A Method and System for IoT Device Group Collaboration and Fault Prediction Based on Digital Twin

Abstract

This invention discloses a method and system for collaborative fault prediction of IoT devices based on digital twins. By using multi-protocol adaptation and intelligent preprocessing mechanisms in the data acquisition unit, it breaks down data barriers between devices, achieves unified access and standardized processing of heterogeneous sensor data, and provides a highly consistent real-time data source for the digital twin model. The digital twin model construction unit uses a real-time synchronization module and a fault feature extraction module to continuously calibrate the mapping relationship between the virtual model and the physical entity, ensuring that the twin accurately reflects the real-time status and potential fault modes of the device group, laying a high-fidelity foundation for collaborative analysis and prediction.

Description

Technical Field

[0001] This invention relates to the field of the Internet of Things (IoT), and in particular to a method and system for IoT device group collaboration and fault prediction based on digital twins. Background Technology

[0002] In terms of IoT device group collaboration, it mainly relies on centralized cloud platforms or distributed edge computing architectures, and achieves task allocation and data synchronization between devices through preset rules or simple machine learning algorithms (such as decision trees or support vector machines).

[0003] In terms of fault prediction, existing technologies mostly use statistical models (such as ARIMA or exponential smoothing) or shallow neural networks (such as BP neural networks) to analyze the historical trends of equipment sensor data in order to identify abnormal patterns.

[0004] Existing systems often design the collaboration and prediction modules separately. The collaboration module focuses on task scheduling while neglecting device health status, while the prediction module analyzes data from individual devices in isolation, failing to leverage digital twin technology to achieve closed-loop feedback between the physical and virtual worlds. This results in wasted resources, high maintenance costs, and an inability to adapt to the real-time changing environment of a group of devices.

[0005] Existing technologies for predicting IoT device failures often limit themselves to isolated analysis of single devices, failing to effectively integrate the collaborative behavior of a group of devices. This results in insufficient prediction accuracy and significant response delays. Traditional methods rely heavily on static historical data modeling, lacking real-time dynamic update mechanisms and failing to capture the cascading failure effects caused by interactions between devices.

[0006] The disconnect between predictive models and collaborative control makes it difficult to achieve proactive intervention and resource optimization after fault warning, thus affecting the overall reliability and efficiency of the system.

[0007] Existing technologies are insufficient in handling device heterogeneity (such as different communication protocols and differences in computing power), which further exacerbates the problems of bias in prediction results and weak generalization ability.

[0008] In summary, a method and system for collaborative and fault prediction of IoT devices based on digital twins is needed to address the shortcomings of existing technologies. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for collaborative fault prediction of IoT devices based on digital twins, aiming to solve the aforementioned problems.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for IoT device group collaboration and fault prediction based on digital twins, comprising the following steps:

[0011] Step S1: Construct a digital twin model of the IoT device group, and continuously collect the operating parameters and abnormal signals of each device through a real-time monitoring module;

[0012] Step S2: Based on the communication logs and interaction event data between devices, construct a device communication topology map and a behavior dependency graph to characterize the collaboration pattern of the device group;

[0013] Step S3: Train the agent model in the simulation environment using a multi-agent reinforcement learning algorithm, quantify and evaluate the collaborative efficiency of the equipment group, and obtain the task completion rate and resource utilization rate indicators.

[0014] Step S4: Combine graph neural networks to extract features from device nodes, simulate the propagation path of faults in the communication topology, and assess the probability of system-level cascading failures.

[0015] Step S5: Based on the collaborative performance evaluation results and fault risk prediction results, dynamically optimize the allocation of computing resources, storage resources and network bandwidth of the device group;

[0016] Step S6: Integrate the output of the digital twin model with the real-time data stream, generate a fault probability report and trigger a graded alarm notification, and iteratively update the prediction algorithm based on feedback from historical fault cases.

[0017] Optionally, step S1 is implemented in the following manner:

[0018] Step A1: Digital twin model construction. Based on the physical structure, functional logic, communication interface and historical operation data of IoT devices, construct a device-level digital twin. Use modeling languages ​​or modeling tools to formally describe the device behavior and synchronize the status with the physical device in real time through the interface.

[0019] Step A2: The monitoring module monitors in real time. The monitoring module is deployed on the IoT device to continuously collect key operating parameters, perform online analysis on the collected data stream, and identify abnormal signals that deviate from the normal benchmark threshold.

[0020] Step A3: Data fusion and model synchronization. The detection module uploads the original sensor data and anomaly marking results to the storage, compares and fuses the real-time data stream of the physical device with the simulation output of the corresponding digital twin model, and updates and synchronizes the fused state information.

[0021] Step A4: Abnormal signal triggering. When the monitoring module detects a significant abnormality within multiple consecutive time windows, it generates an abnormal event and pushes it to the fault prediction unit.

[0022] Optionally, step S2 is implemented in the following manner:

[0023] Step B1: Data acquisition and preprocessing. Real-time acquisition of communication logs and interaction event data from the IoT device group, and unified mapping into structured event tuples;

[0024] Step B2: Construct a dynamic communication topology graph, using IoT devices as nodes and communication behaviors between devices as directed edges, to build a dynamically weighted directed graph. Adjust the connection relationships in the adjacency matrix or graph database through new communication events, and increase the timeliness of the topology through a sliding time window mechanism.

[0025] Step B3: Construct a behavioral dependency graph. Based on the communication topology, mine semantic-level collaborative dependencies, construct a behavioral dependency matrix, identify collaborative patterns, and display the real-time collaborative structure of the device group.

[0026] Optionally, step S3 is implemented in the following manner:

[0027] Step C1: Based on the digital twin model of the IoT device group, build a high-fidelity simulation platform;

[0028] Step C2: Map each IoT device as an intelligent agent with local observation capabilities and independent decision-making capabilities, and construct a multi-objective joint reward function to guide the intelligent agent to learn efficient cooperation strategies, which can be dynamically weighted and adjusted;

[0029] Step C3: Train the multi-agent reinforcement learning model. Use the multi-agent reinforcement learning algorithm to perform multi-round interactive training in a simulation environment, and use the experience replay pool to update the policy network parameters.

[0030] Step C4: Quantitatively evaluate collaborative performance indicators. The multi-agent reinforcement learning model runs agent policies on the test machine, and statistically analyzes key performance indicators such as task completion rate and resource utilization rate. The evaluation results are then output as a structured report.

[0031] Optionally, step S4 is implemented in the following manner:

[0032] Step D1: Construct a dynamic heterogeneous graph with each device as a node and the communication behavior dependencies between devices as edges;

[0033] Step D2: Normalize and encode the original node attributes, mark the abnormal signals output by the real-time monitoring module as additional input features, and apply graph neural networks for feature extraction;

[0034] Step D3: Simulate the fault path. The initial fault point is specified by the implementation of the anomaly monitoring module or the historical fault database. Based on the edge weight and node vulnerability, the probability of the fault propagating to the neighbor is calculated. The fault state is iteratively propagated on the dynamic heterogeneous graph to generate a set of multiple propagation paths.

[0035] Step D4: Assess the probability of system-level cascading failures. For each simulated path, calculate its probability of occurrence, and statistically analyze the impact range of single-point failures, the probability of system collapse, and key indicators of the risk heatmap. Finally, output the weighted cumulative probability of all high-risk paths.

[0036] Optionally, the node attributes in step D1 include: real-time operating status, historical fault records, and equipment type and criticality level;

[0037] Edge attributes include communication protocol type, average latency and packet loss rate, and behavioral dependency strength.

[0038] Optionally, step S5 is implemented in the following manner:

[0039] Step E1: Receive multi-source evaluation input. The system receives two types of input data in real time: collaborative performance evaluation results and fault risk prediction results.

[0040] Step E2: Construct a resource optimization decision model and define the optimization objective function based on the computing power limit, storage capacity limit, total network bandwidth and fault isolation strategy for each machine;

[0041] Step E3: Generate a resource reallocation strategy. Use a hybrid optimization algorithm of heuristic rule engine, integer linear programming and reinforcement learning fine-tuning to solve the resource allocation scheme and output the output instructions for computing resources, storage resources and network loans.

[0042] Optionally, step S6 is implemented in the following manner:

[0043] Step F1: Alignment of multi-source heterogeneous data. Real-time data streams from physical devices and state prediction values ​​output by digital twin models are received in real time. The timestamps of the real-time data streams are aligned with the simulation step size of the digital twin models, and the data is mapped to a unified entity through spatial mapping.

[0044] Step F2: Data fusion, which combines real-time data from physical devices, virtual state parameters generated by the digital twin model, and typical degradation mode vectors from the historical fault feature library to output a high-order joint feature vector;

[0045] Step F3: Fault probability quantification. Input the fused feature vector into the prediction model to calculate the individual equipment failure probability, the group-level failure propagation risk, and the system-level downtime critical index. Perform fault mode recognition, generate a fault probability heatmap with timestamps and a structured fault probability report, and trigger hierarchical alarm notifications.

[0046] Step F4: Iteratively update the prediction algorithm. The model optimization module periodically performs closed-loop verification of prediction performance, uses the gradient descent algorithm to adjust the weight coefficients of the prediction model, and synchronizes the updated model parameters to the digital twin model to optimize the feature selection strategy.

[0047] A digital twin-based IoT device group collaboration and fault prediction system, employing the aforementioned digital twin-based IoT device group collaboration and fault prediction method, includes a digital twin modeling and synchronization module, a communication and behavior relationship modeling module, a multi-agent system performance evaluation module, a graph neural network fault propagation simulation module, a dynamic resource optimization allocation module, and a fault prediction and alarm feedback module.

[0048] The digital twin modeling and synchronization module is used to construct device-level digital twins, formally describe the structure, function, interface and behavior of physical devices, collect device operating parameters and abnormal signals in real time, fuse and synchronize the physical device status with the digital twin model, and trigger abnormal events when continuous abnormalities are detected and push them to the fault prediction unit.

[0049] The communication and behavior relationship modeling module is used to collect and preprocess communication logs and interaction events between devices, construct a dynamic weighted directed communication topology graph to reflect the real-time connection relationship between devices, mine semantic-level collaborative dependencies, construct a behavioral dependency graph, and characterize group collaboration patterns.

[0050] The multi-agent system performance evaluation module is used to map devices as agents in a high-fidelity simulation environment, design a multi-objective joint reward function, guide agents to learn efficient cooperation strategies, use multi-agent reinforcement learning algorithms to train and optimize cooperation strategies, and output quantitative indicators of cooperation performance such as task completion rate and resource utilization.

[0051] The graph neural network fault propagation simulation module is used to construct a dynamic heterogeneous graph containing node attributes and edge attributes. It applies graph neural networks to extract device features and fuses abnormal signals to simulate the propagation path of a fault from its initial point in the topology, assess the risk of system-level cascading failures, and output a risk heatmap and collapse probability.

[0052] The dynamic resource optimization and allocation module receives two types of evaluation results: collaborative performance and fault risk. It establishes an optimization objective function that considers computation, storage, bandwidth limitations, and fault isolation strategies, and uses a hybrid optimization algorithm to generate resource reallocation instructions.

[0053] The fault prediction and alarm feedback module is used to align and fuse real-time data from physical devices, digital twin outputs, and historical fault modes to generate a high-dimensional joint feature vector. This vector is then input into the prediction model to quantify the three-level fault probabilities of individual devices, groups, and the system. The module generates timestamped fault reports and heatmaps, triggers tiered alarms, and iteratively updates the prediction algorithm parameters through closed-loop verification.

[0054] The beneficial effects of this invention are:

[0055] 1. In this invention, the data barriers between devices are broken down through the multi-protocol adaptation and intelligent preprocessing mechanism of the data acquisition unit, realizing unified access and standardized processing of heterogeneous sensor data, providing a highly consistent real-time data source for the digital twin model. The digital twin model construction unit uses the real-time synchronization module and the fault feature extraction module to continuously calibrate the mapping relationship between the virtual model and the physical entity, ensuring that the twin accurately reflects the real-time status and potential fault modes of the equipment group, laying a high-fidelity foundation for collaborative analysis and prediction.

[0056] 2. In this invention, the group collaboration analysis unit deeply analyzes the efficiency bottlenecks and fault propagation chains in the equipment collaboration network through interactive behavior modeling and cascading fault prediction. Combined with the dynamic resource scheduling interface, it realizes preventive resource allocation, which greatly improves the efficiency of group collaboration and reduces the risk of systemic collapse. The fault prediction unit integrates real-time monitoring data and twin model simulation results, uses the data fusion module to improve prediction accuracy, and triggers intervention in advance through the early warning generation interface. The active intervention unit then executes the policy library instructions and dynamically reconfigures resources, changing passive repair to active defense and minimizing downtime losses.

[0057] 3. In this invention, the resource optimization and configuration unit is based on prediction-driven pre-allocation and real-time performance evaluation to realize on-demand elastic scaling and load balancing of computing, storage and network resources, effectively cope with business peaks and reduce redundancy costs. The heterogeneous device adaptation unit relies on protocol conversion and compatibility evaluation to seamlessly integrate devices with different capabilities and protocols. The resource scheduling and coordination module works with the optimization unit to balance performance differences, ensuring efficient collaboration of heterogeneous device groups and avoiding compatibility failures. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0059] Figure 2 This is a schematic diagram of step S1 of the present invention.

[0060] Figure 3 This is a schematic diagram of step S2 of the present invention.

[0061] Figure 4 This is a schematic diagram of step S3 of the present invention.

[0062] Figure 5 This is a schematic diagram of step S4 of the present invention.

[0063] Figure 6 This is a schematic diagram of step S5 of the present invention.

[0064] Figure 7 This is a schematic diagram of step S6 of the present invention.

[0065] Figure 8 This is a schematic diagram of a system structure according to the present invention. Detailed Implementation

[0066] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] like Figures 1 to 7 As shown, a method for IoT device group collaboration and fault prediction based on digital twins includes the following:

[0068] Step S1: Construct a digital twin model of the IoT device group, and continuously collect the operating parameters and abnormal signals of each device through a real-time monitoring module;

[0069] Step S2: Based on the communication logs and interaction event data between devices, construct a device communication topology map and a behavior dependency graph to characterize the collaboration pattern of the device group;

[0070] Step S3: Train the agent model in the simulation environment using a multi-agent reinforcement learning algorithm, quantify and evaluate the collaborative efficiency of the equipment group, and obtain the task completion rate and resource utilization rate indicators.

[0071] Step S4: Combine graph neural networks to extract features from device nodes, simulate the propagation path of faults in the communication topology, and assess the probability of system-level cascading failures.

[0072] Step S5: Based on the collaborative performance evaluation results and fault risk prediction results, dynamically optimize the allocation of computing resources, storage resources and network bandwidth of the device group;

[0073] Step S6: Integrate the output of the digital twin model with the real-time data stream, generate a fault probability report and trigger a graded alarm notification, and iteratively update the prediction algorithm based on feedback from historical fault cases.

[0074] like Figure 8 As shown, an IoT device group collaboration and fault prediction system based on digital twins adopts the IoT device group collaboration and fault prediction method based on digital twins, including a digital twin modeling and synchronization module, a communication and behavior relationship modeling module, a multi-agent system performance evaluation module, a graph neural network fault propagation simulation module, a dynamic resource optimization allocation module, and a fault prediction and alarm feedback module.

[0075] The digital twin modeling and synchronization module is used to construct device-level digital twins, formally describe the structure, function, interface and behavior of physical devices, collect device operating parameters and abnormal signals in real time, fuse and synchronize the physical device status with the digital twin model, and trigger abnormal events when continuous abnormalities are detected and push them to the fault prediction unit.

[0076] The communication and behavior relationship modeling module is used to collect and preprocess communication logs and interaction events between devices, construct a dynamic weighted directed communication topology graph to reflect the real-time connection relationship between devices, mine semantic-level collaborative dependencies, construct a behavioral dependency graph, and characterize group collaboration patterns.

[0077] The multi-agent system performance evaluation module is used to map devices as agents in a high-fidelity simulation environment, design a multi-objective joint reward function, guide agents to learn efficient cooperation strategies, use multi-agent reinforcement learning algorithms to train and optimize cooperation strategies, and output quantitative indicators of cooperation performance such as task completion rate and resource utilization.

[0078] The graph neural network fault propagation simulation module is used to construct a dynamic heterogeneous graph containing node attributes and edge attributes. It applies graph neural networks to extract device features and fuses abnormal signals to simulate the propagation path of a fault from its initial point in the topology, assess the risk of system-level cascading failures, and output a risk heatmap and collapse probability.

[0079] The dynamic resource optimization and allocation module receives two types of evaluation results: collaborative performance and fault risk. It establishes an optimization objective function that considers computation, storage, bandwidth limitations, and fault isolation strategies, and uses a hybrid optimization algorithm to generate resource reallocation instructions.

[0080] The fault prediction and alarm feedback module is used to align and fuse real-time data from physical devices, digital twin outputs, and historical fault modes to generate a high-dimensional joint feature vector. This vector is then input into the prediction model to quantify the three-level fault probabilities of individual devices, groups, and the system. The module generates timestamped fault reports and heatmaps, triggers tiered alarms, and iteratively updates the prediction algorithm parameters through closed-loop verification.

[0081] Specifically, it includes the following:

[0082] The data acquisition unit is used to collect sensor data and operating status information from IoT devices in real time.

[0083] The digital twin model building unit is used to create and maintain a dynamic digital twin model corresponding to the physical device group in virtual space based on the real-time operation status logs fed back from the sensor interface unit of the data acquisition module and the device controller.

[0084] The group collaboration analysis unit is used to analyze the interaction behavior, synergistic effects, and cascading failure risks between devices.

[0085] The fault prediction unit combines a digital twin model and real-time data stream to perform dynamic fault prediction.

[0086] The proactive intervention unit is used to automatically implement intervention strategies and resource reconfiguration when a fault warning is issued;

[0087] The resource optimization and allocation unit is used to optimize the allocation of system resources based on the prediction results.

[0088] Heterogeneous device adapter unit is used to handle the heterogeneity of devices with different communication protocols and computing capabilities.

[0089] The data acquisition unit includes:

[0090] The sensor interface module is used to connect various IoT sensors and read raw data; the sensor interface module supports multiple communication protocols (such as MQTT, CoAP and HTTP) to ensure seamless connection with various heterogeneous sensors.

[0091] The data preprocessing component is responsible for cleaning, filtering, and standardizing the collected data. It uses adaptive filtering algorithms to eliminate noise interference, employs a sliding window cleaning mechanism to remove outliers, and performs rule-based standardization operations to unify data formats and units. This ensures high consistency and reliability of the raw data before it is transmitted to the digital twin model building unit, effectively improving the accuracy of subsequent dynamic modeling while reducing the system's processing load. This provides solid data support for collaborative analysis of equipment groups and fault prediction.

[0092] The data acquisition unit enables real-time, efficient acquisition and preliminary processing of sensor data and operational status information from IoT devices.

[0093] The building blocks of a digital twin model include:

[0094] The model initialization module is used to create an initial digital twin based on historical data and real-time input; a high-fidelity initial model can be quickly built through the model initialization module.

[0095] The real-time synchronization module is used to continuously monitor the real-time data stream acquired from the data acquisition unit and apply machine learning algorithms (such as long short-term memory networks or recurrent neural networks) to dynamically update the parameters of the digital twin model, ensuring that the model accurately maps the changes in the operating status of the physical device group in the virtual space.

[0096] The real-time synchronization module operates as follows:

[0097] H1. Receive and parse the real-time data stream from the data acquisition unit to ensure that the data format is consistent with the model input requirements;

[0098] H2. Apply Long Short-Term Memory (LSTM) or Recurrent Neural Network (RNN) algorithms to perform sequence analysis and feature extraction on the data stream to identify equipment status change patterns.

[0099] H3. Dynamically adjust the key parameters of the digital twin model (such as equipment operating efficiency, energy consumption threshold, etc.) based on the extracted features to achieve incremental updates of model parameters;

[0100] H4. Perform model validation and error correction by comparing the feedback status of physical devices with the model predictions, calculating the error rate, and applying the backpropagation mechanism to fine-tune the model.

[0101] H5 generates model update reports, records parameter change trajectories and synchronization status, and provides real-time modeling basis for the fault prediction unit;

[0102] H6. Continuously repeat the above steps to ensure that the digital twin model reflects the operating status of the physical device group in real time and with high accuracy in the virtual space.

[0103] The fault feature extraction module is used to receive preprocessed data streams from the real-time synchronization module and efficiently identify abnormal patterns and potential fault features in the equipment's operating status through multimodal feature fusion technology (such as deep learning-based convolutional neural networks or time-frequency analysis algorithms).

[0104] The specific working steps of the fault feature extraction module are as follows:

[0105] J1. Obtain structured equipment status sequence data from the real-time synchronization module, combine it with the historical fault database, and apply correlation analysis to filter key feature dimensions (such as vibration spectrum, temperature gradient and energy consumption fluctuation).

[0106] J2. Adaptive wavelet transform or principal component analysis (PCA) algorithms are used to reduce the dimensionality and enhance the features of the multidimensional data stream, and extract high-frequency anomalous signals and steady-state offset patterns.

[0107] J3. Classify and cluster the extracted features using an ensemble learning framework (such as random forest or gradient boosting decision tree) to identify common fault types (such as bearing wear, circuit overload or communication delay) and quantify the fault probability.

[0108] J4. The identification results are converted into standardized fault feature vectors, and compared with the feedback data from physical devices through a cross-validation mechanism to ensure the accuracy and robustness of feature extraction.

[0109] J5. The feature vector is transmitted to the fault prediction unit in real time, and fault correlation data is provided to the group collaborative analysis unit to support cascading risk analysis.

[0110] J6. Based on prediction error feedback, the feature extraction parameters and algorithm weights are dynamically adjusted to ensure that the module maintains high sensitivity and low false alarm rate throughout the entire equipment operation cycle.

[0111] The digital twin model building unit enables collaborative status analysis, fault prediction, and optimization decision support for IoT device groups.

[0112] The group collaboration analysis unit includes:

[0113] The interaction behavior modeling module is used to construct a communication topology and behavior dependency graph between devices. By collecting communication logs and interaction event data between devices, the interaction behavior modeling module constructs a dynamic communication topology graph and behavior dependency matrix, thereby accurately depicting the collaboration mode of a group of devices.

[0114] The synergy evaluation module quantifies the collaborative efficiency of devices through a multi-agent reinforcement learning algorithm. The synergy evaluation module uses a multi-agent reinforcement learning algorithm to train agent models in a simulation environment and quantifies and evaluates the efficiency indicators of device collaboration, such as task completion rate and resource utilization rate.

[0115] The cascading failure prediction module combines graph neural networks to analyze fault propagation paths and system-level risks. It also uses graph neural networks to extract features from equipment nodes, simulates the propagation path of faults in the topology network, and assesses the probability of system-level risks, providing a basis for preventive maintenance.

[0116] The dynamic resource scheduling interface module optimizes the allocation of computing and communication resources for the device group in real time based on the coordination strategy. The dynamic resource scheduling interface module also dynamically allocates computing, bandwidth and storage resources through optimization algorithms based on the coordination strategy and real-time load data to ensure that the device group maintains efficient operation in high-concurrency scenarios.

[0117] With the help of the group collaboration analysis unit, efficient analysis of inter-device collaborative behavior, accurate prediction of fault propagation paths, and resource optimization allocation based on real-time data can be achieved, thereby improving the operating efficiency and reliability of the entire IoT device group and supporting system-level collaborative decision-making and risk prevention and control.

[0118] The fault prediction unit includes:

[0119] The real-time monitoring module is used to continuously track the operating parameters and abnormal signals of IoT devices.

[0120] The data fusion module is responsible for integrating the output of the digital twin model with real-time data streams to improve prediction accuracy. By capturing abnormal signals of operating parameters through the real-time monitoring module and combining the integration of the digital twin model output with real-time data by the data fusion module, the system can identify potential fault points in advance.

[0121] The early warning generation interface automatically generates fault probability reports and tiered alarm notifications; based on the fault probability reports, the early warning generation interface generates tiered alarms to promptly notify maintenance personnel to take intervention measures.

[0122] The model optimization module iteratively updates the prediction algorithm based on feedback from historical failure cases; the model optimization module uses feedback from historical failure cases to continuously iterate the algorithm, thereby improving the adaptability and accuracy of the prediction model.

[0123] The fault prediction unit combines abnormal operating parameter signals captured by the real-time monitoring module, the output of the digital twin model integrated by the data fusion module, and real-time data streams to generate a high-precision fault probability report. It also triggers tiered alarm notifications through the early warning generation interface and uses the model optimization module to iteratively update the prediction algorithm to improve adaptability. Under the action of the fault prediction unit, early and accurate identification and dynamic prediction of potential faults in IoT devices can be achieved, thereby effectively reducing the risk of system downtime, optimizing maintenance strategies, and providing decision-making basis for the proactive intervention unit.

[0124] The specific working steps of the fault prediction unit are as follows:

[0125] K1, the real-time monitoring module collects equipment operating parameters (such as temperature, vibration spectrum, current waveform) through a distributed sensor network, and uses a sliding window anomaly detection algorithm to identify abnormal signals that deviate from the reference threshold;

[0126] After receiving the abnormal signal, the K2 data fusion module performs multi-source heterogeneous data alignment: First, it performs spatiotemporal calibration between the real-time data stream and the predicted state output by the digital twin model. Then, it uses a feature fusion framework based on attention mechanism (such as the Transformer encoder) to integrate physical sensor data, virtual model parameters and historical fault feature library to generate a high-dimensional joint feature vector.

[0127] K3. Apply ensemble learning algorithms (such as the XGBoost and deep belief network combined model) to perform fault mode identification on the fused feature vectors, and quantify and calculate three core indicators: 1) individual equipment failure probability (based on component degradation trajectory analysis), 2) group-level fault propagation risk (relying on the dependency graph output by the group collaborative analysis unit), and 3) system-level downtime critical index (combined with the redundancy resource threshold provided by the resource optimization and allocation unit), and generate a fault probability heatmap with timestamps and a risk classification report.

[0128] K4. The early warning generation interface triggers a dynamic response mechanism based on the risk classification report: 1) For yellow warnings (probability 30%-60%), generate equipment self-test instructions and mark the nodes to be observed; 2) For orange warnings (probability 60%-85%), start the pre-loading of the active intervention unit's contingency plan; 3) For red warnings (probability >85%), immediately push the alarm to the operation and maintenance terminal and release the emergency resource pool.

[0129] K5. The model optimization module periodically performs closed-loop verification of prediction performance: 1) Collect the actual fault occurrence time and type and compare them with the prediction results; 2) Use the confusion matrix to evaluate the recall rate / false alarm rate index; synthesize boundary fault samples through generative adversarial network (GAN) and inject them into the training set; 3) Apply the gradient descent algorithm to adjust the weight coefficients of the prediction model.

[0130] K6. The verification results are fed back to the digital twin model building unit, driving the fault feature extraction module to update the feature selection strategy, forming a self-evolutionary closed loop from data collection to model iteration.

[0131] The active intervention unit includes:

[0132] The strategy execution module is used to parse fault warning signals and trigger a preset intervention strategy library. The strategy execution module automatically calls the corresponding measures in the preset intervention strategy library by parsing the warning signals, such as isolating faulty nodes or activating redundancy mechanisms.

[0133] The resource dynamic reallocation module adjusts the computing, storage, and network resources of the device group in real time based on the fault prediction results; the resource dynamic reallocation module adjusts the allocation of computing, storage, and network resources in real time based on the fault prediction results to ensure that critical tasks are executed first and avoid resource bottlenecks.

[0134] With the active intervention unit, the automatic execution of fault intervention strategies, real-time dynamic optimization of equipment group resources, and rapid recovery and stable maintenance of system operating status can be achieved.

[0135] The resource optimization and allocation unit includes:

[0136] The resource status monitoring module is used to collect real-time data on the computing load, storage utilization, and network bandwidth of the device group. By continuously collecting computing load, storage utilization, and network bandwidth data, the resource status monitoring module provides a real-time benchmark for optimization decisions.

[0137] The prediction-driven pre-allocation module dynamically predicts peak resource demand based on the output of the fault prediction module. It also dynamically predicts peak resource demand based on the fault prediction results, and deploys redundant resources or adjusts allocation strategies in advance to avoid bottlenecks.

[0138] The performance evaluation module generates a resource utilization matrix and an optimized performance analysis report in real time. By generating the resource utilization matrix and the optimized performance analysis report, the performance evaluation module quantifies the configuration effect and drives iterative optimization of strategies, thereby improving overall resource utilization efficiency, reducing operating costs, and enhancing the stability and reliability of system collaborative operation.

[0139] The resource optimization and allocation unit enables dynamic load balancing of computing resources for device groups, elastic capacity adjustment of storage resources, and intelligent bandwidth allocation of network resources.

[0140] Heterogeneous device adapter units include:

[0141] The protocol conversion module is used to automatically identify and convert the communication protocols (such as MQTT, CoAP, HTTP, etc.) of different IoT devices to achieve seamless interconnection; the protocol conversion module automatically handles protocol differences to ensure smooth communication across devices.

[0142] The compatibility assessment module uses historical data analysis and machine learning models to predict compatibility issues and potential conflicts between devices; the compatibility assessment module predicts and mitigates potential conflicts to avoid system interruptions.

[0143] The resource scheduling and coordination module works with the resource optimization and configuration module to dynamically allocate computing and storage resources to balance the performance differences between devices; the resource scheduling and coordination module also works together to optimize resource allocation and balance the needs of high-performance and low-power devices.

[0144] The performance monitoring module collects and analyzes equipment operating indicators in real time and feeds them back to the operation and maintenance monitoring panel for visual diagnosis.

[0145] The heterogeneous device adaptation unit enables seamless collaboration and compatibility assurance between heterogeneous IoT devices; the performance monitoring module provides real-time feedback on operating indicators, supporting rapid diagnosis and adjustment. This effectively reduces integration complexity caused by device heterogeneity, improves group collaboration efficiency, reduces failure rates caused by compatibility issues, and enhances the overall flexibility and scalability of the system.

[0146] It also includes: a visual interactive interface that provides a 3D dynamic twin monitoring view, a collaborative efficiency heatmap, and a fault early warning dashboard, supporting maintenance personnel to remotely adjust collaborative strategies and intervene in maintenance decisions for equipment groups. By setting up the visual interactive interface, comprehensive visual monitoring and interactive control of the equipment group's operating status can be achieved.

[0147] The system configuration management unit is used to define the topology of device groups, configure sensor thresholds, update algorithm model parameters, and manage user operation permissions. By setting up the system configuration management unit, flexible construction of device group architecture, precise definition of monitoring rules, dynamic optimization of prediction models, and comprehensive assurance of system operational security can be achieved.

[0148] With the coordinated efforts of the aforementioned units, the core problems in the management of IoT device groups mentioned in the background technology, such as data silos, delayed fault response, rigid resource scheduling, and poor compatibility of heterogeneous systems, can be effectively solved.

[0149] Specifically, this manifests as follows:

[0150] Data fusion and consistency assurance: By using the multi-protocol adaptation and intelligent preprocessing mechanism of the data acquisition unit, data barriers between devices are broken down, enabling unified access and standardized processing of heterogeneous sensor data, and providing a highly consistent real-time data source for the digital twin model.

[0151] Dynamic and accurate modeling and status synchronization: The digital twin model building unit uses a real-time synchronization module and a fault feature extraction module to continuously calibrate the mapping relationship between the virtual model and the physical entity, ensuring that the twin accurately reflects the real-time status and potential fault modes of the equipment group, laying a high-fidelity foundation for collaborative analysis and prediction.

[0152] Dynamic and accurate modeling and state synchronization include the following steps:

[0153] L1, the real-time synchronization module collects real-time operating data from the physical device group through a multi-protocol adapter, and uses a timestamp alignment mechanism to transmit the data to the virtual model in real time, ensuring that the data flow is consistent with the state of the physical entity;

[0154] L2, the fault feature extraction module analyzes the synchronous data stream based on machine learning algorithms (such as convolutional neural networks or anomaly detection models), identifies features such as abnormal vibration, temperature fluctuations or power consumption deviations during equipment operation, and generates potential fault mode reports.

[0155] L3, the model calibration engine dynamically optimizes the parameters of the digital twin model based on the extracted fault features, including adjusting the physical attribute mapping coefficients and fault prediction weights, to achieve high-precision adaptive updates of the virtual model.

[0156] L4. The status verification unit periodically performs bidirectional comparison tests between the twin and the physical device, evaluates the accuracy of the model through residual analysis, and uses the feedback results to iteratively optimize the calibration process, ultimately ensuring the reliability and real-time performance of collaborative analysis and fault prediction.

[0157] Collaborative optimization and risk prevention: The group collaborative analysis unit deeply analyzes the efficiency bottlenecks and fault propagation chains in the equipment collaboration network through interactive behavior modeling and cascading failure prediction. Combined with the dynamic resource scheduling interface, it realizes preventive resource allocation, which greatly improves the efficiency of group collaboration and reduces the risk of system collapse.

[0158] Predictive maintenance and proactive fault tolerance: The fault prediction unit integrates real-time monitoring data and twin model simulation results, uses the data fusion module to improve prediction accuracy, and triggers intervention in advance through the early warning generation interface; the proactive intervention unit then executes policy library instructions (such as node isolation and redundancy switching) and dynamically reconfigures resources, transforming passive repair into proactive defense, and minimizing downtime losses.

[0159] Resource elasticity and performance enhancement: The resource optimization and allocation unit, based on prediction-driven pre-allocation and real-time performance evaluation, enables on-demand elastic scaling and load balancing of computing, storage and network resources, effectively handling business peaks and reducing redundancy costs.

[0160] Heterogeneous integration and compatibility assurance: The heterogeneous device adaptation unit relies on protocol conversion and compatibility assessment to seamlessly integrate devices with different capabilities and protocols. The resource scheduling and coordination module works together to optimize the balance of performance differences between units, ensuring efficient collaboration of heterogeneous device groups and avoiding compatibility failures.

[0161] In addition, the visual interactive interface provides maintenance personnel with a panoramic monitoring and decision-making intervention entry point, and the system configuration management unit supports flexible policy adjustment and access control, together forming a closed-loop management capability. Ultimately, it realizes end-to-end intelligent closed-loop management of IoT device groups from data perception, status simulation, collaborative optimization, fault prediction to proactive intervention, comprehensively improving the system's reliability, efficiency and scalability.

[0162] The working principle and usage process of this invention: When the device is in use, the data acquisition unit collects sensor data and operating status information of IoT devices in real time. The sensor interface module automatically adapts to sensors with different protocols and reads the raw data. Subsequently, the data preprocessing component cleans, filters and standardizes the raw data to generate a highly consistent and reliable structured data stream.

[0163] The data stream is transmitted to the digital twin model building unit. In the digital twin model building unit, the model initialization module establishes an initial digital twin based on the initial configuration of the equipment and historical data. The real-time synchronization module continuously receives real-time data streams from the data acquisition unit and uses machine learning algorithms to dynamically adjust and update the parameters and status of the twin model to ensure that the virtual model and the physical equipment group maintain a high degree of synchronization. The fault feature extraction module analyzes the real-time data and model status to identify potential abnormal patterns and fault characteristics. The updated high-fidelity twin model is efficiently stored and managed by the model storage and management module, providing a foundation for subsequent analysis.

[0164] The group collaboration analysis unit receives and processes device status and interaction information from the digital twin model building unit. The interaction behavior modeling module builds and updates the communication topology and behavioral dependency graph between devices in real time. The collaboration effect evaluation module uses this graph and multi-agent reinforcement learning algorithm to evaluate the collaborative efficiency of devices (such as task completion rate and resource utilization) in a simulation environment. The cascading failure prediction module combines graph neural network analysis of device node status and topology relationships, simulates possible failure propagation paths, and evaluates the probability of system-level risks. The dynamic resource scheduling interface module generates preliminary resource optimization suggestions based on the collaboration strategy and risk assessment results.

[0165] The fault prediction unit integrates abnormal operating parameter signals captured by the real-time monitoring module, simulation status output by the digital twin model, and risk assessment results from the group collaborative analysis unit. The data fusion module deeply integrates these heterogeneous information sources to improve prediction accuracy; the early warning generation interface automatically generates fault probability reports and graded (such as early warnings and alarms) notifications at the device or system level based on the fusion analysis results; and the model optimization module continuously iterates and updates the prediction algorithm using feedback from historical fault cases.

[0166] When a warning signal (especially a high-level alarm) is received from the fault prediction unit, the proactive intervention unit is immediately activated. The strategy execution module parses the warning information, automatically matches and triggers the corresponding measures in the preset intervention strategy library (e.g., isolating the faulty node, starting backup equipment, adjusting the task flow). The resource dynamic reallocation module adjusts the allocation of computing, storage and network resources of the relevant equipment group in real time according to the predicted fault impact range and strategy requirements, so as to maintain the continuity of critical business and support the execution of intervention strategies.

[0167] During this process, the resource optimization and allocation unit continues to play a role, the resource status monitoring module provides real-time computing load, storage utilization and network bandwidth data of the device group; the prediction-driven pre-allocation module, combined with the output of the fault prediction unit, dynamically predicts future peak resource demand or bottleneck areas, and makes elastic deployment or adjustment of resources in advance; the performance evaluation module analyzes the resource scheduling effect in real time, generates optimization performance reports, drives the iterative optimization of configuration strategies, so as to achieve load balancing and cost-effectiveness maximization.

[0168] For heterogeneous devices connected to the system, the heterogeneous device adaptation unit ensures seamless integration and collaboration; the protocol conversion module automatically handles the conversion of communication protocols between different devices; the compatibility assessment module predicts potential compatibility issues between devices; the resource scheduling and coordination module works with the resource optimization and configuration unit to dynamically adjust resource allocation strategies based on differences in device performance (such as computing power and power consumption); and the performance monitoring module continuously tracks device operating indicators.

[0169] Furthermore, maintenance personnel can use the 3D dynamic twin monitoring view, collaborative efficiency heatmap, and fault early warning dashboard provided by the visual interactive interface to monitor the overall operating status, collaborative efficiency, and risk status of the device group in real time. They can also fine-tune remote collaborative strategies based on the visualized information or manually intervene in automatically generated maintenance decisions. The system configuration management unit supports flexible configuration and management of device topology, sensor thresholds, algorithm parameters, and user permissions. Ultimately, this achieves closed-loop intelligent management from data perception, status synchronization, collaborative analysis, accurate prediction, proactive intervention to resource optimization, greatly improving the operating efficiency, reliability, and maintainability of the IoT device group.

[0170] The sensor interface module connects to various IoT sensors and reads raw data. It supports multiple communication protocols (such as MQTT, CoAP, and HTTP) to ensure seamless connectivity with heterogeneous sensors. The data preprocessing component cleans, filters, and standardizes the collected data. It uses adaptive filtering algorithms to eliminate noise interference, employs a sliding window cleaning mechanism to remove outliers, and performs rule-based standardization to unify data format and units. This ensures high consistency and reliability of the raw data before it is transmitted to the digital twin model building unit, effectively improving the accuracy of subsequent dynamic modeling while reducing system processing load. This provides solid data support for collaborative analysis of device groups and fault prediction.

[0171] The data acquisition unit enables real-time, efficient acquisition and preliminary processing of sensor data and operational status information from IoT devices.

[0172] The model initialization module creates an initial digital twin based on historical data and real-time input; it quickly builds a high-fidelity initial model. The real-time synchronization module dynamically updates the model using machine learning algorithms to match the physical device status; it also dynamically adjusts model parameters to accurately reflect the real-time status of the physical devices. The fault feature extraction module identifies and analyzes potential fault modes; it efficiently identifies and analyzes potential fault characteristics. The model storage and management module efficiently stores, retrieves, and optimizes model data; it optimizes data access and model iteration, thereby improving the collaborative efficiency of the device group, enhancing fault prediction accuracy, and providing reliable decision-making support for the system.

[0173] The digital twin model building unit enables collaborative status analysis, fault prediction, and optimization decision support for IoT device groups.

[0174] The interaction behavior modeling module is used to construct a communication topology and behavior dependency graph between devices. By collecting communication logs and interaction event data between devices, the interaction behavior modeling module constructs a dynamic communication topology graph and behavior dependency matrix, thereby accurately depicting the collaboration mode of a group of devices.

[0175] The synergy evaluation module quantifies the collaborative efficiency of devices through a multi-agent reinforcement learning algorithm. The synergy evaluation module uses a multi-agent reinforcement learning algorithm to train agent models in a simulation environment and quantifies and evaluates the efficiency indicators of device collaboration, such as task completion rate and resource utilization rate.

[0176] The cascading failure prediction module combines graph neural networks to analyze fault propagation paths and system-level risks. It also uses graph neural networks to extract features from equipment nodes, simulates the propagation path of faults in the topology network, and assesses the probability of system-level risks, providing a basis for preventive maintenance.

[0177] The dynamic resource scheduling interface module optimizes the allocation of computing and communication resources for the device group in real time based on the coordination strategy. The dynamic resource scheduling interface module also dynamically allocates computing, bandwidth and storage resources through optimization algorithms based on the coordination strategy and real-time load data to ensure that the device group maintains efficient operation in high-concurrency scenarios.

[0178] With the help of the group collaboration analysis unit, efficient analysis of inter-device collaborative behavior, accurate prediction of fault propagation paths, and resource optimization allocation based on real-time data can be achieved, thereby improving the operating efficiency and reliability of the entire IoT device group and supporting system-level collaborative decision-making and risk prevention and control.

[0179] The real-time monitoring module is used to continuously track the operating parameters and abnormal signals of IoT devices.

[0180] The data fusion module is responsible for integrating the output of the digital twin model with real-time data streams to improve prediction accuracy. By capturing abnormal signals of operating parameters through the real-time monitoring module and combining the integration of the digital twin model output with real-time data by the data fusion module, the system can identify potential fault points in advance.

[0181] The early warning generation interface automatically generates fault probability reports and tiered alarm notifications; based on the fault probability reports, the early warning generation interface generates tiered alarms to promptly notify maintenance personnel to take intervention measures.

[0182] The model optimization module iteratively updates the prediction algorithm based on feedback from historical failure cases; the model optimization module uses feedback from historical failure cases to continuously iterate the algorithm, thereby improving the adaptability and accuracy of the prediction model.

[0183] The fault prediction unit enables real-time anomaly detection of equipment operating status, dynamic simulation of fault propagation paths, and accurate quantitative assessment of system-level risks, thereby providing a reliable basis for preventive maintenance decisions.

[0184] The resource status monitoring module is used to collect real-time data on the computing load, storage utilization, and network bandwidth of the device group. By continuously collecting computing load, storage utilization, and network bandwidth data, the resource status monitoring module provides a real-time benchmark for optimization decisions.

[0185] The prediction-driven pre-allocation module dynamically predicts peak resource demand based on the output of the fault prediction module. It also dynamically predicts peak resource demand based on the fault prediction results, and deploys redundant resources or adjusts allocation strategies in advance to avoid bottlenecks.

[0186] The performance evaluation module generates a resource utilization matrix and an optimized performance analysis report in real time. By generating the resource utilization matrix and the optimized performance analysis report, the performance evaluation module quantifies the configuration effect and drives iterative optimization of strategies, thereby improving overall resource utilization efficiency, reducing operating costs, and enhancing the stability and reliability of system collaborative operation.

[0187] The resource optimization and allocation unit enables dynamic load balancing of computing resources for device groups, elastic capacity adjustment of storage resources, and intelligent bandwidth allocation of network resources.

[0188] The protocol conversion module is used to automatically identify and convert the communication protocols (such as MQTT, CoAP, HTTP, etc.) of different IoT devices to achieve seamless interconnection; the protocol conversion module automatically handles protocol differences to ensure smooth communication across devices.

[0189] The compatibility assessment module uses historical data analysis and machine learning models to predict compatibility issues and potential conflicts between devices; the compatibility assessment module predicts and mitigates potential conflicts to avoid system interruptions.

[0190] The resource scheduling and coordination module works with the resource optimization and configuration module to dynamically allocate computing and storage resources to balance the performance differences between devices; the resource scheduling and coordination module also works together to optimize resource allocation and balance the needs of high-performance and low-power devices.

[0191] The performance monitoring module collects and analyzes equipment operating indicators in real time and feeds them back to the operation and maintenance monitoring panel for visual diagnosis.

[0192] The heterogeneous device adaptation unit enables seamless collaboration and compatibility assurance between heterogeneous IoT devices; the performance monitoring module provides real-time feedback on operating indicators, supporting rapid diagnosis and adjustment. This effectively reduces integration complexity caused by device heterogeneity, improves group collaboration efficiency, reduces failure rates due to compatibility issues, and enhances the overall flexibility and scalability of the system.

[0193] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for group collaboration and fault prediction of Internet of Things (IoT) devices based on digital twins, characterized in that, Includes the following steps: Step S1: Construct a digital twin model of the IoT device group, and continuously collect the operating parameters and abnormal signals of each device through a real-time monitoring module; Step S2: Based on the communication logs and interaction event data between devices, construct a device communication topology map and a behavior dependency map to characterize the collaboration pattern of the device group; Step S3: Train the agent model in the simulation environment using a multi-agent reinforcement learning algorithm, quantitatively evaluate the collaborative efficiency of the equipment group, and obtain the task completion rate and resource utilization rate indicators. Step S4: Combine graph neural networks to extract features from device nodes, simulate the propagation path of faults in the communication topology, and assess the probability of system-level cascading failures. Implemented in the following ways: Step D1: Construct a dynamic heterogeneous graph with each device as a node and the communication behavior dependencies between devices as edges; Step D2: Normalize and encode the original node attributes, mark the abnormal signals output by the real-time monitoring module as additional input features, and apply graph neural networks for feature extraction; Step D3: Simulate the fault path. The initial fault point is specified by the implementation of the anomaly monitoring module or the historical fault database. Based on the edge weight and node vulnerability, the probability of the fault propagating to the neighbor is calculated. The fault state is iteratively propagated on the dynamic heterogeneous graph to generate a set of multiple propagation paths. Step D4: Assess the probability of system-level cascading failures. For each simulated path, calculate its probability of occurrence, and statistically analyze the impact range of single-point failures, the probability of system collapse, and key indicators of the risk heatmap. Finally, output the weighted cumulative probability of all high-risk paths. Step S5: Based on the collaborative performance evaluation results and fault risk prediction results, dynamically optimize the allocation of computing resources, storage resources and network bandwidth of the device group; Step S6: Integrate the output of the digital twin model with the real-time data stream, generate a fault probability report and trigger a graded alarm notification, and iteratively update the prediction algorithm based on feedback from historical fault cases.

2. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 1, characterized in that, Step S1 is implemented in the following manner: Step A1: Digital twin model construction. Based on the physical structure, functional logic, communication interface and historical operation data of IoT devices, construct a device-level digital twin. Use modeling languages ​​or modeling tools to formally describe the device behavior and synchronize the status with the physical device in real time through the interface. Step A2: The monitoring module monitors in real time. The monitoring module is deployed on the IoT device to continuously collect key operating parameters, perform online analysis on the collected data stream, and identify abnormal signals that deviate from the normal benchmark threshold. Step A3: Data fusion and model synchronization. The detection module uploads the original sensor data and anomaly marking results to the storage, compares and fuses the real-time data stream of the physical device with the simulation output of the corresponding digital twin model, and updates and synchronizes the fused state information. Step A4: Abnormal signal triggering. When the monitoring module detects a significant abnormality within multiple consecutive time windows, it generates an abnormal event and pushes it to the fault prediction unit.

3. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 2, characterized in that, Step S2 is implemented in the following manner: Step B1: Data acquisition and preprocessing. Real-time acquisition of communication logs and interaction event data from the IoT device group, and unified mapping into structured event tuples; Step B2: Construct a dynamic communication topology graph, using IoT devices as nodes and communication behaviors between devices as directed edges, to build a dynamically weighted directed graph. Adjust the connection relationships in the adjacency matrix or graph database through new communication events, and increase the timeliness of the topology through a sliding time window mechanism. Step B3: Construct a behavioral dependency graph. Based on the communication topology, mine semantic-level collaborative dependencies, construct a behavioral dependency matrix, identify collaborative patterns, and display the real-time collaborative structure of the device group.

4. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 1, characterized in that, Step S3 is implemented in the following manner: Step C1: Based on the digital twin model of the IoT device group, build a high-fidelity simulation platform; Step C2: Map each IoT device as an intelligent agent with local observation capabilities and independent decision-making capabilities, and construct a multi-objective joint reward function to guide the intelligent agent to learn efficient cooperation strategies, which can be dynamically weighted and adjusted; Step C3: Train the multi-agent reinforcement learning model. Use the multi-agent reinforcement learning algorithm to perform multi-round interactive training in a simulation environment, and use the experience replay pool to update the policy network parameters. Step C4: Quantitatively evaluate collaborative performance indicators. The multi-agent reinforcement learning model runs agent policies on the test machine, and statistically analyzes key performance indicators such as task completion rate and resource utilization rate. The evaluation results are then output as a structured report.

5. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 1, characterized in that, The node attributes in step D1 include: real-time operating status, historical fault records, and equipment type and criticality level; Edge attributes include communication protocol type, average latency and packet loss rate, and behavioral dependency strength.

6. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 1, characterized in that, Step S5 is implemented in the following manner: Step E1: Receive multi-source evaluation input. The system receives two types of input data in real time: collaborative performance evaluation results and fault risk prediction results. Step E2: Construct a resource optimization decision model and define the optimization objective function based on the computing power limit, storage capacity limit, total network bandwidth and fault isolation strategy for each machine; Step E3: Generate a resource reallocation strategy. Use a hybrid optimization algorithm of heuristic rule engine, integer linear programming and reinforcement learning fine-tuning to solve the resource allocation scheme, and output the output instructions for computing resources, storage resources and network bandwidth.

7. The method for IoT device group collaboration and fault prediction based on digital twins according to claim 1, characterized in that, Step S6 is implemented in the following manner: Step F1: Alignment of multi-source heterogeneous data. Real-time data streams from physical devices and state prediction values ​​output by digital twin models are received in real time. The timestamps of the real-time data streams are aligned with the simulation step size of the digital twin models, and the data is mapped to a unified entity through spatial mapping. Step F2: Data fusion, which combines real-time data from physical devices, virtual state parameters generated by the digital twin model, and typical degradation mode vectors from the historical fault feature library to output a high-order joint feature vector; Step F3: Fault probability quantification. Input the fused feature vector into the prediction model to calculate the individual equipment failure probability, the group-level failure propagation risk, and the system-level downtime critical index. Perform fault mode recognition, generate a fault probability heatmap with timestamps and a structured fault probability report, and trigger hierarchical alarm notifications. Step F4: Iteratively update the prediction algorithm. The model optimization module periodically performs closed-loop verification of prediction performance, uses the gradient descent algorithm to adjust the weight coefficients of the prediction model, and synchronizes the updated model parameters to the digital twin model to optimize the feature selection strategy.

8. A system for collaborative operation and fault prediction of IoT devices based on digital twins, employing the method for collaborative operation and fault prediction of IoT devices based on digital twins as described in any one of claims 1-7, characterized in that, It includes modules for digital twin modeling and synchronization, communication and behavior relationship modeling, multi-agent system performance evaluation, graph neural network fault propagation simulation, dynamic resource optimization and allocation, and fault prediction and alarm feedback. The digital twin modeling and synchronization module is used to construct device-level digital twins, formally describe the structure, function, interface and behavior of physical devices, collect device operating parameters and abnormal signals in real time, fuse and synchronize the physical device status with the digital twin model, and trigger abnormal events when continuous abnormalities are detected and push them to the fault prediction unit. The communication and behavior relationship modeling module is used to collect and preprocess communication logs and interaction events between devices, construct a dynamic weighted directed communication topology graph to reflect the real-time connection relationship between devices, mine semantic-level collaborative dependencies, construct a behavioral dependency graph, and characterize group collaboration patterns. The multi-agent system performance evaluation module is used to map devices as agents in a high-fidelity simulation environment, design a multi-objective joint reward function, guide agents to learn efficient cooperation strategies, use multi-agent reinforcement learning algorithms to train and optimize cooperation strategies, and output quantitative indicators of cooperation performance such as task completion rate and resource utilization. The graph neural network fault propagation simulation module is used to construct a dynamic heterogeneous graph containing node attributes and edge attributes. It applies graph neural networks to extract device features and fuses abnormal signals to simulate the propagation path of a fault from its initial point in the topology, assess the risk of system-level cascading failures, and output a risk heatmap and collapse probability. The dynamic resource optimization and allocation module receives two types of evaluation results: collaborative performance and fault risk. It establishes an optimization objective function that considers computation, storage, bandwidth limitations, and fault isolation strategies, and uses a hybrid optimization algorithm to generate resource reallocation instructions. The fault prediction and alarm feedback module is used to align and fuse real-time data from physical devices, digital twin outputs, and historical fault modes to generate a high-dimensional joint feature vector. This vector is then input into the prediction model to quantify the three-level fault probabilities of individual devices, groups, and the system. The module generates timestamped fault reports and heatmaps, triggers tiered alarms, and iteratively updates the prediction algorithm parameters through closed-loop verification.

Images