An intelligent traffic global signal control method

Abstract

This invention discloses a method for intelligent transportation signal control across the entire network, relating to the field of intelligent transportation technology. It aims to address the problem that traditional traffic signal control cannot meet the collaborative optimization needs of intelligent connected vehicles and traffic facilities of different levels across the entire network. The method includes the following steps: S1, extracting multi-dimensional features from traffic data, integrating them to obtain a traffic scene vector, and combining this with a large language model to output a prompt vector; S2, fusing multi-source data, followed by anomaly detection and feature dimensionality reduction, and then performing deep analysis; S3, predicting the traffic flow change trend of the road network, and generating and executing an optimized traffic signal control strategy through a multi-objective optimization decision algorithm and an intersection collaborative control algorithm. This invention improves the traffic efficiency, safety, and intelligence level of urban transportation networks by collaboratively optimizing traffic flow allocation using multiple technologies.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to an intelligent transportation global signal control method. Background Technology

[0002] With the rapid growth of urbanization and motor vehicle ownership, traffic congestion and frequent accidents are becoming increasingly serious problems. Traditional traffic signal control relies on fixed timing schemes or single-point sensing control, which is difficult to adapt to dynamically changing traffic demands (e.g., Chinese patent CN114333370A). Single-vehicle intelligent technology is limited by the sensing range and computing power of onboard sensors, and cannot achieve global road network optimization. Early vehicle-road cooperation only achieved limited information exchange between vehicles and roads, which is difficult to support collaborative decision-making in large-scale connected scenarios. In addition, existing technologies lack the ability to deeply integrate and dynamically decide on multi-source data such as vehicles, roads, and environment in complex environments (such as severe weather and emergencies), and cannot meet the collaborative optimization needs of intelligent connected vehicles and traffic facilities of different levels across the entire domain. Summary of the Invention

[0003] This invention addresses the problem that traditional traffic signal control cannot meet the collaborative optimization needs of intelligent connected vehicles and traffic facilities of different levels across the entire area. It proposes an intelligent traffic signal control method that significantly improves the traffic efficiency, safety, and intelligence level of urban traffic networks.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent traffic signal control method, comprising the following steps: S1: Extract multi-dimensional features from traffic data, integrate them to obtain a traffic scene vector, and combine them with a large language model to output a prompt vector; S2 performs fusion processing on multi-source data, followed by anomaly detection and feature dimensionality reduction, and then performs deep analysis. S3 predicts the traffic flow trend of the road network and generates and executes optimized traffic signal control strategies through multi-objective optimization decision-making algorithms and intersection cooperative control algorithms.

[0005] This invention mainly includes the following processes: 1. Data acquisition: The vehicle-side device collects real-time data on itself and the environment through sensors, and the roadside equipment obtains comprehensive traffic information through multi-sensor fusion; 2. Collaborative transmission: The vehicle-side data is uploaded to the roadside RSU via V2X communication, and after roadside preprocessing, it is transmitted to the cloud via 5G network; 3. Intelligent decision-making: The cloud generates a global control strategy based on LLM and multi-model fusion, and realizes intersection collaboration through optimization algorithms; 4. Closed-loop execution: The strategy is sent to the roadside controller to adjust the traffic lights, and at the same time, it is fed back to the vehicle-side device to realize route guidance and speed optimization.

[0006] The present invention is further configured such that step S1 includes: S11, Obtain features from the traffic scene, including traffic flow sequence, queue length, speed distribution, and event encoding, and integrate the features to obtain a traffic scene vector; S12, based on traffic scene vectors, generates a preliminary control scheme through a large model, and introduces a reinforcement learning module to dynamically adjust the preliminary control scheme; S13: Construct a knowledge graph for the transportation domain, adaptively fuse transportation scene vectors and domain knowledge vectors, and finally output prompt vectors.

[0007] In this technical solution, in order to achieve accurate digital representation of complex traffic scenarios, the present invention adopts a multi-dimensional feature fusion approach to map real-time traffic data into structured natural language prompts and construct traffic scenario vectors.

[0008] The present invention is further configured such that step S2 includes: S21, collect and characterize multi-source data, including vehicle-side data and roadside data; S22, perform spatiotemporal alignment processing on multi-source data, and fuse multi-source data through an attention mechanism to obtain fused data; S23, the roadside edge device performs anomaly detection and feature dimensionality reduction operations to obtain the dimensionality-reduced data; the cloud performs trajectory prediction and conflict warning to complete in-depth analysis.

[0009] In this technical solution, a multimodal heterogeneous data fusion architecture is used to achieve efficient integration of vehicle-side and roadside data through spatiotemporal alignment, feature enhancement, and redundancy elimination.

[0010] The present invention is further configured such that step S3 includes: S31 uses a spatiotemporal graph convolutional network to build a prediction model to predict traffic state vectors at future times; S32 calculates the optimal timing scheme with the objective function of minimizing the total delay time of all vehicles in the network and maximizing traffic efficiency; S33 takes reducing the number of times vehicles stop between adjacent intersections as the core objective and optimizes the phase difference between adjacent intersections to enable vehicles to pass through multiple intersections continuously at a reasonable speed.

[0011] This technical solution constructs a hierarchical, full-domain collaborative decision-making model. This model achieves intelligent control of traffic signals across the entire network through three progressive layers: a traffic situation prediction layer, a multi-objective optimization decision-making layer, and an intersection collaborative control layer. Each layer collaborates with the others, using real-time traffic flow data as a foundation, and through complex algorithms and model calculations, ultimately generates and executes optimized traffic signal control strategies to achieve efficient traffic flow at the road network level.

[0012] The present invention is further configured such that: step S12 includes: the large language model learns from historical traffic data and traffic rules, and then outputs a preliminary control scheme, the preliminary control scheme including a phase switching sequence and a green light duration and a red light duration pre-allocated at a certain time and in a certain direction, the phase switching sequence being able to define the lighting sequence of traffic lights in each direction.

[0013] The present invention is further configured such that step S13 includes: S131, construct a knowledge graph {V,E} for the transportation domain, where V is the set of knowledge nodes and E is the set of relationships between nodes. Encode the knowledge graph into a vector K using a graph embedding algorithm. S132, the hint vector is specifically represented as 1 minus the adaptive weight. Then multiply by the guiding hints generated based on the knowledge graph, plus adaptive weights. Multiply by real-time traffic scene vector .

[0014] In this technical solution, to compensate for the lack of knowledge in the field of transportation in large language models, this system constructs a knowledge graph in the field of transportation.

[0015] The present invention is further configured such that step S22 includes: using a joint Kalman filter algorithm to perform spatiotemporal alignment of multi-source data, and predicting and updating the system state through iteration.

[0016] In this technical solution, joint Kalman filtering is an efficient algorithm for processing multi-sensor data fusion, which predicts and updates the system state through iteration.

[0017] The present invention is further configured such that step S23 includes: S231, an anomaly detection algorithm is used to process the fused data to obtain cleaned data; principal component analysis algorithm is used to perform dimensionality reduction on the cleaned data; S232 uses an LSTM model to predict vehicle trajectories and a convolutional neural network to perform conflict warning analysis based on the predicted vehicle states.

[0018] In this technical solution, in order to ensure data quality and enable more efficient subsequent analysis, anomaly detection and feature dimensionality reduction operations are required.

[0019] The present invention is further configured such that the objective function comprises the sum of three parts: the first part is to minimize the total delay time of all vehicles in the network; the second part is to maximize the traffic capacity of the entire network; and the third part is a policy smoothing penalty term, which can avoid frequent abrupt changes in the traffic light control policy.

[0020] The present invention is further configured such that step S33 includes: using a genetic algorithm to optimize the phase difference parameters, and in the iterative process of the genetic algorithm, through selection, crossover and mutation operations, continuously evolving to generate a better phase difference combination scheme.

[0021] In this technical solution, to improve the search efficiency and optimization effect of the algorithm, the phase difference parameter is optimized by combining a genetic algorithm.

[0022] The intelligent traffic signal control method of the present invention can bring the following beneficial effects: This invention optimizes traffic flow allocation through multi-technology collaboration, thereby improving the efficiency, safety, and intelligence of urban transportation networks. By deeply integrating multi-source data and making dynamic decisions, it meets the collaborative optimization needs of intelligent connected vehicles and transportation facilities of different levels across the entire region. Attached Figure Description

[0023] Figure 1 This is a flowchart of an intelligent traffic signal control method for the entire region according to the present invention.

[0024] Figure 2 This is a schematic diagram of the framework of an intelligent traffic signal control method for the entire region according to the present invention. Detailed Implementation

[0025] Example 1 To address the issue that traditional traffic signal control cannot meet the collaborative optimization needs of intelligent connected vehicles and traffic facilities of different levels across the entire area, this embodiment proposes an intelligent traffic global signal control method, referencing... Figure 1 and Figure 2 It mainly includes the following steps.

[0026] Step S1: First, extract the multidimensional features of traffic data, integrate them to obtain a traffic scene vector, and combine them with the output of a large language model to obtain a prompt vector.

[0027] Step S1 mainly includes the following sub-steps.

[0028] Step S11: Extract the core features of the traffic scene, mainly including traffic flow sequence, queue length, speed distribution, and event encoding. Integrate the above features to obtain the traffic scene vector.

[0029] Step S12: Combining the traffic scene vectors mentioned above, a corresponding preliminary control scheme is generated through a large model, and a reinforcement learning module is introduced to dynamically adjust the preliminary control scheme.

[0030] More specifically, the large language model learns from historical traffic data and traffic rules to output a preliminary control scheme. This preliminary control scheme includes a phase switching sequence and pre-allocated green and red light durations for a specific time and direction. The phase switching sequence defines the lighting order of traffic lights in each direction. The LLM, through learning from historical traffic data and traffic rules, outputs a control strategy that conforms to basic traffic logic.

[0031] Step S13 involves constructing a knowledge graph for the transportation domain, adaptively fusing the transportation scene vector with the domain knowledge vector, and outputting the final prompt vector.

[0032] Step S13 mainly includes the following sub-steps.

[0033] Step S131: Construct a knowledge graph {V,E} for the transportation domain, where V is the set of knowledge nodes and E is the set of relationships between nodes. Encode the knowledge graph into a vector K using a graph embedding algorithm. Step S132, the specific representation of the hint vector is 1 minus the adaptive weights. Then multiply by the guiding hints generated based on the knowledge graph, plus adaptive weights. Multiply by real-time traffic scene vector .

[0034] In this embodiment, in order to achieve accurate digital representation of complex traffic scenarios, the present invention adopts a multi-dimensional feature fusion approach to map real-time traffic data into structured natural language prompts and construct traffic scenario vectors.

[0035] Step S2: Next, the multi-source data is fused, followed by anomaly detection and feature dimensionality reduction, and then deep analysis is performed.

[0036] Step S2 mainly includes the following sub-steps.

[0037] Step S21: Collect and characterize multi-source data, which includes vehicle-side data and roadside data.

[0038] Step S22: Perform spatiotemporal alignment processing on the above multi-source data, and then use an attention mechanism to fuse the multi-source data to obtain the final fused data.

[0039] More specifically, the joint Kalman filter (JKF) is used to achieve spatiotemporal alignment of multi-source data. The joint Kalman filter is an efficient algorithm for processing multi-sensor data fusion, which predicts and updates the system state through iteration.

[0040] Step S23: Perform anomaly detection and feature dimensionality reduction processing at the roadside edge device to obtain dimensionality-reduced data; the operation can perform trajectory prediction and conflict early warning processing to complete in-depth analysis.

[0041] Step S23 mainly includes the following sub-steps.

[0042] Step S231: The fused data is processed using an anomaly detection algorithm to obtain cleaned data; the cleaned data is then subjected to dimensionality reduction using a principal component analysis algorithm. In this embodiment, in order to ensure data quality and enable more efficient subsequent analysis, anomaly detection and feature dimensionality reduction operations are required.

[0043] Step S232: Vehicle trajectory prediction is performed based on the LSTM model, and conflict warning analysis is performed using a convolutional neural network based on the predicted vehicle state.

[0044] In this embodiment, a multimodal heterogeneous data fusion architecture is used to achieve efficient integration of vehicle-side and roadside data through spatiotemporal alignment, feature enhancement, and redundancy elimination.

[0045] Step S3: Again, predict the traffic flow trend of the road network, and generate and execute the optimized traffic signal control strategy through multi-objective optimization decision algorithm and intersection cooperative control algorithm.

[0046] Step S3 mainly includes the following sub-steps.

[0047] Step S31: A prediction model is constructed using a spatiotemporal graph convolutional network to predict the traffic state vector at future times.

[0048] Step S32: Using the objective function of minimizing the total delay time of all vehicles in the network and maximizing traffic efficiency, the optimal timing scheme is calculated.

[0049] The objective function consists of three parts: the first part is to minimize the total delay time of all vehicles in the network; the second part is to maximize the traffic capacity of the entire network; and the third part is a policy smoothing penalty term, which can avoid frequent changes in the traffic light control policy.

[0050] Step S33 aims to reduce the number of times vehicles stop between adjacent intersections by optimizing the phase difference between adjacent intersections, so that vehicles can pass through multiple intersections continuously at a reasonable speed.

[0051] More specifically, step S33 includes the following process: using a genetic algorithm to optimize the phase difference parameters, and continuously evolving to generate a better phase difference combination scheme through selection, crossover, and mutation operations during the iteration process of the genetic algorithm.

[0052] In this embodiment, to improve the search efficiency and optimization effect of the algorithm, the phase difference parameter is optimized by combining a genetic algorithm.

[0053] This embodiment can bring the following improvements and effects.

[0054] 1. Model-Driven Intelligent Decision Making: A hierarchical LLM decision framework is created, decomposing the general large language model (LLM) into a three-tier architecture: a policy planning layer, a domain adaptation layer, and a signal optimization layer. Through domain knowledge graph embedding, professional knowledge such as traffic regulations, intersection topology, and historical traffic flow is structurally injected into the model. Combined with deep reinforcement learning (DRL) algorithms, the system can generate dynamic signal control strategies based on real-time traffic conditions. An innovative decision tree visualization and explanation module is introduced to achieve interpretability of the control logic, solving the problem of the difficulty in verifying traditional black-box models.

[0055] 2. Deep Heterogeneous Data Fusion: A Spatiotemporal Alignment-Attention Fusion (STAF) method is proposed. Through spatiotemporal grid partitioning technology, multimodal information such as video streams, radar point clouds, and vehicle sensor data is uniformly mapped to a high-precision spatiotemporal coordinate system. A data association network is constructed based on a multi-head attention mechanism, dynamically allocating weights for different modal data to achieve millisecond-level latency data processing capabilities. Compared to traditional fusion methods, in complex weather and peak hour scenarios, the effective data utilization rate is improved by 40%, and the accuracy of traffic parameter recognition is increased to 98.7%.

[0056] 3. Global Collaborative Optimization Architecture: A three-tiered decision-making model (cloud-edge-device) is constructed. A global road network optimization engine is deployed in the cloud to generate macro-level traffic control strategies based on historical data and predictive models. Edge nodes dynamically adjust their regions based on real-time sensing data. Terminal devices (roadside units, intelligent traffic lights) execute refined control commands. A spatiotemporal predictive neural network predicts traffic flow distribution for the next 15-30 minutes, and a multi-objective optimization algorithm balances indicators such as traffic efficiency, queue length, and carbon emissions, achieving a 35% increase in road network-level green wave coverage.

[0057] 4. Closed-Loop Intelligent Control System: A closed-loop control process is designed across the entire vehicle-road-cloud chain. V2X communication technology enables bidirectional interaction between real-time vehicle status (speed, location, driving intention) and roadside facilities (traffic lights, sensors) and the cloud platform. Based on feedback control theory, the system uses traffic status data (such as delay time and traffic volume) after strategy execution as reward signals for reinforcement learning, dynamically optimizing decision model parameters. It supports adaptive periodic adjustments and rapid response to emergency events, forming an intelligent evolutionary system of "perception-decision-execution-feedback".

[0058] Example 2 Based on the above embodiment 1, a specific implementation example of an intelligent transportation full-domain signal control method is proposed, including the following steps.

[0059] Step S1: Extract multi-dimensional features from traffic data, integrate them to obtain a traffic scene vector, and combine them with the output of a large language model to obtain a prompt vector.

[0060] In order to achieve accurate digital representation of complex traffic scenarios, this embodiment adopts a multi-dimensional feature fusion approach to map real-time traffic data into structured natural language prompts and construct a traffic scenario description vector.

[0061] Taking a typical four-way intersection as an example, its multidimensional features include the following parts.

[0062] Traffic flow sequence: Traffic flow sequence Represented as It can depict the dynamic traffic flow of each lane. Here, i{1,2,3,4} represents four different directions of travel, j represents the lane number, and n is the total number of lanes in that direction. This represents the number of vehicles queuing and waiting to pass in lane j in the direction i at time t, obtained through video analysis technology or radar detection.

[0063] Queue length: Queue length Represented as This can quantify the degree of vehicle congestion at intersections. Here, m represents the total number of lanes in the i-th direction. This represents the number of vehicles queuing and waiting to pass in lane j in the direction i at time t. Specifically, this is obtained through video analysis technology or radar detection.

[0064] Velocity distribution: Velocity distribution Represented as Statistical methods were used to describe the characteristics of traffic flow speed. Among them, Let be the average speed of all vehicles traveling in the i-th direction at time t. The standard deviation corresponds to the speed, which measures the degree of speed dispersion. The data comes from the fusion calculation of the on-board OBD device and the roadside sensing unit.

[0065] Event coding: Event coding Represented as The system uses binary encoding to record sudden traffic events. Here, k represents the total number of predefined event types (e.g., traffic accidents, road construction, temporary traffic control). ,when The time indicates that the i-th type of event has occurred, and the other indicates that it has not occurred. The information is provided by the traffic management department's real-time notification and the AI ​​event recognition system.

[0066] Integrating the above features, we finally obtain a matrix representation of the traffic scene vector, which represents the traffic scene vector at the intersection at time t. It is a complete digital integration of the intersection's operational status, fully describing the intersection's operational status at time t; the matrix contains the traffic flow sequences for each of the four traffic directions (i=1 to i=4). Queue length velocity distribution and the code for sudden traffic events at time t. This vector encompasses dynamic parameters of traffic flow and information on emergencies at the intersection, comprehensively reflecting the overall operational status of the intersection at time t, and providing basic data support for traffic signal decisions in subsequent large language models.

[0067] After obtaining the traffic scene vector, a two-stage decision-making process (coarse-grained planning and fine-grained optimization) is adopted to achieve intelligent control of traffic signals. By combining the semantic understanding ability of Large Language Model (LLM) and the dynamic optimization ability of reinforcement learning, the accuracy and real-time performance of decision-making are improved.

[0068] Coarse-grained planning mainly includes the following: Based on the traffic scenario vectors mentioned above, and combined with a large language model, a preliminary control scheme is generated through pre-trained traffic control strategies. , which is represented as ,in, Define the lighting sequence of the traffic lights in each direction as the phase switching sequence; and These represent the pre-allocated green light and red light durations for direction i at time t, respectively. The large model can learn from historical traffic data and traffic rules to output control strategies that conform to basic traffic logic.

[0069] Preliminary control plan It includes two parts, one of which is the phase switching sequence. The sequence of traffic lights for each direction of travel must be clearly defined (e.g., first east to west, then south to north); when some lights are pre-assigned for each direction, and These represent the pre-allocated green light duration and red light duration in direction i at time t, respectively. The large language model learns from historical traffic data and traffic rules to generate a preliminary control scheme that conforms to basic traffic logic, providing an initial basis for subsequent fine-grained optimization.

[0070] Fine-grained optimization includes the following process: dynamically adjusting the initial scheme by introducing a reinforcement learning (RL) module; defining the state space S as... The current traffic scenario and initial strategy are taken as the system state; the action space A is... This indicates the duration of the green light in each direction. and red light duration Adjustment range; reward function Employing a multi-objective optimization design, the reward function... for Multiply by the first term, add Multiply by the second term, and add Multiply by the third term; The weighting coefficients (which need to be determined through simulation experiments or training with real-world data) are used to balance different optimization objectives; the first term is... This can minimize traffic congestion at intersections. The sum of the queue lengths in all four directions, the second term is... This can reduce vehicle waiting time. The third term is the sum of vehicle delay times in all directions. It is a phase switching conflict risk assessment value, calculated by analyzing lane flow direction and traffic light phase factors, which can avoid traffic conflicts caused by signal adjustments.

[0071] Subsequently, a knowledge graph KG for the transportation domain is constructed, which is represented as follows: Where V is the set of knowledge nodes (such as traffic regulations, road topology, and historical congestion patterns), and E is the set of relationships between nodes (such as the "intersection-lane" hierarchy and the "accident-delay" causal relationship). The knowledge graph is encoded into a vector K using graph embedding algorithms (such as TransE and GraphSAGE) to achieve a digital representation of domain knowledge.

[0072] During the prompt generation phase, a dynamic prompt fusion strategy is adopted to integrate real-time traffic scene vectors. Domain knowledge vectors Adaptive fusion is performed, and the final input is the cue vector of the large language model. For adaptive weights Multiply by real-time traffic scene vector Add 1 - subtract adaptive weights Multiply by the guiding prompts generated based on the knowledge graph. .

[0073] Among them, the adaptive weights are dynamically adjusted according to the complexity of the traffic scenario (e.g., increasing the weight of real-time data in congestion scenarios and emphasizing knowledge guidance in special event scenarios). To provide guidance based on knowledge graphs, knowledge nodes relevant to the current scenario are retrieved and converted into natural language commands, assisting the large language model in generating control strategies that better align with traffic logic. Through these domain knowledge enhancement and prompt optimization strategies, the large language model can fully leverage traffic expertise and real-time scenario information to output traffic signal control strategies that are more practical, rational, and efficient, laying a solid foundation for subsequent vehicle-road cooperative data processing and cloud-based collaborative decision-making.

[0074] Step S2 involves fusing the multi-source data, followed by anomaly detection and feature dimensionality reduction, and then performing deep analysis.

[0075] Step S2 mainly involves vehicle-road cooperative data processing, starting with multi-source data acquisition and characterization, including vehicle-side data. Include: Vehicle status: Represented as , which represents position, velocity, acceleration, and heading angle; Environmental perception: Represented as This refers to the detection frame and motion parameters of surrounding vehicles / pedestrians; Roadside data Include: Multi-sensor fusion: represented as Represented as ; Traffic situation: Represented as It represents the rate of change of traffic flow and queue length.

[0076] Then, spatiotemporal alignment and data fusion are performed, with joint Kalman filtering used for spatiotemporal data alignment.

[0077] In intelligent transportation scenarios, there are issues with inconsistent timestamps and spatial coordinates in the data collected by vehicle-side sensors (such as radar and cameras) and roadside equipment (such as traffic cameras and geomagnetic sensors). Joint Kalman filtering achieves spatiotemporal alignment through the following steps.

[0078] Prediction phase: based on the previous moment State estimation and system state transition matrix Predict the current moment status :in, Indicates based on Time information Predicting the state at any given moment; Describes the system state from Time's up The transition relationships at different times, such as the velocity and acceleration parameters in the vehicle motion model, are used to map the accurate state estimate of the previous time step to the current time step through the state transition matrix, thus obtaining the preliminary state prediction results.

[0079] Update phase: Combining the measured values ​​at the current moment The predicted state is corrected. equal Add Kalman gain Multiply Subtract the observation matrix Multiply .in, For the updated Time-state estimation is the final result that combines predicted and measured values; Kalman gain is the weight that balances the prediction and measurement values. Its calculation formula depends on the system's covariance matrix and observation noise. This refers to the raw measurement values ​​collected by the sensors at the current moment, such as the vehicle's real-time position and speed. The observation matrix maps the predicted values ​​of the system state space to the measurement space, for example, converting the three-dimensional coordinates of the vehicle into the two-dimensional imaging coordinates of the camera.

[0080] After completing spatiotemporal alignment, to further enhance the feature representation capability of the fused data, an attention mechanism is employed to achieve feature enhancement. Integrating vehicle-side data through attention mechanisms and roadside data , Weights of data features Multiply by data features The tiredness and Indicates the first The data features include vehicle-side data. and roadside data It covers information on vehicle speed and road congestion status. The weights for data features are calculated using a self-attention mechanism. This mechanism dynamically allocates weights by calculating the correlations between different features, enabling the model to focus on key information. For example, in intersection decision-making, if traffic congestion information reported by roadside equipment is detected, the weights of the corresponding features are adjusted accordingly. This will increase, thereby enhancing the proportion of this information in the fused data.

[0081] Subsequently, the roadside edge device performs data preprocessing, mainly anomaly detection and feature dimensionality reduction.

[0082] In intelligent transportation systems, roadside edge devices first acquire fused data, which combines multi-source heterogeneous traffic data collected from roadside sensors (such as cameras and radar), including information on vehicle position, speed, and direction of travel. To ensure data quality and enable more efficient subsequent analysis, anomaly detection and feature reduction operations are required.

[0083] For anomaly detection, specifically through anomaly detection algorithms... The fused data is processed to obtain cleaned data. The anomaly detection algorithm employs methods based on statistical distributions or machine learning, such as Isolation Forest and One-Class SVM, to identify outliers and erroneous data. Specifically, the algorithm analyzes the distribution of each data sample in the feature space; if a sample differs significantly from the majority of other samples, it is identified as an anomaly and removed, thus outputting cleaned data. .

[0084] For feature reduction, since the original data often has high dimensionality, it may contain redundant features and noise, which increases the complexity and time cost of subsequent calculations. Therefore, Principal Component Analysis (PCA) algorithm is used to perform feature reduction on the cleaned data. Dimensionality reduction processing is performed.

[0085] The core idea of ​​the PCA algorithm is to map the original data to a new orthogonal feature space through linear transformation, so that the data has the maximum variance in the new feature dimension, thereby extracting the most representative features. Here, the dimension after dimensionality reduction is set to 1. After PCA processing, the dimensionality-reduced data is obtained. .

[0086] The cloud performs deep analysis. After the data is pre-processed by the roadside edge device, it is transmitted to the cloud. The cloud can then use its powerful computing resources to perform deep analysis, which mainly includes two steps: trajectory prediction and conflict prediction.

[0087] For trajectory prediction, it can transform the dimensionality-reduced data As input, a Long Short-Term Memory (LSTM) network model is used for vehicle trajectory prediction. The LSTM model is a special type of recurrent neural network that can effectively process time-series data. Through memory units and gating mechanisms, it solves the gradient vanishing and gradient exploding problems existing in traditional RNNs, and can effectively capture long-term dependencies in the data. Based on historical data sequences, the model learns the patterns and rules of vehicle movement, and then predicts the vehicle's state at a future time t+1. ,state equal ,in, It includes predicted vehicle position and speed information.

[0088] For conflict warning, it is based on the predicted future state of the vehicle. This paper utilizes Convolutional Neural Networks (CNNs) for conflict early warning analysis. CNNs, through their convolutional, pooling, and fully connected layer structure, can automatically extract spatial features and patterns from data. In traffic scenarios, the predicted vehicle state data is converted into image or tensor forms suitable for CNN processing. The CNN model then analyzes the relative positions and speed changes between different vehicles to determine whether a potential traffic conflict exists at time t, and outputs the conflict assessment result. ,like If the output is true, it indicates a risk of conflict, and a warning should be issued promptly.

[0089] Step S3: Predict the traffic flow trend of the road network, and generate and execute the optimized traffic signal control strategy through a multi-objective optimization decision algorithm and an intersection cooperative control algorithm.

[0090] A hierarchical, full-domain collaborative decision-making model is constructed. This model achieves intelligent control of traffic signals across the entire network through three progressive layers: a traffic situation prediction layer, a multi-objective optimization decision-making layer, and an intersection collaborative control layer. Each layer collaborates with the others, using real-time traffic flow data as a foundation and employing complex algorithms and models to ultimately generate and execute optimized traffic signal control strategies, achieving efficient traffic flow at the road network level.

[0091] To predict traffic flow trends in road networks, a Spatiotemporal Graph Convolutional Network (STGCN) is used to construct the prediction model. To more accurately capture the dynamic changes in traffic flow across time and space, the traffic network is abstracted into a graph structure. , which is represented as In this context, node V represents each traffic intersection, with each node corresponding to an intersection in the actual road network; edge E reflects the traffic relationship between intersections. For example, if there is a main road connecting two intersections, their relationship strength is relatively high.

[0092] Model Input It integrates multi-dimensional traffic characteristic data, including vehicle flow. (Number of vehicles passing through the intersection per unit time), queue length (Length of the queue of vehicles waiting to cross the intersection), speed information (The average speed of vehicles passing through the intersection), that is By alternating between graph convolution and temporal convolution operations, the model can effectively uncover the spatial propagation characteristics and temporal evolution trends of traffic flow. for ; yes Traffic state vector at any given time (including traffic flow, queue length, and speed); Indicates to Time-state vector Intersection Association Matrix Perform the STGCN operation.

[0093] To measure the accuracy of the prediction results, the weighted mean squared error (WMSE) is used as the loss function. WMSE is equal to the weight coefficient of each intersection multiplied by the intersection's weight. Actual traffic flow minus predicted traffic flow The sum of squares, where, The actual traffic flow at intersection i is collected in real time using devices such as geomagnetic sensors and cameras. The corresponding predicted flow rate is output by the STGCN model; The weighting coefficients for each intersection are dynamically adjusted based on factors such as traffic flow volume and importance level. For example, traffic hub intersections, due to their greater impact on the overall road network, will have a higher weight than ordinary intersections, highlighting the importance of accurate predictions for key road sections.

[0094] Multi-objective optimization decision-making includes the following: In comprehensive traffic signal control, multiple interrelated and conflicting optimization objectives must be considered simultaneously. Therefore, an optimization objective function is defined. It consists of three parts: the objective function for Multiply , plus Multiply In addition Multiply .

[0095] in, , , These are weighting coefficients, dynamically adjusted through reinforcement learning or training on historical data to balance the priority of different objectives. For example, they can be increased during evening rush hour. Prioritize reducing vehicle delays. This represents minimizing the total delay time for all vehicles on the network. Let be the delay time for each vehicle at intersection i (the difference between the actual travel time and the ideal travel time). This indicates maximizing the overall vehicle throughput capacity of the network; Let i be the number of vehicles passing through intersection i per unit time. To smooth out the penalty term in the strategy, frequent abrupt changes in traffic light control strategies can be avoided, ensuring a smooth transition of traffic flow. If the green light duration changes too much between two adjacent signal cycles, The value will increase.

[0096] To solve the aforementioned multi-objective optimization problem, this embodiment employs a mixed-integer programming (MIP) method. In practical traffic signal control, two constraints need to be satisfied. The first constraint is... , It is the total number of phases. It is the first The first intersection The duration of the green light for each phase, The first constraint is the maximum duration of the traffic light cycle; this constraint prevents excessively long green lights in one direction from causing congestion in other directions. The second constraint is... , , It is the first The first intersection Red light duration for each phase, It is the minimum green light duration (e.g., 15 seconds, to ensure that slow-moving vehicles can pass). The minimum red light duration is a constraint that ensures vehicle safety and pedestrian crossing needs. The mixed-integer programming algorithm needs to search for the globally optimal timing scheme while satisfying these constraints.

[0097] For intersection cooperative control, a dynamic green wave optimization algorithm is designed to achieve traffic flow coordination between adjacent intersections. This algorithm aims to reduce the number of times vehicles stop between adjacent intersections by optimizing the phase difference between them. This enables vehicles to pass through multiple intersections consecutively at a reasonable speed. Specific optimization objectives... for T is , among which, To pass through adjacent intersections during the study period and The total number of vehicles; Indicates the first The car at the intersection and The number of times a vehicle stops is recorded by onboard sensors or roadside equipment to determine its start-stop status.

[0098] To improve the search efficiency and optimization effect of the algorithm, a genetic algorithm (GA) is used to optimize the phase difference parameters. During the iterative process of the genetic algorithm, through operations such as selection, crossover, and mutation, better phase difference combinations are continuously generated. Specifically, it includes the following six steps.

[0099] Initialize the population: Randomly generate a set of phase difference parameter combinations as the initial population. Each combination represents a possible phase difference configuration for adjacent intersections, for example... Second, Second.

[0100] Fitness assessment: Based on the above optimization objective function, calculate the fitness value for each individual (phase difference combination). The higher the fitness, the fewer times the vehicle stops between adjacent intersections under that phase difference combination.

[0101] Selection operation: Based on fitness values, selection strategies such as roulette wheel selection are used to select the better individuals to enter the next generation. In the roulette wheel selection strategy, individuals with higher fitness have a greater probability of being selected, simulating the survival of the fittest mechanism in nature.

[0102] Crossover operation: Perform a crossover operation on selected individuals, exchanging some gene segments to generate new individuals. For example, swapping some parameters of two phase difference combinations produces a new phase difference scheme.

[0103] Mutation operation: Mutating an individual's genes with a certain probability introduces new parameter combinations. This prevents the algorithm from getting trapped in local optima, for example, by randomly changing the value of a phase difference.

[0104] Iteration Termination: Repeat the above steps until the preset number of iterations or fitness value convergence condition is met, and finally output the optimal phase difference parameter to realize dynamic green wave coordinated control between adjacent intersections and adapt to the traffic demand under different time periods and different traffic flows.

Claims

1. A smart traffic global signal control method, characterized by, Includes the following steps: S1: Extract multi-dimensional features from traffic data, integrate them to obtain a traffic scene vector, and combine them with a large language model to output a prompt vector; S2 performs fusion processing on multi-source data, followed by anomaly detection and feature dimensionality reduction, and then performs deep analysis. S3 predicts the traffic flow trend of the road network and generates and executes optimized traffic signal control strategies through multi-objective optimization decision-making algorithms and intersection cooperative control algorithms.

2. The intelligent traffic signal control method for the entire region according to claim 1, characterized in that, Step S1 includes: S11, Obtain features from the traffic scene, including traffic flow sequence, queue length, speed distribution, and event encoding, and integrate the features to obtain a traffic scene vector; S12, based on traffic scene vectors, generates a preliminary control scheme through a large model, and introduces a reinforcement learning module to dynamically adjust the preliminary control scheme; S13: Construct a knowledge graph for the transportation domain, adaptively fuse transportation scene vectors and domain knowledge vectors, and finally output prompt vectors.

3. A method for intelligent traffic signal control across the entire region according to claim 1 or 2, characterized in that, Step S2 includes: S21, collect and characterize multi-source data, including vehicle-side data and roadside data; S22, perform spatiotemporal alignment processing on multi-source data, and fuse multi-source data through an attention mechanism to obtain fused data; S23, the roadside edge device performs anomaly detection and feature dimensionality reduction operations to obtain the dimensionality-reduced data; the cloud performs trajectory prediction and conflict warning to complete in-depth analysis.

4. The intelligent traffic signal control method for the entire region according to claim 3, characterized in that, Step S3 includes: S31 uses a spatiotemporal graph convolutional network to build a prediction model to predict traffic state vectors at future times; S32 calculates the optimal timing scheme with the objective function of minimizing the total delay time of all vehicles in the network and maximizing traffic efficiency; S33 takes reducing the number of times vehicles stop between adjacent intersections as the core objective and optimizes the phase difference between adjacent intersections to enable vehicles to pass through multiple intersections continuously at a reasonable speed.

5. The intelligent traffic signal control method for the entire region according to claim 2, characterized in that, Step S12 includes: the large language model learns from historical traffic data and traffic rules, and then outputs a preliminary control scheme. The preliminary control scheme includes a phase switching sequence and a pre-allocated green light duration and red light duration in a certain direction at a certain time. The phase switching sequence can define the lighting sequence of the traffic lights in each direction.

6. The intelligent transportation global signal control method according to claim 2, characterized in that, Step S13 includes: S131, construct a knowledge graph {V,E} for the transportation domain, where V is the set of knowledge nodes and E is the set of relationships between nodes. Encode the knowledge graph into a vector K using a graph embedding algorithm. S132, the hint vector is specifically 1 minus the adaptive weight. Then multiply by the guiding hints generated based on the knowledge graph, plus adaptive weights. Multiply by real-time traffic scene vector .

7. The intelligent traffic signal control method for the entire region according to claim 3, characterized in that, Step S22 includes: using a joint Kalman filter algorithm to perform spatiotemporal alignment of multi-source data, and predicting and updating the system state through iteration.

8. The intelligent traffic signal control method for the entire region according to claim 3, characterized in that, Step S23 includes: S231, an anomaly detection algorithm is used to process the fused data to obtain cleaned data; principal component analysis algorithm is used to perform dimensionality reduction on the cleaned data; S232 uses an LSTM model to predict vehicle trajectories and a convolutional neural network to perform conflict warning analysis based on the predicted vehicle states.

9. The intelligent transportation global signal control method according to claim 4, characterized in that, The objective function consists of three parts: the first part is to minimize the total delay time of all vehicles in the network; the second part is to maximize the traffic capacity of the entire network; and the third part is a policy smoothing penalty term, which can avoid frequent changes in the traffic light control policy.

10. A method for intelligent traffic signal control across the entire region according to claim 4 or 9, characterized in that, Step S33 includes: using a genetic algorithm to optimize the phase difference parameters. During the iteration process of the genetic algorithm, through selection, crossover, and mutation operations, a better phase difference combination scheme is continuously generated.

Images