Training Methods for Industrial Anomaly Scenarios Based on Digital Twins and Large AI Models

By constructing a linkage between an industrial digital twin model and a large AI model, and generating and filtering abnormal scenario datasets, the problems of data scarcity and high security risks in existing technologies are solved. This enables efficient and secure training of industrial anomaly identification models, improving the accuracy and robustness of the models.

CN121659134BActive Publication Date: 2026-11-10浙江众合科技股份有限公司
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511681605.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-11-17
Publication Date
2026-11-10
Estimated Expiration
2045-11-17

AI Technical Summary

Technical Problem

Existing technologies rely on real historical anomaly data to train industrial anomaly identification models, resulting in data scarcity, limited scenario coverage, high security risks and costs in real-world testing, and difficulty in generating rich and physically realistic anomaly training data, which affects the accuracy and robustness of the models.

Method used

By constructing an industrial-based digital twin model, defining abnormal scenario types and parameters, using a large AI model to generate parameter control sequences to drive digital twin simulation and construct an abnormal scenario dataset, and using AI evaluation, screening, feedback, and iterative optimization, a target number of training sets are finally generated for downstream task model training.

Benefits of technology

It efficiently generates high-quality industrial anomaly training data, improves the practical application performance of downstream models, reduces data collection and model validation costs, avoids security risks, and improves the accuracy and reliability of models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121659134B_ABST
    Figure CN121659134B_ABST
Patent Text Reader

Abstract

This invention discloses a training method for industrial anomaly scenarios based on digital twins and AI large-scale models, comprising the following steps: Step 1, constructing an industrial basic digital twin model and defining anomaly scenario types and parameters; Step 2, the AI ​​large-scale model generating corresponding parameter control sequences; Step 3, inputting the parameter control sequences into the digital twin model for simulation, recording simulation data to form an anomaly scenario dataset; Step 4, inputting the anomaly scenario dataset back into the AI ​​large-scale model, which evaluates the dataset and selects a portion of the anomaly scenario data to be incorporated into the final training set; Step 5, using the excluded anomaly scenario data as feedback signals to cause the AI ​​large-scale model to readjust the parameter control sequences, repeating steps 3-4 until the target number of final training sets is generated; Step 6, sending the final training set data to downstream task models for training and optimization, thereby improving the performance of the downstream training models in practical applications.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial safety production technology, and in particular to a training method for industrial anomaly scenarios based on digital twins and AI large models. Background Technology

[0002] In the field of industrial safety, the timely identification and handling of equipment anomalies are directly related to production continuity, personnel safety, and control of property losses, and are the core guarantee for the stable operation of industrial systems. Currently, the training of industrial anomaly identification and prediction models mainly relies on historically collected real anomaly data. That is, by accumulating data on actual anomaly events that occur in industrial sites, the models are trained and optimized to achieve the identification and early warning of similar anomalies.

[0003] However, existing training methods based on real historical anomaly data have significant and insurmountable drawbacks, as follows: Major industrial anomalies are characterized by their low incidence and randomness, making it difficult to collect a sufficient quantity and diversity of real-world anomaly data in most industrial scenarios. Insufficient data leads to inadequate model training, making it prone to overfitting and failing to accurately learn the core features of anomalies, thus affecting the model's recognition accuracy.

[0004] Industrial systems operate under complex and variable conditions, and the combinations of different abnormal conditions and the manifestations of extreme scenarios are difficult to fully cover using historical data. Existing methods rely on limited historical data to train models, resulting in insufficient generalization ability and robustness. When faced with unseen abnormal conditions or extreme scenarios, they are prone to false alarms and false negatives.

[0005] To obtain more real anomaly data or verify the effectiveness of the model, anomaly simulation tests need to be conducted in a real industrial environment. This not only requires a high investment in equipment debugging and production downtime costs, but may also cause safety risks such as equipment damage and secondary accidents, seriously threatening personnel safety and production order.

[0006] In recent years, digital twin technology has been able to construct high-fidelity virtual industrial environments, and large AI models have demonstrated powerful capabilities in data generation and complex pattern learning. However, current technologies have not yet effectively integrated the two. When used alone, digital twin technology struggles to generate complex and abnormal scenarios that exceed the limitations of physical mechanism models; when used alone, large AI models generate abnormal data that lacks physical realism and cannot be directly used for model training. This makes it difficult to systematically and efficiently generate rich and effective training data for abnormal scenarios.

[0007] The aforementioned deficiencies directly result in poor training performance of existing industrial anomaly identification and prediction models. In practical applications, these models lack accuracy and robustness, failing to meet the demands of industrial safety production for accurate and rapid response to abnormal events. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies that mainly rely on historical real anomaly data to train industrial anomaly identification models, and fail to effectively integrate digital twins and AI large-scale models. These shortcomings include data scarcity, limited scenario coverage, high safety risks and costs in real-world testing, and difficulty in systematically generating rich and physically realistic anomaly training data. This invention provides an industrial anomaly scenario training method based on digital twins and AI large-scale models. By constructing an industrial basic digital twin model and defining anomaly scenario types and parameters, the AI ​​large-scale model generates corresponding parameter control sequences to drive digital twin simulation to construct anomaly scenario datasets. After AI evaluation and screening, the excluded data is used as feedback for iterative optimization until a target number of training sets are generated. These training sets are then used for downstream task model training and optimization. This achieves the goals of efficiently generating high-quality industrial anomaly training data, improving the practical application performance of downstream models, eliminating dependence on scarce real-world anomaly data, and avoiding the risks of real-world testing.

[0009] The objective of this invention is achieved through the following technical solution: A training method for industrial anomaly scenarios based on digital twins and large AI models includes the following steps: Step 1: Construct an industrial-based digital twin model and define abnormal scenario types and parameters; Step 2: The AI ​​big data model generates the corresponding parameter control sequence based on the abnormal scenario type and parameters; Step 3: Input the parameter control sequence into the digital twin model for simulation, record the simulation data, and form an abnormal scenario dataset; Step 4: Input the abnormal scene dataset into the AI ​​big model. The AI ​​big model evaluates the abnormal scene dataset and selects some abnormal scene data to be incorporated into the final training set. Step 5: Use the filtered and excluded abnormal scene data as feedback signals to make the AI ​​large model readjust the parameter control sequence. Repeat steps 3-4 until the target number of final training sets is generated. Step 6: Send the final training set data to the downstream task model so that the downstream task model can be trained and optimized, thereby improving the performance of the downstream training model in practical applications.

[0010] Preferably, in step 1, the industrial basic digital twin model is a digital twin model constructed based on the physical mechanism, historical operating data and three-dimensional model of the target industrial scenario.

[0011] Preferably, in step 1, the abnormal scenario type and parameters are defined as follows: Obtain the types of abnormal scenarios that need to be trained, and determine the control parameters that drive the digital twin model to change from a normal state to an abnormal state that matches the abnormal scenario type.

[0012] Preferably, the control parameters include equipment structural parameters, operating condition parameters, and environmental interference parameters. The equipment structural parameters include geometric dimensions and material properties. The operating condition parameters include operating speed and load intensity parameters. The environmental interference parameters include temperature, humidity, and physical properties parameters.

[0013] Preferably, in step 3, the simulation data is multimodal fusion data, specifically including one or more combinations of process control data of equipment operation, data collected by virtual sensors, three-dimensional scene image data, and abnormal event text log data.

[0014] Preferably, in step 4, the AI ​​large model evaluates the abnormal scene dataset, specifically as follows: The AI ​​large model evaluates the realism, severity, and relevance of the abnormal scenes in the abnormal scene dataset to the target abnormal scene type and scores them. Several abnormal scene data with higher scores are then incorporated into the final training set.

[0015] Preferably, the AI ​​big model includes at least one of a large language model and a multimodal generative model, and the AI ​​big model incorporates knowledge of industrial physical laws, equipment fault evolution logic, and abnormal pattern recognition.

[0016] Preferably, in step 5, the AI ​​large model readjusts the parameter control sequence, specifically including adjusting one or more of the following: the value range of the parameter control sequence, the parameter combination method, and the abnormal evolution timing logic.

[0017] The beneficial effects of this invention are as follows: This invention, through a complete technical process of constructing an industrial foundation digital twin model and defining anomaly parameters, generating parameter control sequences using a large AI model, driving twin simulation to construct a dataset, AI evaluation and screening, feedback iterative optimization, and training downstream models, efficiently generates massive amounts of physically realistic industrial anomaly scenario training data, completely eliminating the dependence on scarce real anomaly data; it solves the problems of poor generalization ability and high false positive / false negative rates in downstream models due to insufficient training data. It provides these models with a massive, diverse, and secure "real-world" training environment, significantly improving their accuracy and reliability in practical applications.

[0018] By leveraging a digital twin virtual environment to conduct full-process anomaly training and data generation, testing can be performed without the need for real industrial scenarios. This completely avoids safety risks such as equipment damage and secondary accidents that may occur during real-world testing, and significantly reduces the cost of data collection and model verification. Attached Figure Description

[0019] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the present invention; Figure 3 This is a flowchart illustrating the application of this invention to training on coal bunker collapse anomalies. Figure 4 This is a flowchart illustrating the application of the present invention to belt misalignment training; Figure 5 This is a flowchart of the application of the present invention to turnout anomaly training. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] Example: Training methods for industrial anomaly scenarios based on digital twins and large AI models, such as Figures 1-2 As shown, it includes the following steps: Step 1: Construct an industrial-based digital twin model and define abnormal scenario types and parameters; Step 2: The AI ​​big data model generates the corresponding parameter control sequence based on the abnormal scenario type and parameters; Step 3: Input the parameter control sequence into the digital twin model for simulation, record the simulation data, and form an abnormal scenario dataset; Step 4: Input the abnormal scene dataset into the AI ​​big model. The AI ​​big model evaluates the abnormal scene dataset and selects some abnormal scene data to be incorporated into the final training set. Step 5: Use the filtered and excluded abnormal scene data as feedback signals to make the AI ​​large model readjust the parameter control sequence. Repeat steps 3-4 until the target number of final training sets is generated. Step 6: Send the final training set data to the downstream task model so that the downstream task model can be trained and optimized, thereby improving the performance of the downstream training model in practical applications.

[0025] In step 1, the industrial foundation digital twin model is a digital twin model constructed based on the physical mechanism, historical operating data, and three-dimensional model of the target industrial scenario.

[0026] In step 1, the abnormal scenario type and parameters are defined, specifically as follows: Obtain the types of abnormal scenarios that need to be trained, and determine the control parameters that drive the digital twin model to change from a normal state to an abnormal state that matches the abnormal scenario type.

[0027] The control parameters include equipment structural parameters, operating condition parameters, and environmental interference parameters. The equipment structural parameters include geometric dimensions and material properties. The operating condition parameters include operating speed and load intensity parameters. The environmental interference parameters include temperature, humidity, and physical properties parameters.

[0028] In step 3, the simulation data is multimodal fusion data, specifically including one or more combinations of process control data of equipment operation, data collected by virtual sensors, three-dimensional scene image data, and abnormal event text log data.

[0029] In step 4, the AI ​​large model evaluates the abnormal scene dataset, specifically as follows: The AI ​​large model evaluates the realism, severity, and relevance of the abnormal scenes in the abnormal scene dataset to the target abnormal scene type and scores them. Several abnormal scene data with higher scores are then incorporated into the final training set.

[0030] The AI ​​big model includes at least one of a large language model and a multimodal generative model. The AI ​​big model incorporates knowledge of industrial physical laws, equipment fault evolution logic, and abnormal pattern recognition.

[0031] In step 5, the AI ​​large model readjusts the parameter control sequence, specifically including adjusting one or more of the following: the range of values ​​for the parameter control sequence, the parameter combination method, and the abnormal evolution timing logic.

[0032] Specifically, this example is used to train a model for identifying and warning of coal bunker collapse anomalies. Through the linkage of digital twins and a large AI model, it simulates the physical process of pressure imbalance and bunker wall instability caused by changes in coal quality or improper operation, ultimately leading to a bunker collapse. Figure 3 As shown, it includes the following steps: Constructing a digital twin model: Based on parameters such as the geometric dimensions, structural materials, and coal powder characteristics (density, friction coefficient) of the target coal bunker, construct its digital twin in a high-fidelity simulation platform. This model needs to be able to accurately calculate and output the dynamic changes of key physical quantities such as pressure distribution within the bunker, material level height, and stress on the supporting structure.

[0033] Define abnormal parameters and interfaces: Identify the core parameters that trigger coal bin collapse, such as the internal friction angle, cohesion, moisture content, and unloading rate of the coal. Set up externally driven data interfaces for these parameters in the digital twin platform.

[0034] The AI-powered large-scale model generates abnormal operating conditions: Instructions are sent to the finely tuned industrial large-scale model, such as generating 10 combinations of coal quality and operating parameters that could lead to severe coal bin collapse. The parameters must be within a reasonable but dangerous range. The AI ​​large-scale model generates parameter sequences based on its learned knowledge, for example: {Internal friction angle: 25°, cohesion: 1.2 kPa, moisture content: 24%, coal unloading rate: 280 t / h}.

[0035] Simulation and Data Acquisition Driven: The AI-generated parameter set is injected into the digital twin model via an interface to initiate high-speed physical simulation. Sensor data throughout the process, 3D stress cloud animations of key nodes, and event logs marked as breaches are collected simultaneously to form a multimodal training data sample.

[0036] Simulation Result Evaluation and Screening: The simulation results (such as the final pressure change curve and coal flow outflow image) are submitted to the AI ​​model for evaluation. The AI ​​judges the realism and effectiveness of the simulation based on its knowledge and provides a score and optimization suggestions. High-scoring samples are selected and added to the database, while the defect descriptions of low-scoring samples are fed back to step 3 for optimization of the next generation.

[0037] Iteration and Model Training: Repeat steps 3-5 to construct a large-scale, high-quality warehouse collapse anomaly dataset. Use this dataset to train a warehouse collapse early warning model based on a Temporal Convolutional Network (TCN) and attention mechanism. This model can predict the probability of a warehouse collapse occurring within a future period, using real-time sensor data as input.

[0038] Specifically, this example is used to train a detection and diagnosis model for abnormal belt misalignment in belt conveyors. Through virtual-real linkage, it simulates various scenarios where the belt's centerline deviates during operation due to factors such as idler roller failure, tension imbalance, and uneven material loading. For example... Figure 4 As shown, it includes the following steps: Constructing a digital twin model: Based on the mechanical design, belt properties, and drive parameters of the target belt conveyor, construct its dynamic digital twin. The model should be able to simulate the lateral vibration and belt misalignment during operation, and output virtual images and misalignment sensor data.

[0039] Define abnormal parameters and interfaces: Identify the core parameters that induce deviation, such as idler roller offset angle, roller non-parallelism, material drop point deviation, and tension force. Set up control interfaces for these parameters in the digital twin platform.

[0040] The AI-generated large-scale model simulates an abnormal operating condition: It indicates a progressive deviation fault caused by the combined effects of head roller tilt and middle drop point offset, providing 5 sets of parameters. The AI ​​generates parameters such as {roller tilt angle: 1.5°, drop point deviation: 10% bandwidth, tension: 80% of rated}.

[0041] Drive simulation and data acquisition: Parameters are injected into the model and simulation is performed. Data sequences of belt misalignment during operation, image frames captured by virtual cameras at the drive and tension ends, and vibration sensor data are acquired.

[0042] Simulation result evaluation and selection: The deviation curve and video clips are submitted to the AI ​​large model for evaluation. The AI ​​determines whether the deviation process is continuous and conforms to mechanical principles, and selects the most representative samples.

[0043] Iteration and Model Training: After generating rich samples through multiple iterations, a vision and data fusion deviation detection model (such as CNN+LSTM structure) is trained to achieve early deviation identification and fault cause classification based on real-time video stream and sensor data.

[0044] Specifically, this example is used to train a diagnostic model for abnormal railway turnout switching. Through virtual-real linkage, it simulates turnout switching failure scenarios caused by foreign objects in the switch rail, abnormal friction of the slide plate, insufficient switching force, etc. For example... Figure 5 As shown, it includes the following steps: Constructing a digital twin model: Based on the target turnout's model, geometric dimensions, installation parameters, and motor characteristics, construct an integrated digital twin of its mechanical-hydraulic-control system. The model must be able to simulate the contact process between the switch rail and the stock rail with high precision and output data such as conversion force, conversion displacement curves, and current curves in real time.

[0045] Define abnormal parameters and interfaces: Identify the core parameters that induce conversion anomalies, such as the friction coefficient of the slide plate, the resistance of the switch rail, the hydraulic pressure, and the motor power. Set the adjustment interfaces for these parameters in the platform.

[0046] The AI-generated large-scale model generates abnormal operating conditions: It indicates that the model generates three fault parameters that could cause the turnout to jam during switching, requiring the switching force curve to show a significant peak followed by a sharp drop. AI-generated parameters include: {friction coefficient increases to 0.25, simulating 5mm obstacle resistance, oil pressure: 90% of rated value}.

[0047] Simulation and Data Acquisition: Inject parameters and start the simulation. Acquire the full-process conversion force-displacement curve and motor current curve for each conversion attempt, and generate a three-dimensional state snapshot of the turnout jamming position.

[0048] Simulation result evaluation and screening: The conversion force curve and state snapshot are submitted to the AI ​​large model for evaluation. The AI ​​determines whether the abnormal pattern conforms to the real physical laws and screens out the valid samples.

[0049] Iteration and Model Training: After accumulating a sufficient number of samples, a turnout fault diagnosis model based on a one-dimensional convolutional network (1D-CNN) is trained so that it can quickly diagnose the root cause of turnout jamming based on the actual switching force and current curves.

[0050] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0051] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A training method for industrial anomaly scenarios based on digital twins and large AI models, characterized by: Includes the following steps: Step 1: Construct an industrial-based digital twin model and define abnormal scenario types and parameters; Step 2: The AI ​​big data model generates the corresponding parameter control sequence based on the abnormal scenario type and parameters; Step 3: Input the parameter control sequence into the digital twin model for simulation, record the simulation data, and form an abnormal scenario dataset; Step 4: Input the abnormal scene dataset into the AI ​​big model. The AI ​​big model evaluates the abnormal scene dataset and selects some abnormal scene data to be incorporated into the final training set. Step 5: Use the filtered and excluded abnormal scene data as feedback signals to make the AI ​​large model readjust the parameter control sequence. Repeat steps 3-4 until the target number of final training sets is generated. Step 6: Send the final training set data to the downstream task model so that the downstream task model can be trained and optimized, thereby improving the performance of the downstream task model in practical applications.

2. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 1, characterized in that, In step 1, the industrial foundation digital twin model is a digital twin model constructed based on the physical mechanism, historical operating data, and three-dimensional model of the target industrial scenario.

3. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 2, characterized in that, In step 1, the abnormal scenario type and parameters are defined, specifically as follows: Obtain the types of abnormal scenarios that need to be trained, and determine the control parameters that drive the digital twin model to change from a normal state to an abnormal state that matches the abnormal scenario type.

4. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 3, characterized in that, The control parameters include equipment structural parameters, operating condition parameters, and environmental interference parameters. The equipment structural parameters include geometric dimensions and material properties. The operating condition parameters include operating speed and load intensity parameters. The environmental interference parameters include temperature, humidity, and physical properties parameters.

5. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 3 or 4, characterized in that, In step 3, the simulation data is multimodal fusion data, specifically including one or more combinations of process control data of equipment operation, data collected by virtual sensors, three-dimensional scene image data, and abnormal event text log data.

6. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 1, characterized in that, In step 4, the AI ​​large model evaluates the abnormal scene dataset, specifically as follows: The AI ​​large model evaluates the realism, severity, and relevance of the abnormal scenes in the abnormal scene dataset to the target abnormal scene type and scores them. Several abnormal scene data with higher scores are then incorporated into the final training set.

7. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 1, characterized in that, The AI ​​big model includes at least one of a large language model and a multimodal generative model. The AI ​​big model incorporates knowledge of industrial physical laws, equipment fault evolution logic, and abnormal pattern recognition.

8. The industrial anomaly scenario training method based on digital twin and AI large model according to claim 1, characterized in that, In step 5, the AI ​​large model readjusts the parameter control sequence, specifically including adjusting one or more of the following: the range of values ​​for the parameter control sequence, the parameter combination method, and the abnormal evolution timing logic.

Citation Information

Patent Citations

  • Digital twin distribution network fault scene generation method and system based on reinforcement learning

    CN116933619A

  • Neural network model training method based on virtual reality and simulation and related equipment

    CN120337786A