A Neural Network-Based Digital Twin Visualization Modeling Method and System

Abstract

The present invention relates to the field of digital twin modeling, and discloses a digital twin visualization modeling method and system based on a neural network, including collecting multi-dimensional perception data of a target physical system, performing format unification and normalization processing on the original sensor data through a heterogeneous data fusion module, and constructing a high-dimensional input feature set for neural network modeling in combination with a structured embedding algorithm; performing a stability pre-evaluation on the constructed input feature set, and screening core features by using a perturbation sensitivity analysis mechanism; introducing a dynamic residual feedback mechanism to perform enhanced training on a preliminary twin model, and modeling the spatio-temporal dependence relationship of different components through a graph neural network; tracking the stability index of the visualization performance in real time during the model training process; and interpreting according to the multi-dimensional visualization results output by the final twin model in combination with an industrial scenario semantic rule base. The present invention has the advantage of improving the practicability of the twin model in an industrial scenario.

Description

Technical Field

[0001] The present invention relates to the field of digital twin modeling, and specifically to a digital twin visualization modeling method and system based on neural networks. Background Art

[0002] In the process of digital twin modeling of industrial systems, neural networks are widely used in the modeling of physical entity behaviors and state prediction due to their powerful non-linear fitting ability and adaptive learning characteristics. The current mainstream methods usually combine sensor data and use feedforward neural networks, convolutional neural networks or recurrent neural networks to build virtual models corresponding to actual devices or systems. However, there is generally a problem in the prior art that in the model training stage, neural networks are highly sensitive to changes in the input feature distribution. Once there are minor offsets or noise disturbances in the sensor data in the actual environment, it may lead to a significant decrease in the prediction accuracy of the twin model. This lack of robustness to feature perturbations limits the long-term stable operation and visual credible display of digital twin systems in complex industrial sites, especially in scenarios where high-frequency dynamic visual interactions are required. Therefore, it is necessary to design a digital twin visualization modeling method and system based on neural networks to improve the practicality of twin models in industrial scenarios. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a digital twin visualization modeling method and system based on neural networks, which has the advantage of improving the practicality of twin models in industrial scenarios and solves the problems in the above background art.

[0004] To achieve the above object of improving the practicality of twin models in industrial scenarios, the present invention provides the following technical solutions: A digital twin visualization modeling method based on neural networks, comprising the following steps:

[0005] Collect multi-dimensional perception data of the target physical system, perform format unification and normalization processing on the original sensor data through a heterogeneous data fusion module, and construct a high-dimensional input feature set for neural network modeling in combination with a structured embedding algorithm;

[0006] Perform a stability pre-evaluation on the constructed input feature set, screen core features by using a perturbation sensitivity analysis mechanism, and extract system behavior features in combination with a temporal convolutional network to generate a preliminary twin model structure;

[0007] Introduce a dynamic residual feedback mechanism to enhance the training of the preliminary twin model, model the spatio-temporal dependence relationship of different components through a graph neural network, and combine a perturbation injection training strategy;

[0008] During the model training process, the stability index of the visualization performance is tracked in real time, and the parameter settings of the visualization interaction layer are dynamically adjusted according to the index feedback;

[0009] According to the multi-dimensional visualization results output by the final twin model, combined with the industrial scenario semantic rule base for interpretation, an interpretability analysis report is automatically generated.

[0010] Preferably, the process of constructing a high-dimensional input feature set for neural network modeling by combining the structured embedding algorithm is as follows:

[0011] Standardize the format of the raw data collected from multiple sensors; normalize the data after unified format; according to the structural characteristics of the target physical system, use the structured embedding algorithm to express the relationship between each sub-component in the system in the form of a vector; fuse the structured information with the time series data so that the input features can reflect both the static structural information and the dynamic behavior pattern of the system; align and splice the structured embedding vector and the sensor data in the time dimension; finally, form a high-dimensional input feature set containing structure, behavior, and context semantics.

[0012] Preferably, the process of screening core features by using the perturbation sensitivity analysis mechanism is as follows:

[0013] Without introducing any perturbation, input the high-dimensional input feature set into the preliminarily constructed neural network model to obtain the system output as the benchmark response result; for each input feature, apply a small perturbation in turn, and the perturbation range can be set according to the numerical range of the feature; record the output change caused by each perturbation, and construct a feature-output change mapping relationship: for each feature, statistically calculate the output change amplitude corresponding to the perturbation, and calculate the perturbation sensitivity index.

[0014] Preferably, the process of combining the temporal convolutional network to extract system behavior features and generate the structure of the preliminary twin model is as follows:

[0015] Organize the core features screened by the perturbation sensitivity analysis into an input tensor in the time series format, retaining the time context information;

[0016] Construct a temporal convolutional network based on one-dimensional convolution;

[0017] Input the time series features into the temporal convolutional network model:

[0018] Each layer extracts the feature changes within the local time window through the convolutional kernel;

[0019] The output is the high-level behavior feature representation corresponding to each time step;

[0020] Pool the temporal behavior features output by the temporal convolutional network to generate a unified representation of the entire input sequence as the behavior encoding for predicting the state.

[0021] Preferably, the process of enhancing the training of the preliminary twin model by introducing a dynamic residual feedback mechanism is as follows:

[0022] Input the filtered input feature sequence into the preliminarily constructed digital twin model to obtain the behavior prediction output; compare the model prediction output with the actual system observation value to calculate the residual; use the historical residual sequence as a new type of input signal, introduce residual-related features to re-encode and form a residual feedback vector; construct a feedback module, fuse the residual feedback vector with the original features of the main model, and generate an enhanced input feature representation.

[0023] Preferably, the process of modeling the spatio-temporal dependence relationship of different components through a graph neural network and combining a perturbation injection training strategy is as follows:

[0024] Model the target physical system as a graph structure; introduce weights to represent the strength of the edges; assign feature vectors to each node; use a graph neural network for information propagation and aggregation, and each node updates its own state by aggregating the features of adjacent nodes.

[0025] Preferably, the process of dynamically adjusting the parameter settings of the visual interaction layer according to the index feedback is as follows:

[0026] Embed a visual performance evaluation module in the twin modeling system to continuously monitor the following key stability indicators;

[0027] Integrate a performance collection script through the front-end rendering engine;

[0028] Regularly feed the collected indicators back to the back-end control logic;

[0029] Use dynamic thresholds to judge and evaluate the visual performance trend;

[0030] Establish an index-parameter mapping rule library;

[0031] Automatically update the parameter configuration of the interaction layer according to the feedback result.

[0032] Preferably, the process of interpreting in combination with the industrial scenario semantic rule library and automatically generating an interpretability analysis report is as follows:

[0033] Bind the visual structured data to the entities in the industrial semantic rule library;

[0034] Enable a rule-based semantic engine for matching;

[0035] The matching result generates event tags;

[0036] Bind the identified events to the original data to construct the content of the analysis report.

[0037] Preferably, a multi-scale model fusion strategy is adopted to improve the prediction accuracy of the twin model. The process is as follows:

[0038] Input the input data with different resolutions into multiple sub-models respectively;

[0039] Each sub-model is independently trained and outputs local prediction results;

[0040] Through a weighted fusion strategy, the prediction results of different sub-models are combined to form a global prediction output;

[0041] According to the weighted output, optimize the accuracy and stability of the twin model.

[0042] A digital twin visualization modeling system based on a neural network, comprising:

[0043] Feature construction module: perform unified format conversion and normalization processing on multi-source sensor data, and generate a high-dimensional input feature set required by the neural network through a structured embedding algorithm;

[0044] Behavior extraction module: perform perturbation sensitivity analysis on the input feature set to screen core features, and extract the key behavior patterns of the system through a temporal convolutional network;

[0045] Enhanced training module: introduce a dynamic residual feedback mechanism to improve the model accuracy, and model the spatio-temporal dependencies between system components through a graph neural network, and combine the perturbation injection strategy to enhance the model robustness;

[0046] Adaptive regulation module: monitor the visualization stability index during the training process, and dynamically adjust the parameters of the interaction layer according to the feedback;

[0047] Report generation module: combine the output results of the twin model with the industrial scenario rule base for semantic understanding, and automatically generate an interpretability analysis report.

[0048] Compared with the prior art, the present invention provides a digital twin visualization modeling method and system based on a neural network, having the following beneficial effects:

[0049] The present invention realizes the accurate modeling of a physical system by collecting multi-dimensional perception data and performing heterogeneous data fusion to construct a high-dimensional input feature set. By analyzing the perturbation sensitivity and using a temporal convolutional network to extract system behavior features, a preliminary twin model structure is generated, and the spatio-temporal dependence modeling ability of the model is enhanced through a dynamic residual feedback mechanism and a graph neural network, further improving the prediction accuracy and robustness of the model. The visualization stability of the model is tracked in real time and the parameters of the interaction layer are dynamically adjusted. Combining with the interpretability analysis report generated by the industrial scenario semantic rule base, it ensures that the multi-dimensional visualization results output by the model are easy to understand and support decision-making. It improves the modeling accuracy, real-time performance, and interpretability of the digital twin system, effectively supporting prediction and decision-making in complex industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the method of the present invention;

[0051] Figure 2 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment

[0053] Please refer to Figure 1 As shown, a neural network-based digital twin visualization modeling method described in an embodiment of the present invention includes the following steps:

[0054] S1: Collect multi-dimensional perception data of the target physical system, perform format unification and normalization processing on the original sensor data through a heterogeneous data fusion module, and construct a high-dimensional input feature set for neural network modeling in combination with a structured embedding algorithm.

[0055] The process of constructing a high-dimensional input feature set for neural network modeling in the above S1 in combination with a structured embedding algorithm is as follows:

[0056] Standardize the format of the raw data collected from multiple sensors, unify the timestamp, sampling frequency, data type, etc., to ensure the comparability of heterogeneous data on a unified time scale; perform normalization on the data with unified format to eliminate the influence of dimensionality between different data dimensions and improve the stability and convergence speed of neural network training; according to the structural characteristics of the target physical system, use the structured embedding algorithm to express the relationship between each sub-component in the system in the form of vectors; utilize techniques such as sliding windows and time decay functions to fuse the structured information with time series data, so that the input features can simultaneously reflect the static structural information and dynamic behavior patterns of the system; align and splice the structured embedding vectors and sensor data in the time dimension to generate multi-dimensional, context-aware composite input features, constituting the high-dimensional input tensor required for neural network modeling; finally form a multi-dimensional feature set containing structure, behavior, and context semantics, providing a rich information basis for subsequent feature screening, time series modeling, and residual feedback training.

[0057] Through the heterogeneous data fusion module, perform format unification and normalization processing on the raw data from different types of sensors, effectively eliminate the structural differences and scale inconsistencies between data, and fully mine the spatial structure and functional association information of the target physical system in combination with the structured embedding algorithm, constructing a high-dimensional input feature set with both time series characteristics and system topology semantics, significantly improving the accuracy and generalization ability of the neural network to model the system behavior, and providing a high-quality input basis for the stable construction of the subsequent twin model.

[0058] S2: Conduct a pre-evaluation of the stability of the constructed input feature set, use the perturbation sensitivity analysis mechanism to screen out the core features, and extract the system behavior features in combination with the temporal convolutional network to generate the initial twin model structure.

[0059] The process of screening out the core features by using the perturbation sensitivity analysis mechanism in S2 is as follows:

[0060] Without introducing any perturbations, input the high-dimensional input feature set into the preliminarily constructed neural network model to obtain the system output as the benchmark response result; for each input feature, apply a small perturbation in turn, and the perturbation range can be set according to the numerical range of the feature to ensure that the perturbation is representative and out of control; input the input feature set with a single perturbation into the model respectively, record the output changes caused by each perturbation, and construct a feature-output change mapping relationship: for each feature, statistically calculate the output change amplitude corresponding to its perturbation, and calculate the perturbation sensitivity index, such as: mean output deviation, normalized change rate, variance change rate, etc., these indexes reflect the influence degree of the feature on the model output; according to the calculated sensitivity indexes, sort all the features. Select the top several features with higher sensitivity as the core input feature set, and eliminate redundant or less influential features on the output.

[0061] In S2, the process of combining a temporal convolutional network to extract system behavior features and generate a preliminary digital twin model structure is as follows:

[0062] Organize the core features screened by perturbation sensitivity analysis into an input tensor in time series format, retaining temporal context information;

[0063] Construct a temporal convolutional network based on one-dimensional convolution, including:

[0064] Causal convolution: Ensure that the output depends only on the current and past time steps, satisfying the temporal causality of the actual system.

[0065] Dilated convolution: Adopt an exponentially increasing dilation rate to expand the receptive field to capture long-term dependencies.

[0066] Residual connection: Introduce a skip connection structure to alleviate the training difficulty of deep networks and improve stability;

[0067] Input the time series features into the temporal convolutional network model:

[0068] Each layer extracts the feature changes within a local time window through a convolutional kernel;

[0069] Stack multiple layers to achieve comprehensive modeling of short-term and long-term dynamic patterns;

[0070] The output is the high-level behavior feature representation corresponding to each time step;

[0071] Pool the temporal behavior features output by the temporal convolutional network to generate a unified representation of the entire input sequence, serving as the behavior encoding of the current or predicted state of the system;

[0072] Use the above behavior encoding as the backbone modeling basis, combine the input-output mapping relationship, and generate a preliminary digital twin model structure. The model has the following characteristics: It can capture the behavior patterns of the system evolving over time; it can serve as the infrastructure for subsequently introducing graph neural networks and residual feedback mechanisms; it provides a basis for the dynamic response of system state changes at the visualization level. The output of the preliminary digital twin model includes: Prediction of the current system state; Short-term trend prediction; Identification of key behavior events.

[0073] Through the stability pre-evaluation of the constructed high-dimensional input feature set, use the perturbation sensitivity analysis mechanism to screen out the core features that have a significant impact on system behavior, effectively reducing redundant dimensions and noise interference; further combine the temporal convolutional network to extract key temporal behavior patterns, realize the efficient modeling of the system's dynamic response, and thus generate a preliminary digital twin model with good temporal perception ability and structural scalability, improving the adaptability and prediction accuracy of the model for complex working conditions.

[0074] S3: Introduce a dynamic residual feedback mechanism to enhance the training of the preliminary twin model, model the spatio-temporal dependence relationships of different components through a graph neural network, and combine a perturbation injection training strategy.

[0075] The process of introducing a dynamic residual feedback mechanism to enhance the training of the preliminary twin model in the above S3 is as follows:

[0076] Input the filtered input feature sequence into the preliminarily constructed digital twin model (system behavior features extracted based on a temporal convolutional network) to obtain the predicted output of the system state or behavior; compare the model predicted output with the actual system observation value to calculate the residual, that is, the prediction error; use the historical residual sequence as a new type of input signal, introduce residual-related features to re-encode and form a residual feedback vector to enhance the model's perception ability of prediction error changes; construct a feedback module, fuse the residual feedback vector with the original features of the main model to generate an enhanced input feature representation: dynamically adjust the feature weights; focus on compensating for dimensions or time periods with large deviations in model prediction; adopt a joint loss function to perform end-to-end training on the model after fusing the residual feedback. The dynamic residual mechanism is updated in each training iteration, gradually guiding the model to correct prediction blind spots, improving the global fitting ability and stability; continuously monitor the residual evolution trend of the model on the validation set. When the residual distribution tends to be stable or drops to a set threshold, it is regarded as the convergence of the enhanced training, and a digital twin model with optimized performance is obtained.

[0077] The process of modeling the spatio-temporal dependence relationships of different components through a graph neural network and combining a perturbation injection training strategy in the above S3 is as follows:

[0078] Model the target physical system as a graph structure, where: nodes represent the key components or sensing units of the system; edges represent the spatial connection relationships, functional coupling relationships, or data correlations between them; introduce weights or attributes to represent the strength of the edges; assign a feature vector to each node, including: sensor input data of each component; temporal behavior features (from the extraction results of the temporal convolutional network); residual feedback vectors; use a graph neural network for information propagation and aggregation: each node updates its own state through the feature aggregation of adjacent nodes; the aggregation method adopts algorithms such as GCN, GAT, GraphSAGE, etc.; support multi-layer propagation to capture a wider system dependence structure;

[0079] Introduce a temporal extension method, including:

[0080] Dynamic graph modeling: Nodes and edges change over time;

[0081] Spatial-temporal graph convolutional network: Combine convolution on the graph structure to extract time series dependencies;

[0082] Modeling the evolution of node states: Introduce a recurrent unit in the time dimension to represent the change of state over time.

[0083] Construct multiple types of perturbations for training enhancement, including:

[0084] Input perturbation: Add noise or simulate abnormal fluctuations to the sensor data;

[0085] Structure perturbation: Perturb the edge weights in the graph structure to simulate connection failures or structural changes;

[0086] Time perturbation: Adjust the rhythm of the input sequence to simulate non-stationary dynamics.

[0087] Generate multiple training copies through perturbations to enhance the model's robustness to abnormal inputs and structural changes, and improve its generalization ability; Combine the outputs of the perturbed version and the original version, and design corresponding training objectives: Use contrastive loss functions or consistency regularization terms; Ensure the stability and consistency of the model's output under perturbation conditions.

[0088] The technical solution of this embodiment is as follows: Based on the preliminarily constructed digital twin model, introduce a dynamic residual feedback mechanism to monitor the model prediction error in real time and feedback the residual information to the model input to dynamically adjust the feature weights and optimize the learning path; At the same time, construct the graph structure between the key components of the system, use graph neural networks to model their spatio-temporal dependence relationships, and capture the complex structural couplings and behavioral associations inside the system; Combine the perturbation injection training strategy to enhance the diversity and robustness of the training data through perturbations to the input data, graph structure, and time dimension, and further improve the model's adaptability and generalization performance under complex working conditions. It realizes the error self-adaptive adjustment ability, system structure relationship modeling ability, and anti-interference robustness of the digital twin model when facing high-dimensional dynamic data, improves the model's sensitive response ability and prediction accuracy to multi-source data changes, makes the modeling results closer to the real system behavior, and is suitable for the dynamic perception, behavior prediction, and visual analysis requirements in high-complexity industrial scenarios. Embodiment

[0089] As Figure 1 shown, a digital twin visualization modeling method based on neural networks further includes the following steps:

[0090] S4: During the model training process, continuously track the stability indicators of the visualization performance, and dynamically adjust the parameter settings of the visualization interaction layer according to the indicator feedback.

[0091] The process of dynamically adjusting the parameter settings of the visualization interaction layer according to the indicator feedback in S4 is as follows:

[0092] Embed a visualization performance evaluation module in the twin modeling system, and continuously monitor the following key stability indicators:

[0093] Frame rate: Measure the smoothness of the visualization rendering;

[0094] Response delay: Reflects the response time after an interactive operation;

[0095] Error heat distribution consistency: Compares the distribution deviation between the prediction result and the actual data in the visualization;

[0096] User interaction frequency and area heat: Reflects the areas of user attention and operation habits;

[0097] Color jitter rate: Measures the stability of color changes in the visualization.

[0098] Integrate the performance collection script through the front-end rendering engine;

[0099] Regularly feedback the collected metrics to the back-end control logic;

[0100] Use methods such as sliding window statistics and dynamic threshold judgment to evaluate the visualization performance trend.

[0101] Establish a mapping rule library between metrics and parameters to implement a feedback-driven automatic parameter adjustment mechanism, for example:

[0102] FPS lower than the threshold: Reduce the rendering precision and simplify graphic elements;

[0103] Increasing interaction delay: Reduce the dynamic refresh frequency and use a local refresh strategy;

[0104] Inconsistent visual distribution of errors: Adjust the color mapping strategy or enhance the prediction error prompt;

[0105] Frequent interaction in the hot spot area: Increase the resolution of this area or the frequency of label prompts.

[0106] The system automatically updates the parameter configuration of the interaction layer according to the feedback results, such as: view resolution (switching between high / low precision modes); color gradient range (improving recognition); rendering refresh period (balancing low latency and low resource consumption); animation switch and simplified logic (improving interaction efficiency); data loading method (full volume vs incremental vs cache refresh).

[0107] S5: According to the multi-dimensional visualization results output by the final twin model, combined with the industrial scenario semantic rule library for interpretation, automatically generate an interpretability analysis report.

[0108] The process of interpreting and automatically generating an interpretability analysis report by combining with the industrial scenario semantic rule library in S5 is as follows:

[0109] Bind the visualized structured data with the entities in the industrial semantic rule library, such as:

[0110] Data field → Physical quantity (temperature → heat exchange unit);

[0111] Trend change → semantic event (temperature continuously rising → cooling failure trend);

[0112] Spatial region → equipment or process section name.

[0113] Enable a rule-based semantic engine to match the following patterns:

[0114] Abnormal trend recognition: For example, an indicator continuously rises by 20% within 15 minutes;

[0115] Status change detection: For example, the current pressure exceeds the safety upper limit;

[0116] Multi-index joint analysis: For example, temperature + vibration + flow anomaly → failure risk;

[0117] Generate event tags for the matching results, such as warning: possible pump failure, recommended maintenance: heat exchanger efficiency decline.

[0118] Bind the identified events to the original data to construct the content of the analysis report, including:

[0119] Overview section: The time range of this round of analysis, the system coverage;

[0120] Abnormal analysis: List all detected abnormal events and their data supports;

[0121] Semantic interpretation: Refer to the rule library description to explain the meaning of the event and possible causes;

[0122] Impact assessment: Deduce the potential impact path by combining the operation map of the equipment or system;

[0123] Suggested measures: Call the expert rules or experience library to provide operation suggestions, risk levels, handling time limits, etc.

[0124] Use the preset template + dynamic filling method to automatically generate text with strong readability:

[0125] Example: "At 14:35 on May 11, 2025, it was detected that the temperature at the outlet of the heat exchanger continued to rise (+8.2°C, duration 12 minutes), which has exceeded the upper limit threshold. Combining with the semantic rule R_302, it is judged as abnormal cooling efficiency. It is recommended to check the cooling water flow and the operation status of the pump body."

[0126] Support generating reports combining text and graphics, embedding chart screenshots, highlighted abnormal point diagrams, process trends, etc.

[0127] By combining the multi-dimensional visualization results output by the digital twin model with the semantic rule library of industrial scenarios, automatic semantic interpretation and intelligent event recognition of model prediction data can be achieved. It can accurately identify key abnormal trends, operating condition changes and potential failure risks, and generate structured, graphic and text-based interpretable analysis reports in natural language, thereby greatly improving the efficiency of understanding and operability of model results, and strengthening the system's intelligent diagnosis and decision-making support capabilities. Example

[0128] See also Figure 2 As shown, a digital twin visualization modeling system based on a neural network according to an embodiment of the present invention includes:

[0129] Feature construction module: It converts and normalizes multi-source sensor data into a unified format, and generates the high-dimensional input feature set required by the neural network through a structured embedding algorithm;

[0130] Behavior extraction module: performs perturbation sensitivity analysis on the input feature set to screen the core features, and extracts the key behavior patterns of the system through a temporal convolutional network;

[0131] Enhanced training module: Introducing a dynamic residual feedback mechanism to improve model accuracy, and using graph neural networks to model the spatiotemporal dependencies between system components, combined with a disturbance injection strategy to enhance model robustness;

[0132] Adaptive control module: monitors visual stability indicators during training and dynamically adjusts the parameters of the interaction layer based on feedback;

[0133] Report generation module: Combines the output results of the twin model with the industrial scenario rule library for semantic understanding and automatically generates an explainable analysis report.

[0134] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0135] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital twin visualization modeling method based on a neural network, characterized in that, It includes the following steps: Collect multi-dimensional perception data of the target physical system, unify the format and normalize the original sensor data through the heterogeneous data fusion module, and construct a high-dimensional input feature set for neural network modeling by combining the structured embedding algorithm; Conduct a pre-evaluation of the stability of the constructed input feature set, screen the core features using the perturbation sensitivity analysis mechanism, extract the system behavior features by combining the temporal convolutional network, and generate a preliminary twin model structure; Introduce a dynamic residual feedback mechanism to enhance the training of the preliminary twin model, model the spatio-temporal dependence relationship of different components through the graph neural network, and combine the perturbation injection training strategy; During the model training process, real-time track the stability index of the visualization performance, and dynamically adjust the parameter settings of the visualization interaction layer according to the index feedback; According to the multi-dimensional visualization results output by the final twin model, combine the industrial scenario semantic rule base for interpretation, and automatically generate an interpretable analysis report; The process of screening core features using the perturbation sensitivity analysis mechanism is as follows: Without introducing any perturbation, input the high-dimensional input feature set into the preliminarily constructed neural network model to obtain the system output as the benchmark response result; for each input feature, apply a small perturbation in turn, and the perturbation range is set according to the numerical range of the feature; record the output change caused by each perturbation, and construct a feature-output change mapping relationship: for each feature, statistically analyze the output change amplitude corresponding to the perturbation, and calculate the perturbation sensitivity index; The process of modeling the spatio-temporal dependence relationship of different components through the graph neural network and combining the perturbation injection training strategy is as follows: Model the target physical system as a graph structure; introduce weights to represent the strength of the edges; assign feature vectors to each node; use the graph neural network for information propagation and aggregation, and each node updates its own state by aggregating the features of adjacent nodes.

2. The digital twin visualization modeling method based on a neural network according to claim 1, wherein The process of constructing a high-dimensional input feature set for neural network modeling by combining the structured embedding algorithm is as follows: Standardize the format of the original data collected from multiple sensors; normalize the data after unifying the format; according to the structural characteristics of the target physical system, use the structured embedding algorithm to express the relationship between each sub-component in the system in vector form; fuse the structured information with the time series data so that the input features can reflect both the static structure information and the dynamic behavior pattern of the system; Align and splice the structured embedding vector and the sensor data in the time dimension; finally form a high-dimensional input feature set containing structure, behavior, and context semantics.

3. A digital twin visualization modeling method based on a neural network according to claim 1, characterized in that, The process of extracting the system behavior features by combining the temporal convolutional network and generating a preliminary twin model structure is as follows: Organize the core features screened by the perturbation sensitivity analysis into an input tensor in the time series format, and retain the time context information; Construct a temporal convolutional network based on one-dimensional convolution; Input the time series features into the temporal convolutional network model: Each layer extracts the feature changes within the local time window through the convolutional kernel; The output is the high-level behavior feature representation corresponding to each time step; Pool the temporal behavior features output by the temporal convolutional network to generate a unified representation of the entire input sequence as the behavior encoding for predicting the state.

4. A digital twin visualization modeling method based on a neural network according to claim 1, characterized in that The process of enhancing the initial twin model by introducing a dynamic residual feedback mechanism is as follows: Input the filtered input feature sequence into the initially constructed digital twin model to obtain the behavioral prediction output; compare the model prediction output with the actual system observation value to calculate the residual; use the historical residual sequence as a new type of input signal, introduce residual-related features to re-encode and form a residual feedback vector; construct a feedback module to fuse the residual feedback vector with the original features of the main model to generate an enhanced input feature representation.

5. A digital twin visualization modeling method based on a neural network according to claim 1, characterized in that, The process of dynamically adjusting the parameter settings of the visualization interaction layer according to the indicator feedback is as follows: Embed a visualization performance evaluation module in the twin modeling system to continuously monitor the following key stability indicators; Integrate a performance collection script through the front-end rendering engine; Regularly feedback the collected indicators to the back-end control logic; Use a dynamic threshold to judge and evaluate the visualization performance trend; Establish an indicator-parameter mapping rule library; Automatically update the parameter configuration of the interaction layer according to the feedback result.

6. A digital twin visualization modeling method based on a neural network according to claim 1, characterized in that, The process of generating an interpretable analysis report by combining with the industrial scenario semantic rule library is as follows: Bind the visualized structured data to the entities in the industrial semantic rule library; Enable a rule-based semantic engine for matching; The matching result generates event tags; Bind the identified events to the original data to construct the content of the analysis report.

7. A digital twin visualization modeling method based on a neural network according to claim 1, characterized in that The process of adopting a multi-scale model fusion strategy to improve the prediction accuracy of the twin model is as follows: Input the input data with different resolutions into multiple sub-models respectively; Each sub-model is independently trained and outputs local prediction results; Through a weighted fusion strategy, combine the prediction results of different sub-models to form a global prediction output; Optimize the accuracy and stability of the twin model according to the weighted output.

8. A digital twin visualization modeling system based on a neural network, applied to the method according to any one of claims 1-7, characterized in that, Including: Features Construction module: perform unified format conversion and normalization processing on multi-source sensor data, and generate a high-dimensional input feature set required by the neural network through a structured embedding algorithm; Behavior extraction module: perform perturbation sensitivity analysis on the input feature set to screen core features, and extract the key behavior patterns of the system through a temporal convolutional network; Enhanced training module: introduce a dynamic residual feedback mechanism to improve the model accuracy, and model the spatio-temporal dependencies between system components through a graph neural network, and combine the perturbation injection strategy to enhance the model robustness; Adaptive regulation module: monitor the visualization stability indicators during the training process, and dynamically adjust the parameters of the interaction layer according to the feedback; Report generation module: combine the output results of the twin model with the industrial scenario rule library for semantic understanding, and automatically generate an interpretable analysis report.

Images