Unplanned operation early warning method and system fusing multi-source data mining

Abstract

This invention discloses a method and system for early warning of unplanned operations based on multi-source data mining. The method collects vehicle trajectory, work plan, and electronic fence data. After cleaning and map matching preprocessing of the trajectories, a trajectory representation learning model is used to map the trajectory sequence into a low-dimensional embedding vector. Based on this vector, abnormal behavior patterns are identified through clustering. Simultaneously, dynamic risk indicators are constructed using real-time data to predict the future state of vehicles. The abnormal patterns, risk indicators, and prediction results are input into a fuzzy logic decision-maker for multi-source fusion, outputting a comprehensive risk level and triggering a tiered early warning. A reinforcement learning mechanism is used to dynamically optimize the decision-maker parameters based on the early warning effect. Finally, the early warning information is pushed to the management platform and feedback is received, forming a closed-loop management system. This invention achieves accurate and adaptive early warning for unplanned operations, significantly improving the intelligent level of operational safety supervision.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and big data analysis technology, specifically relating to a method and system for early warning of unplanned operations that integrates multi-source data mining. Background Technology

[0002] With the rapid development and widespread application of the Internet of Things (IoT) and GPS technologies, vehicle trajectory data is experiencing explosive growth, providing an unprecedented data foundation for the monitoring and management of operational vehicle behavior. In enclosed or semi-enclosed operating areas such as ports, mines, and large construction sites, effective early warning of unplanned vehicle operations is crucial for ensuring operational safety and improving operational efficiency. Unplanned operations, such as vehicles deviating from their predetermined routes, entering unauthorized areas, or operating outside of planned times, can not only lead to resource waste and scheduling chaos but also potentially cause serious safety accidents.

[0003] Currently, monitoring of vehicle operation behavior mainly relies on traditional technologies: one is static area alarms based on geofences, which trigger notifications when a vehicle enters or leaves a preset geofence; the other is simple time comparison based on the work plan, verifying whether the vehicle appears at the designated location within the planned time. However, these methods have significant limitations: firstly, they are mostly reactive alarms, unable to provide early warning and intervention for potential, ongoing unplanned behaviors; secondly, the data between systems are isolated, failing to effectively integrate the spatiotemporal characteristics of trajectories, the business logic of the work plan, and the spatial constraints of the geofence for joint analysis, resulting in low warning accuracy and a high false alarm rate.

[0004] In recent years, although some studies have attempted to improve the intelligence level of trajectory analysis by using data mining and machine learning techniques, such as using clustering algorithms to discover abnormal trajectory patterns or using recurrent neural networks for trajectory representation learning, these methods still fall short when faced with complex operational scenarios. On the one hand, most trajectory similarity measurement methods have high computational complexity, making it difficult to adapt to the real-time processing needs of massive trajectory data. On the other hand, existing methods generally lack the ability to model the unique business rules, vehicle behavioral intentions, and dynamic game relationships between the system and vehicles in operational scenarios, resulting in insufficient accuracy and foresight in their early warning results.

[0005] Therefore, there is an urgent need in this field for an unplanned operation early warning solution that can deeply integrate multi-source data, accurately quantify operational risks, and possess self-learning and adaptive capabilities, in order to overcome the shortcomings of existing technologies and meet the growing demand for intelligent safety supervision. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an early warning scheme for unplanned operations that can deeply integrate multi-source data, accurately quantify operational risks, and has self-learning and adaptive capabilities.

[0007] In a first aspect, embodiments of this application provide a method for early warning of unplanned jobs that integrates multi-source data mining, the method comprising:

[0008] Collect vehicle trajectory data, work plan data, and electronic fence data; perform cleaning, noise reduction, and map matching preprocessing on the trajectory data.

[0009] The preprocessed trajectory sequence is mapped into a low-dimensional trajectory embedding vector using a trajectory representation learning model.

[0010] Based on the low-dimensional trajectory embedding vector, typical operation patterns and abnormal behavior patterns are identified through clustering algorithms;

[0011] Based on real-time trajectory data, work plans, and electronic fence data, dynamic risk indicators are constructed, and time-series prediction models are used to predict the future motion state and position of vehicles.

[0012] The abnormal behavior pattern recognition results, dynamic risk indicators, and vehicle future behavior prediction results are input into the fuzzy logic decision-maker for multi-source information fusion. The fuzzy logic decision-maker outputs the comprehensive risk level of unplanned operations and triggers graded early warning accordingly.

[0013] A reinforcement learning mechanism is adopted to generate a reward signal based on the accuracy and false alarm rate of historical early warning results, and the fuzzy logic decision-maker is dynamically optimized based on the reward signal;

[0014] The early warning information generated based on the optimized decision-making strategy is pushed to the vehicle dispatching and operation management platform, and the platform is received to provide feedback on the handling of the situation, thus forming a closed-loop management system.

[0015] Secondly, embodiments of this application provide an unplanned job early warning system integrating multi-source data mining, applied to the unplanned job early warning method integrating multi-source data mining as described in the first aspect, the system comprising:

[0016] The data acquisition and preprocessing module is used to collect vehicle trajectory data, work plan data and electronic fence data, and to perform cleaning, noise reduction and map matching preprocessing on the trajectory data.

[0017] The trajectory representation learning module is used to map the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector using a trajectory representation learning model.

[0018] The behavior pattern recognition module is used to identify typical work patterns and abnormal behavior patterns based on the low-dimensional trajectory embedding vector through a clustering algorithm.

[0019] The risk prediction module is used to construct dynamic risk indicators based on real-time trajectory data, work plans and electronic fence data, and to use time-series prediction models to predict the future motion state and position of vehicles.

[0020] The multi-source information fusion decision module is used to input the abnormal behavior pattern recognition results, dynamic risk indicators and vehicle future behavior prediction results into the fuzzy logic decision-maker for multi-source information fusion. The fuzzy logic decision-maker outputs the comprehensive risk level of unplanned operations and triggers graded early warning accordingly.

[0021] The decision optimization module is used to generate a reward signal based on the accuracy and false alarm rate of historical early warning results using a reinforcement learning mechanism, and to dynamically optimize the fuzzy logic decision-maker based on the reward signal.

[0022] The early warning execution and feedback module is used to push early warning information generated based on the optimized decision-making strategy to the vehicle dispatching and operation management platform, and receive handling feedback information from the platform to form a closed-loop management.

[0023] Thirdly, embodiments of this application provide an electronic device, including:

[0024] processor;

[0025] Memory used to store processor-executable instructions;

[0026] The processor is configured to implement the unplanned job early warning method as described in the first aspect when executing the instructions.

[0027] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the unplanned job early warning method based on multi-source data mining as described in the first aspect.

[0028] Compared with existing technologies, this invention has the following significant advantages: it realizes the transformation from post-event alarm to pre-event warning, significantly improving the timeliness of warnings; through multi-source data fusion and deep analysis, it greatly improves the accuracy of warnings; it has self-learning and adaptive capabilities, enabling continuous optimization of warning performance; it supports cross-regional knowledge sharing, promoting the collective intelligent evolution of the system; and it is applicable to complex and ever-changing actual operating environments, possessing wide applicability and good scalability. This invention effectively solves the problems of warning lag, high false alarm rate, and poor adaptability existing in existing unplanned operation warning systems, providing strong technical support for the safety supervision of operating vehicles. Attached Figure Description

[0029] Figure 1 This is a schematic flowchart of an unplanned job early warning method that integrates multi-source data mining, provided as an embodiment of this application.

[0030] Figure 2 The architecture diagram of the unplanned job early warning system that integrates multi-source data mining provided in this application.

[0031] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0033] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0034] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Example 1

[0036] Figure 1 This is a schematic flowchart of an unplanned job early warning method that integrates multi-source data mining, provided as an embodiment of this application. Figure 1 As shown, a method for early warning of unplanned jobs that integrates multi-source data mining includes:

[0037] Step 1: Collect vehicle trajectory data, work plan data, and electronic fence data. Perform cleaning, noise reduction, and map matching preprocessing on the trajectory data.

[0038] Specifically, vehicle trajectory data refers to a sequence of data recorded by devices such as GPS / BeiDou, containing timestamps and latitude / longitude coordinates, used to describe the vehicle's movement path. Work plan data refers to pre-defined task arrangements, including planned work vehicles, work areas, and start and end times. Electronic fence data refers to virtual boundaries defined by geographic coordinates, used to define restricted areas, danger zones, or specific work areas. By cleaning to remove obvious location drift points (such as a point suddenly jumping into a river), smoothing trajectories through noise reduction, and correcting the original GPS points to the actual road network through map matching, clean and accurate data is provided for subsequent analysis.

[0039] Step 2: Use a trajectory representation learning model to map the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector.

[0040] Specifically, trajectory representation learning models refer to a type of deep learning model. Low-dimensional trajectory embedding vectors are fixed-length numerical vectors output by the model. This vector is a digital fingerprint or feature encoding of the trajectory. It compresses and abstracts complex trajectory sequences into a mathematical vector rich in semantic information. This vector can capture the shape, direction, and pattern of the trajectory, allowing computers to efficiently measure the similarity between trajectories by calculating the distance between vectors, such as Euclidean distance.

[0041] Specifically, in this embodiment, the step of mapping the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector using a trajectory representation learning model includes: adaptively selecting a model branch based on the spatial characteristics of the trajectory. Here, spatial characteristics refer to whether the environment in which the vehicle moves is open free space or a road network space constrained by a road network. Free space includes places like ports, mining areas, and construction sites, where vehicles can move relatively freely without a fixed path. For example, the navigation trajectory of a ship in a port area. Road network space includes urban roads and highways, where vehicle movement is restricted to the road network. For example, the driving trajectory of a taxi in a city. This is the model's scheduling center. Its function is to perform scene judgment and assign the most suitable feature extractor to different types of trajectories to achieve accurate modeling.

[0042] For free-space trajectories, a parallel model AG-TSM based on Gated Recurrent Units (GRUs) and a self-attention mechanism is used for encoding. The GRU module captures the continuous spatial features of the trajectory, while the self-attention mechanism focuses on the key structural features. The parallel model (AG-TSM) is a hybrid model specifically designed for free-space trajectories. The Gated Recurrent Unit (GRU) captures the continuity and temporal dependencies of the trajectory. The GRU acts like a memory unit, processing each point in the trajectory sequentially, understanding the continuous process of vehicle movement, such as speed changes and basic direction. It excels at learning patterns of movement from one point to the next. For example, in a port, the trajectory of a gantry crane might include long periods of straight-line movement and short turns. The GRU can effectively model such continuous movement patterns.

[0043] For trajectories within the road network constraint space, the RAG-TSM model, based on a graph neural network (GNN), is used for encoding. This model integrates road network topology and vehicle trajectory sequence information through structure-aware and location-aware learning to generate road network-aware trajectory embedding vectors. The graph neural network model (RAG-TSM) is specifically designed for road network spatial trajectories. GNNs are used to learn the road network as a graph structure. In this graph, road intersections are nodes, and road segments are edges. The core operation of GNNs is message passing, where each node aggregates information from its neighbors to update its representation. Structure awareness refers to GNNs learning the local connectivity relationships of nodes within the road network. Location awareness means the model also considers the absolute or relative geographical location of nodes within the entire road network, preventing two nodes with similar topologies but geographically distant locations from generating the same representation. This branch utilizes GNNs to model the underlying road network topology, ensuring that the generated trajectory vectors not only contain the sequence information of the trajectory points themselves but also deeply encapsulate the constraint information of the road network. For example, it understands that making a U-turn on a highway is a highly unusual behavior, while it is relatively common in urban grid roads.

[0044] The final output of the parallel model or graph neural network model is aggregated into a single, fixed-dimensional low-dimensional trajectory embedding vector through a global pooling layer.

[0045] Through the above steps, for trajectories in free space, AG-TSM is used to understand them simultaneously from two dimensions: temporal flow GRU and spatial structure self-attention. For trajectories in road network space, RAG-TSM is used to interpret the trajectory more accurately by understanding its inherent structure, namely the road network GNN. This targeted design enables the trajectory representation learning model to generate high-quality, information-rich trajectory embedding vectors in different scenarios, laying a solid technical foundation for subsequent anomaly detection and risk warning.

[0046] Step 3: Based on the low-dimensional trajectory embedding vector, a clustering algorithm is used to identify typical operation patterns and abnormal behavior patterns.

[0047] Specifically, clustering algorithms are unsupervised machine learning methods, such as DBSCAN. Typical work patterns refer to the dense, large clusters formed after clustering, representing common and compliant behaviors. For example, driving from area A to area B along a main road. Anomalous behavior patterns refer to the sparse, small clusters or outliers formed after clustering, representing rare and suspicious behaviors. It unsupervisedly discovers inherent behavioral patterns in data and automatically identifies anomalous segments that deviate from the norm. It does not require pre-defining what constitutes an anomaly; instead, it finds anomalies through the distribution of the data itself.

[0048] Specifically, in this embodiment, the step of identifying typical work patterns and abnormal behavior patterns based on the low-dimensional trajectory embedding vector using a clustering algorithm includes:

[0049] A density-based clustering algorithm is used to initially cluster the low-dimensional trajectory embedding vectors to identify high-density core behavioral pattern clusters and low-density outliers. The density-based clustering algorithm refers to algorithms like DBSCAN. Its core idea is to group high-density areas (areas where points are closely clustered) into clusters, while considering points in low-density areas as outliers. Core behavioral pattern clusters refer to large groups of tightly clustered points formed after clustering. These represent common, compliant, and reproducible operational behaviors. Outliers are isolated trajectory vectors that do not belong to any cluster after clustering. They represent rare and unusual behaviors. Suspicious objects are quickly and automatically filtered from massive trajectories. This step only focuses on the distribution of trajectory vectors in the vector space; those that do not belong to any cluster are marked first.

[0050] The trajectory segments corresponding to the outliers are then subjected to a secondary matching analysis with preset electronic fence areas and work plan time windows. The trajectory segments corresponding to the outliers refer to the original trajectories or trajectory segments represented by the anomalous vectors identified in the previous step. The electronic fence area refers to the restricted areas, danger zones, or designated work zones pre-defined on the electronic map. The work plan time window refers to when a specific task should occur as specified in the plan. Suspicious objects are then further examined. This step places the initially screened anomalous trajectories back into a real business scenario, using spatial rule-based electronic fences and time rule-based work plans as benchmarks to evaluate them.

[0051] If a trajectory segment is identified as an outlier, and its spatial extent overlaps with any electronic fence area exceeding a first threshold, or its occurrence time falls outside the scheduled work window, then the behavior pattern represented by that trajectory segment is determined to be an abnormal unplanned work mode. This step establishes clear anomaly determination logic; only when a trajectory simultaneously meets the criteria of an outlier and violates spatial or temporal rules is it formally determined to be an abnormal unplanned work mode.

[0052] The above steps implement an intelligent process from coarse screening to precise judgment through a three-step method: Step 1: Clustering: Utilizing machine learning, all unique candidates are efficiently and unbiasedly identified from massive amounts of data. Step 2: Matching: Linking technical discoveries with business knowledge fences and plans to provide a basis for judgment. Step 3: Judgment: Combining the results of the first two steps, explicit business rules are applied to accurately identify the actual violations. This method greatly reduces the false positives that pure clustering may cause and the false negatives that pure rule-based systems may cause, achieving high-precision early warning for unplanned operations.

[0053] Step four: Based on real-time trajectory data, work plans, and electronic fence data, construct dynamic risk indicators and use time-series prediction models to predict the future motion state and position of vehicles.

[0054] Specifically, dynamic risk indicators are a series of quantified risk values, such as path intrusion index, time conflict index, and area intrusion index. Time series prediction models are models used to predict future sequences, such as Seq2Seq, a recurrent neural network based on an encoder-decoder architecture. The focus has shifted from past tense analysis to future tense early warning. It not only assesses the risk of the vehicle's current state but, more importantly, predicts the vehicle's location and state in the next minute, thereby determining whether there is a trend towards unplanned operations.

[0055] Specifically, in this embodiment, the construction of dynamic risk indicators based on real-time trajectory data, work plans, and electronic fence data, and the prediction of the future motion state and position of vehicles using a time-series prediction model, specifically includes:

[0056] Part 1: Construction of Dynamic Risk Indicators. This part defines three core risk quantification indicators, which are measured from three dimensions: spatial, temporal, and access-based.

[0057] The path intrusion index is calculated as a function of the lateral distance from the vehicle's current position to its planned operating path and the difference between its current heading angle and the expected heading angle towards the planned path. The planned operating path refers to the ideal or prescribed route pre-set for the vehicle. The lateral distance is the vertical distance from the vehicle's current position to the planned path. The heading angle difference is the angular difference between the vehicle's current direction of travel and the expected direction of travel on the planned path at that point. This quantifies the risk of the vehicle deviating from its planned route. It considers not only how far it has deviated but also whether the vehicle intends to return to its original path. A vehicle that is far from its planned path and has deviated from it is at higher risk than a vehicle that is far from its planned path but is turning back.

[0058] The time conflict index is calculated as the absolute value of the difference between the estimated arrival time of a vehicle at the next critical task node and the planned time, weighted by the vehicle's current speed and the remaining distance between nodes. Critical task nodes are important time or location points in the plan, such as arriving at the loading area or leaving the unloading area. The estimated arrival time is calculated based on the vehicle's current speed, location, and remaining path. This step quantifies the risk of task schedule disruptions and is used to determine whether a vehicle can complete its planned task on time, preventing the subsequent task chain from stalling due to the delay of one vehicle.

[0059] The system calculates an intrusion index based on the relationship between the vehicle's predicted future location and the geofence boundary. If the predicted trajectory point enters an unauthorized geofence area, an alarm is triggered. The predicted future location of the vehicle is the position of the vehicle at a series of future time points, output by the time-series prediction model described below. An unauthorized geofence area is a geofence where the vehicle has not been authorized to enter. This enables proactive intrusion warnings. This is the most critical function; instead of waiting for the vehicle to intrude before triggering an alarm, it predicts the tendency and possibility of intrusion before it occurs, buying valuable time for intervention.

[0060] Part Two: Operation of the Time-Series Prediction Model. The vehicle's historical trajectory sequence, including position, speed, and heading angle, is input into a recurrent neural network based on an encoder-decoder architecture. The encoder encodes the historical sequence, and the decoder, based on this encoded state, autoregressively predicts the vehicle's motion state and position sequence within a predetermined future time period. The encoder-decoder architecture recurrent neural network can be a Seq2Seq model with LSTM or GRU. The encoder is responsible for understanding and compressing the historical sequence. The decoder is responsible for generating the future sequence. When generating the future sequence, the decoder uses the prediction output of the previous time step as the input for the next time step, progressively extrapolating the future.

[0061] Step 5: Input the abnormal behavior pattern recognition results, dynamic risk indicators and vehicle future behavior prediction results into the fuzzy logic decision-maker for multi-source information fusion. The fuzzy logic decision-maker outputs the comprehensive risk level of unplanned operations and triggers a graded early warning accordingly.

[0062] Specifically, the fuzzy logic decision-maker is an intelligent decision-making system that handles uncertainty and fuzzy concepts. It doesn't simply judge yes / no, but rather assesses risk as high / medium / low. It comprehensively analyzes various pieces of evidence obtained in previous steps, such as abnormal behavior patterns, risk indicators, and future predictions, ultimately outputting an easily understandable comprehensive risk level, such as low, medium, or high, and determining whether to trigger an early warning and the level of warning.

[0063] Specifically, in this embodiment, in step five, the fuzzy logic decision-maker performs multi-source information fusion, which is implemented through a dynamic game-enhanced digital twin reasoning framework. The operation of this framework includes:

[0064] S1: Constructing a digital twin of the work scenario. Based on real-time vehicle trajectory data, work plans, and electronic fences, a digital twin scenario containing a vehicle dynamic model, work rule constraints, and an environmental semantic map is synchronously constructed in virtual space.

[0065] Specifically, a digital twin is a virtual digital model that perfectly corresponds to the work site, vehicles, and rules in the physical world. It is a dynamic, computable virtual copy. The vehicle dynamic model describes the vehicle's physical characteristics, such as acceleration, braking, and cornering performance. Work rule constraints refer to the business logic defined in the virtual world, such as one-way streets, speed limits, and loading zone priorities. The environmental semantic map contains not only geographical information but also semantic information, such as "this is a danger zone," "this is a highway," and "this is an intersection." A virtual platform for advanced analysis and simulation is created. It integrates scattered real-time data trajectories, plans, and fences into a unified, computable virtual environment, providing a stage for subsequent behavioral extrapolation and game simulation.

[0066] S2: Perform intent-driven multimodal behavior inference within the digital twin. In the digital twin, the low-dimensional trajectory embedding vector is input to a pre-trained behavior decoder to generate the vehicle's potential operational intent. Based on this intent, and combined with the temporal prediction model, multiple possible future behavior trajectory sequences are inferred in parallel within the twin space. The behavior decoder is a pre-trained neural network model whose function is to interpret the intent behind the trajectory vector. Based on the inferred intent, multiple, rather than a single, possible future behavior trajectories are simulated in the digital twin. This step translates the low-dimensional trajectory vector—a behavioral feature—into a specific behavioral intent, and based on this intent, combined with the prediction model, multiple future behavior paths are simulated in parallel within the digital twin.

[0067] S3: Conflict Detection and Risk Assessment Based on Game Theory. This involves dynamically simulating the vehicle's deduced behavioral sequence against a standardized operating procedure model defined in a digital twin, calculating the strategy utility value between the vehicle's behavior and the standardized procedure. The standardized operating procedure model refers to the ideal, safe standard operating procedures defined in the digital twin. Dynamic game simulation involves adversarially simulating the vehicle's deduced behavioral sequence against the standard procedure to see how they interact. The strategy utility value, calculated through game simulation, quantifies the degree of alignment or conflict between the vehicle's behavior and the system's expectations. It recreates the future in a virtual environment, quantifying conflict risk. It no longer statically compares indicators but dynamically simulates the interaction between vehicle behavior and system rules, thus more accurately assessing the consequences and risk levels of behavior. For example, the system simulates the deduced behavioral sequence of entering the oncoming lane to overtake against the standardized procedure. The simulation finds that this behavior will cause a severe conflict with an oncoming vehicle as predicted by the rules, thus calculating a very low strategy utility value and high conflict risk.

[0068] S4: The fuzzy logic decision-maker uses the policy utility value as its core input, integrates other risk indicators, assesses whether the behavior is cooperative, neutral, or adversarial, and outputs a comprehensive risk level accordingly. This is the decision endpoint of the entire framework. The fuzzy logic decision-maker integrates the core evidence policy utility value generated by the game simulation with all other risk indicators such as path intrusion index and area intrusion index, ultimately outputting an easily understandable, comprehensive risk level such as low, medium, or high.

[0069] Specifically, the steps for conflict detection and risk assessment based on game theory are implemented using a hierarchical stochastic game model, including:

[0070] A first-layer signaling game is constructed, where the potential operational intent is treated as a private signal sent by the vehicle to the system. The system updates its prior beliefs about the vehicle's behavior type based on this signal. Signaling game is a game theory model characterized by information asymmetry: one vehicle possesses private information about its true intent and sends signals to the other system through its behavior. The potential operational intent is the vehicle's true purpose, such as compliant operation, fault repair, or intentional violation. This is the vehicle's private information, which the system cannot directly know. The private signal is the behavioral pattern exhibited by the vehicle through its actual trajectory, which is considered a signal about its true intent transmitted to the system. Prior beliefs are the initial probability estimates of the vehicle's behavior type made by the system before observing a specific signal. The role of this layer is to allow the system to infer the vehicle's true intent, continuously updating and revising its judgment of the vehicle's private intent type by observing the vehicle's behavioral signals.

[0071] A second-layer stochastic game is constructed, modeling the interaction between the vehicle and the system as a sequential decision-making process within the time domain of the digital twin's deduction. A stochastic game is a dynamic game model where participants, the vehicle and the system, make sequential decisions across a series of states. Each decision affects the next state and yields a corresponding payoff. The sequential decision-making process involves the vehicle and system making decisions based on the current state at each step of the deduction, leading to state transitions and the next round of decisions. The supervisory strategy refers to the measures the system can take in each state, such as simply observing, issuing a warning, or notifying the dispatcher. In the digital twin, the future multi-step interaction process between the vehicle and the system is dynamically deduced. This layer extends a single conflict judgment into a dynamic decision-making process that considers the impact of multiple future steps. For example, in the deduction, the current state is that the vehicle is heading towards a danger zone. The vehicle's action space might be {accelerate, maintain speed, turn away}. The system's action space might be {observe, warn}. Choosing different action combinations will lead to different next states, such as approaching the danger zone more closely or moving away, and yielding different immediate payoffs.

[0072] Payoff Function and State Transition Design: The payoff function is designed to grant positive payoffs when vehicle actions align with its stated operational intent and plan, and impose high negative payoffs when deceptive behavior occurs. State transition probabilities are jointly determined by the vehicle dynamics model and the uncertainty of the operational environment. The payoff function is a mathematical function used to quantify the benefits or penalties gained by the vehicle and system after taking a certain combination of actions in a given state. Deceptive behavior refers to a vehicle's actual actions being inconsistent with the intention signals it conveyed in the first-level signaling game. For example, a vehicle conveys a compliance signal but accelerates into a danger zone. Rules and objectives are set for the game. The payoff function defines what constitutes good behavior that grants positive payoffs and what constitutes bad behavior that incurs negative payoffs. The state transition mechanism defines the uncertainty of the game environment.

[0073] Optimal Strategy Solving and Risk Quantification: The Monte Carlo Tree Search algorithm is used to solve the Bayesian-Nash equilibrium of the stochastic game online. By calculating the deviation between the actual deduced path and the equilibrium strategy, and combining this with the final expected payoff of the path, the overall risk level is quantified. The Bayesian-Nash equilibrium represents the optimal combination of strategies adopted by both vehicles and the system, assuming that the constantly updated beliefs originate from the first layer. It represents the most likely rational behavioral path given the information. Monte Carlo Tree Search is an algorithm used to efficiently find near-optimal solutions in complex decision spaces. Deviation refers to the degree of difference between the vehicle's actual deduced behavioral path and the theoretically most rational behavioral path equilibrium strategy. Finding the most likely behavioral path and quantifying the risk through comparison is the decision output of the entire model. The system quantifies risk by calculating the deviation: the greater the deviation from rational behavior, the more abnormal and unpredictable the vehicle's behavior, and the higher the risk.

[0074] The hierarchical stochastic game model in this step implements a sophisticated computational process from intention inference to behavior deduction and risk quantification: The first layer is signal game, which addresses the question of what the vehicle might intend to do, dynamically updating the judgment of vehicle type. The second layer is stochastic game, which addresses the question of what will happen in the future if this happens, simulating multi-step dynamic interactions. The final solution: by comparing actual behavior with rational expectations, the degree of abnormality and risk level of the behavior is accurately quantified. This makes risk assessment no longer based on simple rule thresholds, but on a complex cognitive model that considers motivation, future, and rationality, greatly improving the system's intelligence and judgment accuracy.

[0075] Furthermore, the hierarchical stochastic game model is extended into a meta-game framework with strategy evolution capabilities, which achieves the system's self-evolution and adversarial adaptation in the following ways:

[0076] Part 1: Constructing the Policy Population and Meta-Game Layer. In the digital twin, a diverse population of vehicle behavior policies is maintained, including typical policies mined from historical data and adversarial policies synthesized through a policy generation network. Above the existing hierarchical stochastic game, a meta-game layer is introduced. This layer treats the various Nash equilibria of the lower-level games as higher-order policies and evaluates the robustness and generalization ability of these equilibria when dealing with the entire policy population.

[0077] The strategy population is a policy library containing a wide variety of vehicle behavior strategies. It includes not only common strategies learned from historical data, such as compliant driving and regular overtaking, but also edge case strategies and adversarial strategies synthesized by the strategy generation network that may not yet exist in reality or are difficult to identify. The meta-game layer introduces a higher-level game on top of the existing hierarchical stochastic game. Instead of analyzing single interactions, it treats the multiple Nash equilibrium solutions of the lower-level game—that is, the optimal regulatory strategy of the system in different scenarios—as new pieces to evaluate which regulatory strategy performs most stably and reliably when facing the ever-changing strategy population. This step provides the system with a rich pool of hypothetical adversaries and establishes a training ground for evaluating regulatory strategies. The strategy population simulates all types of adversary vehicles that may be encountered, while the meta-game layer is used to select and train the all-around champion regulatory strategy best suited to deal with these adversaries.

[0078] For example, the strategy population includes strategies such as: strategy A (normal following), strategy B (aggressive lane-cutting), strategy C (pretending to comply and then suddenly violating regulations), and strategy D (slowly infiltrating the edge by exploiting sensor blind spots). Meta-game evaluation: The meta-game layer tests the current system's regulatory strategies X and Y. It finds that strategy X is effective against A and B, but can be deceived by C and D; while strategy Y has good defensive effects against A, B, C, and D. Therefore, strategy Y has stronger robustness and generalization ability.

[0079] Part Two: Establishing a Dynamic Environment and Co-evolution Mechanism. The environmental parameters of the digital twin are dynamically adjusted based on historical game results, simulating changes in work plans, sudden electronic fence deployments, and weather and road condition changes; the system's monitoring strategy and vehicle behavior strategy form a co-evolution relationship in this dynamic environment.

[0080] In this context, the dynamic environment refers to the fact that environmental parameters of the digital twin, such as work plans, geofence settings, and road conditions, are no longer fixed but dynamically adjusted based on historical game results to simulate changes in the real world. The co-evolutionary relationship is an arms race between the system's regulatory strategy and vehicle behavior strategies. The system updates its strategy to suppress newly emerging anomalous behaviors; the behavioral strategy population evolves accordingly, attempting to find loopholes in the new regulatory strategy. Simulating a constantly changing real world drives the system and its adversaries to co-evolve. This ensures that the system does not become obsolete in a static environment and can adapt to new operational rules and unforeseen circumstances. For example, historically, vehicles often illegally travel from point A to point B. After the system upgrades its regulatory strategy, it blocks this path. As a result, new adversarial strategies emerge in the strategy population, such as detouring from point A to point C to reach point B, and their adaptability increases. In the next round of training, the system needs to learn how to identify and warn of this new path.

[0081] Part Three: Implementing Early Risk Warning Based on Group Effects. The fuzzy logic decision-maker introduces a group distribution index; when a novel adversarial strategy is found to have a rapidly increasing fitness in the strategy population at the meta-game layer, a targeted early strategic warning is triggered; the system periodically uses an evolutionary algorithm to conduct offline co-evolution training on the strategy population and the supervisory strategy, and updates the most robust supervisory strategy set obtained from the training online into the hierarchical random game model.

[0082] The population distribution index doesn't just look at the behavior of a single vehicle, but focuses on the popularity and fitness of a particular adversarial strategy within the entire strategy population. Early strategic warning involves issuing a strategic-level alert when a new and effective adversarial strategy begins to gain popularity in the virtual environment, even if it's rare in the real world. This step achieves strategic-level threat warning and proactive defense. This allows the system to detect threats before a new type of violation breaks out on a large scale and update its defense strategies in advance. For example, in offline co-evolution training at the meta-game layer, the system discovered a new strategy that exploits low visibility during rain or snow to commit violations, with its fitness rapidly increasing from 5% to 30% in the strategy population. The system immediately triggers an early strategic warning, indicating: "A new violation pattern that may exploit severe weather has been detected; it is recommended to strengthen relevant regulatory rules in advance." The system then uses an evolutionary algorithm to train a regulatory strategy that can counter this new strategy and updates it online into the running hierarchical stochastic game model. When rain or snow actually occurs, the system is already prepared.

[0083] The meta-game framework in this step infuses the entire early warning system with evolutionary and adaptive capabilities. The strategy population provides diverse hypothetical adversaries. The meta-game layer serves as the arena for selecting the strongest commanders. The dynamic environment simulates a constantly changing battlefield. The co-existence mechanism drives a relentless arms race. Early warning achieves a leap from tactical to strategic early warning. Ultimately, this framework transforms the system from a static expert system requiring manual rule updates into an adaptive immune system capable of autonomously identifying threats, training, and upgrading itself, thus enabling long-term and effective responses to unknown future security challenges.

[0084] Step six involves employing a reinforcement learning mechanism to generate reward signals based on the accuracy and false alarm rates of historical early warning results. These reward signals are then used to dynamically optimize the fuzzy logic decision-maker. Through this reinforcement learning mechanism, the system achieves self-learning and continuous improvement. It automatically adjusts the parameters and rules within the fuzzy logic decision-maker based on historical feedback reward signals from actual early warning effects, resulting in increasingly accurate future decisions and reduced false alarms and missed alarms.

[0085] Specifically, in step six, a reinforcement learning mechanism is used for dynamic optimization, specifically through a hierarchical reinforcement learning framework based on curriculum learning and adversarial reward shaping. This framework includes:

[0086] Part 1: Hierarchical Policy Optimization Structure. This includes a bottom-level policy network responsible for fine-tuning the membership function parameters in the fuzzy logic decision-maker, with its action space being the continuous adjustment of the parameters; and an upper-level meta-policy network responsible for selecting and weighting the fuzzy inference rules in the fuzzy logic decision-maker, with its action space being the discrete decision-making on the rule base structure.

[0087] The underlying policy network is a neural network responsible for fine-tuning. Its task is to precisely adjust the parameters of the membership function in the fuzzy logic decision-maker. In fuzzy logic, the membership function defines the degree to which an input value, such as a path intrusion index of 0.7, belongs to a certain fuzzy concept, such as high risk. Adjusting its parameters adjusts the tightness of the risk assessment. The upper-level meta-policy network is a neural network responsible for strategic decision-making. Its task is to select or combine the most suitable fuzzy inference rules from a rule base and determine their weights. Fuzzy inference rules are the logical core of the decision-maker, typically in the form of IF, THEN. Decomposing the complex optimization problem into two levels with different granularities enables more accurate and stable learning. The underlying network performs continuous, fine-grained parameter adjustments, while the upper-level network performs discrete, macroscopic rule structure adjustments.

[0088] Part Two: Adversarial Reward Shaping Mechanism. A discriminator network is introduced, aiming to distinguish between warning decisions generated by the system policy and ideal decisions labeled by experts. The reinforcement learning agent receives adversarial reward signals from this discriminator network in addition to environmental reward signals. The discriminator network is a model similar to the discriminator in a Generative Adversarial Network (GAN). Its goal is to learn to distinguish between warning decisions generated by the current system policy and ideal decisions labeled by experts. This provides the reinforcement learning agent with richer and higher-quality learning signals. Besides the rewards from the environment itself, such as bonus points for accurate warnings, this mechanism provides an additional guidance signal that mimics experts, guiding the system's decision-making style to continuously align with human experts and avoiding learning strange tactics that, while obtaining environmental rewards, are illogical.

[0089] Part Three: Course-Driven Training Process. A training course is constructed, progressing from simple to complex. In the initial stage, training is conducted on high-determinism scenarios in historical data, gradually increasing to difficult scenarios that include complex adversarial behavior and noisy data. At each stage of the course, a reinforcement learning algorithm based on maximum entropy is used for training.

[0090] Among these, curriculum-based learning is a training strategy that mimics the human learning process, progressing from easy to difficult. Maximum entropy-based reinforcement learning algorithms, such as SAC, are advanced algorithms that encourage agents to explore. Their core idea is to maintain the stochastic entropy of the policy as much as possible while pursuing high rewards, avoiding premature entrapment in local optima. This addresses the problems of unstable and inefficient reinforcement learning training. By designing courses from simple to complex, the agent first builds a solid foundation before tackling more challenging problems, making the training process smoother and more efficient. The maximum entropy algorithm ensures that all possibilities are fully explored during the learning process, thus increasing the chance of finding the globally optimal policy.

[0091] For example, in the initial stage, training is conducted on high-risk or low-risk, high-certainty scenarios selected from historical data that are distinctive, undisputed, and allow the system to learn to walk. In the advanced stage, complex and challenging scenarios that are difficult to judge, contain noisy data, or even involve adversarial behavior are gradually introduced, allowing the system to learn to run and cope with emergencies. In the exploration and safeguard stage, throughout the process, the maximum entropy algorithm encourages the system to occasionally try some unconventional parameter adjustments or rule selections to prevent it from handling all problems in a rigid way.

[0092] A hierarchical reinforcement learning framework is used to design a sophisticated autonomous driving optimization system for the core decision-maker: the hierarchical structure enables refined and stable parameter learning. Adversarial rewards introduce expert experience, accelerating the learning process and improving decision quality. Curriculum learning ensures the robustness and efficiency of the training process. This allows the fuzzy logic decision-maker to autonomously and continuously learn from historical experience, becoming increasingly intelligent and accurate, ultimately forming a self-improving intelligent decision-making core that requires minimal human intervention.

[0093] Step seven involves pushing the early warning information generated based on the optimized decision-making strategy to the vehicle dispatching and operation management platform, and receiving handling feedback information from the platform to form a closed-loop management system. This ensures that each early warning and intervention by the system becomes a measurable, attributable, and reusable knowledge unit, and through intelligent distribution and federated learning, achieves optimal allocation of management resources and cross-system collective evolution.

[0094] Specifically, in step seven, the formation of closed-loop management is achieved through an intelligent feedback execution system based on causal inference and maximizing impact. This system includes:

[0095] The causal effect assessment and strategy attribution module: After pushing out a warning, the system continuously tracks the changes in the vehicle's behavior trajectory after receiving the warning and constructs a counterfactual reasoning model. This model quantifies the net causal effect of the warning intervention on the vehicle's behavior and attributes the behavioral change to the contribution of a specific component in the decision-making system. The counterfactual reasoning model is an advanced model for causal inference. Its core question is: what would have happened to the vehicle's behavior if the warning hadn't been issued? The net causal effect is the change in vehicle behavior caused purely by the warning, after removing other factors. Contribution refers to attributing the merit or fault of the behavioral change to a specific component in the decision-making system, such as a fuzzy logic decision-maker or a game theory model. This module accurately assesses the true effect of the warning and identifies specific directions for system optimization. It solves the problem that traditional systems can only see whether an alarm has been issued but cannot assess whether the alarm was effective. For example, if the system issues a warning to a vehicle approaching a danger zone, and the vehicle subsequently turns away, the model infers that without the warning, there was a 90% probability that the vehicle would have entered the danger zone. Therefore, the net causal effect of this warning is that it prevented the intrusion, demonstrating a significant effect. Further analysis revealed that this accurate early warning was primarily due to the game-theoretic reasoning framework successfully identifying the vehicle's potential intentions. This conclusion was recorded to strengthen the decision-making weight of this component.

[0096] The adaptive distribution and impact maximization module dynamically calculates the optimal distribution path based on the risk level of the warning, the vehicle sensitivity inferred through causal inference, and the real-time load of each node on the dispatch platform, using an influence propagation model. It prioritizes routing warning information to key dispatch nodes most likely to trigger effective intervention. Vehicle sensitivity refers to the degree of a vehicle's response to a warning, derived from historical causal inference. Some vehicles respond, while others ignore it. The influence propagation model is a model that simulates how information spreads and has an impact in social or organizational networks. Here, it is used to simulate how warning information is transmitted within the dispatch platform and ultimately triggers intervention. Key dispatch nodes refer to the dispatchers or management units most likely, most authorized, or most efficient in handling the warning. This achieves accurate delivery of warning information and optimal allocation of management resources. It ensures that the most important warnings are delivered to the most appropriate handlers with the highest priority, thereby maximizing the actual intervention effect of the warning. For example, if the system generates a high-risk warning, it will not broadcast it to all dispatchers. The influence propagation model calculates that dispatcher A is handling other tasks, and history shows that he responds slowly to this type of warning. Dispatcher B is currently available and is responsible for this area. Historically, he has demonstrated extremely high efficiency in handling such alerts, and the target vehicles are highly sensitive to his instructions. The system prioritizes routing the alert information to Dispatcher B's terminal, thereby triggering intervention as quickly and effectively as possible.

[0097] Cross-platform knowledge synchronization and model federated updates: After receiving feedback from the scheduling platform, the system encapsulates the anonymized intervention effect-causal effect data pairs into a knowledge package; while ensuring privacy, the knowledge package is then federated and synchronized with similar systems deployed in other operational areas.

[0098] The knowledge package is a structured data packet containing anonymized information about the entire chain from early warning content to intervention actions, net causal effects, and component attribution. Federated knowledge synchronization is a distributed machine learning technique. Multiple systems, such as early warning systems deployed in different ports or mining areas, exchange learned knowledge (model parameters or knowledge packages) without sharing local data, collectively improving model performance. This step breaks down data silos, enabling collective intelligent evolution where experience from one source benefits all, while strictly protecting privacy and trade secrets. For example, the system deployed at Tianjin Port discovered a new pattern of illegal operations utilizing tidal time and verified effective strategies to address it through closed-loop management. The system encapsulated this experience into a knowledge package without specific vehicle or company information. This knowledge package was synchronized with similar systems at Shanghai and Qingdao Ports via a federated learning platform. Upon receiving this knowledge package, the systems at Shanghai and Qingdao Ports, although having never encountered such violations before, had their internal models already acquired the ability to identify and respond to this new pattern, achieving cross-regional collaborative awareness of security situations.

[0099] The intelligent feedback execution system elevates closed-loop management to a whole new level: from simply providing feedback to truly understanding it—through causal inference, it truly comprehends the value of each warning; from casting a wide net to targeting precisely—by maximizing impact, it ensures warnings translate into concrete actions; and from self-learning to co-evolution—through federated synchronization, it achieves collective experience growth across regions and scenarios. Ultimately, this transforms the entire system from a standalone tool into a network of interconnected, wisdom-sharing, and continuously evolving collective intelligence, significantly enhancing its long-term value and applicability.

[0100] Example 2

[0101] like Figure 2 As shown, this application provides an architecture diagram of an unplanned operation early warning system that integrates multi-source data mining, applied to the unplanned operation early warning system that integrates multi-source data mining as described in Embodiment 1, including: a data acquisition and preprocessing module 210, a trajectory representation learning module 220, a behavior pattern recognition module 230, a risk prediction module 240, a multi-source information fusion decision module 250, a decision optimization module 260, and an early warning execution and feedback module 270.

[0102] The data acquisition and preprocessing module 210 is used to acquire vehicle trajectory data, work plan data and electronic fence data, and to perform cleaning, noise reduction and map matching preprocessing on the trajectory data.

[0103] The trajectory representation learning module 220 is used to map the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector using a trajectory representation learning model.

[0104] The behavior pattern recognition module 230 is used to identify typical work patterns and abnormal behavior patterns based on the low-dimensional trajectory embedding vector through a clustering algorithm.

[0105] The risk prediction module 240 is used to construct dynamic risk indicators based on real-time trajectory data, work plans and electronic fence data, and to use time-series prediction models to predict the future motion state and position of vehicles.

[0106] The multi-source information fusion decision module 250 is used to input the abnormal behavior pattern recognition results, dynamic risk indicators and vehicle future behavior prediction results into the fuzzy logic decision-maker for multi-source information fusion. The fuzzy logic decision-maker outputs the comprehensive risk level of unplanned operations and triggers graded early warning accordingly.

[0107] The decision optimization module 260 is used to generate a reward signal based on the accuracy and false alarm rate of historical early warning results using a reinforcement learning mechanism, and to dynamically optimize the fuzzy logic decision-maker based on the reward signal.

[0108] The early warning execution and feedback module 270 is used to push early warning information generated based on the optimized decision-making strategy to the vehicle dispatching and operation management platform, and receive handling feedback information from the platform to form a closed-loop management.

[0109] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0110] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0111] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0112] In one embodiment of this application, the program operating in the electronic device may be a program that controls the Central Processing Unit (CPU) to achieve the functions of the above-described embodiments of the present invention, thus enabling the computer to function. Information processed by these systems is then temporarily stored in Random Access Memory (RAM) during processing, and subsequently stored in various ROMs such as Read Only Memory (ROM) and Flash ROM, or in Hard Disk Drives (HDDs). It is then read, corrected, and written by the CPU as needed.

[0113] It should be noted that a portion of the electronic device described in the above embodiments can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0114] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0115] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0116] Furthermore, the electronic device in the above embodiments can also be implemented as a system group composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0117] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for early warning of unscheduled work by fusing multi-source data mining, characterized in that, The method comprises the following steps: Collecting vehicle trajectory data, work plan data and electronic fence data, and performing cleaning, denoising and map matching preprocessing on the trajectory data; Using a trajectory representation learning model to map the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector; Based on the low-dimensional trajectory embedding vector, a clustering algorithm is used to identify typical work modes and abnormal behavior patterns; Based on real-time trajectory data, work plans and electronic fence data, dynamic risk indicators are constructed, and a time series prediction model is used to predict the future motion state and position of the vehicle, including: Constructing a path intrusion index based on the lateral distance from the current position of the vehicle to the planned work path and the difference between the current heading angle and the expected heading angle; Constructing a time conflict index based on the absolute value of the difference between the time at which the vehicle is expected to arrive at the next key work node and the planned time, combined with the current speed of the vehicle and the remaining distance between nodes; Constructing a region intrusion index, which triggers an alarm when the vehicle's future predicted position enters an unauthorized electronic fence region; Inputting a historical trajectory sequence containing position, speed and heading angle into a recurrent neural network based on an encoder-decoder architecture, encoding the historical sequence through the encoder, and the decoder predicting the motion state and position sequence of the vehicle in the future preset time period based on the encoded state; The abnormal behavior pattern recognition result, dynamic risk indicator and vehicle future behavior prediction result are input into a fuzzy logic decision maker for multi-source information fusion, and the fuzzy logic decision maker outputs the comprehensive risk level of unplanned work, and triggers a hierarchical warning accordingly; Using a reinforcement learning mechanism to generate a reward signal based on the accuracy and false alarm of historical warning results, and dynamically optimizing the fuzzy logic decision maker based on the reward signal; The warning information generated based on the optimized decision strategy is pushed to the vehicle dispatching and work management platform, and feedback information from the platform is received, forming a closed-loop management. 2.The unplanned job pre-warning method of fusing multi-source data mining according to claim 1, wherein, The step of using a trajectory representation learning model to map the preprocessed trajectory sequence into a low-dimensional trajectory embedding vector comprises: Adaptively selecting a model branch according to the spatial characteristics of the trajectory; For trajectories in free space, a parallel model based on a gated recurrent unit and a self-attention mechanism is used for encoding, and the output of the parallel model is globally pooled to form the low-dimensional trajectory embedding vector; For trajectories in road network constrained space, a graph neural network model is used for encoding, and the output of the graph neural network model is globally pooled to form the low-dimensional trajectory embedding vector. 3.The unplanned job pre-warning method of fusing multi-source data mining according to claim 1, characterized in that, The step of identifying typical work modes and abnormal behavior patterns based on the low-dimensional trajectory embedding vector using a clustering algorithm comprises: Using a density-based clustering algorithm to initially cluster the low-dimensional trajectory embedding vector to identify high-density core behavior pattern clusters and low-density outliers; Matching and analyzing the trajectory segments corresponding to the outliers with the preset electronic fence regions and work plan time windows; When a trajectory segment is identified as an outlier, and its spatial range overlaps with any electronic fence region beyond a first threshold, or its occurrence time is outside the operation plan time window, the trajectory segment is determined as an abnormal unplanned operation mode. 4.The unplanned job pre-warning method of fusing multi-source data mining according to claim 1, wherein, The step of inputting the abnormal behavior mode identification result, dynamic risk indicator and vehicle future behavior prediction result into a fuzzy logic decision maker for multi-source information fusion is implemented through a digital twin reasoning framework enhanced by dynamic game, and includes the following steps: A digital twin of the operation scene is constructed, and a digital twin scene including a vehicle dynamic model, operation rule constraints and an environment semantic map is constructed in a virtual space based on real-time vehicle trajectory data, operation plans and electronic fences; Intention-driven multi-modal behavior deduction is performed in the digital twin, the low-dimensional trajectory embedding vector is input into a pre-trained behavior decoder to generate a potential operation intention, and based on the intention, a time series prediction model is used to deduce a plurality of possible behavior trajectory sequences in the twin space in parallel; Conflict detection and risk assessment are performed based on game theory, the behavior sequence deduced by the vehicle is dynamically simulated with a standardized operation process model defined in the digital twin through game theory, and a strategy utility value between the vehicle behavior and the standardized process is calculated; The strategy utility value is taken as a core input, combined with other risk indicators, and the behavior attribute is evaluated and a comprehensive risk level is output through the fuzzy logic decision maker.

5. The method of claim 4, wherein, The step of conflict detection and risk assessment based on game theory is implemented through a hierarchical stochastic game model, including the following steps: A first layer of signaling game is constructed, the potential operation intention is taken as a private signal sent by the vehicle to the system, and the system updates the prior belief of the vehicle behavior type according to the signal; A second layer of stochastic game is constructed, and the interaction between the vehicle and the system is modeled as a sequential decision process within the deduction time domain of the digital twin; A reward function is designed, which gives a positive reward when the vehicle action conforms to its operation intention and plan, and applies a negative reward when fraudulent behavior occurs; A state transition probability is designed, which is determined by the vehicle dynamics model and the uncertainty of the operation environment; A Monte Carlo tree search algorithm is used to solve the Bayesian-Nash equilibrium of the stochastic game online; The deviation degree of the actual deduced path from the equilibrium strategy is calculated, and the comprehensive risk level is quantified by combining the final expected reward of the path. 6.The unplanned work pre-warning method of fusing multi-source data mining according to claim 5, wherein, The hierarchical stochastic game model is extended to a meta-game framework with strategy evolution capability, and the system evolution and antagonistic adaptation are realized through the following ways: A diversified vehicle behavior strategy population is maintained in the digital twin, including typical strategies mined from historical data and antagonistic strategies synthesized by a strategy generation network; A meta-game layer is introduced on top of the hierarchical stochastic game, taking the Nash equilibrium solution of the lower layer game as a high-order strategy, and evaluating the robustness and generalization ability of these equilibrium solutions to the entire strategy population; The environment parameters of the digital twin are dynamically adjusted according to the historical game results, simulating operation plan changes, sudden electronic fence settings and environmental changes; The system regulation strategy and the vehicle behavior strategy form a co-evolution relationship in the dynamic environment. When the meta-game layer detects that the fitness of a new type of adversarial strategy in the strategy population is rapidly rising, trigger targeted early strategic early warning; Periodically use evolutionary algorithms to perform offline co-evolution training on the strategy population and regulatory strategies, and update the most robust regulatory strategy set obtained by training to the hierarchical stochastic game model online.

7. The method of claim 1, wherein, The step of dynamically optimizing the fuzzy logic decision maker using a reinforcement learning mechanism is implemented through a hierarchical reinforcement learning framework based on curriculum learning and adversarial reward shaping, including: A bottom-layer strategy network is constructed to fine-tune the membership function parameters in the fuzzy logic decision maker, and its action space is the continuous adjustment amount of the parameters; An upper-layer meta-strategy network is constructed to select and weight the fuzzy inference rules in the fuzzy logic decision maker, and its action space is the discrete decision on the rule base structure; A discriminator network is introduced to distinguish between system warning decisions and ideal decisions labeled by experts, providing an adversarial reward signal for the reinforcement learning agent; A training curriculum from simple to complex is constructed, with the initial stage training in a high-deterministic historical data scenario, and gradually increasing to a difficult scenario containing complex adversarial behavior and noisy data; At each stage of curriculum learning, a maximum entropy-based reinforcement learning algorithm is used for training to maintain the exploratory and diversity of the strategy. 8.The unplanned job pre-warning method of fusing multi-source data mining according to claim 1, wherein, The step of pushing the warning information generated based on the optimized decision strategy to the vehicle dispatching and operation management platform, and receiving disposal feedback information from the platform to form a closed-loop management, is implemented through an intelligent feedback execution system based on causal inference and impact maximization, including: After pushing the warning information, the vehicle behavior trajectory changes are continuously tracked, and an counterfactual reasoning model is constructed to quantify the net causal effect of the warning intervention on the vehicle behavior; Attributing the behavior change to the contribution of a specific component in the decision system; According to the risk level of the warning, the vehicle sensitivity inferred by the causal inference, and the real-time load of the dispatching platform node, the optimal distribution path is dynamically calculated through an influence propagation model; The warning information is preferentially routed to the key dispatching node that is most likely to trigger effective intervention actions; After receiving the disposal feedback from the dispatching platform, the desensitized intervention effect-cause and effect data pair is packaged as a knowledge package; On the premise of ensuring privacy, the knowledge package is synchronized with the same system deployed in other operation areas in a federated manner.

9. A non-scheduled work early warning system fusing multi-source data mining, applied to the non-scheduled work early warning method fusing multi-source data mining according to any one of claims 1 to 8, characterized in that, The system comprises: A data acquisition and preprocessing module for acquiring vehicle trajectory data, operation plan data and electronic fence data, and preprocessing the trajectory data by cleaning, denoising and map matching; A trajectory representation learning module for mapping the preprocessed trajectory sequence to a low-dimensional trajectory embedding vector using a trajectory representation learning model; An abnormal behavior recognition module for identifying typical operation modes and abnormal behavior patterns based on the low-dimensional trajectory embedding vector through a clustering algorithm; A risk prediction module for constructing dynamic risk indicators based on real-time trajectory data, operation plans and electronic fence data, and predicting the future motion state and position of the vehicle using a time series prediction model, including: A path invasion index is constructed based on the lateral distance from the current position of the vehicle to the planned work path and the difference between the current heading angle and the expected heading angle; A time conflict index is constructed based on the absolute value of the difference between the time at which the vehicle is expected to arrive at the next key work node and the planned time, combined with the current speed of the vehicle and the remaining distance between nodes; A region invasion index is constructed to trigger an alarm when the vehicle enters an unauthorized electronic fence region in the future predicted position; A history trajectory sequence including position, speed and heading angle is input into a recurrent neural network based on an encoder-decoder architecture, the history sequence is encoded by the encoder, and the decoder predicts the motion state and position sequence of the vehicle in the future preset time period based on the encoded state; A multi-source information fusion decision module is used to input the abnormal behavior pattern recognition result, dynamic risk index and vehicle future behavior prediction result into a fuzzy logic decision maker for multi-source information fusion, and the fuzzy logic decision maker outputs the comprehensive risk level of the unplanned work, and triggers a graded warning accordingly; A decision optimization module is used to adopt a reinforcement learning mechanism to generate a reward signal according to the accuracy and false alarm of the historical warning result, and to dynamically optimize the fuzzy logic decision maker based on the reward signal; A warning execution and feedback module is used to push the warning information generated based on the optimized decision strategy to the vehicle scheduling and work management platform, and to receive the disposal feedback information from the platform, forming a closed-loop management.

Images