reasoning analysis framework, deeply integrating expert experience with model diagnostic results to construct an interpretable fault diagnosis knowledge system and enhance the reliability of diagnostic results. A collaborative decision-making mechanism based on deep reinforcement learning enables the continuous accumulation and optimization of knowledge. The system continuously enhances decision-making capabilities through knowledge transfer, providing experts with more accurate diagnostic recommendations and significantly improving fault diagnosis efficiency.

[0065] In an optional embodiment, causal reasoning analysis is performed on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model. A dynamic time warping algorithm is used to align the expert annotation results and the diagnostic results in time and space. The causal relationship between the variables of the aligned data is identified based on a conditional independence test to construct an initial causal network. The initial causal network is modified based on expert prior knowledge and the strength of the causal relationship is quantified to obtain weighted causal knowledge. The weighted causal knowledge is encoded as triples and stored in the blockchain system to form an expert knowledge base, including: Based on the adaptive fault diagnosis model, causal reasoning analysis is performed on the expert annotation results and the diagnosis results to obtain initial variable association data; based on the initial variable association data, a dynamic time warping algorithm is used to construct a cumulative distance matrix, and the optimal alignment path between the expert annotation results and the diagnosis results is calculated according to the cumulative distance matrix to obtain first time-aligned data, and spatial feature matching is performed on the first time-aligned data through a multi-scale feature pyramid to obtain second spatiotemporal alignment data, and the first spatiotemporal alignment data and the second spatiotemporal alignment data are weightedly combined to form a spatiotemporal alignment dataset; Based on the spatiotemporal aligned dataset, conditional mutual information between variable pairs is calculated to obtain a mutual information matrix, a Fisher-Z test is used to perform conditional independence judgment on the mutual information matrix to obtain an independence judgment result, and the causal direction between variables is identified based on the independence judgment result to construct an initial causal network; expert prior knowledge is converted into network constraints, the structure of the initial causal network is optimized according to the network constraints to obtain a modified causal network, and a Bayesian network learning method is used to calculate the weight strength of each causal edge in the modified causal network to obtain weighted causal knowledge; The causal relationship in the weighted causal knowledge is encoded as a triple representation, where each triple contains a head entity, a tail entity and a relationship attribute, and the relationship attribute includes the causal relationship type and weight information; the blockchain system is used to encapsulate the triple representation into a block transaction unit, and the knowledge validity and version information of the block transaction unit are verified through a smart contract. The block transaction unit that passes the verification is written into the blockchain system in a time sequence to form an expert knowledge base.

[0066] First, a causal reasoning analysis was performed between the expert annotation results and the diagnostic results of the adaptive fault diagnosis model. By comparing and analyzing the time series characteristics and fault characteristics of the two data sets, key variables were extracted as initial variable association data. For example, in the case of fault diagnosis of an industrial production line, the expert annotation results included 10 key parameters, such as temperature, pressure, and vibration, and the diagnostic model output included the corresponding fault prediction values.

[0067] Next, a dynamic time warping algorithm is used for spatiotemporal alignment. First, a cumulative distance matrix is constructed, where the matrix elements represent the Euclidean distances between the corresponding time points of the two data sets. Taking temperature as an example, the expert annotation values are 35°C, 36°C, and 34°C, while the diagnostic values are 34.5°C, 35.8°C, and 34.2°C. This distance matrix is then calculated. Based on the distance matrix, a dynamic programming method is used to find the optimal alignment path, resulting in the first temporal alignment. Spatial feature matching is then performed using a multi-scale feature pyramid, extracting and matching features at different scales to obtain the second spatiotemporal alignment. Finally, the two alignment results are weighted and combined, with the weights adjusted according to the actual application scenario, to obtain the final spatiotemporal aligned dataset.

[0068] A conditional independence test was performed on the aligned dataset. The conditional mutual information between pairs of variables was calculated, and a mutual information matrix was constructed. For example, using temperature and pressure as an example, their mutual information values were calculated conditionally on other variables. The Fisher-Z test was used to determine conditional independence between variables. If the mutual information value was less than a set threshold, the variables were considered conditionally independent. Based on the independence test results, the causal direction between the variables was determined, and an initial causal network was constructed.

[0069] The network is modified by incorporating expert prior knowledge. Expert experience is converted into network constraints, such as the physical law that rising temperature inevitably leads to rising pressure. Based on these constraints, the network structure is optimized, removing or adding causal edges. A Bayesian network learning method is used to calculate the weight of each causal edge, with weight values ranging from 0 to 1, indicating the strength of the causal relationship.

[0070] Finally, causal knowledge is encoded as triples and stored on the blockchain. Each triple contains a head entity (such as temperature), a tail entity (such as pressure), and relationship attributes (causal type and weight value). Triples are encapsulated into block transactions, and their validity is verified by smart contracts. The contract checks the integrity, consistency, and version information of the knowledge. Once verified, it is written to the blockchain in a chronological order, forming a traceable expert knowledge base.

[0071] The solution of this application can: Dynamic time warping and multi-scale feature matching enable precise alignment of heterogeneous data, improving the accuracy of subsequent causal analysis and resolving the spatial and temporal inconsistencies between expert annotation and model diagnosis. A causal network is constructed using conditional independence tests combined with expert prior knowledge, ensuring data-driven objectivity while incorporating the subjective cognition of expert experience, enhancing the reliability and interpretability of causal relationship identification. Storing weighted causal knowledge in the form of triples in a blockchain system enables decentralized knowledge management and version traceability, ensuring the security, consistency, and scalability of the knowledge base and providing reliable knowledge support for the continuous optimization of fault diagnosis models.

[0072] Figure 2 This is a structural diagram of the multi-source heterogeneous data acquisition and remote collaborative analysis system for smart stations according to an embodiment of the present invention. Figure 2 As shown, the system includes: The first unit is used to collect multi-source heterogeneous data of smart stations, standardize the multi-source heterogeneous data to generate standard data; use multi-sensor fusion self-calibration laser scanning technology to build a real-time adaptive three-dimensional model of the station, and semantically map the adaptive three-dimensional model with the BIM model to establish an equipment geometry and attribute information library; use improved WebGL technology to build a digital twin visualization platform, map and integrate the standard data with the real-time adaptive three-dimensional model of the station in real time to form a digital twin model with dynamic update capabilities; use a distributed database cluster to store the standard data, and establish a real-time data update mechanism based on incremental synchronization with the digital twin model; The second unit is used to construct a self-evolving feature library containing equipment parameter abnormality features, system performance degradation features and safety risk warning features based on the digital twin model, and use an improved graph convolutional neural network to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; based on the equipment geometry and attribute information library, a knowledge graph equipment group analysis model integrating causal reasoning is established, and the dynamic correlation influence between devices is calculated through the graph neural network to obtain equipment association data; a transfer learning method is used to combine the abnormal features and the equipment association data to train an adaptive fault diagnosis model, and a multi-level evaluation system is established according to the diagnosis results of the adaptive fault diagnosis model to classify the degree of abnormality, and the graded alarm information is pushed to the remote expert terminal through a distributed message queue; The third unit is used to build a remote collaboration platform that supports weak network environments after receiving the graded alarm information, and call the digital twin model in the remote collaboration platform to realize real-time collaboration among multiple experts; use intelligent layered SVG vector graphics technology to support real-time labeling by multiple people, perform causal reasoning analysis on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model, and establish an expert knowledge base; integrate a collaborative decision-making recommendation system based on deep reinforcement learning based on the expert knowledge base to provide experts with diagnostic solution suggestions; record the diagnostic process through blockchain technology and ensure that the operation traces cannot be tampered with; feed back the diagnostic results confirmed by the experts to the digital twin model, and continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

[0073] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0074] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0075] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The method for multi-source heterogeneous data collection and remote collaborative analysis of smart stations is characterized by: include: Collect multi-source heterogeneous data from smart stations and standardize the multi-source heterogeneous data to generate standard data; A real-time adaptive 3D model of the site is constructed using multi-sensor fusion self-calibration laser scanning technology, and the adaptive 3D model is semantically mapped with the BIM model to establish an equipment geometry and attribute information library. A digital twin visualization platform is constructed using improved WebGL technology, and the standard data is mapped and integrated with the real-time adaptive 3D model of the site in real time to form a digital twin model with dynamic update capabilities. A distributed database cluster is used to store the standard data, and a real-time data update mechanism based on incremental synchronization is established with the digital twin model. Based on the digital twin model, a self-evolving feature library containing device parameter anomaly features, system performance degradation features, and safety risk warning features is constructed, and an improved graph convolutional neural network is used to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; A knowledge graph device group analysis model integrating causal reasoning is established based on the device geometry and attribute information database, and the dynamic correlation influence between devices is calculated through a graph neural network to obtain device association data. A transfer learning method is used to combine the abnormal characteristics and the device association data to train an adaptive fault diagnosis model. A multi-level evaluation system is established based on the diagnosis results of the adaptive fault diagnosis model to classify the degree of abnormality, and graded alarm information is pushed to remote expert terminals through a distributed message queue. After receiving the graded alarm information, a remote collaboration platform supporting a weak network environment is established, and the digital twin model is called in the remote collaboration platform to realize real-time collaboration among multiple experts; intelligent layered SVG vector graphics technology is used to support real-time labeling by multiple people, and causal reasoning analysis is performed on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model to establish an expert knowledge base; a collaborative decision-making recommendation system based on deep reinforcement learning is integrated with the expert knowledge base to provide experts with diagnostic solution suggestions; the diagnostic process is recorded through blockchain technology to ensure that the operation traces cannot be tampered with; the diagnostic results confirmed by the experts are fed back to the digital twin model to continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

2. The method according to claim 1, characterized in that The digital twin visualization platform is constructed using improved WebGL technology, and the standard data is mapped and integrated with the real-time adaptive three-dimensional model of the station in real time to form a digital twin model with dynamic update capability, including: Improve the vertex shader of the WebGL rendering pipeline to a parallel computing architecture to achieve batch preprocessing capabilities for scene data; introduce a hierarchical caching mechanism in the fragment shading stage and optimize the depth test algorithm; design a dynamic vertex buffer management mechanism to achieve dynamic scheduling of vertex data, and form an optimized basic rendering pipeline based on this dynamic adjustment; build an adaptive shader compilation system based on this optimized basic rendering pipeline, dynamically generate and cache GLSL shader code based on scene complexity, and output a rendering engine that supports real-time dynamic compilation; Utilizing the rendering engine supporting real-time dynamic compilation, the three-dimensional models of the equipment within the station are merged and processed through instanced rendering technology to generate a basic static scene; the basic static scene is spatially divided into an octree to obtain a hierarchical grid structure, and the models in the hierarchical grid structure are removed in real time based on a frustum clipping algorithm to output an optimized scene grid; based on the distance relationship between the camera and each model in the optimized scene grid, a multi-level precision model library and a distance threshold mapping table are established, and a real-time adaptive three-dimensional model of the station is generated by dynamically adjusting the model precision level; Construct a device identifier index table to establish a unique mapping relationship between externally collected standard data and the corresponding equipment in the real-time adaptive three-dimensional model of the station; hierarchically process the mapped standard data into a geometric parameter layer, an operating status layer, and an alarm information layer to form a hierarchical data structure; implement incremental push of the hierarchical data structure based on the WebSocket protocol, and smooth the data update frequency to output a real-time data stream; use a distributed state manager to establish a binding relationship between the real-time data stream and the device status, and store the scene status information in the corresponding state tree according to the hierarchical data structure; Based on the observer mode, a dynamic association is established between the state tree and the real-time data stream to realize automatic status update when the data changes; using the rendering engine that supports real-time dynamic compilation, the material and effect of the equipment in the real-time adaptive three-dimensional model of the station are adjusted in real time according to the latest status information in the state tree, the standard data and the three-dimensional model are mapped and integrated in real time, and the initial digital twin model is output; a two-way data synchronization mechanism is constructed to realize real-time data interaction between the initial digital twin model and the state tree, and the user operation records are synchronously updated to the state tree, continuously triggering the rendering engine to update the screen, and finally generating a complete digital twin model with dynamic interaction and real-time visualization capabilities.

3. The method according to claim 1, characterized in that Using an improved graph convolutional neural network to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; Based on the device geometry and attribute information database, a knowledge graph device group analysis model integrating causal reasoning is established. The dynamic correlation influence between devices is calculated through the graph neural network to obtain device association data including: A temporal self-attention module is designed in the convolutional layer, including a multi-head attention mechanism to extract time-scale features and a position encoding unit to maintain temporal information, outputting enhanced temporal features. An adaptive gating mechanism is designed based on the enhanced temporal features, which dynamically adjusts the propagation of information between nodes by updating the gate to retain historical information and resetting the gate to filter redundant information. The gated inter-node information is input into a multi-scale feature fusion module to obtain fused features. A feature pyramid is used to extract hierarchical features, and residual connections are used to retain the original information. Causal constraints and feature compression are applied to the fused features, and graph adversarial training is used to improve robustness, resulting in an improved graph convolutional neural network. The fused features are dynamically filtered using the adaptive gating mechanism to output standardized device spatiotemporal features; the standardized device spatiotemporal features are input into the improved graph convolutional neural network, and the causal perception module is used to causally constrain the standardized device spatiotemporal features to ensure the causal consistency of the features and obtain causal relationships; the features are compressed based on a hierarchical graph pooling strategy to obtain node-level anomaly representations; a variational autoencoder is constructed to encode and reconstruct the node-level anomaly representations, and the device anomaly features are output through optimization using a combined loss function of reconstruction error and KL divergence; Utilizing the device anomaly characteristics to calculate the time delay mutual information between nodes; determining the optimal time delay between nodes based on maximizing the mutual information; fusing the causal relationship with device attribute information to construct an initial knowledge graph; utilizing the causal perception module of the improved graph convolutional neural network to dynamically fuse the device anomaly characteristics with the initial knowledge graph; achieving knowledge compression through the hierarchical graph pooling strategy to obtain a device group causal analysis model; The device anomaly characteristics are input into the device group causal analysis model, and the dynamic coupling strength between nodes is calculated through the improved graph convolutional neural network; device groups are hierarchically clustered based on the dynamic coupling strength; an anomaly propagation network is constructed by combining the device anomaly characteristics and the causal relationship; the anomaly propagation path is analyzed using the temporal self-attention module; the dynamic coupling strength, the optimal delay and the causal relationship are integrated in multiple dimensions to output the final device association data.

4. The method according to claim 1, wherein A transfer learning method is used to combine the abnormal features with the device-related data to train an adaptive fault diagnosis model. A multi-level evaluation system is established based on the diagnosis results of the fault diagnosis model to classify the degree of abnormality. The graded alarm information is pushed to the remote expert terminal through a distributed message queue, including: The abnormal features are fused with the device-related data to construct a multimodal feature; historical data with sufficient fault labels is used as the source domain feature, and the new device data to be diagnosed is used as the target domain feature. The multimodal features are adaptively weighted and fused using an attention mechanism to obtain an adaptive fusion feature; The adaptive fusion feature is used to train an adaptive fault diagnosis model. The maximum mean difference between the source domain feature and the target domain feature is calculated and the difference is minimized to reduce the feature distribution difference. A feature discriminator is constructed to distinguish the feature distributions of the source domain feature and the target domain feature to obtain a feature discrimination result. Based on the feature discrimination result, an adversarial learning method is used to train a feature extractor to generate a domain-invariant feature representation. The expert knowledge in the source domain feature is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model. Input the diagnostic results of the adaptive fault diagnosis model into a multi-level evaluation system, perform a device-level quantitative score on individual devices based on fault severity parameters, development rate parameters, and propagation risk parameters, input the device-level quantitative score into a system-level evaluation unit, calculate the importance of the faulty device using weighted network centrality, and evaluate the system impact by combining the device-related data; The degree of impact of the system is input into a distributed message queue for hierarchical caching. The fault mode similarity is calculated for the evaluation result of the hierarchical cache based on time correlation and spatial correlation, and the alarm information with similar fault modes is merged. The secondary alarms in the merged alarm information are filtered using alarm suppression rules to obtain filtered hierarchical alarm information, and the filtered hierarchical alarm information is pushed to the remote expert terminal through a publish-subscribe model.

5. The method according to claim 4, characterized in that The adaptive fault diagnosis model is trained using the fusion features, the feature distribution difference is reduced by calculating the maximum mean difference between the source domain features and the target domain features and minimizing the difference value, a feature discriminator is constructed to distinguish the feature distributions of the source domain features and the target domain features to obtain a feature discrimination result, an adversarial learning method is used to train a feature extractor based on the feature discrimination result to generate a domain-invariant feature representation, and the expert knowledge in the source domain features is transferred to the domain-invariant feature representation through a progressive knowledge distillation strategy to obtain the adaptive fault diagnosis model, including: The fusion feature is used to train an adaptive fault diagnosis model, the source domain feature and the target domain feature are mapped to a reproducing kernel Hilbert space, the empirical mean of the source domain feature and the target domain feature in the reproducing kernel Hilbert space is calculated to obtain distribution difference data, and a multi-kernel maximum mean difference method is used to perform adaptive kernel function optimization on the distribution difference data to obtain a feature distribution difference value; A dual-branch feature discriminator is constructed to discriminate the feature distributions of the source domain features and the target domain features, the network parameters of the feature discriminator are updated by fixing the feature extractor parameters to obtain a feature discrimination result, the network parameters of the feature discriminator are updated by fixing the feature discriminator parameters based on the feature discrimination result, a gradient penalty term is introduced into the feature discriminator, and a spectral normalization technique is used to control the discrimination ability; The feature extractor constructs an optimization objective function based on classification loss, adversarial loss and maximum mean difference loss, and adopts an adaptive weight balancing strategy to iteratively optimize the optimization objective function to generate a domain-invariant feature representation; The expert knowledge in the source domain features is divided into a multi-level knowledge representation according to the degree of abstraction, and the multi-level knowledge representation is temperature-adaptively adjusted to obtain soft label data. The soft label data and the multi-level knowledge representation are input into a progressive knowledge distillation network, and a curriculum learning strategy is used to transfer the expert knowledge to the domain-invariant feature representation to obtain an adaptive fault diagnosis model.

6. The method according to claim 1, characterized in that Adopting intelligent layered SVG vector graphics technology to support real-time annotation by multiple people, performing causal reasoning analysis on expert annotation data and the diagnostic results of the adaptive fault diagnosis model, and establishing an expert knowledge base; Based on the expert knowledge base, a collaborative decision-making recommendation system integrating deep reinforcement learning is used to provide experts with diagnostic solution suggestions including: Fault data containing time series data, feature data, and parameter data is acquired, and the fault data is converted into SVG graphic elements using intelligent layered SVG vector graphics technology to construct a data visualization base layer. An expert annotation operation layer is constructed on the data visualization base layer to support experts in sequentially selecting fault areas to obtain fault area data, marking fault parameters to obtain fault parameter data, and annotating causal relationships to obtain causal link data. An operation conversion algorithm is used to perform consistency processing on the concurrent causal link data of multiple experts, and the processed data is synchronized to the interactive response layer via WebSocket to form the expert annotation results. Causal reasoning analysis is performed on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model. The expert annotation results and the diagnostic results are aligned in time and space using a dynamic time warping algorithm. The causal relationship between the variables of the aligned data is identified based on a conditional independence test to construct an initial causal network. The initial causal network is modified based on expert prior knowledge and the strength of the causal relationship is quantified to obtain weighted causal knowledge. The weighted causal knowledge is encoded as triples and stored in the blockchain system to form an expert knowledge base. Based on the expert knowledge base, a deep reinforcement learning collaborative decision-making system is constructed. The fault scenario characteristics are integrated with historical diagnostic records to form a state vector. According to the causal relationship in the expert knowledge base, prior probabilities are assigned to diagnostic actions to construct an action space. A recurrent neural network is used to capture the temporal dependencies of the decision sequence to establish a double Q network. Through knowledge distillation, the experience of the expert knowledge base is transferred to the policy network. The input fault scenario is analyzed and diagnostic suggestions are generated and pushed to the experts.

7. The method according to claim 6, characterized in that Performing causal reasoning analysis on the expert annotation results and the diagnostic results of the adaptive fault diagnosis model, using a dynamic time warping algorithm to perform spatiotemporal alignment on the expert annotation results and the diagnostic results, identifying the causal relationship between variables in the aligned data based on a conditional independence test to construct an initial causal network, modifying the initial causal network based on expert prior knowledge and quantifying the strength of the causal relationship to obtain weighted causal knowledge, encoding the weighted causal knowledge into triples and storing them in the blockchain system to form an expert knowledge base, including: Based on the adaptive fault diagnosis model, causal reasoning analysis is performed on the expert annotation results and the diagnosis results to obtain initial variable association data; based on the initial variable association data, a dynamic time warping algorithm is used to construct a cumulative distance matrix, and the optimal alignment path between the expert annotation results and the diagnosis results is calculated according to the cumulative distance matrix to obtain first time-aligned data, and spatial feature matching is performed on the first time-aligned data through a multi-scale feature pyramid to obtain second spatiotemporal alignment data, and the first spatiotemporal alignment data and the second spatiotemporal alignment data are weightedly combined to form a spatiotemporal alignment dataset; Based on the spatiotemporal aligned dataset, conditional mutual information between variable pairs is calculated to obtain a mutual information matrix, a Fisher-Z test is used to perform conditional independence judgment on the mutual information matrix to obtain an independence judgment result, and the causal direction between variables is identified based on the independence judgment result to construct an initial causal network; expert prior knowledge is converted into network constraints, the structure of the initial causal network is optimized according to the network constraints to obtain a modified causal network, and a Bayesian network learning method is used to calculate the weight strength of each causal edge in the modified causal network to obtain weighted causal knowledge; The causal relationship in the weighted causal knowledge is encoded as a triple representation, where each triple contains a head entity, a tail entity and a relationship attribute, and the relationship attribute includes the causal relationship type and weight information; the blockchain system is used to encapsulate the triple representation into a block transaction unit, and the knowledge validity and version information of the block transaction unit are verified through a smart contract. The block transaction unit that passes the verification is written into the blockchain system in a time sequence to form an expert knowledge base.

8. A smart station multi-source heterogeneous data acquisition and remote collaborative analysis system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect multi-source heterogeneous data of the smart station and standardize the multi-source heterogeneous data to generate standard data; A real-time adaptive 3D model of the site is constructed using multi-sensor fusion self-calibration laser scanning technology, and the adaptive 3D model is semantically mapped with the BIM model to establish an equipment geometry and attribute information library. A digital twin visualization platform is constructed using improved WebGL technology, and the standard data is mapped and integrated with the real-time adaptive 3D model of the site in real time to form a digital twin model with dynamic update capabilities. A distributed database cluster is used to store the standard data, and a real-time data update mechanism based on incremental synchronization is established with the digital twin model. The second unit is used to build a self-evolving feature library containing equipment parameter abnormality features, system performance degradation features, and safety risk warning features based on the digital twin model, and use an improved graph convolutional neural network to perform spatiotemporal correlation analysis on the standard data to obtain abnormal features; A knowledge graph device group analysis model integrating causal reasoning is established based on the device geometry and attribute information database, and the dynamic correlation influence between devices is calculated through a graph neural network to obtain device association data. A transfer learning method is used to combine the abnormal characteristics and the device association data to train an adaptive fault diagnosis model. A multi-level evaluation system is established based on the diagnosis results of the adaptive fault diagnosis model to classify the degree of abnormality, and graded alarm information is pushed to remote expert terminals through a distributed message queue. The third unit is used to build a remote collaboration platform that supports weak network environments after receiving the graded alarm information, and call the digital twin model in the remote collaboration platform to realize real-time collaboration among multiple experts; use intelligent layered SVG vector graphics technology to support real-time labeling by multiple people, perform causal reasoning analysis on the expert labeling data and the diagnosis results of the adaptive fault diagnosis model, and establish an expert knowledge base; integrate a collaborative decision-making recommendation system based on deep reinforcement learning based on the expert knowledge base to provide experts with diagnostic solution suggestions; record the diagnostic process through blockchain technology and ensure that the operation traces cannot be tampered with; feed back the diagnostic results confirmed by the experts to the digital twin model, and continuously optimize the self-evolving feature library and the adaptive fault diagnosis model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images