A multi-agent collaborative decision-making system based on digital twinning and knowledge graph

Abstract

This application relates to the field of industrial intelligent control technology, and in particular to a multi-agent collaborative decision-making system based on digital twins and knowledge graphs. The system includes: a data acquisition module for acquiring multimodal heterogeneous data, preprocessing it, and then connecting it to a digital twin base to obtain twin state data; a knowledge graph reasoning module for performing semantic reasoning analysis on the twin state data to obtain anomaly detection results and semantic feature expressions; a perception agent for generating semantic vectors based on the semantic feature expressions; a decision-making agent for generating a global scheduling scheme based on the semantic vectors; a simulation and deduction module for dynamically simulating the global scheduling scheme to obtain simulation and deduction results; and a system that sends the system to the execution agent for execution when preset operating constraints are met, otherwise feeds back to the decision-making agent for iterative optimization until the conditions are met before execution. This application uses dynamic simulation and preset operating constraint verification of the global scheduling scheme to improve the accuracy, stability, and operational safety of decision-making.

Description

Technical Field

[0001] This application relates to the field of industrial intelligent control technology, and in particular to a multi-agent collaborative decision-making system based on digital twins and knowledge graphs. Background Technology

[0002] In many fields such as modern industrial production, energy dispatch, intelligent transportation, and complex equipment operation and maintenance, as the scale of systems continues to expand and the operating environment becomes increasingly complex, the demand for intelligent decision-making driven by multi-source heterogeneous data is increasing.

[0003] In existing technologies, decision-making methods for complex systems mainly include rule-based expert systems and data-driven single-model prediction methods. Specifically, rule-based expert systems typically construct rule bases manually to determine system states and formulate control strategies. However, these methods inherently rely on discrete rule matching mechanisms and lack the ability to model deep semantic relationships between multi-source data. This makes it difficult to achieve globally consistent decision-making reasoning when facing complex coupling relationships and dynamically changing scenarios, and also results in high rule maintenance costs and poor scalability. Meanwhile, data-driven machine learning or deep learning methods, while capable of uncovering potential patterns from historical data, often model single tasks or single data modalities, lacking a unified semantic space for fusing and expressing multimodal information. This makes it difficult for different modules to effectively share and collaborate, thus limiting the system's overall decision-making capabilities in cross-task and cross-level scenarios. For example, patent application publication number CN118798494A (classification number G06Q) provides a smart production control method and system for the garment industry based on data analysis; patent application publication number CN121212859A (classification number G06Q) provides a smart rural management platform and data processing method; and patent application publication number CN121724394A (classification number G06Q) provides an interactive deduction method and system for production command cases. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a multi-agent collaborative decision-making system based on digital twins and knowledge graphs, which solves the technical problems of poor coordination consistency among multiple agents and insufficient decision-making accuracy and reliability in dynamic environments in the prior art.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted in this application include:

[0008] This application provides a multi-agent collaborative decision-making system based on digital twins and knowledge graphs, comprising: a data acquisition module for acquiring multimodal heterogeneous data, preprocessing the multimodal heterogeneous data, and then connecting it to a digital twin base to obtain twin state data; a knowledge graph reasoning module for performing semantic reasoning analysis on the twin state data based on a pre-constructed domain knowledge graph to obtain anomaly detection results and corresponding semantic feature expressions; a perceptual agent for generating semantic vectors based on the semantic feature expressions; a decision-making agent for generating a global scheduling scheme based on the semantic vectors; and a simulation deduction module for virtually executing and dynamically simulating the global scheduling scheme in the digital twin base to obtain simulation deductions. The results include: the simulation results include information on changes in equipment operating status; if the simulation results meet preset operating constraints, the global scheduling scheme is sent to the execution agent; if the simulation results do not meet the preset operating constraints, the simulation results are fed back to the decision agent, which iteratively optimizes the global scheduling scheme based on the semantic vector and the simulation results, generates a new global scheduling scheme, and inputs it back into the simulation module for verification until the simulation results meet the preset operating constraints; the execution agent receives the global scheduling scheme sent by the simulation module and controls the equipment corresponding to the global scheduling scheme.

[0009] Optionally, in some embodiments of this application, the pre-constructed domain knowledge graph specifically includes: a device knowledge sub-graph, used to structurally express device types, device identifiers, operating parameters, state attributes, and topological relationships between devices; an operating state evolution sub-graph, used to model device state nodes, state transition relationships, state triggering conditions, and operating condition information to characterize the dynamic evolution process of device operating states, wherein the state nodes correspond to device identifiers; and a fault mode propagation sub-graph, used to model fault types, fault characteristics, fault root causes, and fault propagation relationships to characterize the fault occurrence process and its impact diffusion path, wherein the fault-related objects correspond to device identifiers; wherein the sub-graphs are connected and fused across graphs through device identifiers to form a domain knowledge graph.

[0010] Optionally, in some embodiments of this application, the knowledge graph reasoning module is further configured to: map the twin state data to the device knowledge subgraph and the operating state evolution subgraph in the domain knowledge graph to generate a semantic alignment relationship between the twin state data and the graph entities, and construct an initial state association graph based on the semantic alignment relationship; perform state consistency analysis on the current operating state and historical state sequence of the device in the initial state association graph based on the operating state evolution subgraph to generate state offset features, and identify whether there are abnormal state changes based on the state offset features; when there are abnormal state changes, perform fault mode matching and propagation path deduction on the state offset features based on the fault mode propagation subgraph to determine potential faults. The system identifies the types of potential faults and their impact ranges. Based on these potential fault types and their impact ranges, it maps the impact ranges in the device knowledge sub-graph based on device identifiers and their topological relationships to determine the set of target devices experiencing anomalies. It then determines the degree of fault impact of each target device based on the deviation between its operating state and a preset normal state. Anomaly detection results are generated using the set of target devices and their corresponding degree of fault impact, where the anomaly detection results include the target device identifiers and their corresponding degree of fault impact information. The anomaly detection results are mapped to the initial state association graph, and semantic association update processing is performed on the initial state association graph based on device nodes and state nodes to generate semantic feature representations.

[0011] Optionally, in some embodiments of this application, the perceptual agent is further configured to perform feature encoding processing on the semantic feature expression based on the semantic feature expression to obtain a semantic vector.

[0012] Optionally, in some embodiments of this application, the decision-making agent is an agent constructed based on a pre-trained semantic decision-making model; wherein, the pre-trained semantic decision-making model is obtained through supervised learning based on historical operating data, the historical operating data includes historical semantic vector samples and historical global scheduling schemes corresponding to the historical semantic vector samples, the historical semantic vector samples are generated by multi-source heterogeneous data after synchronization by a digital twin base and semantic reasoning processing of a knowledge graph, the historical global scheduling scheme is used to represent the device control strategy under different operating scenarios and abnormal states; the decision-making agent is used to perform feature mapping processing on the semantic vectors to obtain semantic embedding representations; based on the semantic embedding representations, similarity calculation is performed in the semantic feature space composed of historical semantic vector samples to determine the set of historical semantic samples corresponding to the current semantic vector; based on the historical global scheduling schemes corresponding to the set of historical semantic samples, the historical global scheduling schemes are subjected to policy weighted fusion to generate a global scheduling scheme.

[0013] Optionally, in some embodiments of this application, the decision agent is further configured to: perform reasoning calculations on the semantic vector using the pre-trained semantic decision model to obtain the semantic similarity relationship between the semantic vector and historical semantic vector samples; determine a set of historical semantic vector samples corresponding to the current semantic vector from the historical semantic vector samples based on the semantic similarity relationship; obtain a set of corresponding historical global scheduling schemes based on the set of historical semantic vector samples; perform weighted fusion processing on the set of historical global scheduling schemes based on the semantic similarity relationship to obtain a fused global scheduling scheme expression; and generate a global scheduling scheme based on the fused global scheduling scheme expression.

[0014] Optionally, in some embodiments of this application, the preset operating constraints include: the operating parameters and status attributes of each device are within a preset safety threshold range; the device state transition process conforms to the preset state transition relationship and state triggering conditions; and the potential fault impact range does not exceed the preset risk range.

[0015] Optionally, in some embodiments of this application, the decision agent is further configured to: map the simulation results to the domain knowledge graph to obtain feedback semantic feature expressions; perform semantic correction processing on the semantic vector based on the feedback semantic feature expressions to generate updated semantic vectors; input the updated semantic vectors into the decision agent, calculate the semantic similarity relationship between the updated semantic vectors and historical semantic vector samples through the pre-trained semantic decision model to determine the corresponding historical semantic vector sample set; obtain the corresponding historical global scheduling scheme set based on the historical semantic vector sample set; and perform weighted fusion processing on the historical global scheduling scheme set based on the updated semantic vectors and the semantic similarity relationship to generate a new global scheduling scheme.

[0016] Optionally, in some embodiments of this application, the simulation results are mapped to the domain knowledge graph to obtain feedback semantic feature expressions associated with device nodes and state nodes. Specifically, this includes: parsing the simulation results into operational state change data and state transition results for each device during the simulation process; based on the device identifier, associating and matching the operational state change data with device nodes in the device knowledge subgraph of the domain knowledge graph to determine device-level semantic associations; based on the state transition results, matching the state transition relationships with state nodes and state transition relationships in the operational state evolution subgraph to determine state-level semantic associations; and generating the feedback semantic feature expressions based on the device-level semantic associations and state-level semantic associations.

[0017] Optionally, in some embodiments of this application, semantic correction processing is performed on the semantic vector based on the feedback semantic feature expression to generate an updated semantic vector. Specifically, this includes: matching the feedback semantic feature expression with the semantic vector to determine the semantic portion of the semantic vector corresponding to the abnormal device node and the state node; determining the abnormal change direction based on the deviation relationship between the state change result of the corresponding device node in the feedback semantic feature expression and a preset normal state, wherein the abnormal change direction includes a state elevation direction, a state depreciation direction, or a state shift direction; dynamically adjusting the weights of the corresponding semantic portion in the semantic vector based on the feedback semantic feature expression to correct the semantic vector, wherein the weight adjustment direction is determined by the abnormal change direction; and inputting the adjusted semantic vector into the decision-making agent for recalculation to obtain an updated semantic vector.

[0018] (III) Beneficial Effects

[0019] This application provides a multi-agent collaborative decision-making system based on digital twins and knowledge graphs. By preprocessing multimodal heterogeneous data during the data acquisition stage and connecting it to a digital twin platform, unified twin state data is obtained. Semantic reasoning analysis of the twin state data is then performed using a knowledge graph reasoning module. This enables semantic modeling and anomaly detection of the operational state of complex systems, thus avoiding the inaccurate analysis caused by missing data semantics in traditional methods. Simultaneously, a perceptual agent generates semantic vectors based on the semantic feature expression, and a decision-making agent generates a global scheduling scheme based on these semantic vectors, enabling multi-source information to be effectively integrated. Fusion and decision-making are completed within a unified semantic space, improving the consistency and accuracy of collaborative decision-making. Furthermore, the global scheduling scheme is virtually executed and dynamically simulated in the digital twin platform through a simulation and deduction module. Based on the simulation and deduction results, the global scheduling scheme is constrained, verified, and optimized. This allows the decision-making process to be verified and iteratively optimized before actual execution. Even in complex dynamic environments, the generated global scheduling scheme can still meet the preset operating constraints. Then, the executing intelligent agent controls the corresponding devices, thereby achieving closed-loop optimization of the entire decision-making process and significantly improving the reliability, stability, and decision-making accuracy of the system. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of a multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to an embodiment of this application;

[0021] Figure 2 This is a flowchart illustrating the process of obtaining a feedback semantic feature representation according to an embodiment of this application;

[0022] Figure 3This is a flowchart illustrating the process of generating and updating semantic vectors according to one embodiment of this application. Detailed Implementation

[0023] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0024] In related technologies, existing methods for collaborative decision-making problems in complex systems can be mainly categorized into two types:

[0025] The first category is rule-driven expert system decision-making schemes. These schemes involve manually constructing rule bases or knowledge bases to logically judge the system's operational status and generate control strategies based on preset rules. However, this approach inherently relies on fixed rule matching mechanisms and lacks the ability to model deep semantic relationships between multimodal heterogeneous data. When faced with complex systems involving multiple devices and multiple constraints, it struggles to achieve globally consistent decision-making under a unified semantic framework. Furthermore, rule construction and maintenance heavily depend on expert experience, resulting in high update costs, poor scalability, and a tendency for decision lags or conflicts in dynamically changing environments.

[0026] The second category is data-driven single-model prediction and decision-making solutions. These solutions typically employ machine learning or deep learning models to train and model historical operational data, generating scheduling or control strategies based on the model's output. While this approach can uncover potential patterns from data, it often focuses on a single task or data modality, lacking a unified semantic expression capability for multimodal heterogeneous data. Information between different modules is difficult to effectively integrate and share, making it challenging to support cross-level and cross-task collaborative decision-making. Furthermore, these methods usually make decisions directly based on model output, lacking a pre-verification mechanism for the decision results. In complex scenarios, this may lead to decision results that do not meet actual operational constraints, thus affecting the system's security and reliability.

[0027] To address this, this application provides a multi-agent collaborative decision-making system based on digital twins and knowledge graphs. This system preprocesses multimodal heterogeneous data during the data acquisition phase and connects it to a digital twin platform to obtain twin state data. A knowledge graph reasoning module then performs semantic reasoning analysis on the twin state data to obtain anomaly detection results and corresponding semantic feature expressions, thereby achieving unified semantic modeling of the complex system's operational state. Furthermore, a perceptual agent generates semantic vectors based on the semantic feature expressions, and a decision-making agent generates a global scheduling scheme based on these semantic vectors, enabling the fusion and collaborative decision-making of multi-source information within a unified semantic space. Simultaneously, a simulation module performs virtual execution and dynamic simulation of the global scheduling scheme within the digital twin platform. Based on the simulation results, the global scheduling scheme undergoes constraint verification and feedback optimization. If the simulation results do not meet preset operational constraints, iterative optimization is performed until the conditions are met, at which point the executing agent takes control. Even in complex dynamic environments and under multiple constraints, the system maintains stability, consistency, and high reliability in the decision-making process, while also possessing good versatility and scalability.

[0028] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Figure 1 This is a schematic diagram of the structure of a multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to an embodiment of this application. Figure 1 As shown, the multi-agent collaborative decision-making system based on digital twins and knowledge graphs includes: a data acquisition module, used to acquire multimodal heterogeneous data, and after preprocessing the multimodal heterogeneous data, it is connected to the digital twin base to obtain twin state data; in this embodiment, the multimodal heterogeneous data includes video image data, used to reflect road traffic conditions and vehicle behavior information; sensor detection data, including traffic detection data such as traffic flow, vehicle speed, and occupancy rate; equipment operation status data, including traffic light timing status and controller operation status; traffic business data, including historical traffic flow data, congestion records, and traffic event data; and environmental data, including meteorological data, road construction information, and emergency event information. The preprocessing of the multimodal heterogeneous data in this embodiment includes denoising, time alignment, and feature fusion. Denoising removes abnormal and redundant data, time alignment unifies the time scale of different data sources, and feature fusion maps different modal data to a unified feature space.

[0029] The system comprises a knowledge graph reasoning module for performing semantic reasoning analysis on the digital twin state data based on a pre-constructed domain knowledge graph, obtaining anomaly detection results and corresponding semantic feature expressions; a perceptual agent for generating semantic vectors based on the semantic feature expressions; a decision agent for generating a global scheduling scheme based on the semantic vectors; a simulation and deduction module for virtually executing and dynamically simulating the global scheduling scheme in the digital twin base, obtaining simulation and deduction results; wherein the simulation and deduction results include information on changes in device operating status; if the simulation and deduction results meet preset operating constraints, the global scheduling scheme is sent to the execution agent; if the simulation and deduction results do not meet the preset operating constraints, the simulation and deduction results are fed back to the decision agent, which iteratively optimizes the global scheduling scheme based on the semantic vectors and the simulation and deduction results, generates a new global scheduling scheme, and inputs it back into the simulation and deduction module for verification until the simulation and deduction results meet the preset operating constraints; and an execution agent for receiving the global scheduling scheme sent by the simulation and deduction module and controlling the device corresponding to the global scheduling scheme.

[0030] In this embodiment, the equipment operating state change information includes, but is not limited to: state attribute change sequences, operating parameter change sequences, and state evolution process data for each device at different simulation time steps. Specifically, the state attribute change sequences characterize the device's state switching during the simulation process, such as switching from an initial normal operating state to a congestion optimization state, a balanced control state, or an anomaly suppression state; the operating parameter change sequences characterize the changes in key control parameters of the device under different states, such as the dynamic adjustment process of signal cycle, green light duration, red light duration, and phase ratio parameters of signal control devices; and the state evolution process data characterizes the continuous change trajectory of the device's state over time and its corresponding state transition path.

[0031] The following section provides a detailed explanation of the system using a specific application scenario, such as intelligent transportation operation and scheduling: In urban transportation systems, multiple data sources are involved, including road monitoring equipment, vehicle operation data, traffic light control systems, and meteorological information. During operation, these sources continuously generate multimodal heterogeneous data such as traffic flow, vehicle speed, occupancy rate, signal timing status, and traffic event information. The data acquisition module collects this multimodal heterogeneous data uniformly, performs preprocessing such as noise reduction, time synchronization, and format conversion, and then connects it to a digital twin platform to construct twin state data corresponding to the actual traffic network, thereby achieving real-time mapping of urban traffic operation status. Based on this, the knowledge graph reasoning module performs semantic reasoning analysis on the twin state data based on a pre-constructed traffic domain knowledge graph. For example, by combining the road network topology, the relationships between road segments, and historical congestion propagation patterns, it identifies abnormal congestion, traffic accidents, or decreased traffic efficiency, and outputs anomaly detection results and corresponding semantic feature expressions. The perceptual agent further generates semantic vectors based on the semantic features, enabling data from different road segments and of different types to be expressed and fused in a unified semantic space. Subsequently, the decision-making agent, based on the semantic vectors, comprehensively considers road network traffic efficiency, signal control constraints, and traffic safety requirements to generate a global scheduling scheme, such as optimizing signal timing schemes at multiple intersections, dynamically adjusting lane usage strategies, or implementing traffic diversion control measures. The simulation module virtually executes and dynamically simulates the global scheduling scheme in a digital twin platform, obtaining simulation results, including information on changes in traffic flow, average travel time, queue length, and congestion propagation trends in each road segment, as well as changes in equipment operating status. When the simulation results meet the preset operational constraints (such as road capacity constraints, signal control safety constraints, and overall congestion index thresholds), the global scheduling scheme is sent to the execution agent, which then controls the actual traffic control equipment (such as traffic light controllers and guidance screens). When the simulation results do not meet the preset operational constraints, the simulation results are fed back to the decision agent. Based on the semantic vector and the simulation results, the decision agent iteratively optimizes the global scheduling scheme and re-enters it into the simulation module for verification until the constraints are met before execution, thereby preventing unreasonable scheduling strategies from directly affecting the actual traffic system.

[0032] As can be seen from the above process, the technical solution of this application achieves unified semantic modeling of traffic operation status by integrating multimodal heterogeneous traffic data into a digital twin platform and combining it with knowledge graph reasoning. Through the collaboration of perceptual and decision-making agents, multi-source traffic information can be fused and used for decision-making within a unified semantic space, improving the consistency and globality of traffic scheduling. Simultaneously, the scheduling scheme is virtually verified through a simulation module and iteratively optimized based on feedback, ensuring thorough verification of the scheme before actual execution, thus constructing a closed-loop collaborative decision-making mechanism. This not only effectively reduces the risks associated with information fragmentation or unverified decisions in traditional methods but also significantly improves the efficiency, stability, and accuracy of road network operation and scheduling decisions in complex and dynamic traffic environments, demonstrating significant practical application value.

[0033] In some embodiments of this application, the pre-constructed domain knowledge graph specifically includes:

[0034] The device knowledge subgraph is used to structurally represent device types, device identifiers, operating parameters, status attributes, and topological relationships between devices. For example, in the device knowledge subgraph, various traffic facilities in an urban transportation system are used as modeling objects, such as traffic lights, road segment detectors, cameras, and roadside communication units. By structurally representing the device types, device identifiers, operating parameters (such as traffic light timing cycles and detector sampling frequencies), status attributes (such as running, faulty, and offline), and topological relationships between devices (such as relationships between adjacent intersections and upstream / downstream road segments), a semantic network at the traffic device level is formed. For instance, between intersection A and intersection B on a main road, the traffic flow dependency can be represented through topological relationships, providing a foundation for subsequent congestion propagation analysis.

[0035] The operational state evolution sub-graph is used to model equipment state nodes, state transition relationships, state triggering conditions, and operating condition information to characterize the dynamic evolution process of equipment operating states. The state nodes correspond to equipment identifiers. For example, the operational state evolution sub-graph uses equipment operating state nodes as the core to model the state change process of traffic equipment at different time scales, such as state nodes like "normal traffic state," "congestion state," and "abnormal signal timing state." State transition relationships are constructed, such as "normal → congestion" and "congestion → recovery." State triggering conditions are also introduced, such as traffic flow exceeding a threshold, average queue length exceeding a set value, or abnormal signal cycle information. This sub-graph can depict the dynamic evolution process of the traffic system over time. For example, during the morning rush hour, an intersection gradually evolves from a normal traffic state to a congestion state due to a continuous increase in traffic flow, further affecting the traffic efficiency of downstream road sections.

[0036] The Fault Mode Propagation (FMT) sub-graph is used to model fault types, fault characteristics, root causes, and fault propagation relationships to characterize the fault occurrence process and its impact diffusion path. The fault-related objects correspond to equipment identifiers. For example, the FMT sub-graph models potential anomalies and faults in the traffic system, such as traffic light controller malfunctions, detector data loss, and sudden road accidents. It provides a structured representation of fault characteristics, fault types, and potential root causes, and constructs fault propagation relationships. For instance, a traffic light malfunction at an intersection may lead to vehicle congestion in adjacent road sections, subsequently causing regional congestion to spread. This sub-graph allows for modeling the propagation path of traffic anomalies from local equipment failure to regional traffic impact, enabling interpretable analysis of the traffic risk diffusion process.

[0037] In this system, the various sub-graphs are connected and fused across graphs via device identifiers to form a domain knowledge graph. Specifically, the device knowledge sub-graph, the operational state evolution sub-graph, and the fault mode propagation sub-graph are connected and fused across graphs via device identifiers. This allows the structural relationships, state evolution processes, and fault propagation impacts of the same traffic equipment across different dimensions to be expressed in a unified knowledge graph. For example, a traffic light device has its topological location and parameter information in the device knowledge sub-graph, reflects its state change process in the operational state evolution sub-graph, and is associated with its fault impact range in the fault propagation sub-graph, thus forming a complete multi-dimensional semantic expression system. Therefore, this application can achieve unified semantic modeling of multi-source heterogeneous traffic data in a traffic system, enabling the fusion and expression of device information, operational state, and fault propagation relationships within the same knowledge framework, thereby significantly improving the system's ability to understand and reason about complex traffic conditions. Meanwhile, this structure can support the early identification and analysis of traffic congestion evolution and fault propagation paths, providing a high-quality semantic foundation for subsequent digital twin-based simulation and collaborative decision-making, effectively improving the accuracy, foresight, and overall operational efficiency of traffic scheduling.

[0038] In some embodiments of this application, the knowledge graph reasoning module is further configured to map the twin state data to the device knowledge subgraph and the operational state evolution subgraph in the domain knowledge graph, so as to generate a semantic alignment relationship between the twin state data and the graph entities, and to construct an initial state association graph based on the semantic alignment relationship; more specifically, the twin state data obtained from the digital twin base is mapped to the device knowledge subgraph and the operational state evolution subgraph in the domain knowledge graph. For example, the real-time timing data of a traffic light at an intersection, the traffic flow data collected by the detector, and the road segment traffic status information are associated with the corresponding "traffic light device node" and "road segment node" and their historical state evolution sequence in the knowledge graph, thereby generating a semantic alignment relationship between the twin state data and the graph entities, and constructing an initial state association graph, so that the physical traffic system forms a computable semantic mapping structure in the graph.

[0039] Based on the aforementioned operational state evolution sub-map, a state consistency analysis is performed on the current operational state and historical state sequence of the equipment in the initial state association diagram to generate state offset features. These features are then used to identify whether any abnormal state changes exist. For example, for a traffic light at an intersection, its normal timing cycle and historical stable operating range can serve as a reference benchmark. When the currently acquired signal cycle deviates significantly from historical patterns, or when there is a sudden change in the coordination relationship with adjacent intersections, state offset features can be generated through state consistency analysis, such as "abnormal signal timing offset" or "abnormally shortened green light time," and based on this, whether any abnormal state changes exist can be identified.

[0040] When abnormal state changes occur, fault mode matching and propagation path deduction are performed on the state offset features based on the fault mode propagation sub-map to determine the potential fault type and its impact range. For example, when a traffic light at an intersection experiences a control anomaly, fault modes such as "signal controller failure" or "communication interruption" can be matched. Combined with the preset fault propagation relationship in the map, it can be deduced that the anomaly may spread to upstream and downstream road sections, thereby causing traffic backlog or regional congestion on a main road, thus determining the potential fault type and its impact range.

[0041] Based on the potential fault types and their impact range, the impact range is mapped in the device knowledge sub-graph based on device identifiers and their topological relationships to determine the target device set where anomalies occur; for example, multiple traffic lights at intersections, detectors on adjacent road sections, and guidance screens, etc., which are affected. Furthermore, by calculating the deviation between the current operating state and the preset normal state of each device in the target device set, such as the abnormal fluctuation range of traffic flow, the growth rate of queue length, or the percentage decrease in traffic speed, the degree of fault impact corresponding to each target device is determined, achieving a quantitative expression of the impact of the anomaly.

[0042] Based on the deviation between the operating status of each target device in the target device set and the preset normal state, the degree of fault impact of the corresponding target device is determined. In a smart transportation scenario, for example, a main road in a city consists of multiple intersections and road segments. When the knowledge graph reasoning module identifies a traffic light control anomaly and determines the target device set, the operating status of each target device can be further quantitatively analyzed. For example, for intersection traffic light A in the target device set, its current green light time has decreased from 60 seconds to 30 seconds compared to the preset normal state, a deviation of 50%; for detector B on the adjacent road segment, the detected traffic flow has increased by 40% compared to the historical average, deviating from the normal operating range; for traffic light C at the downstream intersection, its queue length growth rate is significantly higher than the preset threshold. By calculating the deviation between the above operating statuses and the preset normal state, the degree of fault impact of each target device can be determined. For example, traffic light A is "high impact level", detector B is "medium impact level", and traffic light C is "medium-high impact level", thereby achieving differentiated quantitative expression of the abnormal impact of different devices.

[0043] Anomaly detection results are generated based on the target device set and their corresponding fault impact levels. The anomaly detection results include the target device identifier and its corresponding fault impact level information. For example, the output results may include structured information such as "Signal Light A - Impact Level 0.8", "Detector B - Impact Level 0.6", and "Signal Light C - Impact Level 0.7", so that the anomaly detection results not only include the device identifier, but also its corresponding impact intensity information, thereby realizing the transformation from qualitative anomaly judgment to quantitative anomaly assessment.

[0044] The anomaly detection results are mapped to the initial state association graph, and the initial state association graph is semantically updated based on device nodes and state nodes to generate semantic feature expressions. For example, in the device node layer of the traffic knowledge graph, traffic light A node is marked as a high-impact anomaly node, and its "timing anomaly state" is labeled in the state node layer. At the same time, detector B and traffic light C nodes are labeled with corresponding impact levels, and combined with their topological relationships with upstream and downstream road segments, the state propagation path is visualized, updated, and semantically enhanced. Through this process, the initial state association graph not only reflects whether the device is abnormal, but also intuitively expresses the intensity of the anomaly's impact and its diffusion relationship in the road network, thereby generating semantic feature expressions with enhanced semantic expression capabilities.

[0045] As can be seen from the above process, the embodiments of this application, through a processing chain of operational state deviation degree, fault impact degree quantification, structured anomaly result generation, and graph semantic update, expand the anomaly detection result from a single device-level judgment to a multi-device, multi-level impact assessment result. Compared with the traditional detection method that only outputs "whether it is abnormal", it can significantly improve the richness of anomaly information expression and decision support capability. At the same time, by writing the anomaly result back to the initial state association graph, the graph is dynamically updated, enabling the system to continuously reflect the changes in traffic operation status, thereby improving the accuracy and interpretability of subsequent simulation and collaborative decision-making.

[0046] Optionally, in some embodiments of this application, the perceptual agent is further used to perform feature encoding processing on the semantic feature expression to obtain a semantic vector. In intelligent transportation scenarios, the process of the perceptual agent performing feature encoding processing based on the semantic feature expression can be understood as further converting the "structured semantic information" output by the knowledge graph reasoning module into a "vectorized representation" that can be directly used by the decision-making agent. For example, in an urban road network, the knowledge graph reasoning module has already output a semantic feature expression for a certain area. This expression may include information such as "Traffic light A has a high-impact timing anomaly", "Detector B's traffic flow deviates by 40%", "Road segment C has a moderate to high congestion risk", and its corresponding equipment topology and anomaly propagation path. Although these semantic feature expressions already have clear semantic meanings, they still exist in graph structure or symbolic form, making it difficult to directly use them for subsequent global scheduling calculations and optimizations. At this point, the perceptual agent performs feature encoding on the semantic feature expression. For example, it maps device identification, anomaly type, impact level, topological relationship strength, and state evolution characteristics to a high-dimensional semantic space. Through an encoding model, the combined relationship of "traffic light A—high-impact anomaly—downstream congestion propagation" is transformed into a semantic vector representation. This semantic vector can simultaneously integrate device attribute information, state offset information, and spatial topological relationship information. For example, it can be represented as different dimensions in a high-dimensional vector corresponding to "anomaly strength dimension," "spatial propagation dimension," and "temporal evolution dimension," respectively, thereby achieving a unified expression of complex traffic semantic information.

[0047] Through the above process, for example in the morning rush hour traffic scenario, a certain main road area may have multiple abnormal semantic features at the same time, such as abnormal signal timing, local traffic surge and increased congestion propagation trend. The perceptual agent can encode these scattered semantic features into a unified semantic vector, so that it can comprehensively reflect the overall traffic operation status of the area, instead of relying on a single device or a single indicator for description.

[0048] Specifically, the decision-making agent is an agent built based on a pre-trained semantic decision-making model; this decision-making agent does not directly rely on real-time rules or human experience to make decisions, but is trained in advance using a large amount of historical traffic operation data, so that it has the ability to "directly infer scheduling strategies from traffic semantic states".

[0049] The pre-trained semantic decision model is obtained through supervised learning based on historical operational data. This historical operational data includes historical semantic vector samples and corresponding historical global scheduling schemes. The historical semantic vector samples are generated from multi-source heterogeneous data after synchronization with a digital twin platform and semantic reasoning using a knowledge graph. The historical global scheduling scheme is used to represent equipment control strategies under different operational scenarios and abnormal states. For example, in an urban traffic management system, historical operational data is accumulated over a long period. This historical operational data is not simply raw traffic flow or video data, but structured data formed after processing by a digital twin platform and a knowledge graph reasoning module. The historical semantic vector samples originate from multi-source heterogeneous traffic data, such as intersection camera video data, geomagnetic detector traffic flow data, traffic light timing data, and meteorological and traffic event data. After time synchronization and state alignment by the digital twin platform, this data is combined with a knowledge graph to perform semantic reasoning on equipment relationships, road network topology relationships, and traffic state evolution relationships, ultimately generating historical semantic vectors that can express the "semantics of traffic operation status." For example, a historical sample might correspond to a semantic vector representing "congestion on main road A during morning rush hour + signal timing imbalance at intersection B + continuous increase in queue length on downstream roads." Simultaneously, each set of historical semantic vector samples corresponds to a historical global scheduling scheme, which characterizes the actual equipment control strategy employed in that traffic operation scenario or abnormal state. For instance, in the "congestion scenario on main road during morning rush hour," the historical global scheduling scheme might include a combination of strategies such as "optimizing the green wave signal timing," "increasing the proportion of traffic released on main roads," and "diverting traffic to side roads." In the "local blockage scenario caused by a traffic accident," it might correspond to a control strategy of "fixed local signal timing + detour path guidance + regional traffic restriction." Based on this, by using a large number of "historical semantic vector samples—historical global scheduling schemes" as supervised learning data to train a pre-trained semantic decision model, the model learns the mapping relationship between different traffic semantic states and optimal or suboptimal scheduling strategies. For example, the model gradually learns that when the semantic vector indicates "increasing congestion spread trend," it should tend to generate diversion and signal coordination optimization strategies; when the semantic vector indicates "local equipment failure + sudden drop in capacity," it should prioritize generating emergency fixed timing and route guidance strategies. Through this approach, the pre-trained semantic decision model in this application can fully absorb scheduling experience from different traffic operation scenarios and abnormal states during the training phase, enabling the model to directly infer the corresponding global scheduling scheme based on the input semantic vector without relying on manual rules during actual operation.

[0050] The decision-making agent performs feature mapping on the semantic vectors to obtain semantic embedding representations. For example, in an urban traffic network, the knowledge graph reasoning module and the perception agent have already generated semantic vectors representing the current traffic state. These semantic vectors may contain semantic information such as "intensified congestion on main road A," "abnormal signal timing at intersection B," "congestion queue length continuing to increase downstream," and "congestion propagation trend in area C." Although this information already possesses clear traffic semantic meanings, it still exists in a structured or semi-structured form, making it difficult to directly use for efficient matching calculations with historical scenarios. In this case, the decision-making agent performs feature mapping on the semantic vectors, for example, by using a mapping layer or encoding layer in a pre-trained semantic decision model to map the semantic vectors, which contain multi-dimensional information such as equipment status, topological relationships, and anomaly intensity, into a unified semantic embedding space. In this process, discrete semantic features such as "intensified congestion on main road A," "abnormal signal timing," and "intensified congestion propagation trend" are converted into continuous high-dimensional semantic embedding representations, making them computable and comparable in the numerical space. For example, after feature mapping, the current traffic status no longer exists as an independent semantic label, but is represented as a semantic embedding vector, in which different dimensions implicitly encode information such as "congestion severity characteristics", "spatial propagation characteristics", "temporal evolution characteristics" and "equipment malfunction impact characteristics", so that the embedding vector can fully express the comprehensive operating status of the current traffic system.

[0051] Based on the semantic embedding representation, similarity calculation is performed in the semantic feature space composed of historical semantic vector samples to determine the set of historical semantic samples corresponding to the current semantic vector. For example, during the morning rush hour in a certain city, the current road network status is encoded into a semantic embedding representation, which comprehensively reflects multi-dimensional traffic semantic features such as "signal flow on main road A increases significantly," "signal timing deviation at intersection B leads to queue growth," and "downstream road segment C shows an enhanced trend of congestion diffusion." This embedding vector is then input into the historical semantic feature space and similarity calculation is performed with a large number of labeled historical semantic vector samples, for example, using cosine similarity or vector distance measurement. Through similarity calculation, the system can identify several sets of historical semantic samples that are highly similar to the current traffic status. For example: one is "a congestion scenario on main roads during the morning rush hour in a certain year," characterized by overloaded main road traffic but a basically normal signal system; the second is "a scenario of local congestion caused by abnormal signal timing," characterized by unbalanced signal timing at intersections leading to a rapid increase in queue length; and the third is "a scenario of regional congestion diffusion," characterized by downstream road segments becoming congested step by step and spreading to the surrounding road network.

[0052] Each historical semantic sample corresponds to a historical global scheduling scheme. For example, the strategy corresponding to the "morning rush hour congestion scenario" may include green wave optimization and main road priority release control, the "signal anomaly scenario" corresponds to signal timing reconstruction and local fixed cycle control, and the "congestion spread scenario" corresponds to regional diversion guidance and entrance flow restriction control, etc.

[0053] Based on the historical global scheduling schemes corresponding to the historical semantic sample set, the historical global scheduling schemes are fused using a strategy weighting method to generate a global scheduling scheme. For example, for the current scenario, since it simultaneously exhibits mixed characteristics of "main road congestion" and "signal timing anomalies," the "morning peak congestion strategy" and "signal anomaly control strategy" will be given higher weights, while the "regional diffusion strategy" will have a relatively lower weight. Finally, by weighting and combining multiple historical global scheduling schemes, a global scheduling scheme suitable for the current traffic condition is generated, such as fusing them into a comprehensive control strategy of "main road green wave optimization + intersection signal reconstruction + local dynamic diversion."

[0054] As can be seen from the above process, the technical solution of this application enables the current traffic state to accurately match similar historical operating scenarios by performing similarity retrieval in the semantic feature space, thereby fully reusing existing scheduling experience and avoiding complete reliance on real-time rule design; at the same time, by performing policy weighted fusion of multiple historical global scheduling schemes, the generated control strategy is no longer limited to a single historical experience, but integrates the optimal policy features of multiple similar scenarios, thereby significantly improving the adaptability and robustness of scheduling schemes in complex traffic scenarios and improving the overall operating efficiency and stability of the road network.

[0055] In some embodiments of this application, the decision-making agent is further configured to: perform reasoning calculations on the semantic vectors using the pre-trained semantic decision model to obtain the semantic similarity relationship between the semantic vectors and historical semantic vector samples; determine a set of historical semantic vector samples corresponding to the current semantic vector from the historical semantic vector samples based on the semantic similarity relationship; obtain a set of corresponding historical global scheduling schemes based on the set of historical semantic vector samples; and perform weighted fusion processing on the set of historical global scheduling schemes based on the semantic similarity relationship to obtain a fused global scheduling scheme expression. In this embodiment, the weighted fusion processing on the set of historical global scheduling schemes based on the semantic similarity relationship refers to using the similarity between each historical semantic vector sample and the current semantic vector as a weight to perform weighted calculations on the control parameters in the corresponding historical global scheduling scheme, including signal timing parameters, road segment traffic weights, and diversion control coefficients, thereby obtaining a unified parameterized scheduling expression, and generating a global scheduling scheme based on this expression.

[0056] Based on the fused global scheduling scheme expression, a global scheduling scheme is generated. For example, during morning and evening rush hours in a city, the current road network generates a semantic vector through perception and knowledge graph reasoning. This semantic vector represents a comprehensive traffic state such as "continuous increase in traffic flow on main road A," "signal timing deviation at intersection B leading to queue growth," and "congestion spreading trend in downstream road segment C." Subsequently, a pre-trained semantic decision model performs reasoning calculations on this semantic vector to obtain its semantic similarity relationship with historical semantic vector samples. For example, the similarities with three typical historical traffic scenarios are calculated to be 0.82, 0.76, and 0.69, respectively. The historical semantic vector sample with the highest similarity may correspond to a "morning rush hour main road congestion scenario," the second to a "signal timing anomaly leading to local congestion scenario," and the third to a "congestion spreading to the regional road network scenario." Based on this semantic similarity relationship, a set of historical semantic vector samples corresponding to the current semantic vector is determined from historical data, thus forming multiple historical reference scenarios similar to the current traffic state. Further, based on this set of historical semantic vector samples, a corresponding set of historical global scheduling schemes is obtained. For example, the historical scheduling scheme for a "main road congestion scenario" might be "green wave optimization + main road priority," while a "signal anomaly scenario" might correspond to "signal cycle reconstruction + local fixed timing control," and a "congestion diffusion scenario" might correspond to "regional diversion + entrance flow restriction control." Subsequently, based on semantic similarity, the aforementioned set of historical global scheduling schemes is weighted and fused. For instance, scenarios with a similarity of 0.82 have higher weights, and their control strategies dominate the fusion result; while scenarios with lower similarity have lower strategy weights, serving only as supplementary optimization factors. In this way, multiple historical scheduling schemes are fused under a unified weight system to form a fused global scheduling scheme expression, which comprehensively reflects the optimal control experience under different historical traffic scenarios. Finally, based on this fused global scheduling scheme expression, a global scheduling scheme for the current traffic state is generated, such as outputting a combined control strategy of "main road green wave optimization + dynamic signal cycle reconstruction + local diversion guidance + regional entrance flow restriction."

[0057] As can be seen from the above process, the technical solution of this application introduces a semantic similarity-driven historical sample retrieval mechanism, enabling the system to accurately locate the historical operation scenario most relevant to the current traffic state. At the same time, by performing weighted fusion of multiple historical global scheduling schemes based on similarity weights, the generated scheduling strategy no longer relies on a single historical experience, but integrates the optimal strategy features of multiple similar traffic scenarios, thereby significantly improving the generalization ability and robustness of scheduling schemes in complex traffic environments, and improving the road network operation efficiency and emergency response capability.

[0058] In other embodiments of this application, in intelligent transportation scenarios, such as when a main road in a city and its adjacent road network experience localized congestion accompanied by abnormal signal timing during the morning rush hour, the system first determines a set of historical semantic vector samples matching the current traffic state from historical operational data based on the semantic similarity relationship between the current semantic vector and historical semantic vector samples. These samples may include several typical historical traffic operation states such as "morning rush hour main road congestion scenario," "abnormal signal timing scenario," and "regional congestion spread scenario." Each of these historical semantic vector samples corresponds to a pre-recorded historical global scheduling scheme, used to characterize the actual equipment control strategy adopted in the corresponding traffic scenario. Based on this, the system assigns weights to the set of historical semantic vector samples according to the semantic similarity relationship between the current semantic vector and each historical semantic vector sample. For example, the "main road congestion scenario" with high similarity receives a higher weight, the "abnormal signal timing scenario" with moderate similarity receives a moderate weight, and the "regional congestion spread scenario" with low similarity receives a lower weight. Subsequently, based on the aforementioned weights, the corresponding historical global scheduling schemes are weighted and fused. For example, signal timing parameters, arterial road traffic priority parameters, branch road release ratio parameters, and diversion control intensity parameters from different historical schemes are weighted and calculated to form a fused scheduling expression within a unified parameter space. For instance, in the above scenario, after weighted fusion, a comprehensive control strategy is obtained with a signal cycle of approximately 107 seconds, a significant increase in arterial road traffic priority, a corresponding decrease in branch road release ratio, and the introduction of an appropriate regional diversion control intensity. This allows the fused global scheduling scheme to simultaneously address multiple objectives, including improving arterial road traffic efficiency, correcting signal timing anomalies, and suppressing congestion spread. In this way, this application can collaboratively fuse multiple historical traffic operation experiences driven by semantic similarity to generate a unified scheduling scheme suitable for the current complex traffic conditions, thereby significantly improving the decision-making adaptability, robustness, and overall operational efficiency of the traffic system in dynamic environments.

[0059] In another embodiment of this application, an urban road traffic scheduling scenario is used as an example. During the morning rush hour at 08:15, a city experiences abnormal traffic conditions on its main roads and adjacent road networks. A digital twin platform is used to map the traffic network in real time, and a knowledge graph reasoning module is used to perform semantic reasoning analysis on multi-source heterogeneous traffic data to obtain a semantic vector representing the current traffic state. This semantic vector indicates that the current traffic state is characterized by a 45% increase in traffic flow on main road A compared to normal levels, showing a significant trend of worsening congestion; abnormal signal timing at intersection B, with a significantly shorter green light duration; and a continuously increasing queue length on downstream road segment C, indicating a trend of congestion spreading downstream. Based on these characteristics, the current traffic state is characterized as a complex traffic state with both localized congestion and signal imbalance, and a moderate risk of congestion spreading. On this basis, the decision-making agent, based on a pre-trained semantic decision model, calculates the similarity of the current semantic vector in the historical semantic feature space and determines the set of historical semantic samples most similar to the current traffic state from the historical semantic vector samples. Specifically, the system identified three representative historical operational scenarios, including a morning rush hour main road congestion scenario, a signal timing anomaly scenario, and a congestion spread control scenario. These historical semantic samples each correspond to different semantic similarity weights. The main road congestion scenario has the highest similarity, with a weight of 0.5; the signal timing anomaly scenario has the second highest similarity, with a weight of 0.3; and the congestion spread control scenario has a relatively low similarity, with a weight of 0.2.

[0060] Furthermore, each historical semantic sample corresponds to a pre-recorded historical global scheduling scheme. Specifically, the scheduling scheme for the main road congestion scenario includes increasing the green light priority ratio on the main road to 0.80, decreasing the priority ratio on side roads to 0.20, and extending the signal cycle to 120 seconds, with the control objective being to improve the traffic efficiency of the main road. The scheduling scheme for the signal timing anomaly scenario includes balanced priority control for main roads and side roads and a 90-second signal cycle configuration, with the control objective being to correct the signal timing anomaly. The scheduling scheme for the congestion spread control scenario includes dynamically allocating the priority ratio for main roads and side roads and a 100-second signal cycle configuration, with the control objective being to suppress the downstream spread of congestion. After obtaining the above set of historical global scheduling schemes, the decision-making agent performs weighted fusion processing on each historical scheduling scheme based on semantic similarity. Specifically, using semantic similarity as weight, the signal cycle parameter, main road priority ratio parameter, and side road priority ratio parameter in each historical scheduling scheme are weighted and calculated to obtain a unified fused scheduling parameter. The signal cycle, after weighted fusion, is calculated to be 107 seconds. The fusion result for the release weight of the main road is 0.67, and the fusion result for the release weight of the branch road is 0.33.

[0061] Based on the fused scheduling parameters, a global scheduling scheme is generated. This scheme simultaneously considers multiple control objectives, including improving the traffic efficiency of main roads, correcting signal timing anomalies, and suppressing congestion spread. Finally, the executing agent maps the global scheduling scheme to the actual traffic control equipment, adjusting traffic light control parameters. For example, the green light time on main roads is adjusted to approximately 72 seconds, dynamic periodic correction control is implemented at intersections with signal timing anomalies, and the release ratio of downstream branch roads is reduced, along with traffic diversion prompts on guidance screens. This achieves coordinated and optimized control of the overall road network operation.

[0062] In practical applications of this application, the preset operational constraints include: the operating parameters and status attributes of each device are within a preset safety threshold range; the device state transition process conforms to preset state transition relationships and state triggering conditions; and the potential impact range of a fault does not exceed a preset risk range. Specifically, regarding constraints on device operating parameters and status attributes, for example, in an urban intersection signal control system, the green light duration, red light duration, and signal cycle of traffic lights must all be within a preset safety threshold range. For instance, the green light duration must not be lower than the minimum traffic demand threshold to avoid severe vehicle queuing, nor must it exceed the road capacity limit to prevent the risk of intersection conflicts. Simultaneously, the traffic flow, occupancy rate, and average vehicle speed of road segment detectors must also be within a reasonable range. For example, if the traffic flow abnormally exceeds the road design capacity threshold or the vehicle speed is consistently lower than the minimum traffic threshold, it is considered that the safety constraints are not met.

[0063] Regarding constraints on equipment state transition processes, for example, a traffic light system should, under normal circumstances, follow a state transition path of "normal timing state → congestion optimization state → restoration of equilibrium state," and state changes must meet corresponding state triggering conditions, such as traffic flow continuously exceeding a set threshold, queue length continuously increasing, or intersection delay time increasing. If the simulation results show that the signal state frequently jumps between "high traffic flow state" and "low traffic flow state" within a short period of time, or that the state transition changes without meeting the preset triggering conditions, it indicates that it does not conform to the preset state transition relationship and therefore does not meet the operational constraints.

[0064] Regarding constraints on the potential impact range of faults, for example, when a signal controller at an intersection malfunctions, its impact should primarily be limited to that intersection and its upstream and downstream adjacent road segments. If simulation results show that the signal control anomaly will cause severe congestion on multiple main roads across the region simultaneously, or even significantly reduce the overall traffic capacity of the regional road network, then the fault propagation range exceeds the preset risk range and is deemed not to meet the operational constraints. Conversely, if the fault impact can be limited to a local road segment and controlled through traffic diversion or signal optimization, it is considered to be within an acceptable risk range.

[0065] Specifically, the decision-making agent is further configured to: map the simulation results to the domain knowledge graph to obtain feedback semantic feature expressions; perform semantic correction processing on the semantic vector based on the feedback semantic feature expressions to generate updated semantic vectors; input the updated semantic vectors into the decision-making agent, calculate the semantic similarity relationship between the updated semantic vectors and historical semantic vector samples through the pre-trained semantic decision model to determine the corresponding historical semantic vector sample set; obtain the corresponding historical global scheduling scheme set based on the historical semantic vector sample set; and perform weighted fusion processing on the historical global scheduling scheme set based on the updated semantic vectors and the semantic similarity relationship to generate a new global scheduling scheme.

[0066] In this embodiment, the scheduling of a composite road network consisting of urban expressways and urban arterial roads is used as an example. During the morning rush hour, a city implements a global scheduling plan for arterial road A and its adjacent intersections. The simulation module virtually executes and dynamically simulates this scheduling plan in a digital twin platform, obtaining simulation results. These results show that after implementing green wave optimization and priority release strategies, while the traffic efficiency of arterial road A improves in the short term, a continuous accumulation of queue length occurs at the downstream intersection B. Simultaneously, the traffic flow on branch road C is excessively compressed, leading to a decline in the service level of local branch roads. Furthermore, by mapping the simulation results to the device knowledge sub-graph and the operational state evolution sub-graph in the domain knowledge graph, feedback semantic feature expressions can be obtained, including semantic information such as "the capacity of the arterial road is improved, but downstream queue overflows," "excessive flow restriction on branch roads leads to local saturation and imbalance," and "the state of local intersections shifts from equilibrium to congestion." Based on the above feedback semantic feature expressions, the decision-making agent performs semantic correction processing on the original semantic vector. For example, the optimistic semantic feature "the main road priority strategy is effective" in the original semantic vector is corrected by introducing negative feedback semantic factors such as "increased downstream queuing overflow risk" and "insufficient branch road supply," thereby generating an updated semantic vector. This updated semantic vector can more accurately reflect the actual operational impact of the current scheduling strategy at the system level, adjusting the semantic representation from a single local optimization state to a corrected state expression that takes into account the global road network balance.

[0067] Subsequently, the updated semantic vector is re-inputted into the decision-making agent. A pre-trained semantic decision model is used to recalculate the updated semantic vector, recalculating its semantic similarity relationship with historical semantic vector samples in the historical semantic feature space. Due to the introduction of simulation feedback correction information, the updated semantic vector is closer to the complex abnormal operating states in history, such as "overly strong optimization of the main road leading to downstream congestion spread" and "traffic restriction on branch roads causing local imbalance," thus redefining the corresponding set of historical semantic vector samples. Further, based on this set of historical semantic vector samples, a corresponding set of historical global scheduling schemes is obtained. For example, this set may include different historical global scheduling schemes such as "scheduling strategy that prioritizes the main road but weakens the green wave intensity," "balanced control scheme that introduces dynamic compensation for branch road traffic release," and "correction strategy that limits the main road traffic release gain and enhances downstream traffic flow capacity." Based on this, the decision-making agent performs weighted fusion processing on the set of historical global scheduling schemes based on the updated semantic vector and the semantic similarity relationship. Among them, the "balanced control scheme" with a high similarity to the current modified semantic vector has a larger weight, while the scheme related to the original over-traffic artery priority strategy has a correspondingly lower weight. Thus, different historical scheduling strategies are recombined and weighted at the parameter level to generate a new global scheduling scheme expression.

[0068] Ultimately, the new global scheduling scheme in the simulation environment is characterized by: moderately reducing the green wave intensity of the main road, improving the traffic compensation capacity of the branch road, and optimizing the signal timing structure of the downstream intersection, thereby effectively suppressing the spread of congestion downstream while ensuring the traffic efficiency of the main road, and achieving balanced optimization of the overall operation of the road network.

[0069] Preferably, see Figure 2 The simulation results are mapped to the domain knowledge graph to obtain feedback semantic feature expressions associated with device nodes and state nodes. Specifically, this includes: parsing the simulation results into operational state change data and state transition results for each device during the simulation process; based on the device identifier, associating and matching the operational state change data with device nodes in the device knowledge subgraph of the domain knowledge graph to determine device-level semantic associations; based on the state transition results, matching the state transition relationships with state nodes and state transition relationships in the operational state evolution subgraph to determine state-level semantic associations; and generating the feedback semantic feature expressions based on the device-level semantic associations and state-level semantic associations.

[0070] In this embodiment, taking urban road traffic signal coordinated control as an example, the process of mapping simulation results to a domain knowledge graph and generating feedback semantic feature expressions is explained. After simulating the global scheduling scheme, the simulation module outputs simulation results, which record the operational state changes and state transition results of each traffic device during the execution of the scheduling strategy. For example, in the simulation of main road A and its adjacent intersections, the signal controller of main road A is in the "main road priority state" at the initial stage of the simulation. Subsequently, due to the continuous increase in the queue length of downstream intersection B, its state transitions from "main road priority state" to "balanced timing state" at simulation time step t1, and is further adjusted to "congestion suppression state" at t2. At the same time, the signal control device of branch road C transitions from "normal release state" to "restricted release state", and its green light release ratio continues to decrease to ensure the traffic capacity of the main road. Based on the above simulation results, the operational state change data and state transition results of each device during the simulation process are first analyzed to obtain a structured device-level operational state sequence and state evolution sequence. Among them, the operating state change data is used to characterize the changes in control parameters of each traffic signal control device at different simulation time steps, and the state transition results are used to characterize the switching process and triggering conditions of the device from one operating state to another.

[0071] Furthermore, based on the device identification information, the operational status change data is associated and matched with the device knowledge subgraph in the domain knowledge graph. For example, the device identification of the main road A signal controller is mapped to the corresponding device node in the knowledge graph, and its operational parameters such as green light timing ratio and cycle duration at different simulation times are aligned with the attributes of the device node, thereby establishing a device-level semantic association. Through this process, the state evolution trajectory of the device during the simulation, from "high-priority passage device" to "balanced control device," can be clearly defined.

[0072] Simultaneously, based on the state transition results, the change path of the equipment state during the simulation process is matched with the sub-graph of the operational state evolution. For example, the state transition sequence of "main road priority state → balanced timing state → congestion suppression state" is matched with predefined state nodes and state transition edges in the knowledge graph to identify whether the state change conforms to the preset state evolution law. If the matching result shows that the state transition corresponds to the "dynamic suppression control path triggered by congestion diffusion", then a state-level semantic association relationship is further established to characterize the evolution logic and triggering mechanism of the equipment's operational state in the time dimension.

[0073] Based on this, a feedback semantic feature expression is generated by integrating the device-level semantic association relationships and the state-level semantic association relationships. This feedback semantic feature expression not only includes the specific changes in the operating parameters of each traffic signal control device during the simulation process, but also includes the matching results between its state evolution path and the standard state transition relationship in the knowledge graph, thereby fully characterizing the execution effect of the current global scheduling scheme in the digital twin environment and its impact on the road network operation status.

[0074] Optionally, see Figure 3 The semantic vector is semantically corrected based on the feedback semantic feature expression to generate an updated semantic vector. Specifically, this includes: matching the feedback semantic feature expression with the semantic vector to determine the semantic parts in the semantic vector corresponding to abnormal device nodes and state nodes; determining the abnormal change direction based on the deviation between the state change result of the corresponding device node in the feedback semantic feature expression and a preset normal state, wherein the abnormal change direction includes a state increase direction, a state decrease direction, or a state shift direction; dynamically adjusting the weights of the corresponding semantic parts in the semantic vector based on the feedback semantic feature expression to correct the semantic vector, wherein the weight adjustment direction is determined by the abnormal change direction; and inputting the adjusted semantic vector into the decision-making agent for recalculation to obtain an updated semantic vector.

[0075] In this embodiment, taking urban road traffic signal cooperative control as an example, the process of semantic correction processing of semantic vectors based on feedback semantic feature expression is explained. After completing the simulation and generating the feedback semantic feature expression, the system obtains semantic feedback information about the execution effect of the current global scheduling scheme. For example, the feedback semantic feature expression shows that the signal controller of main road A exhibits the phenomenon of "the main road priority is too high, causing the queue at downstream intersection B to continue to accumulate" during execution, while the branch road C exhibits the state change of "the release ratio is too low, causing the traffic capacity of local road sections to decrease". This feedback semantic feature expression also includes the operating state change information at the device node level and the evolution deviation information at the state node level.

[0076] Based on the aforementioned feedback semantic features, they are first matched with the current semantic vector to determine the semantic components in the semantic vector that correspond to the abnormal device nodes and status nodes. For example, semantic segments in the semantic vector such as "main road priority upgrade," "branch road flow control," and "signal cycle extension" are matched with device nodes such as main road A, intersection B, and branch road C involved in the feedback information, thereby identifying the semantic subspace in the current semantic vector that is directly related to the abnormal operation results.

[0077] Furthermore, based on the deviation between the state change results of the corresponding device nodes in the feedback semantic feature expression and the preset normal state, the direction of abnormal changes is determined. For example, in this embodiment, the state of main road A shifts from the preset normal "balanced traffic state" to "excessive priority traffic state", which is a state increase direction, that is, the traffic priority of the main road is too high; intersection B evolves from the normal "stable queuing state" to "continuous queue growth state", which is a state shift direction, that is, shifting from a stable state to a congestion expansion state; and branch road C decreases from the normal "basic traffic state" to "restricted traffic state", which is a state decrease direction, that is, the traffic capacity is excessively compressed.

[0078] After determining the direction of abnormal changes, the weights of the corresponding semantic parts in the semantic vector are dynamically adjusted based on the feedback semantic feature expression. For example, for the semantic part related to the "main road priority control strategy," since its corresponding state shows an upward trend (excessive priority), its weight is reduced; for the semantic part related to the "branch road flow restriction control strategy," since its corresponding state shows a downward trend (excessive capacity restriction), its weight is appropriately increased; for the semantic part related to the "regional coordination control strategy," since its corresponding state shows a deflection direction, it is adjusted with moderate intensity according to the magnitude of the deflection to enhance its constraining effect in subsequent decisions.

[0079] Through the aforementioned dynamic weight adjustment process, the original semantic vector is structurally modified from a state of "prioritizing control of main roads" to a semantic expression of "coordinated and balanced control of main roads and branch roads," thus obtaining an updated semantic vector. Subsequently, the updated semantic vector is re-inputted into the decision-making agent, and the pre-trained semantic decision model is used for further reasoning and calculation. This enables the system to recalculate semantic similarity relationships and optimize the global scheduling scheme based on the corrected semantic state.

[0080] The above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A multi-agent collaborative decision-making system based on digital twins and knowledge graphs, characterized in that, The system is applied to intelligent transportation operation and scheduling scenarios, and the system includes: The data acquisition module is used to acquire multimodal heterogeneous data, and after preprocessing the multimodal heterogeneous data, it is connected to the digital twin base to obtain twin state data. The knowledge graph reasoning module is used to perform semantic reasoning analysis on the twin state data based on a pre-built domain knowledge graph to obtain anomaly detection results and corresponding semantic feature expressions. A perceptual agent is used to generate semantic vectors based on the semantic feature representation; A decision-making agent is used to generate a global scheduling scheme based on the semantic vector; The simulation and deduction module is used to virtually execute and dynamically simulate the global scheduling scheme in the digital twin base to obtain simulation and deduction results; wherein, the simulation and deduction results include equipment operating status change information; If the simulation results meet the preset operational constraints, the global scheduling scheme will be sent to the execution agent. If the simulation results do not meet the preset operating constraints, the simulation results are fed back to the decision agent. The decision agent iteratively optimizes the global scheduling scheme based on the semantic vector and the simulation results, generates a new global scheduling scheme, and inputs it back into the simulation module for verification until the simulation results meet the preset operating constraints. An execution agent is used to receive the global scheduling scheme sent by the simulation and inference module and control the devices corresponding to the global scheduling scheme; The pre-built domain knowledge graph specifically includes: equipment knowledge subgraph, operating state evolution subgraph, and fault mode propagation subgraph; the fault mode propagation subgraph is used to model fault types, fault characteristics, fault root causes, and fault propagation relationships to characterize the fault occurrence process and its impact diffusion path, wherein the fault-related objects correspond to the equipment identifiers; The knowledge graph reasoning module is further configured to map the twin state data to the device knowledge subgraph and the operating state evolution subgraph in the domain knowledge graph to generate semantic alignment relationships between the twin state data and graph entities, and construct an initial state association graph based on the semantic alignment relationships; based on the operating state evolution subgraph, perform state consistency analysis on the current operating state and historical state sequence of the device in the initial state association graph to generate state offset features, and identify whether there are abnormal state changes based on the state offset features; when there are abnormal state changes, perform fault mode matching and propagation path deduction on the state offset features based on the fault mode propagation subgraph to determine the potential fault type and its impact range; Based on the potential fault types and their impact range, an impact range mapping is performed in the device knowledge subgraph based on device identifiers and their topological relationships to determine the target device set where anomalies occur. Based on the degree of deviation between the operating state of each target device in the target device set and the preset normal state, the fault impact degree of the corresponding target device is determined. Anomaly detection results are generated using the target device set and its corresponding fault impact degree, wherein the anomaly detection results include target device identifiers and their corresponding fault impact degree information. The anomaly detection results are mapped to the initial state association graph, and semantic association update processing is performed on the initial state association graph based on device nodes and state nodes to generate semantic feature expressions.

2. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 1, characterized in that, The equipment knowledge subgraph is used to structurally represent equipment types, equipment identifiers, operating parameters, status attributes, and topological relationships between equipment. The operating state evolution sub-graph is used to model the equipment state nodes, state transition relationships, state triggering conditions and operating condition information to characterize the dynamic evolution process of the equipment operating state, wherein the state nodes correspond to the equipment identifier. Among them, the sub-graphs are connected and merged across graphs through device identifiers to form a domain knowledge graph.

3. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 2, characterized in that, The perceptual agent is also used to perform feature encoding processing on the semantic feature expression based on the semantic feature expression to obtain a semantic vector.

4. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 3, characterized in that, The decision-making agent is an agent constructed based on a pre-trained semantic decision-making model; The pre-trained semantic decision model is obtained through supervised learning based on historical operating data. The historical operating data includes historical semantic vector samples and historical global scheduling schemes corresponding to the historical semantic vector samples. The historical semantic vector samples are generated from multi-source heterogeneous data after synchronization by a digital twin base and semantic reasoning of a knowledge graph. The historical global scheduling scheme is used to characterize the equipment control strategies under different operating scenarios and abnormal states. The decision-making agent is used to perform feature mapping on the semantic vector to obtain a semantic embedding representation; Based on the semantic embedding representation, similarity calculation is performed in the semantic feature space composed of historical semantic vector samples to determine the set of historical semantic samples corresponding to the current semantic vector; Based on the historical global scheduling schemes corresponding to the historical semantic sample set, the historical global scheduling schemes are fused by policy weighting to generate a global scheduling scheme.

5. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 4, characterized in that, The decision-making agent is also used to perform reasoning calculations on the semantic vector through the pre-trained semantic decision model to obtain the semantic similarity relationship between the semantic vector and historical semantic vector samples. Based on the semantic similarity relationship, a set of historical semantic vector samples corresponding to the current semantic vector is determined from the historical semantic vector samples; Based on the historical semantic vector sample set, obtain the corresponding historical global scheduling scheme set; Based on the semantic similarity relationship, the set of historical global scheduling schemes is weighted and fused to obtain the fused global scheduling scheme expression; Based on the fused global scheduling scheme expression, a global scheduling scheme is generated.

6. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 5, characterized in that, The preset operating constraints include: the operating parameters and status attributes of each device are within the preset safety threshold range; the device state transition process conforms to the preset state transition relationship and state triggering conditions; and the potential fault impact range does not exceed the preset risk range.

7. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 6, characterized in that, The decision-making agent is also used to map the simulation results to the domain knowledge graph to obtain feedback semantic feature expressions; Based on the feedback semantic feature expression, the semantic vector is semantically corrected to generate an updated semantic vector; The updated semantic vector is input into the decision agent, and the semantic similarity relationship between the updated semantic vector and the historical semantic vector samples is calculated through the pre-trained semantic decision model to determine the corresponding set of historical semantic vector samples. Based on the historical semantic vector sample set, obtain the corresponding historical global scheduling scheme set; Based on the updated semantic vector and the semantic similarity relationship, the historical global scheduling scheme set is weighted and fused to generate a new global scheduling scheme.

8. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 7, characterized in that, The simulation results are mapped to the domain knowledge graph to obtain feedback semantic feature expressions associated with device nodes and state nodes, specifically including: The simulation results are analyzed into data on the changes in the operating status of each device and the state transition results during the simulation process; Based on the device identifier, the data on changes in operating status are associated and matched with device nodes in the device knowledge subgraph of the domain knowledge graph to determine device-level semantic association relationships; Based on the state transition results, the state transition relationships are matched with the state nodes and state transition relationships in the running state evolution sub-graph to determine the state-level semantic association relationships. Based on the device-level semantic association and the state-level semantic association, the feedback semantic feature expression is generated.

9. The multi-agent collaborative decision-making system based on digital twins and knowledge graphs according to claim 8, characterized in that, Based on the feedback semantic feature representation, the semantic vector is semantically corrected to generate an updated semantic vector, specifically including: The feedback semantic features are matched with the semantic vector to determine the semantic part of the semantic vector that corresponds to the abnormal device node and the status node. Based on the deviation relationship between the state change result of the corresponding device node in the feedback semantic feature expression and the preset normal state, the abnormal change direction is determined, wherein the abnormal change direction includes the state increase direction, the state decrease direction or the state offset direction. Based on the feedback semantic feature expression, the weights of the corresponding semantic parts in the semantic vector are dynamically adjusted to correct the semantic vector, wherein the direction of weight adjustment is determined by the direction of abnormal change; The adjusted semantic vector is input into the decision agent for recalculation to obtain an updated semantic vector.

Images