Highway digital intelligent operation monitoring system

Abstract

The invention discloses a digital intelligent operation monitoring system for an expressway. The system comprises a data acquisition module, a feature extraction and fusion module, a space-time analysis and prediction module, an anomaly detection module and an intelligent decision module. Wherein the feature extraction and fusion module fuses multi-source heterogeneous data by using a collaborative mechanism of a graph neural network and an auto-encoder; the space-time analysis and prediction module generates a prediction result containing a congestion diffusion probability by using a Fourier neural operator, and provides reference data for anomaly detection; and the intelligent decision-making module inputs the prediction result as an environment model of a reinforcement learning agent, and dynamically adjusts a reward function according to the predicted congestion diffusion probability, so as to realize prospective autonomous decision-making. According to the invention, through deep cooperation and feedback closed loop among the modules, the traffic operation efficiency and the safety emergency capability are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technologies of intelligent transportation and artificial intelligence, and in particular to a digital intelligent operation and monitoring system for highways that integrates multiple artificial intelligence technologies such as graph neural networks, autoencoders, Fourier neural operators, reinforcement learning, and hybrid expert models. Background Art

[0002] The mileage of expressways continues to grow, and the road network structure is becoming increasingly complex. However, traditional expressway operation monitoring models face significant challenges. First, tidal traffic flows during holidays can easily cause large-scale, long-distance congestion. Existing manual dispatch and passive response measures are often delayed and inefficient. Second, frequent inclement weather such as fog, ice, and snow have become a major cause of serious traffic accidents, making traditional monitoring methods unable to provide accurate and early warning of such disasters.

[0003] The root of these challenges lies in the fact that although existing technologies have been applied in some aspects, such as using graph neural networks (GNNs) for traffic flow prediction and using reinforcement learning (RL) for signal optimization, their application in the specific scenario of highway operation and management is often scattered and isolated, failing to form a complete solution for collaborative operation, resulting in inherent defects that are difficult to overcome.

[0004] These deficiencies are primarily manifested in the following ways: First, there are "information silos" at the data level. Data from disparate sources and formats, such as ETC gantries, weather stations, and video surveillance, is difficult to effectively integrate, resulting in a lack of a holistic perspective in decision-making analysis. Second, the prediction layer lacks foresight. Systems often remain limited to describing the current state, unable to accurately predict potential future congestion or risks caused by severe weather, leading to passive and delayed decision-making. Finally, the decision-making layer lacks intelligence. Whether it is daily traffic management or emergency response to traffic accidents, it still relies heavily on static rules or manual experience, making it impossible to achieve dynamic, optimal coordinated control. Therefore, a next-generation intelligent monitoring architecture is urgently needed that can deeply integrate multiple technologies to form a closed loop of "perception-prediction-decision-making" to fundamentally improve the operational efficiency and safety emergency response capabilities of highways. Summary of the Invention

[0005] In view of this, an embodiment of the present application provides a digital intelligent operation monitoring system for a highway, which at least partially solves the problems existing in the prior art.

[0006] An embodiment of the present application provides a digital intelligent operation and monitoring system for highways, characterized by comprising: a data acquisition module for acquiring multi-source heterogeneous data including at least traffic flow data, roadside equipment status data, and environmental data of a highway network; a feature extraction and fusion module for receiving the multi-source heterogeneous data, modeling the topology of the traffic network using a graph neural network to generate topological structure features, aligning and compressing the multi-source heterogeneous data using an autoencoder to generate dynamic content features, and fusing the topological structure features with the dynamic content features to generate traffic status features; a spatiotemporal analysis and prediction module for receiving the traffic status features and processing them using a Fourier neural operator to generate a traffic status prediction result including a congestion diffusion probability and roadside equipment status data; an anomaly detection module for receiving the traffic status features and determining whether the equipment status deviates from the normal operating range based on the roadside equipment status data to generate an anomaly alarm; and an intelligent decision-making module for using the traffic status prediction result as an environmental model input for a reinforcement learning agent, wherein the reinforcement learning agent autonomously learns and generates a traffic control strategy based on the traffic status prediction result.

[0007] Optionally, the feature extraction and fusion module fuses the topological structure feature and the dynamic content feature in the following manner: The topological structure features are embedded into the feature reconstruction path of the autoencoder as a priori structure in the form of an adjacency matrix to guide the dynamic content features to preserve road connectivity and turning relationships during compression and reconstruction.

[0008] Optionally, the reward function of the reinforcement learning agent in the intelligent decision-making module is dynamically adjusted in combination with the congestion diffusion probability in the traffic state prediction result to guide the traffic control strategy to be optimized towards high traffic efficiency.

[0009] Optionally, the anomaly detection module further detects an abnormal event based on at least one of the following judgment criteria: The reconstruction error from the autoencoder in the feature extraction and fusion module; or The prediction deviation between the traffic state prediction result from the spatiotemporal analysis and prediction module and the actual observation value.

[0010] Optionally, the spatiotemporal analysis and prediction module is further configured to: Receive external spatiotemporal constraints including geographic information data, time features, and event calendars; Based on the traffic state characteristics and through a gated fusion unit, the fusion weight of the external spatiotemporal constraints is dynamically adjusted, and the weighted external spatiotemporal constraints are fused with the traffic state characteristics to generate enhanced fusion features containing multi-source context information for subsequent spatiotemporal prediction.

[0011] Optionally, the spatiotemporal analysis and prediction module is configured to receive a feedback signal as a dynamic intervention signal to perform real-time correction on the traffic status prediction result; the feedback signal includes the traffic control strategy generated by the intelligent decision-making module and / or the abnormality alarm generated by the abnormality detection module.

[0012] Optionally, the anomaly detection module further adopts a hybrid expert model to classify the abnormal events, wherein the hybrid expert model includes a gating network and multiple expert networks: The gating network uses a multi-layer perceptron and a softmax output mechanism to perform nonlinear transformation on the traffic state features to output an expert activation probability vector; The plurality of expert networks are dynamically selected according to the expert activation probability vector to perform abnormal event classification, and the plurality of expert networks include at least four of the following: Equipment failure expert network, whose input includes roadside equipment signal loss rate and communication interruption duration; The passenger flow surge expert network, whose input includes the passenger flow growth value per unit time; Congestion situation expert network, whose input includes the traffic growth value of the section and the upstream queue length; A network of severe weather experts whose input includes environmental data such as visibility, slippery or icy road conditions, etc. Traffic accident expert network, whose input includes instantaneous vehicle speed reduction rate and spatial propagation range.

[0013] Optionally, the intelligent decision-making module further adopts a hybrid expert model as a strategy routing, which is used to select and activate a matching control strategy from a strategy pool including conventional control strategies for handling daily traffic diversion and emergency response strategies for dealing with major abnormal events based on the abnormal alarm or the traffic status prediction results.

[0014] Optionally, the intelligent decision-making module uses a decision engine based on multi-agent deep reinforcement learning to generate the traffic control strategy, and the decision engine includes multiple agents, wherein each agent is responsible for an independent traffic section or area, and the action space of the agent covers at least one of adjusting variable speed limit values, opening and closing entrance ramps, issuing traffic guidance information, or dispatching emergency rescue resources.

[0015] Optionally, the system further includes an application layer module, and the application layer module is configured to: receiving the traffic control strategy from the intelligent decision-making module and visually presenting it on a three-dimensional digital twin system; A human-machine collaborative interface is provided, which includes at least one of a real-time monitoring screen for globally displaying the operating status of the expressway network, a road network management and control platform for issuing and tracking traffic control instructions, or an emergency management platform for rapid response.

[0016] The embodiments of this application only relate to a digital intelligent highway operation monitoring system. This system can be executed by a general-purpose electronic device equipped with a processor and memory, or it can be deployed as a computer program product on a server, workstation, or terminal device. When the processor on the device executes the program code stored in the memory, the various steps described in this application are completed.

[0017] The digital intelligent operation and monitoring system for highways provided in the embodiment of the present application effectively solves the problems of data silos, delayed decision-making, and inefficient emergency response in the prior art, and has significant beneficial effects.

[0018] Specifically, by building a collaborative "fusion-prediction-decision-making" architecture, this system can deeply integrate multi-source heterogeneous data from highway networks, breaking down "information silos." By utilizing Fourier neural operators and reinforcement learning models, it enables forward-looking predictions of traffic congestion and proactive intelligent control, transforming passive response into active intervention. Furthermore, by introducing a parallel hybrid expert model, it can accurately and rapidly identify and classify different types of emergencies, such as traffic accidents and severe weather, and collaboratively generate intelligent emergency response strategies, significantly improving highway operational efficiency and safety response capabilities.

[0019] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0021] Figure 1 The core module flow chart of a digital intelligent operation and monitoring system for highways provided in an embodiment of the present application is schematically shown.

[0022] Figure 2 The hierarchical architecture and data flow diagram of a digital intelligent operation and monitoring system for a highway provided in an embodiment of the present application are schematically shown. DETAILED DESCRIPTION

[0023] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0024] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0025] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will appreciate that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement an apparatus and / or practice a method. Additionally, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this apparatus and / or practice this method.

[0026] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0027] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0028] The present disclosure provides a digital intelligent highway operation and monitoring system. The system provided in this embodiment can be executed by a computing device, which can be implemented as software or a combination of software and hardware, and can be integrated into a server, terminal device, or the like.

[0029] It should be noted that the system and method proposed in the present invention can be applied to a variety of traffic operation scenarios. The following will be explained using highway operation scenarios and urban public transportation operation scenarios as examples, but this does not constitute a limitation on the scope of application of the present invention.

[0030] See also Figure 1 , schematically illustrates a core module flow diagram of a highway digital intelligent operation monitoring system provided by an embodiment of the present application. In a preferred embodiment, the system may include: Data acquisition module 100. The data acquisition module 100 performs operations to acquire multi-source heterogeneous data of a traffic network, wherein the multi-source heterogeneous data at least includes traffic flow data of a highway network, roadside equipment status data, and environmental data.

[0031] It's important to note that this step is the cornerstone of the entire system, aiming to comprehensively perceive the real-time status of the transportation network from multiple dimensions, providing high-quality basic data support for subsequent steps such as feature extraction, spatiotemporal analysis and prediction, and intelligent decision-making. This multi-source, heterogeneous data can be acquired through various methods, including but not limited to various sensors deployed in the transportation network, business system databases, and third-party APIs.

[0032] In a specific embodiment, the obtained multi-source heterogeneous data is specifically composed of the following: Traffic flow data collection for highway networks: Using equipment deployed at key locations such as roads and intersections, including but not limited to cameras, radar, and geomagnetic coils, real-time data is collected on dynamic traffic flow, average speed, vehicle type distribution, and lane occupancy. This data also includes ETC gantries' cross-sectional flow, average speed within each section, and vehicle type data; video surveillance reveals lane occupancy and queue lengths; and ramp coil flow data. It's important to note that this data directly reflects the real-time state of traffic flow and serves as the core input for subsequent spatiotemporal analysis and prediction modules.

[0033] Roadside Equipment Status Data Collection: This system monitors the operating status of traffic infrastructure hardware, including but not limited to traffic lights, cameras, and sensors, to determine whether the equipment is functioning properly and whether data transmission is stable. This includes the online status and display content of variable message boards, the status of camera video streams, and the power supply and network connectivity of weather stations. This data is crucial for ensuring the stable operation of the entire monitoring system and serves as a crucial basis for the subsequent anomaly detection module to identify equipment failures.

[0034] Environmental Data Collection: Deployed weather stations, road condition sensors, and other equipment provide real-time access to external environmental information that significantly impacts traffic safety and efficiency, including weather conditions, visibility, road temperature, humidity, icing coefficients (collected by road sensors), and slippery conditions. It's important to note that these factors, such as slippery roads caused by rain or snow, can directly impact the safety parameters considered by the system's decision-making model.

[0035] The feature extraction and fusion module 200 receives the multi-source heterogeneous data, models the topological structure of the traffic network through a graph neural network to generate topological structure features; uses an autoencoder to align and compress the multi-source heterogeneous data to generate dynamic content features; and fuses the topological structure features with the dynamic content features to generate traffic status features.

[0036] It should be noted that this step aims to address the core "data silo" problem in existing technologies by converting the data from multiple sources and formats acquired in step 101 into high-quality, standardized feature representations that can be understood and used by subsequent artificial intelligence (AI) models. In a preferred embodiment, the workflow of this module can be divided into three closely connected stages: data preprocessing, multi-model collaborative feature extraction, and cross-modal feature fusion.

[0037] In a more specific embodiment, GNN is used for spatial feature extraction. The system first abstracts the highway network into a graph structure, in which toll booths, interchange hubs, service areas, etc. serve as nodes of the graph and highway sections serve as edges. Through a GNN model (such as GCN or GraphSAGE), the features of each node and its neighboring nodes (such as flow rate and vehicle speed) are aggregated to effectively capture the spatial dependencies of traffic flow in the road network topology and generate topological structure features. Autoencoders are used to compress time series and state features. For example, a long short-term memory network autoencoder (LSTM-Autoencoder) can be designed to compress the continuous one-hour vehicle speed time series data collected by a roadside camera into a low-dimensional dynamic content feature vector while retaining its core time series change pattern.

[0038] Optionally, before the multi-model collaborative feature extraction, the module first performs a data preprocessing phase. It should be noted that this preprocessing phase is intended to ensure the quality and consistency of the input data. Specifically, it may include: (1) cleaning and standardization, such as using Z-Score standardization combined with dynamic threshold filtering to eliminate sensor noise (e.g., geomagnetic vehicle detector error ≤ 5%) and unify the dimensions of each indicator; (2) spatiotemporal alignment, such as using the GPS clock synchronization protocol to achieve timestamp alignment of multi-source data (error < 100ms) and resolve spatial coordinate differences between different sensors through GIS projection transformation.

[0039] Optionally, the feature extraction and fusion module uses a deep coupling architecture of graph neural network (GNN) and autoencoder to perform multi-model collaborative feature extraction. It should be noted that this collaborative architecture is intended to capture the core information of the spatial and temporal dimensions in traffic data respectively. In a more specific embodiment, GNN is used for spatial feature extraction. The system constructs a knowledge graph of the traffic network, abstracts intersections, road sections, etc. into nodes, and converts traffic flow, traffic light status, etc. into edge weights. Preferably, the GraphSAGE sampling strategy is adopted to enable each node to aggregate its multi-layer (for example, 3-layer) neighbor information to effectively capture complex spatial dependencies, thereby generating the topological structure features. Experiments show that the model used in the present invention can improve the accuracy of spatial feature extraction by 18.7% compared to traditional CNN in the task of road congestion prediction. In a more specific embodiment, the autoencoder is used for time series and state feature compression. For example, a long short-term memory autoencoder (LSTM-Autoencoder) architecture can be designed to compress historical time series data (e.g., the past seven days) into a lower-dimensional latent space (e.g., 16 dimensions), while keeping the decoder's reconstruction error low (e.g., MSE < 0.03). Alternatively, an attention mechanism can be introduced to automatically increase feature weights for key periods such as morning and evening rush hours. Furthermore, a parallel variational autoencoder (VAE) can be used to specifically process device log data, achieving a probabilistic representation of device health status through KL divergence constraints. Its fault prediction F1-score can reach 0.89, thereby generating the aforementioned dynamic content features.

[0040] Optionally, the feature extraction and fusion module fuses the topological structure features with the dynamic content features by embedding the topological structure features as a prior structure in the form of an adjacency matrix into the feature reconstruction path of the autoencoder. It should be noted that this cross-modal feature fusion stage aims to achieve semantic alignment of multi-source features. The traffic topology structure constructed by the GNN is embedded as a prior structure input in the form of an adjacency matrix into the feature reconstruction path of the autoencoder. The technical principle is to inject topological information representing the global spatial structure into the learning process of the dynamic content features, thereby guiding the dynamic content features to preserve road connectivity and turning relationships during compression and reconstruction. In a more preferred embodiment, a cross-attention mechanism can be used to explicitly establish the correlation matrix between the spatial features extracted by the GNN and the temporal features extracted by the autoencoder. In addition, a gated recurrent unit (GRU) can be used to dynamically adjust the fusion weight of spatial and temporal features based on real-time traffic status trends (for example, in congested scenarios, the weight of temporal features can be increased to 0.7).

[0041] Ultimately, the module outputs a traffic state feature tensor (e.g., a 128-dimensional tensor) that integrates multi-level semantic information. This tensor can further include: base layer information (such as lane-level traffic flow), pattern layer information (such as encoding of periodic congestion patterns), and context layer information (such as device health index).

[0042] The spatiotemporal analysis and prediction module 300 receives the traffic status features and processes them using a Fourier neural operator to generate a traffic status prediction result including a congestion diffusion probability and roadside equipment status data.

[0043] It should be noted that this step is a prerequisite for implementing the "active intelligent" decision-making of the present invention and is responsible for forward-looking analysis. In a preferred embodiment, the input of this module presents multimodal features, specifically including: the traffic state features from the feature extraction and fusion module (specifically, they can be in the form of a structured feature tensor, such as a 128-dimensional tensor containing spatial, temporal, and contextual features); external spatiotemporal constraints; and dynamic intervention signals.

[0044] Optionally, the spatiotemporal analysis and prediction module is further configured to receive external spatiotemporal constraints, including geographic information data, temporal features, and an event calendar; dynamically adjust the fusion weights of the external spatiotemporal constraints based on the traffic state features through a gated fusion unit; and fuse the weighted external spatiotemporal constraints with the traffic state features to generate enhanced fused features containing multi-source contextual information for subsequent spatiotemporal prediction. It should be noted that this optional feature is intended to enable the prediction model to intelligently understand the complex context of the real world. The fusion process preferably includes three steps: vector space alignment, spatiotemporal dimension expansion, and dynamic feature fusion. For example, different types of external constraints (such as categorical data such as "road grade" and continuous data such as "speed limit") are mapped to the same vector dimensions as traffic flow features through embedding and fully connected layers. The dimensions are then unified through broadcasting and concatenation operations. Finally, the gated fusion unit dynamically and nonlinearly fuses the external information based on real-time traffic conditions, significantly improving prediction accuracy and robustness.

[0045] Optionally, the system also includes a feedback loop, wherein the spatiotemporal analysis and prediction module is further configured to receive feedback signals from the intelligent decision-making module and warning information from the anomaly detection module as the dynamic intervention signal to perform real-time correction on the traffic status prediction result.

[0046] In a more specific embodiment, the module's workflow can adopt a three-stage process: "spatiotemporal decomposition - feature encoding - prediction generation." In the spatiotemporal data decomposition stage, the spatial dimension can be decomposed based on the road network topology using graph convolutional networks (GCNs) or one-dimensional convolutional neural networks (1D-CNNs) to separately process arterial and ramp flows. Time series can then be decomposed into trend, fluctuation, and residual terms using methods such as wavelet transforms to capture long-term patterns and short-term shocks. In the feature encoding and enhancement stage, a spatiotemporal joint attention mechanism can be designed to calculate the association weights between spatial and temporal dimensions using the QKV matrix, and interpretability constraints can be introduced to enhance the confidence of model decisions. In the prediction generation and optimization stage, the Fourier Neural Operator (FNO) is the core method for processing the fused features and the dynamic intervention signal. FNO treats traffic flow as the solution of a partial differential equation (PDE) and efficiently computes it in the frequency domain using Fourier transforms. It can be designed as a hierarchical operator network consisting of Fourier transform layers, spectral filters, and inverse Fourier transform layers. Multi-task learning can also be used at this stage to not only predict traffic flow but also assist in predicting indicators such as congestion duration and queue length. Simultaneously, methods such as Monte Carlo Dropout can be used to quantify uncertainty and provide confidence assessments for the prediction results.

[0047] The module's resulting traffic status predictions include spatiotemporal predictions (e.g., lane-level traffic flow) and key indicator predictions (e.g., average speed error <3 km / h) for multiple future time windows (e.g., 15, 30, and 60 minutes), along with congestion diffusion probability. The generated roadside equipment status data can include benchmarks for the normal operating range of key equipment, such as track wear, signal system reliability, and charging station availability.

[0048] The anomaly detection module 500 receives the traffic status characteristics and, based on the roadside equipment status data, determines whether the equipment status deviates from the normal operating range to generate an anomaly alarm. It should be noted that this step operates in parallel with the spatiotemporal analysis and prediction described in step 103 and is responsible for real-time monitoring of the system's health status. In a preferred embodiment, the input data of the anomaly detection module exhibits multi-source heterogeneous features, and may specifically include: the traffic status prediction results and the roadside equipment status data output by the spatiotemporal analysis and prediction module; the traffic status characteristics and real-time streaming data output by the feature extraction and fusion module; and meteorological data and event data obtained through an external data interface.

[0049] In a more specific embodiment, the workflow of this module can be divided into three stages: anomaly identification, classification and risk assessment, and output: Abnormal identification stage: This stage aims to accurately identify abnormal signals from massive data. Optionally, the abnormality detection module further detects abnormal events based on at least one of the following judgment criteria: The reconstruction error of the autoencoder in the feature extraction and fusion module; or the prediction deviation between the traffic state prediction result and the actual observation value from the spatiotemporal analysis and prediction module. It should be noted that this feature reflects the high reuse and collaborative verification of information within the system. When the reconstruction error of the autoencoder exceeds the preset threshold, it means that the current input data deviates from the normal data distribution learned by the model; when the prediction deviation of the Fourier neural operator exceeds the dynamic threshold, it means that the actual evolution of the system deviates from the predicted normal trajectory. Both of these "deviation" signals can serve as powerful anomaly indicators. In a more preferred embodiment, it can also be based on the detection of graph structure. Specifically, the graph neural network is used to continuously monitor the changes in the topological structure of the traffic network to identify dynamic patterns such as congestion propagation and fault diffusion. For example, when the traffic volume on a certain section of road suddenly increases, and the traffic volume on its adjacent sections also shows a coordinated upward trend, the system can determine that it is a congestion propagation event.

[0050] Classification and risk assessment stage: After an anomaly is identified, this stage is responsible for conducting an in-depth analysis of the nature and severity of the anomaly. Optionally, the anomaly detection module further uses a mixture of experts (MoE) model to classify the abnormal events. It should be noted that the technical principle of this "expert consultation" internal structure is to decompose the complex full set classification task into multiple simpler and more focused subtasks. In a preferred embodiment, the hybrid expert model includes a gated network and multiple expert networks. The gated network extracts key discriminant features through nonlinear transformations (such as ReLU activation) and outputs a weight vector consistent with the number of experts to dynamically select the most matching expert network to perform fine classification. The expert network can be specially trained to identify specific types of anomalies.

[0051] In a preferred embodiment, the hybrid expert model (MoE) comprises a gating network and multiple parallel expert networks. The gating network receives fused traffic state features and outputs a weight vector, which is used to dynamically select the most appropriate expert network for the current scenario. These multiple expert networks include at least the following four types: An equipment failure expert network: This network focuses on identifying issues with the roadside equipment itself. For example, if the received equipment status data for a particular ETC gantry indicates that its "communication interruption duration" exceeds a preset threshold (e.g., 5 minutes), the network is activated and outputs a "device offline" alarm.

[0052] Congestion Situation Expert Network: This network identifies congestion caused by the natural evolution of traffic flow. Its inputs include the cross-sectional flow rate growth rate for a specific road segment and the length of the upstream vehicle queue. If it detects a sustained high flow rate increase and a queue length exceeding 1 kilometer, the network outputs a "Flow-Type Congestion" event.

[0053] Severe Weather Expert Network: This network focuses on identifying risks posed by sudden environmental changes. For example, if the received environmental data indicates that visibility on a certain road section is less than 50 meters, or the road icing coefficient exceeds a dangerous threshold, the network is activated and outputs a "fog risk" or "road icing" alert.

[0054] Traffic Accident Expert Network: This network is used to identify incidents caused by abnormal vehicle behavior. Its inputs can be instantaneous vehicle speed and location data. If it detects that the speeds of multiple vehicles on a road section drop suddenly from 100 km / h to 0 within a short period of time (e.g., within 5 seconds), forming a spatial cluster, the network will output a "suspected traffic accident" alert.

[0055] Output phase: This module ultimately generates the anomaly alert and can further output the analysis results in a structured format. In a preferred embodiment, this output may include: a list of anomaly events with event descriptions, evidence chains, and recommended actions; and a risk assessment report containing a risk matrix, trend forecasts, and optimization recommendations. This output information provides key input for the intelligent decision-making module and application layer.

[0056] The intelligent decision-making module 400 executes the operation of using the traffic state prediction result as the environmental model input of the reinforcement learning agent, and enabling the reinforcement learning agent to autonomously learn and generate a traffic control strategy based on the traffic state prediction result.

[0057] It should be noted that this step is performed by the system's "brain"—the intelligent decision-making module 400—which is responsible for generating the final control strategy. In a preferred embodiment, the module's input data exhibits multimodal characteristics, specifically including: traffic status predictions and equipment status predictions from the spatiotemporal analysis and prediction module 300; abnormal event lists and risk level assessments from the anomaly detection module 500; and external resource constraints, such as capacity configuration parameters and a policy rule engine. The module's workflow can be divided into phases of situational awareness, solution generation, and solution evaluation, forming a closed-loop iterative mechanism.

[0058] Optionally, the intelligent decision-making module generates the traffic control strategy using a decision engine based on multi-agent deep reinforcement learning (MADRL). The decision engine comprises multiple agents, each responsible for a separate traffic section or area, and each agent's action space encompasses at least one of adjusting variable speed limits, opening and closing entrance ramps, issuing traffic guidance information, or dispatching emergency rescue resources. It should be noted that a multi-agent architecture is a preferred solution for solving large-scale, distributed traffic network control problems. Compared to a single agent that must process the dimensionally exploding state space of the entire network, MADRL decomposes the complex problem. Each agent only needs to focus on the local state within its area of ​​responsibility and exchange information with neighboring or key agents through a shared communication channel (e.g., an attention mechanism), significantly improving the real-time and scalability of decision-making.

[0059] In an embodiment based on multi-agent deep reinforcement learning (MADRL), the entire highway network is divided into multiple regions, each of which is responsible for an independent agent. The action space of each agent is specifically defined as a set of executable highway control instructions, covering at least one of the following: adjusting the variable speed limit value: the agent can decide to lower the speed limit value of the variable speed limit sign on the road segment it is responsible for from the current value (e.g. 120 km / h) to a preset level (e.g. 100 km / h or 80 km / h) to actively smooth the traffic flow. Turn on / off the ramp: the agent can decide to temporarily turn off the signal light of an entrance ramp to limit the traffic flow into the main line and relieve the congestion pressure on the main line. Publish traffic guidance information: the agent can generate text information (e.g. '5 km ahead of the accident, please detour from XX exit') and instruct the variable information board along the line to publish it. Dispatch emergency rescue resources: upon receiving the 'traffic accident' alert, the agent can automatically calculate the optimal driving path from the nearest tow truck or ambulance base to the accident point and submit it as a recommended solution to the command center.

[0060] Optionally, the reward function of the reinforcement learning agent in the intelligent decision-making module combines the congestion diffusion probability in the traffic state prediction result for dynamic adjustment to guide the traffic control strategy to optimize towards high traffic efficiency. It should be noted that the core of this dynamic reward function design is to decompose a long-term, global optimization goal (such as road network traffic efficiency) into an instant reward signal at each time step. Specifically, after the agent performs an action, the system evaluates the immediate impact of the action on the future congestion diffusion probability and gives corresponding positive or negative rewards. This mechanism enables the learning goal of the agent to be adjusted in real time based on the prediction of the future, thereby guiding it to learn more forward-looking control strategies.

[0061] Optionally, the intelligent decision-making module further adopts a mixed expert model (MoE) as a policy router to select and activate a matching control strategy from a policy pool containing both regular control strategies for handling daily traffic diversion and emergency response strategies for dealing with major abnormal events based on the abnormal alarm or the traffic state prediction result.

[0062] It should be noted that this two-stage decision-making mechanism is an important innovation of the present application. It first uses the reinforcement learning framework to continuously learn and optimize, generating and storing optimal strategies in the policy pool under various scenarios; then, when real-time decision-making is needed, the mixed expert model plays the role of an 'intelligent dispatcher', quickly analyzes the current situation, and selects and applies the most suitable strategy from the policy pool. This solves the problem of a single model being difficult to balance regular optimization and rapid emergency response.

[0063] In a preferred embodiment including a post-processing step, the module's resulting traffic control strategy may include at least one of: vehicle control commands for specific train numbers (such as adding temporary or shuttle trains, or skipping stops), or passenger flow control measures for key stations (such as setting flow thresholds or gate opening strategies). Furthermore, the module can output a decision-making effect prediction report, which can be used to quantitatively evaluate the impact of the implemented plan on key performance indicators (KPIs).

[0064] Optionally, the feature extraction and fusion module fuses the topological structure features and the dynamic content features in the following manner: the topological structure features are embedded into the feature reconstruction path of the autoencoder as a priori structure in the form of an adjacency matrix to guide the dynamic content features to preserve road connectivity and turning relationships during compression and reconstruction.

[0065] It should be noted that this collaborative fusion mechanism is one of the core innovations of the present invention in solving the "data island" problem. It aims to overcome the shortcomings of existing technologies that simply splice or shallowly fuse spatial and temporal features, resulting in information loss and insufficient correlation. The technical principle is that by using the adjacency matrix generated by the graph neural network (GNN) and representing the global spatial structure as a guide or constraint term in the autoencoder model (especially its decoder), the autoencoder can "take into account" the topological relationship of the data location when compressing and reconstructing the data content.

[0066] In a specific embodiment, this means that when the autoencoder reconstructs the traffic flow time series data of a certain intersection A, its loss function not only considers the similarity between the reconstructed data and the original data (for example, minimizing the mean square error), but also introduces a regularization term related to the adjacency matrix. This regularization term penalizes reconstruction results that contradict the physical connection relationship of intersection A (for example, connected to intersections B and C, but not connected to intersection Z). In this way, the system ensures that the final generated dynamic content feature tensor is not just a simple encoding of the time series, but rather that key static topological information such as road connectivity and turning relationships is injected into the encoding process, achieving true and deep spatiotemporal feature fusion, providing a high-quality foundation for subsequent more accurate predictions and decisions.

[0067] Optionally, the reward function of the reinforcement learning agent in the intelligent decision-making module is dynamically adjusted in combination with the congestion diffusion probability in the traffic state prediction result to guide the traffic control strategy to be optimized towards high traffic efficiency.

[0068] It should be noted that this feature is one of the core mechanisms for achieving the forward-looking, proactive decision-making of the present invention, aiming to overcome the drawback of existing reinforcement learning methods, which typically calculate rewards based on the current environmental state, resulting in a certain lag in their decision-making. By using the congestion diffusion probability, which includes future evolution trends and is generated by the spatiotemporal analysis and prediction module, as a key input for constructing the reward function, the present invention shifts the learning objective of the reinforcement learning agent from simply optimizing the current state to optimizing a projected future state, thereby achieving a qualitative shift from "passive response" to "active intervention."

[0069] In one specific embodiment, the dynamic adjustment mechanism can be implemented as follows: when the reinforcement learning agent performs an action at a certain time step (for example, adjusting the timing of a set of traffic lights), the system does not immediately calculate a reward based on current traffic flow changes. Instead, the action is used as a variable and input into the spatiotemporal analysis and prediction module, which re-evaluates the immediate impact of the action on the probability of congestion spreading over a future period (for example, 15 minutes). If the action significantly reduces the predicted congestion risk, the agent receives a large positive reward; conversely, if the action is likely to exacerbate future congestion, it receives a negative reward. This reward mechanism based on future predictions can guide the agent to avoid strategies that are beneficial in the short term but harmful in the long term.

[0070] In this way, the agent can learn a more forward-looking control strategy aimed at improving global high-throughput efficiency. This high-throughput efficiency can be reflected in optimizing one or more key performance indicators (KPIs), including but not limited to reduced average travel delays, increased network throughput, or reduced platform congestion. Experiments have shown that this prediction-driven decision-making loop can increase network throughput by 22% while accelerating policy convergence by 40%, highlighting the innovative value of multi-model collaboration.

[0071] Optionally, the anomaly detection module further detects abnormal events based on at least one of the following judgment criteria: a reconstruction error from the autoencoder in the feature extraction and fusion module; or a prediction deviation between the traffic status prediction result from the spatiotemporal analysis and prediction module and the actual observation value.

[0072] It's important to note that this feature demonstrates the system's high degree of internal information reuse and collaborative verification capabilities, aiming to improve the accuracy and reliability of anomaly identification through multi-dimensional judgment. Rather than relying on a single threshold, it employs cross-validation across two dimensions: "data distribution" and "time series evolution."

[0073] In a specific embodiment, the technical principles are as follows: Detection based on reconstruction error: During training, the autoencoder in the feature extraction and fusion module learns the inherent patterns and distribution of massive amounts of normal data. When a new data sample is input, the autoencoder attempts to compress and reconstruct it. If the data falls within the normal range, the difference between the reconstructed data and the original data (i.e., the reconstruction error) will be very small. Conversely, if the data is abnormal due to reasons such as equipment failure or abnormal passenger flow, it will deviate from the learned normal data distribution, preventing the autoencoder from effectively reconstructing it, resulting in a reconstruction error exceeding a preset threshold. This error signal serves as a powerful anomaly indicator, used to identify anomalies at the device or data source level.

[0074] Detection based on prediction deviation: The Fourier Neural Operator (FNO) in the spatiotemporal analysis and prediction module is responsible for predicting the evolution trajectory of the traffic system under normal circumstances. When an unforeseen real-world event (such as a sudden traffic accident or traffic control) occurs, the observed traffic state will deviate significantly from the normal trajectory predicted by the FNO (i.e., prediction deviation). When this prediction deviation exceeds a dynamically adjusted threshold, the system determines that a sudden change in traffic flow has occurred.

[0075] By combining at least one of the above judgment bases, the anomaly detection module can examine the system status from different dimensions, greatly enhancing the detection capability of various known and unknown anomalies.

[0076] Optionally, the spatiotemporal analysis and prediction module is further configured to: receive external spatiotemporal constraints including geographic information data, time features and event calendars; dynamically adjust the fusion weights of the external spatiotemporal constraints based on the traffic status features and through a gated fusion unit, and fuse the weighted external spatiotemporal constraints with the traffic status features to generate enhanced fusion features containing multi-source contextual information for subsequent spatiotemporal predictions.

[0077] It's important to note that this optional feature aims to overcome the limitations of traditional prediction models, which rely solely on internal highway network traffic flow data and fail to perceive the complex context of the real world. By introducing and intelligently incorporating external spatiotemporal constraints, the accuracy and robustness of the prediction model can be significantly improved under abnormal conditions.

[0078] In a specific embodiment, the fusion process may preferably include three steps: "vector space alignment - spatiotemporal dimension expansion - dynamic feature fusion".

[0079] Vector space alignment: This step aims to unify external constraint data of different types and dimensions into the same feature space as the traffic state features. For example, categorical geographic information such as "road grade" can be mapped into a continuous vector using an embedding layer. Continuous values ​​such as "speed limit" and "slope" can be projected onto the target dimension using normalization and a fully connected layer. Similarly, temporal features such as "weekdays / holidays" can be processed in a similar manner.

[0080] Spatiotemporal dimension expansion: For geographic information features that only contain spatial dimensions, or global time features that only contain time dimensions, the system will expand their dimensions to match the traffic state feature tensor (for example, the dimensions are T×N×F, where T is the time step, N is the number of spatial nodes, and F is the feature dimension) through broadcasting operations.

[0081] Dynamic Feature Fusion: After all features are aligned and concatenated, the core of this invention lies in dynamic fusion via the gated fusion unit. This gate acts as an intelligent switch, dynamically and nonlinearly adjusting the fusion weights of external spatiotemporal constraints based on real-time traffic status. For example, when the system detects normal traffic flow, external events (such as "large-scale events") may be given a lower weight. However, when traffic concentration is abnormal, the gate automatically increases the weight of "large-scale events," enabling the model to make more realistic predictions.

[0082] Through the above mechanism, the system can generate an enhanced fusion feature that contains multi-source contextual information, providing richer and more comprehensive input for subsequent precise spatiotemporal predictions.

[0083] Optionally, the spatiotemporal analysis and prediction module is configured to receive a feedback signal as a dynamic intervention signal to perform real-time correction on the traffic status prediction result; the feedback signal includes the traffic control strategy generated by the intelligent decision-making module and / or the abnormality alarm generated by the abnormality detection module.

[0084] It should be noted that this closed-loop feedback mechanism is one of the core features of this invention that distinguishes it from traditional one-way pipeline systems. It enables the system to dynamically self-correct and adapt based on the execution of its own decisions and sudden changes in the environment, thus forming a more robust and intelligent organic whole.

[0085] In a specific embodiment, the feedback mechanism includes two key information return paths: Decision feedback from the intelligent decision-making module: When the intelligent decision-making module generates and issues a new traffic control strategy (for example, adjusting the traffic light timing at a key intersection), this instruction is not only the system's final output; it also serves as the feedback signal, re-input into the spatiotemporal analysis and prediction module. Its technical significance lies in enabling the prediction model to understand in real time the changes caused by the system's own interventions to the traffic environment. In the next round of predictions, these changes are used as new known conditions for calculation, avoiding inaccurate predictions caused by the system's own decisions and achieving "self-reflection" and "self-consistency" of the system.

[0086] Alarm feedback from the anomaly detection module: When the anomaly detection module identifies an unexpected event (e.g., a traffic accident or road congestion) and generates an anomaly alarm, this alarm information is also immediately input into the spatiotemporal analysis and prediction module as the feedback signal. This technical significance lies in enabling the prediction model to instantly switch from conventional prediction mode to emergency prediction mode. For example, after receiving an accident alarm, the Fourier neural operator can immediately re-analyze the traffic flow evolution in the affected area, thereby accurately predicting the spread and duration of the congestion that may be caused by the accident.

[0087] Through the dual feedback mechanism, the spatiotemporal analysis and prediction module receives these signals as dynamic intervention signals and integrates them with the latest traffic status characteristics, thereby achieving real-time correction of traffic status prediction results, ensuring that the entire system always maintains the highest accuracy and adaptability in a dynamically changing environment.

[0088] Optionally, the anomaly detection module further adopts a hybrid expert model to classify the abnormal events, and the hybrid expert model includes a gated network and multiple expert networks: the gated network adopts a multi-layer perceptron and a softmax output mechanism to perform a nonlinear transformation on the traffic state characteristics to output an expert activation probability vector; the multiple expert networks are dynamically selected according to the expert activation probability vector to perform abnormal event classification, and the multiple expert networks include at least four of the following: an equipment failure expert network, whose input includes the roadside equipment signal loss rate and the communication interruption duration; a passenger flow surge expert network, whose input includes the passenger flow growth value per unit time; a traffic accident expert network, whose input includes the instantaneous vehicle speed drop rate and the spatial propagation range; a congestion situation expert network, whose input includes the road section flow growth value and the upstream queue length; a severe weather expert network, whose input includes environmental data such as visibility, road slippery or icing coefficient.

[0089] It should be noted that the technical principle of this "expert consultation" internal structure is to decompose the complex full-set classification task into multiple simpler and more focused sub-tasks. By allowing professional sub-models to handle professional anomaly types, the accuracy of detection and classification can be significantly improved.

[0090] In a specific embodiment, the gating network, acting as an intelligent scheduler, may consist of a multi-layer perceptron (MLP) and a softmax activation function. It receives traffic state features (e.g., a 128-dimensional fused feature tensor) and may incorporate spatiotemporal context (e.g., time tags and road segment levels) as input. Its computational logic extracts key discriminative features through nonlinear transformations (e.g., ReLU activation), ultimately outputting a vector of expert activation probabilities (e.g., [0.1, 0.7, 0.2]) that corresponds to the number of expert networks. Each value in this vector represents the probability, or weight, of assigning the current task to the corresponding expert network. Experts with higher weights contribute more to the final classification result. For example, if the speed on a road segment is detected to have dropped by 80% within 30 seconds during peak hours, the gating network will assign at least 80% of the weight to the traffic accident expert network.

[0091] The multiple expert networks are dynamically selected based on the expert activation probability vectors output by the gating network to perform the final fine-grained classification. Each expert network is specifically trained to identify a specific type of anomaly, and its input and internal structure are optimized for a specific task, for example: Equipment Failure Expert Network: Its inputs may include statistical features such as the signal loss rate and communication interruption duration of roadside sensors. In a preferred embodiment, this expert network employs a structure consisting of a one-dimensional convolutional layer (for extracting timing fluctuations) and a fully connected layer to focus on identifying anomalies such as equipment signal interruptions and data jumps.

[0092] Passenger flow surge expert network: Its inputs may include spatiotemporal sequence features such as unit-time passenger flow growth, passenger flow density, and transfer pressure. In a preferred embodiment, this expert network employs a "long short-term memory (LSTM)" layer (to capture temporal trends) + an attention mechanism (to focus on peak periods) structure to excel at distinguishing between normal holiday passenger flow and abnormal surges caused by emergencies.

[0093] Traffic Accident Expert Network: Its inputs can include features such as the instantaneous rate of vehicle speed reduction, frequency of sudden braking, and the spatial spread of abnormal events. It can also be combined with road network topology information extracted by a graph neural network. In a preferred embodiment, this expert network employs a "graph convolutional layer (for modeling spatial dependencies) + fully connected layer" structure to focus on identifying abnormal congestion with spatial spread characteristics caused by accidents such as collisions and breakdowns.

[0094] Congestion Situation Expert Network: This expert network focuses on identifying congestion anomalies caused by common factors (such as peak commuting times and bottlenecks) but whose evolution may exceed expectations. Its inputs may include spatiotemporal sequence features such as the growth in traffic flow across the road section, queue lengths at upstream key nodes (such as ramps and toll booths), and the difference between the current congestion length and the historical average for the same period. In a preferred embodiment, this expert network employs a "Gated Recurrent Unit (GRU) layer + Fully Connected Layer" architecture. As an efficient variant of LSTM, GRU is equally adept at capturing the evolutionary trends of time series and can accurately simulate the complete process of congestion formation, intensification, and dissipation. By modeling the speed and scope of congestion evolution, this expert network can effectively distinguish between "benign congestion that is easing" and "serious congestion that is at risk of getting out of control."

[0095] Severe Weather Expert Network: This expert network is specifically trained to identify global or regional traffic anomalies directly or indirectly caused by severe weather conditions (such as dense fog, heavy rain, and snow). Its input includes not only traffic flow data but also, more importantly, environmental sensor data, such as visibility (in meters) provided by weather stations along the highway, road surface slipperiness or iciness detected by road sensors, and crosswind speed. In a preferred embodiment, this expert network employs a "multimodal fusion attention network" architecture. This network first processes traffic flow data and environmental data separately using parallel feature extractors. It then utilizes an attention mechanism to dynamically learn which environmental factors have the greatest impact on traffic conditions under different weather conditions (for example, visibility is given the highest weight in foggy weather). This approach enables the expert network to accurately identify traffic anomalies caused by severe weather, which are distinguished from typical congestion by widespread slowdowns or surges in accident risk.

[0096] Through deep feature mining by dedicated experts and dynamic scheduling of the gating network, the hybrid expert model can accurately match anomaly types with detection requirements, thereby significantly improving the accuracy of anomaly classification and response efficiency.

[0097] Optionally, the intelligent decision-making module further adopts a hybrid expert model as a strategy routing, which is used to select and activate a matching control strategy from a strategy pool including conventional control strategies for handling daily traffic diversion and emergency response strategies for dealing with major abnormal events based on the abnormal alarm or the traffic status prediction results.

[0098] It's important to note that this two-stage decision-making mechanism is a key innovation of this invention, aiming to address the technical challenge of a single decision-making model's inability to simultaneously address both refined optimization of daily operations and rapid, efficient response to emergencies. By introducing a Mixture of Experts (MoE) model as an intelligent "policy router," this invention effectively decouples complex decision-making tasks, enabling adaptive and specialized processing for diverse scenarios.

[0099] In a specific embodiment, the working principle is as follows: when the intelligent decision-making module receives input from an upstream module (such as the abnormal alarm or the traffic status prediction result), the hybrid expert model serving as the strategy routing first analyzes the current situation characteristics.

[0100] If the current situation is judged to be a normal operating state, the model will activate and call the corresponding normal scheduling expert or strategy module from the strategy pool, for example, a rule-based algorithm designed to optimize capacity matching, or a module for optimizing energy-sensitive driving organization during off-peak hours.

[0101] Conversely, if a high-level abnormal alarm is detected, the model will activate an emergency response expert or policy module, for example, a module driven by the reinforcement learning agent for dynamic path planning or emergency resource scheduling.

[0102] The strategy pool is a collection of pre-generated or continuously updated strategies, containing validated or optimized control strategy templates for different scenarios. Through this "routing first, execution later" mechanism, the system ensures the most appropriate decision logic is invoked in every scenario, significantly improving decision-making relevance, responsiveness, and overall operational intelligence.

[0103] Optionally, the intelligent decision-making module uses a decision engine based on multi-agent deep reinforcement learning to generate the traffic control strategy, and the decision engine includes multiple agents, wherein each agent is responsible for an independent traffic section or area, and the action space of the agent covers at least one of adjusting variable speed limit values, opening and closing entrance ramps, issuing traffic guidance information, or dispatching emergency rescue resources.

[0104] It should be noted that this feature aims to overcome the challenges of traditional single-agent reinforcement learning in large-scale, distributed transportation networks, which often result from the dimensional explosion of state and action spaces, leading to training difficulties, slow convergence, and decision delays. This invention employs a decision engine based on multi-agent deep reinforcement learning (MADRL) to decompose complex global decision-making tasks into multiple parallel, collaborative subtasks.

[0105] In one specific embodiment, the decision engine's organizational structure is as follows: the entire traffic network is divided into multiple independent or semi-independent traffic segments or zones, each of which is managed by a dedicated agent. This "divided-area-governed" model allows each agent to focus only on the local state information within its responsible zone, significantly reducing computational complexity. The agent's state space may include, but is not limited to, factors such as passenger flow demand, vehicle location, and traffic light status within its responsible zone.

[0106] Although individual agents are distributed at the execution level, they can exchange information with neighboring or key agents through a shared communication channel (e.g., an attention mechanism) and learn within a centralized training framework (i.e., a "centralized training, decentralized execution" model) to ensure that local decisions are executed towards the global optimal goal. In this way, the decision-making engine of this invention significantly improves the real-time performance and scalability of decision-making, as well as its practical application in complex transportation networks.

[0107] Optionally, the traffic control strategy specifically includes at least one of the following: vehicle command commands for specific train numbers, the vehicle command commands including adding temporary trains, shuttle trains or skip-stop operations; or passenger flow control measures for key stations, the passenger flow control measures including flow limiting threshold settings or gate opening strategies.

[0108] It's important to note that this feature aims to transform the abstract strategies generated by the intelligent decision-making module into a concrete set of instructions that can be executed in actual traffic operations. It clarifies that the final output of the system can directly affect traffic elements in the physical world, thereby achieving end-to-end closed-loop control.

[0109] In a specific embodiment, the traffic control strategy may be expressed in two or more of the following forms: Vehicle Commands: These commands directly impact vehicle scheduling and operation. For example, based on predicted peak passenger flow, the intelligent decision-making module can automatically generate "additional trains" or "shuttle trains" for specific routes or train numbers to increase capacity. Alternatively, if a fault or congestion is detected on the route ahead, a "skip stop" command can be generated to ensure efficient operation on subsequent routes.

[0110] Passenger flow control measures: These instructions are primarily used to manage passenger flow at key nodes, such as large transfer stations. For example, if the system predicts a significant passenger flow at a key station in the near future, it can automatically generate instructions to set flow control thresholds or adjust gate opening strategies to proactively mitigate passenger flow shocks and ensure order and safety within the station.

[0111] By concretizing the control strategy into the above-mentioned executable instructions, the present invention not only provides an intelligent analysis and decision-making engine, but also provides a complete solution that can seamlessly connect with the existing dispatching and command system and has actual operational intervention capabilities, thus having significant practical value.

[0112] See also Figure 2 , schematically illustrates a hierarchical architecture of a highway digital intelligent operation and monitoring system provided by an embodiment of the present application. In a preferred embodiment, the system may include the following modules from bottom to top: Data acquisition module 100 This module is the cornerstone of the entire system, responsible for acquiring multi-source heterogeneous data from the transportation network. In a specific embodiment, the multi-source heterogeneous data collected by this module includes, but is not limited to: traffic flow data collection from the highway network; roadside equipment status data collection; and environmental data collection. Real-time streaming data collection: In addition to the aforementioned pre-aggregated data, this module can also directly collect high-frequency raw data streams, such as continuous speed and position data from on-board sensors and real-time passenger counts from platforms, to support millisecond-level exception responses.

[0113] External Data Interface: This module may also include an external data interface for accessing third-party APIs to obtain event data related to traffic accidents, large-scale events, construction road closures, etc. from sources including but not limited to social media and news websites.

[0114] Feature extraction and fusion module 200: This module receives the multi-source heterogeneous data from the data acquisition module 100. Its core task is to solve the "data island" problem and generate unified traffic status features.

[0115] In a preferred embodiment, the workflow of this module includes: data preprocessing stage; multi-model collaborative feature extraction stage; cross-modal feature fusion stage; heterogeneous space mapping; dynamic weight fusion; and context enhancement.

[0116] Finally, the module outputs a traffic state feature tensor that integrates multi-level semantic information.

[0117] The spatiotemporal analysis and prediction module 300 is a prerequisite for realizing the "active intelligent" decision-making of the present invention. It is responsible for performing forward-looking analysis, receiving the traffic status features from the feature extraction and fusion module 200, and processing them using the Fourier neural operator (FNO).

[0118] In a preferred embodiment, the input of this module includes three categories of core data: structured feature tensors from the feature extraction and fusion module 200 (e.g., a 128-dimensional tensor containing spatial, temporal, and contextual features); external spatio-temporal constraints; and dynamic intervention signals.

[0119] Optionally, the external spatio-temporal constraints include geographic information data (e.g., road class, speed limit), temporal features (e.g., weekday / holiday identification), and event calendar (e.g., large-scale activities, construction plans). This application adopts a preferred three-step process of "vector space alignment - spatio-temporal dimension expansion - dynamic feature fusion" to fuse these external information. For example, different types of external spatio-temporal constraints are mapped to the same vector dimension as traffic flow features through embedding layers and fully connected layers, then dimension is unified through broadcast and concatenation operations, and finally the fusion weight of external information is dynamically adjusted according to the traffic state through a gating fusion unit.

[0120] Optionally, the dynamic intervention signals include but are not limited to feedback signals from the intelligent decision-making module 400 (e.g., signal timing adjustment instructions) and alarm information from the anomaly detection module 500. This module receives these signals to achieve real-time correction of the prediction model.

[0121] In a preferred embodiment, the workflow of this module can adopt a three-stage processing flow of "spatio-temporal decomposition - feature encoding - prediction generation".

[0122] In the spatio-temporal data decomposition stage, the spatial dimension can be decomposed based on the road network topology using graph convolution network (GCN) or one-dimensional convolutional neural network (1D-CNN), etc.; and the time series can be decomposed into trend, fluctuation and residual terms using wavelet transform, etc.

[0123] In the feature encoding and enhancement stage, a spatio-temporal joint attention mechanism can be designed to calculate the correlation weight of spatial and temporal dimensions through QKV matrix, and an interpretability constraint can be introduced.

[0124] In the prediction generation and optimization stage, the core is to process the fused features and the dynamic intervention signals using FNO. FNO treats traffic flow as the solution of partial differential equation (PDE), which is calculated efficiently in the frequency domain through Fourier transform, and can be designed as a hierarchical operator network containing Fourier transform layer, spectral filter and inverse Fourier transform layer. This stage can also use multi-task learning to predict traffic flow while assisting in predicting congestion duration, queue length, etc. At the same time, methods such as Monte Carlo Dropout can be used for uncertainty quantification to provide confidence evaluation for the prediction results.

[0125] This module finally generates two categories of core data: Traffic state prediction results: These results include forward-looking predictions of traffic flow, average vehicle speed, congestion diffusion probability, and other information for multiple time windows (e.g., 15, 30, and 60 minutes) in the future. Their primary purpose is to provide environmental model input to the intelligent decision-making module 400 to formulate proactive control strategies.

[0126] Traffic status reference baseline: It should be noted that this baseline is a reference data set generated by the spatiotemporal analysis and prediction module to represent the state that the traffic system should be in under normal circumstances without abnormal events. It is essentially a prediction of the "ideal normal state" and its main purpose is to provide a judgment scale for the parallel anomaly detection module 500. The anomaly detection module 500 compares the real-time traffic status characteristics with this "traffic status reference baseline". Once the deviation between the actual value and the baseline value exceeds a preset threshold, it can be determined that an abnormal event has occurred. This baseline can specifically include a traffic flow baseline and a roadside equipment status baseline.

[0127] Anomaly detection module 500 This module operates in parallel with the spatiotemporal analysis and prediction module 300 and is responsible for real-time monitoring of the system health status.

[0128] In a preferred embodiment, the module's input data exhibits multi-source heterogeneous features, encompassing spatiotemporal data, equipment status, and business indicators. Specifically, its input may include: the traffic status prediction results and roadside equipment status data output by the spatiotemporal analysis and prediction module 300; traffic status features and real-time streaming data output by the feature extraction and fusion module 200; and meteorological data included in the event data and environmental data.

[0129] The workflow of this module can be divided into several stages: Abnormality identification stage: This stage receives the traffic status characteristics and determines whether the equipment status deviates from the normal operating range based on the roadside equipment status data. Preferably, the judgment basis may include: Detection based on reconstruction error: Detection is performed using the reconstruction error of the autoencoder in the feature extraction and fusion module. When the reconstruction error exceeds a preset threshold, it is determined to be abnormal.

[0130] Prediction deviation detection: Detection is based on the predicted deviation between the traffic state prediction results from the spatiotemporal analysis and prediction module and the actual observed value. When the predicted deviation exceeds a dynamic threshold, an early warning can be triggered.

[0131] Graph-based detection: Utilizing the graph neural network (GNN) in the feature extraction and fusion module, the system monitors changes in the topology of traffic networks to identify patterns such as congestion propagation and fault diffusion. For example, if traffic volume on a certain road section suddenly increases, and traffic volume on adjacent sections also shows an upward trend, this can be identified as a congestion propagation event.

[0132] Classification and risk assessment phase: After an anomaly is identified, a mixture of experts (MoE) is used to classify the anomaly. In a more specific embodiment, the MoE includes a gating network and multiple expert networks. The gating network extracts key discriminant features through nonlinear transformations (such as ReLU activation) and outputs a weight vector consistent with the number of experts. Experts with higher weights contribute more to the final classification. For example, if a road section is detected to have a significant drop in speed within a short period of time, the gating network will assign a higher weight to the traffic accident expert network. The expert network can be specifically trained to identify specific types of anomalies, such as: Equipment Fault Expert Network: This network uses structures such as one-dimensional convolutional layers to focus on identifying anomalies such as device signal interruptions and data jumps.

[0133] Surge Expert Network: This network uses a structure such as a long short-term memory (LSTM) layer and is good at distinguishing between normal high passenger flow and abnormal surges.

[0134] Traffic Accident Expert Network: This network utilizes graph convolutional layers and other structures to identify abnormal congestion caused by collisions, breakdowns, and other factors. After classification, the system also comprehensively considers the severity, speed of transmission, and difficulty of handling the incident to assign a risk level (for example, a five-level risk scale). Leveraging FNO's spatiotemporal deduction capabilities, the system then assesses the spread of its impact.

[0135] Output stage: This module ultimately generates an abnormality alert and can further output a list of abnormal events containing structured fields such as event type, occurrence time, location, and impact range, as well as a risk assessment report containing risk matrix, trend forecast, optimization suggestions, etc.

[0136] It should be noted that the module can output abnormal alarm information on the visual interface. In a preferred embodiment, the module integrates multimedia elements such as maps, charts, and videos to display the operating status of the traffic network, the distribution of abnormal events, risk levels and other information in real time and outputs them on the real-time monitoring screen, and supports multi-level drilling and interactive analysis operations; and provides functions such as timeline playback of abnormal events, evidence chain display, and disposal process recording in a certain event tracing panel, supporting event review and experience summary. It also includes a risk warning pop-up mechanism. When a high-level risk event is detected, a warning window will automatically pop up to display event details, disposal suggestions, contact personnel and other information to ensure a quick response.

[0137] Intelligent Decision Module 400 This module is the "brain" of the system, responsible for generating the final control strategy. In a preferred embodiment, the module's input data exhibits multimodal characteristics, including traffic status predictions and equipment status predictions from the spatiotemporal analysis and prediction module 300, abnormal event lists and risk level assessments from the anomaly detection module 500, and external resource constraints (such as capacity allocation and policy rules). The module's workflow can be divided into three stages: situational awareness, solution generation, and solution evaluation. Situational cognition stage: This stage uses GNN to model the relationship between the input spatiotemporal prediction results and abnormal events, construct a global state diagram, and use autoencoders to compress the high-dimensional state space to generate input representations that can be processed by the decision model.

[0138] Solution generation stage: This stage is the core of decision-making.

[0139] Optimally, a Mixed of Experts (MoE) model is introduced as a policy routing framework, activating the corresponding sub-expert network based on the current context. For example, in the event of a major incident, emergency response experts are activated, initiating a dynamic path planning module driven by reinforcement learning. During daily operations, routine scheduling experts are activated. During off-peak hours, energy efficiency optimization experts are activated, invoking the FNO to optimize energy-sensitive traffic organization.

[0140] Crucially, traffic control strategies are generated based on a multi-agent deep reinforcement learning (MADRL) framework. Each agent corresponds to an independent traffic segment or area. The agent's state space can include passenger demand and vehicle location, while its action space can encompass operations such as adjusting variable speed limits, opening and closing on-ramps, issuing traffic guidance information, or dispatching emergency rescue resources.

[0141] Proactive learning and evaluation: The traffic state prediction results are used as the environmental model input of each agent, and its reward function is dynamically adjusted according to the congestion diffusion probability to guide the agent to learn forward-looking strategies aimed at improving global traffic efficiency.

[0142] In a specific embodiment, the reward function of the reinforcement learning agent is designed as a weighted sum of multiple optimization objectives to guide it to learn a strategy that can improve the overall operational efficiency of the highway. The reward function R can be defined as: R = w1*(T_avg)+w2*(Q_len)+w3*(A_resp) Here, T_avg represents the average speed of the road network, which we wish to maximize; Q_len represents the length of the mainline congestion queue, which we wish to minimize; and A_resp represents the emergency response time for accidents, which we also wish to minimize. w1, w2, and w3 are the weights for each of these factors. By optimizing this reward function, which is directly linked to core highway operational metrics, the agent can learn to make more forward-thinking decisions aimed at improving overall traffic efficiency and safety.

[0143] Output and feedback: This module ultimately generates and outputs traffic control strategies. These strategies can include: vehicle command commands for specific train numbers (e.g., adding temporary trains or skipping stops); passenger flow control measures for key stations (e.g., setting flow thresholds and gate opening strategies); or emergency response plans (e.g., enabling backup routes in the event of equipment failure, bus connection scheduling, and information release and filing).

[0144] At the same time, the traffic control strategy is output as a feedback signal to the spatiotemporal analysis and prediction module, forming a decision-making-prediction closed loop.

[0145] Application layer module 600 This module is the user interface of the system, responsible for collecting all processing results and presenting them in a friendly manner. In a preferred embodiment, the system also includes an application layer module, which is configured to: Receiving and Fusion: Receives the traffic control strategy and decision-making effect predictions from the intelligent decision-making module 400, the abnormal event list and risk assessment report from the anomaly detection module 500, and the real-time traffic status from the spatiotemporal analysis and prediction module 300. "Real-time traffic status" here refers to a data set representing the current, most realistic operating conditions of the highway network. Its primary purpose is to provide a real-time data source for visualization interfaces such as the real-time monitoring screen of the application layer module 600, enabling "what you see is what you get" monitoring of operational status. This "real-time traffic status" may include, but is not limited to, real-time traffic flow data and real-time equipment status data acquired by the data acquisition module 100 and initially processed by the feature extraction and fusion module 200. It is clearly distinguished in terms of function and connotation from "traffic status prediction results" and "traffic status reference baselines."

[0146] Visualization: The integrated information is visualized in a three-dimensional digital twin system. For example, a three-dimensional digital twin of a transportation network, built using GIS and BIM technologies, maps the operational status of the physical world in real time and displays KPI trends and risk distribution through dynamic heat maps and multi-dimensional chart linkage.

[0147] The output stage includes three types of information: business application, decision support, and user interaction, which directly serve the daily operation and emergency management of highways.

[0148] The business application output provides a human-computer collaborative interface: the interface may specifically include at least one of the following: Real-time monitoring screen: Based on GIS and digital twin technology, the 3D visualization presents the real-time operation status of the entire highway network, displays the congestion level of each section through dynamic heat maps, and marks the precise location of all abnormal events (such as accidents and severe weather).

[0149] Road network control platform: Dispatchers can use this platform to issue control instructions generated by the intelligent decision-making module with one click, such as "implement an 80km / h speed limit on the K100-K110 section of the G2 Expressway" and track the execution status and effects of the instructions in real time.

[0150] Emergency Management Platform: When a major accident occurs, the platform automatically pops up an early warning window and integrates a plan library and resource library (such as the locations of tow trucks and maintenance teams along the route) to provide dispatchers with intelligent disposal plan recommendations and resource scheduling suggestions.

[0151] Data Analysis Reports: Regularly generate analytical reports on topics such as operational performance, risk trends, and equipment status to support management decision-making. Reports include visual charts, data tables, and text interpretations for easy reading and sharing.

[0152] The decision support output may specifically include at least one of the following: Intelligent recommendation plan: Based on historical data and real-time status, it provides dispatchers with multiple decision-making options, such as "Plan A: Adjust departure intervals, and expect a 5% increase in punctuality" and "Plan B: Activate spare vehicles, and expect a 10% increase in transport capacity."

[0153] Simulation tools: Provide a digital twin environment for the transportation network, allowing dispatchers to test the effects of different solutions in virtual scenarios, such as changes in passenger flow distribution after adjusting the flow limit threshold, and the increase in transportation capacity after activating spare vehicles.

[0154] The user interaction output may specifically include at least one of the following: Personalized workstations: Customize your work interface and function menu based on your role (dispatcher, administrator, analyst, etc.) and permissions. For example, the dispatcher interface focuses on real-time monitoring and command operations, while the analyst interface emphasizes data analysis and report generation.

[0155] Interactive Operation Log: This log records all user operations, such as command issuance, parameter adjustments, and interface switching, and supports operation backtracking and auditing. The log is presented in a timeline format, allowing users to quickly locate key operation nodes.

[0156] The embodiment of the present application provides a digital intelligent operation monitoring system for highways, which realizes end-to-end intelligent monitoring of traffic operations through a collaborative architecture of "fusion-prediction-decision-making". First, through the deep collaboration of graph neural networks and autoencoders, multi-source heterogeneous data is converted into a unified feature representation containing spatial topology and dynamic content, fundamentally breaking the "information island" and providing a high-quality data foundation for global analysis. Secondly, the system uses Fourier neural operators to perform long-term spatiotemporal predictions and uses the prediction results as the environmental model input of the reinforcement learning agent to build a forward-looking decision-making closed loop, which transforms the system decision-making from passive rule-driven to active, intelligent optimization based on future predictions. In addition, by introducing a parallel hybrid expert model for accurate anomaly detection and linking it with the reinforcement learning decision core, an integrated emergency response from second-level detection to rapid response is achieved. Through the synergistic effect of the above-mentioned key technical links, the system of the present invention overcomes the limitations of traditional methods, realizes comprehensive perception, accurate prediction, active decision-making and efficient emergency response of traffic operations, and significantly improves the operating efficiency and safety of the transportation system.

[0157] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0158] References herein to "one embodiment," "an embodiment," or "one or more embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Furthermore, please note that instances of the phrase "in one embodiment" do not necessarily all refer to the same embodiment.

[0159] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0160] In the claims, any reference signs placed between brackets shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A digital intelligent operation monitoring system for highways, characterized by: include: A data acquisition module is used to acquire multi-source heterogeneous data including at least traffic flow data of the highway network, roadside equipment status data and environmental data; A feature extraction and fusion module is configured to receive the multi-source heterogeneous data, model the topological structure of the highway network through a graph neural network to generate topological structure features, and perform feature alignment and compression on the multi-source heterogeneous data using an autoencoder to generate dynamic content features; fusing the topological structure feature with the dynamic content feature to generate a traffic state feature; a spatiotemporal analysis and prediction module, configured to receive the traffic state characteristics and process them using a Fourier neural operator to generate a traffic state prediction result including a congestion diffusion probability and roadside equipment status data; an anomaly detection module, configured to receive the traffic status characteristics and determine, based on the roadside equipment status data, whether the equipment status deviates from a normal operating range, so as to generate an anomaly alarm; The intelligent decision-making module is used to use the traffic state prediction result as the environmental model input of the reinforcement learning agent, and the reinforcement learning agent autonomously learns and generates a traffic control strategy based on the traffic state prediction result.

2. The system according to claim 1, wherein: The feature extraction and fusion module fuses the topological structure features and the dynamic content features in the following manner: The topological structure features are embedded into the feature reconstruction path of the autoencoder as a priori structure in the form of an adjacency matrix to guide the dynamic content features to preserve road connectivity and turning relationships during compression and reconstruction.

3. The system according to claim 1, wherein: The reward function of the reinforcement learning agent in the intelligent decision-making module is dynamically adjusted in combination with the congestion diffusion probability in the traffic state prediction result to guide the traffic control strategy to be optimized towards high traffic efficiency.

4. The system according to claim 1, wherein: The anomaly detection module further detects an abnormal event based on at least one of the following judgment criteria: The reconstruction error from the autoencoder in the feature extraction and fusion module; or The prediction deviation between the traffic state prediction result from the spatiotemporal analysis and prediction module and the actual observation value.

5. The system according to claim 1, wherein: The spatiotemporal analysis and prediction module is further configured to: Receive external spatiotemporal constraints including geographic information data, time features, and event calendars; Based on the traffic state characteristics and through a gated fusion unit, the fusion weight of the external spatiotemporal constraints is dynamically adjusted, and the weighted external spatiotemporal constraints are fused with the traffic state characteristics to generate an enhanced fusion feature containing multi-source context information for subsequent spatiotemporal prediction.

6. The system according to claim 1, wherein: The spatiotemporal analysis and prediction module is configured to receive a feedback signal as a dynamic intervention signal to perform real-time correction on the traffic status prediction result; the feedback signal includes the traffic control strategy generated by the intelligent decision-making module and / or the abnormality alarm generated by the abnormality detection module.

7. The system according to claim 4, wherein: The anomaly detection module further uses a hybrid expert model to classify the abnormal events. The hybrid expert model includes a gating network and multiple expert networks: The gating network uses a multi-layer perceptron and a softmax output mechanism to perform nonlinear transformation on the traffic state features to output an expert activation probability vector; The plurality of expert networks are dynamically selected according to the expert activation probability vector to perform abnormal event classification, and the plurality of expert networks include at least four of the following: Equipment failure expert network, whose input includes roadside equipment signal loss rate and communication interruption duration; The passenger flow surge expert network, whose input includes the passenger flow growth value per unit time; Congestion situation expert network, whose input includes the traffic growth value of the section and the upstream queue length; A network of severe weather experts whose input includes environmental data such as visibility, slippery or icy road conditions, etc. Traffic accident expert network, whose input includes instantaneous vehicle speed reduction rate and spatial propagation range.

8. The system according to claim 1, wherein: The intelligent decision-making module further adopts a hybrid expert model as a strategy routing, which is used to select and activate a matching control strategy from a strategy pool including conventional control strategies for handling daily traffic diversion and emergency response strategies for dealing with major abnormal events based on the abnormal alarm or the traffic status prediction result.

9. The system according to claim 1, wherein: The intelligent decision-making module uses a decision engine based on multi-agent deep reinforcement learning to generate the traffic control strategy, and the decision engine includes an action space covering at least one of adjusting the variable speed limit value, opening and closing the entrance ramp, issuing traffic guidance information or dispatching emergency rescue resources.

10. The system according to any one of claims 1 to 9, characterized in that The system further includes an application layer module, wherein the application layer module is configured to: receiving the traffic control strategy from the intelligent decision-making module and visually presenting it on a three-dimensional digital twin system; A human-machine collaborative interface is provided, which includes at least one of a real-time monitoring screen for globally displaying the operating status of the expressway network, a road network management and control platform for issuing and tracking traffic control instructions, or an emergency management platform for rapid response.

Images