Traffic signal control method and system based on vehicle and road cloud multi-modal data fusion

Abstract

The invention relates to the technical field of signal devices, and discloses a traffic signal control method and system based on vehicle-road cloud multi-modal data fusion, and the method comprises the steps: collecting multi-modal traffic data synchronously in real time through a vehicle-end sensor, road-side sensing equipment and a cloud Internet platform; fusing the heterogeneous data by adopting a space-time alignment algorithm, and constructing a standardized space-time feature matrix; traffic flow prediction is carried out based on a multi-layer space-time diagram neural network trained by a federated learning mechanism, and a signal control instruction is generated through reinforcement learning and a multi-objective optimization model; and issuing the green wave parameter, the dynamic timing scheme and the cross-domain coordination strategy to a roadside signal machine through the cloud edge coordination architecture to execute control. The problems that in the prior art, low-delay private network communication cannot be achieved, the data fusion efficiency is low, unmanned driving is not supported, and the deployment cost is high are solved, and the purposes of low-delay communication, high reliability and low risk are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal device, and particularly relates to a traffic signal control method and system based on vehicle-road cloud multi-modal data fusion. BACKGROUND

[0002] Current urban traffic signal control generally faces three core challenges: traditional schemes rely on single-point detection devices (such as ground coils and fixed cameras) to obtain local traffic data, which can only cover the traffic state within a 200-meter range of the intersection, and the data update frequency is low, which cannot realize real-time sensing of vehicle trajectory, driving intention and other dynamic information. The data obtained by these single-point detection devices has single dimension, which can only reflect the basic information such as the number of vehicles passing through, and cannot obtain the speed change, following distance and other important details of traffic signal control, making it difficult to fully depict the complex traffic operation state. The fixed timing scheme based on historical data (such as four-phase timing control) has a long update cycle (usually ≥1 hour), which is difficult to cope with dynamic scenarios such as tidal traffic and sudden accidents. This fixed timing scheme lacks dynamic adaptability, and when the traffic flow changes, it is easy to cause vehicles in some directions to wait for a long time, while the road resources in another direction are idle, aggravating the congestion at the intersection and reducing the road traffic efficiency. Independent deployment of intelligent traffic systems in small and medium-sized cities requires a large amount of funds for server clusters, storage devices and algorithm platform construction, and the annual operation and maintenance cost accounts for a high proportion. Due to the lack of cloud collaboration capability, cross-regional signal coordination needs to deploy additional special communication equipment, further increasing the cost and leading to insufficient equipment coverage.

[0003] Although the existing vehicle-road cloud scheme realizes data interconnection, it has key technical shortcomings: the multi-modal data fusion efficiency is low. The time and space alignment error of vehicle-end GPS trajectory (latitude and longitude sequence), roadside radar point cloud (polar coordinates) and Internet traffic (grid congestion index) is more than 30 seconds, and the traffic flow prediction accuracy after feature fusion is only 75%. Different modal data has significant differences in feature dimension, data structure and semantic expression, and traditional fusion methods are difficult to effectively mine the potential association between data, leading to insufficient understanding of complex traffic scenarios by the model. At the same time, the data cleaning and noise reduction processing in the fusion process is insufficient, which further affects the accuracy and reliability of traffic flow prediction and signal control decision.

[0004] For example, the Chinese patent with publication number CN118785114B provides the following technical solution, a vehicle-road cooperation system and method based on a double-intelligent private network and a storage medium are disclosed. The system includes a vehicle terminal system installed on a vehicle, a roadside subsystem, and a cloud platform subsystem. The roadside subsystem includes a base station applied to the double-intelligent private network. The vehicle terminal system collects real-time driving data of vehicles in the road network and sends it to the cloud platform subsystem through the base station. The roadside subsystem collects real-time road perception information in the road network and sends it to the cloud platform subsystem through the base station. The cloud platform subsystem dynamically identifies the obstacle areas that affect vehicle driving in the road network and the target vehicles passing through the obstacle areas based on the received driving data and road perception information, creates vehicle driving parameters for each target vehicle, and sends the vehicle driving parameters to the vehicle terminal system installed on each target vehicle through the base station. Therefore, the application can apply vehicle-road cooperation to the field of autonomous driving while reducing the risk of vehicle driving. However, the above-mentioned vehicle-road cooperation system and method based on a double-intelligent private network and a storage medium cannot dynamically plan routes, rely on manual preset maps and calibration parameters, cannot automatically identify the import flow direction of vehicles at any intersection, and the system fails when the vehicle deviates from the preset route. SUMMARY

[0005] The present application solves the problems of low delay private network communication, low data fusion efficiency, no support for autonomous driving, and high deployment cost in the prior art, and proposes a traffic signal control method and system for vehicle-road cloud multi-modal data fusion, achieving the purposes of low delay communication, high reliability, and low risk.

[0006] Furthermore, the present application compresses the spatio-temporal alignment error of multi-source heterogeneous data through a spatio-temporal alignment algorithm, solves the problem of low prediction accuracy caused by data fragmentation in traditional solutions, and improves the accuracy of global traffic situation awareness; combines the enhanced spatio-temporal graph neural network based on federated learning with reinforcement learning dynamic timing, reduces the traffic flow prediction time, and realizes fast dynamic iteration of the timing scheme, breaks through the bottleneck of fixed timing that cannot respond to sudden traffic, and reduces intersection delay.

[0007] To achieve the above-mentioned purposes, the present application adopts the following technical solutions: A traffic signal control method for vehicle-road cloud multi-modal data fusion, comprising: Real-time synchronous acquisition of multi-modal traffic data by vehicle terminal sensors, roadside perception equipment, and a cloud internet platform; Fusion of heterogeneous data using a spatio-temporal alignment algorithm to construct a standardized spatio-temporal feature matrix; Traffic flow prediction by a multi-layer spatio-temporal graph neural network trained based on a federated learning mechanism, generation of a signal control scheme by reinforcement learning and a multi-objective optimization model; The green wave parameters, the dynamic timing scheme and the cross-domain coordination strategy are issued to a roadside signal machine for execution control through a cloud edge coordination architecture.

[0008] Real-time synchronous acquisition ensures the timeliness of multi-source data, the spatio-temporal alignment algorithm solves the problem of heterogeneous data fusion, the large model trained by federated learning improves the prediction accuracy, reinforcement learning and multi-objective optimization realize dynamic decision-making, cloud edge coordination issuance guarantees low-latency execution of control instructions, forming an efficient closed-loop system.

[0009] A kind of traffic signal control system of vehicle-road cloud multi-modal data fusion, including vehicle end perception module and roadside intelligent module and cloud platform module;The vehicle end perception module is connected with roadside intelligent module by communication link, real-time upload vehicle spatio-temporal trajectory data and local traffic flow matrix;The roadside intelligent module deploys edge computing node, processes multi-modal sensor data by spatio-temporal calibration and feature extraction, generates intersection state vector, and compresses and uploads to cloud platform module.

[0010] Vehicle end module provides high-precision vehicle dynamics, roadside module realizes intersection-level feature processing, communication link guarantees real-time interaction of data, edge computing node supports localized processing, and provides structured input for cloud decision-making.

[0011] As preferred, the synchronous acquisition of multi-modal traffic data includes: using an extended Kalman filter algorithm to fuse positioning and inertial data through a vehicle terminal to generate millimeter-level precision vehicle spatio-temporal trajectory; real-time acquisition of engine start-stop signals and speed sensor data through a vehicle interface, identification of parking events based on predefined speed thresholds and duration; real-time reception of six-dimensional dynamic information of surrounding vehicles through a vehicle-road communication interface, construction of a local traffic flow matrix; acquisition of environmental data through temperature and humidity sensors and raindrop sensors.

[0012] Millimeter-level trajectory provides accurate vehicle dynamics, predefined threshold identifies parking events to support emission optimization, six-dimensional dynamic information constructs a microscopic traffic flow model, and environmental data perception enables special weather strategy adaptation, fully breaking through the limitations of traditional single-point detection.

[0013] As preferred, the use of a spatio-temporal alignment algorithm to fuse heterogeneous data includes: designing a multi-layer fusion architecture to perform spatio-temporal calibration on vehicle end data and roadside data and Internet data, unifying the coordinate systems and time stamps of different modal data, and minimizing spatio-temporal errors; processing heterogeneous data through feature extraction methods, including vehicle end features and roadside features and Internet features, to construct a standardized feature vector covering multiple dimensions; wherein the vehicle end features include average speed and parking frequency density, the roadside features include traffic volume and queue length, and the Internet features include real-time congestion index and historical traffic deviation.

[0014] The multi-layer architecture spatio-temporal calibration eliminates the differences in coordinates and time sequences, the feature extraction standardizes the multi-source heterogeneous data, the vehicle-side and roadside internet features are cooperatively outputted to macro and micro traffic situations, and high-quality inputs are provided for the model.

[0015] Preferably, the construction of the standardized spatio-temporal feature matrix includes calculating the correlation weight between road network nodes based on historical traffic flow mutual information, and constructing a directed and weighted road network correlation graph; the spatial features are processed by using a graph convolution operation, the graph convolution complexity is calculated and optimized by using a polynomial recursion; a time series processing method is used to extract the dynamic change features of the traffic flow by using a gated recurrent unit and a self-attention mechanism, and a standardized multi-dimensional spatio-temporal feature matrix is constructed.

[0016] The mutual information correlation weight strengthens the dependence of the key path, the polynomial recursion optimizes the calculation efficiency of the spatial features, the gated recurrent unit and the attention mechanism cooperatively capture the time sequence rules of the flow, and the standardized matrix improves the generalization ability of the model.

[0017] Preferably, the multi-layer spatio-temporal graph neural network trained based on the federated learning mechanism performs traffic flow prediction, which includes: using a hierarchical federated learning framework to perform local model training on the edge node, injecting privacy protection noise and dynamically adjusting the training batch size based on real-time data flow; performing global model aggregation on the cloud server, using the actual number of intersections in each region as a weighting factor to calculate global parameters, and introducing a parameter verification mechanism to filter abnormal values; the model difference is evaluated by using the Euclidean distance, and an adaptive adjustment mechanism including increasing the training rounds and transfer learning is triggered.

[0018] The hierarchical framework realizes distributed learning, the privacy noise guarantees data security, the dynamic batch optimizes resource utilization, the weighted aggregation strengthens the influence of high-flow areas, the parameter verification ensures the stability of the model, and the adaptive mechanism accelerates the convergence in special scenarios.

[0019] Preferably, the signal control scheme generated by the reinforcement learning and the multi-objective optimization model includes: identifying high-delay road segments based on road segment delay indexes, and creating green wave corridors; constructing a multi-objective optimization function including minimizing total delay, number of stops, and speed deviation, and using a phase difference coordination mechanism to constrain adjacent intersections; using a decomposition iteration algorithm, combining historical traffic flow data to establish a prior distribution, and dynamically updating parameters by fusing real-time data.

[0020] The delay index triggers accurate identification of key road segments, the multi-objective function balances the traffic efficiency and environmental protection demand, accurately locates high-delay road segments, balances the traffic efficiency and environmental protection demand, and realizes rapid and dynamic convergence of parameters.

[0021] As preferred, the signal control scheme generated by the multi-objective optimization model through reinforcement learning further comprises: constructing a state space based on real-time traffic conditions, the state space including real-time flow and delay index and traffic period type; wherein the state space further comprises phase state and queue length and road segment average speed and weather state; defining an action space including adjustable signal control parameters, including signal cycle adjustment amount and adjacent direction phase difference adjustment amount and priority coefficient of straight and turning phases; designing a reward function, evaluating the effect of action space output signal control parameters through weighted combination of delay reduction reward and stop number reduction reward and capacity improvement reward and policy stability reward, and dynamically adjusting the weight coefficients of each reward item using a fuzzy logic controller; using a deep deterministic policy gradient algorithm, constructing a policy network and a value network, using the state space as input, the action space as output, and the reward function as optimization objective, to generate a dynamic timing scheme.

[0022] The fine reinforcement learning control is realized, the optimal timing strategy is output through full-dimensional state perception and flexible parameter adjustment, and a dynamic reward mechanism is combined.

[0023] As preferred, the control executed by the roadside signal machine under the cloud edge collaborative architecture comprises the following steps: generating green wave band parameters, a dynamic timing scheme and cross-domain collaborative strategy control instructions through a cloud platform, processing by a mobile edge computing node, realizing the collaboration of cloud model decision and edge real-time execution; the control instructions are sent to the roadside signal machine through a message protocol to perform phase switching and timing adjustment; the green wave operation state is visualized and monitored and intelligently diagnosed, and an optimization report is generated based on deviation analysis.

[0024] The control instructions are ensured to be efficiently landed, the decision-making delay is shortened through edge cloud collaboration, and closed-loop optimization is realized through visualized diagnosis.

[0025] As preferred, the cloud platform module comprises a data center and a large model engine, the data center receives and fuses trajectory data of a vehicle end perception module, state vectors of a roadside intelligent module and Internet traffic data, and constructs a space-time database; the large model engine processes the fused feature vectors based on a space-time graph neural network, and outputs green wave parameters, a dynamic timing scheme and collaborative strategy instructions, and the decision instructions of the large model engine are sent to the roadside intelligent module through a low-latency link to drive a signal control device to execute the dynamic timing scheme.

[0026] The data center integrates multiple sources of information to construct a space-time database, the large model engine processes fused features to generate a control strategy, and a low-latency link ensures that instructions are quickly sent, so that an intelligent closed loop from data analysis to signal control is realized.

[0027] Compared with the prior art, the present application has the following advantages.

[0028] 1. The present application constructs a full-factor perception network with millisecond-level response through the deep collaboration of vehicle-road-cloud multi-source data. The dynamic timing algorithm based on reinforcement learning optimizes signal control in real time, quickly adjusts phase length and cycle when traffic flow suddenly changes, effectively deals with tidal traffic and sudden congestion. Combined with intelligent green wave coordination technology, the system significantly improves the continuity of trunk road traffic, greatly reduces the number of vehicle stops and intersection delays, realizes the leap-forward growth of regional traffic efficiency, and makes the traffic flow run more smoothly and efficiently.

[0029] 2. The present application relies on a multi-modal fusion core algorithm, breaks through the spatio-temporal alignment bottleneck of heterogeneous data, and compresses the feature fusion error of vehicle-end trajectory, roadside perception and Internet data to a very low level. The hierarchical federated learning framework enhances the model generalization ability through a dynamic parameter aggregation mechanism while ensuring data privacy and security, and ensures high-precision prediction and decision-making ability when deployed across regions. This technology system significantly improves the accuracy of analyzing complex road network traffic conditions, providing strong and reliable intelligent support for signal control.

[0030] 3. The present application adopts a cloud-based large model and edge lightweight inference collaborative architecture to realize a SaaS service mode with almost zero localization deployment. Cloud-based centralized training supports elastic resource scheduling, and edge nodes only need to perform efficient data compression and real-time inference, greatly reducing bandwidth and hardware requirements. This mode reduces the initial investment of small and medium-sized cities to a small fraction of traditional solutions, significantly reduces operation and maintenance costs, and supports visual interactive operation, greatly shortens the deployment time of emergency scenarios, and promotes the rapid popularization and application of intelligent transportation technology. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The present application is a whole flow chart of a traffic signal control method for vehicle-road-cloud multi-modal data fusion.

[0032] Figure 2 The present application is a flow chart of a spatio-temporal alignment algorithm for fusing heterogeneous data in a traffic signal control method for vehicle-road-cloud multi-modal data fusion.

[0033] Figure 3 The present application is a parameter diagram of a defined action space for a traffic signal control method for vehicle-road-cloud multi-modal data fusion.

[0034] Figure 4 The present application is a parameter diagram of a reward function for a traffic signal control method for vehicle-road-cloud multi-modal data fusion.

[0035] Figure 5 The present application is a parameter diagram of a state space for a traffic signal control method for vehicle-road-cloud multi-modal data fusion. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical scheme and advantages of the present disclosure clearer, the embodiments of the present disclosure will be described in further detail below with reference to the drawings. The proportions of various components are not drawn according to true proportions, and the proportions and sizes shown in the drawings should not limit the essential technical scheme of the present disclosure. These embodiments do not describe all the details, and the present disclosure is not limited to the specific embodiments described.

[0037] Referring to Figures 1-5 As shown in the drawings, a traffic signal control method based on vehicle-road-cloud multi-modal data fusion comprises the following steps: Real-time synchronous acquisition of multi-modal traffic data by vehicle-end sensors, road-side perception devices and cloud-end Internet platforms; Fusion of heterogeneous data by using a space-time alignment algorithm to construct a standardized space-time feature matrix; Traffic flow prediction by a multi-layer space-time graph neural network trained based on a federated learning mechanism, generation of a signal control scheme by reinforcement learning and a multi-objective optimization model; Downlink of green wave parameters, dynamic timing schemes and cross-domain coordination strategies to road-side signal machines for execution control through a cloud-edge collaborative architecture.

[0038] A traffic signal control system based on vehicle-road-cloud multi-modal data fusion comprises a vehicle-end perception module, a road-side intelligent module and a cloud-end platform module. The vehicle-end perception module is connected to the road-side intelligent module through a communication link and uploads vehicle space-time trajectory data and local traffic flow matrix in real time. The road-side intelligent module deploys an edge computing node, processes multi-modal sensor data through space-time calibration and feature extraction, generates an intersection state vector and compresses and uploads it to the cloud-end platform module.

[0039] The present disclosure focuses on the field of intelligent transportation engineering and cloud computing technology, and proposes a traffic signal control large model SaaS system and method based on a vehicle-road-cloud collaborative architecture. The scheme integrates vehicle terminal sensor data, road-side perception device collected information and Internet platform dynamic data to build a multi-modal data fusion system. Based on the cloud-end deployed traffic signal control large model, the innovative application of intelligent software as a service (SaaS) is realized. The system builds a full-link closed-loop technical system of "data acquisition - model decision - service delivery", and solves the dynamic self-adaptive problem in urban road network traffic signal control.

[0040] Its application scenarios cover mainline green wave coordination control, regional traffic signal coordination optimization, and emergency response in complex traffic scenarios such as traffic accidents and large-scale activities, which can effectively improve the efficiency of urban traffic operation and promote the intelligent and refined development of traffic management. The technology system relies on advanced cloud computing, big data and artificial intelligence technology, and comprehensively uses multi-modal data fusion algorithm to extract and analyze the features of vehicle sensor data, roadside sensing device information and cloud traffic big data; through the time and space sequence prediction model, based on historical traffic flow, real-time traffic and prediction data, a traffic flow trend model is constructed; using reinforcement learning optimization strategy, dynamically adjust the signal timing scheme according to the traffic state of different periods and different intersections, and finally through the deep mining and analysis of massive multi-source heterogeneous data, provide a more intelligent and efficient solution for traffic signal control. In the context of the increasingly serious problem of urban traffic congestion, its research and application are of great significance to improve the efficiency of urban traffic operation and improve the travel experience of residents.

[0041] As Figure 1 With Figure 2 In an embodiment, Figure 1 is the overall flowchart of a traffic signal control method of the vehicle-road cloud multi-modal data fusion of the application, Figure 2 is the flowchart of the time and space alignment algorithm of the traffic signal control method of the vehicle-road cloud multi-modal data fusion of the application. First, real-time synchronization of multi-modal traffic data is collected through vehicle-end sensors, roadside sensing devices and cloud internet platforms, including generating vehicle time and space trajectory by fusing positioning and inertial data through extended Kalman filtering algorithm through vehicle terminal, identifying parking events by real-time acquisition of engine start-stop signals and speed sensor data through vehicle interface, constructing local traffic flow matrix by real-time receiving of dynamic information of surrounding vehicles through vehicle-road communication interface, and collecting environmental data through temperature and humidity sensors and raindrop sensors.

[0042] Then, the time and space alignment algorithm is used to fuse heterogeneous data, a multi-layer fusion architecture is designed for time and space calibration processing to unify the coordinate system and timestamp to minimize the time and space error, the vehicle-end features such as average speed and parking frequency density of road section, roadside features such as flow and queue length, and internet features such as real-time congestion index and historical flow deviation are processed through feature extraction method to construct standardized feature vector. Then, the standardized time and space feature matrix is constructed, the correlation weight between road network nodes is calculated based on historical traffic flow mutual information to construct a directed weighted road network correlation graph, the spatial features are processed by graph convolution operation and the complexity of graph convolution is calculated by polynomial recursion, and the dynamic change features of traffic flow are extracted by using time series processing method through gated recurrent unit and self-attention mechanism.

[0043] Subsequently, the application trains a multi-layer spatio-temporal graph neural network based on a federated learning mechanism to perform traffic flow prediction, adopts a hierarchical federated learning framework to perform local model training on an edge node, injects privacy protection noise, and dynamically adjusts the training batch size, performs global model aggregation on a cloud server, calculates global parameters using the actual number of intersections in each region as a weighting factor, and introduces a parameter verification mechanism to filter abnormal values, and triggers an adaptive adjustment mechanism including increasing the training rounds and transfer learning by evaluating the model difference through the Euclidean distance.

[0044] After that, the application generates a signal control scheme through reinforcement learning and a multi-objective optimization model, identifies high delay road segments based on road segment delay indices to create green wave corridors, constructs a multi-objective optimization function including minimizing total delay and number of stops and speed deviation, and uses a phase difference coordination mechanism to constrain adjacent intersections, uses a decomposition iteration algorithm to establish a prior distribution combined with historical traffic flow data and dynamically updates parameters through real-time data fusion; at the same time, the application constructs a state space based on real-time traffic conditions, including real-time flow and delay index, traffic period type, phase state, queue length, road average speed, and weather state, defines an action space including signal cycle adjustment amount and adjacent direction phase difference adjustment amount, and priority coefficient of straight and turning phases, designs a reward function to evaluate action effect through weighted combination of delay reduction reward, stop number reduction reward, capacity improvement reward, and policy stability reward, uses a deep deterministic policy gradient algorithm to construct a policy network and a value network to generate a dynamic timing scheme.

[0045] Finally, the application delivers the green wave parameters, dynamic timing scheme, and cross-domain coordination strategy to roadside signal machines for execution control through a cloud edge coordination architecture, and generates control instructions through a cloud platform to use mobile edge computing nodes to process and realize the coordination of cloud model decision and edge real-time execution, the control instructions are delivered to roadside signal machines through a message protocol to perform phase switching and timing adjustment, and the green wave operation state is visualized and intelligently diagnosed to locate faults based on deviation analysis and generate an optimization report.

[0046] As Figures 3 to 5 shown in an embodiment, Figure 3 is a parameter diagram for defining an action space of a traffic signal control method of the application, Figure 4 is a parameter diagram of a reward function of a traffic signal control method of the application, Figure 5 is a parameter diagram of a state space of a traffic signal control method of the application. First, the application constructs a state space based on real-time traffic conditions, which includes real-time flow and delay index, traffic period type, and in addition, phase state, queue length, road average speed, and weather state.

[0047] Then, an action space is defined, which contains adjustable signal control parameters, including signal cycle adjustment amount, adjacent direction phase difference adjustment amount and priority coefficient of straight and turning phases.

[0048] Next, a reward function is designed, which evaluates the effect of the action space output signal control parameters through a weighted combination of delay reduction reward, parking frequency reduction reward, traffic capacity improvement reward and policy stability reward, and dynamically adjusts the weight coefficients of each reward item using a fuzzy logic controller.

[0049] Finally, a deep deterministic policy gradient algorithm is used to construct a policy network and a value network, using the state space as input, the action space as output, and the reward function as the optimization target to generate a dynamic timing plan.

[0050] The application also designs a vehicle-road cloud multi-modal data fusion traffic signal control system, which includes a vehicle-end perception module, a roadside intelligent module and a cloud platform module. The vehicle-end perception module is connected to the roadside intelligent module through a communication link and uploads vehicle spatio-temporal trajectory data and local traffic flow matrix in real time. The roadside intelligent module deploys edge computing nodes, processes multi-modal sensor data through spatio-temporal calibration and feature extraction, generates intersection state vectors, and compresses and uploads them to the cloud platform module. The application also relates to a green wave line operation effect diagnosis method, which includes: dynamically drawing a time-distance graph and displaying green wave pass rate; calculating speed deviation, phase difference deviation and delay deviation and comparing them with preset thresholds; locating faults based on a Bayesian network and generating an optimization report.

[0051] The cloud platform module includes a data hub and a large model engine. The data hub receives and fuses trajectory data from the vehicle-end perception module, state vectors from the roadside intelligent module and internet traffic data, and constructs a spatio-temporal database. The large model engine processes fused feature vectors based on a spatio-temporal graph neural network, outputs green wave parameters, dynamic timing plans and coordination strategy instructions, and sends decision instructions to the roadside intelligent module through a low-latency link to drive signal control devices to execute dynamic timing plans.

[0052] In another embodiment, the application uses the following specific scheme: The system architecture design includes a vehicle-road cloud coordination layer, which specifically includes: 1. Vehicle-end perception module A global dynamic data acquisition system is constructed to achieve high-precision perception of vehicle state: Multi-source data collection: C-V2X vehicle terminal (compliant with 3GPP Rel-16 standard) integrates U-Blox M8N GPS module (positioning accuracy ≤2.5m) and IMU unit, fuses positioning and inertial data through extended Kalman filtering algorithm, generates millimeter-level precision vehicle space-time trajectory, specifically vehicle trajectory is composed of time stamp t i , corresponding latitude and longitude coordinate sequence (x i , y i ) and contains n data points. Upload frequency 10Hz. Real-time acquisition of engine start-stop signal through vehicle OBD interface, combined with speed sensor data (accuracy ±1km / h), identify parking events (defined as speed ≤5km / h and duration ≥10 seconds) through state machine algorithm, count the number of parking per minute. Combined with environmental monitoring equipment such as temperature and humidity sensors, raindrop sensors, real-time collection of weather and environmental data. When detecting heavy weather such as rain and fog, automatically adjust data collection frequency and signal control strategy, such as reducing speed threshold, extending pedestrian crossing time, improving safety and reliability of traffic control in special environment.

[0053] Net interaction capability: realize V2V communication through PC5 interface, real-time acquisition of relative position, speed, steering intention and other 6-dimensional information of vehicles within 500 meters, build local traffic flow matrix, provide micro-dynamic data support for cooperative control.

[0054] 2. Roadside intelligent module Deploy edge computing nodes (MEC) and multi-modal sensor arrays to achieve accurate monitoring of intersection micro-situations: Edge computing processing: three layers of data processing: ① Space-time calibration: through radar-camera fusion calibration algorithm, unify camera pixel coordinates and radar polar coordinates to UTM projection coordinate system, spatial error ≤5cm, time synchronization error ≤10ms; ② Feature extraction: build 16-dimensional intersection state vector S inter , including flow Q, queue length L, delay index D (based on through time difference calculation of license plate recognition), current phase P, average speed V avg , time period type H (weekday / weekend), weather state W; ③ Data compression: reduce minute-level data to 10-dimensional through Fourier transform, upload to cloud through MQTT protocol (QoS=1), bandwidth occupancy ≤200kbps.

[0055] 3. Cloud platform module Build a three-layer architecture of "data platform - model engine - application service" to realize global data processing and intelligent decision-making: Data Hub: Fusing multi-source data to form a spatio-temporal database, including: ① Static data: Road geometry parameters (number of lanes, intersection spacing, speed limit value), signal control basic configuration (phase sequence, initial timing scheme); ② Dynamic data: Vehicle trajectory (100 million vehicles processed daily), real-time roadside state (thousand-level intersection second-level update), Internet data (Gaode real-time traffic API, update frequency 30 seconds).

[0056] Large model engine: Adopting the "feature fusion - spatio-temporal reasoning - decision generation" architecture: ① Feature fusion layer: Through 8-head self-attention mechanism to fuse vehicle trajectory embedding vector (128 dimensions), roadside state vector (16 dimensions), and Internet congestion index (5 dimensions), output unified feature vector, capturing cross-modal data correlation; ② Spatio-temporal reasoning layer: Improved STGNN model (12-layer graph convolution + 8-layer LSTM), using Chebyshev graph convolution to extract intersection spatial correlation features, processing time series through GRU units, achieving 15-minute traffic flow prediction, with an average absolute error ≤8%; ③ Decision generation layer: Output three types of control instructions: green wave parameters (speed range, phase difference matrix), dynamic timing scheme (cycle length, phase length vector), and cross-domain coordination strategy (adjacent region signal coordination priority), with a decision delay ≤100ms.

[0057] The multi-modal data fusion algorithm specifically includes: 1. Spatio-temporal feature matrix construction method Design a three-layer fusion architecture to realize the standardized processing of heterogeneous data: In the system parameter setting, the time step T is defined as 60, corresponding to an actual time span of 30 minutes, and each time step is 30 seconds, i.e. the system samples and analyzes the road network state at intervals of 30 seconds; N represents the total number of nodes in the road network, which are used as the basic unit to build the traffic network model. The feature vector dimension F=18 collected by the system covers three sources of data: vehicle, roadside, and Internet: Vehicle features: include road segment average speed v seg , which represents the average driving speed of vehicles on a specific road segment; parking frequency density s den is calculated by the formula parking density = total parking frequency S / (road length L seg × statistical time T), where S is the total number of parking times in the statistical period, L seg is the road length, and T is the total statistical time, which reflects the stagnation of road traffic.

[0058] Roadside features: traffic flow q in The number of vehicles entering a specific area per unit of time; the queue length l records the length of the vehicle queue waiting for traffic at an intersection or road segment; the delay index d quantifies the time loss of vehicles during traffic; and the phase state P represents the current working state of the traffic light.

[0059] Internet features: real-time congestion index c web Integrating multi-source data such as Internet maps to intuitively reflect the congestion level of the road network; the historical same-period traffic deviation AQ is calculated by the formula traffic deviation = (current traffic q now - historical same-period traffic q his ) / historical same-period traffic, where q now is the current traffic, q his is the historical same-period traffic, and this parameter is used to capture abnormal fluctuations in traffic.

[0060] 2. Spatio-temporal graph neural network optimization model (M-STGCN) An improved spatio-temporal graph convolutional network is proposed to enhance the modeling capability of spatio-temporal correlation: Spatial graph convolutional layer: To accurately depict the relationship between road network topology and traffic propagation, a directed and weighted road network association graph G = (V, E, A) is constructed, where the node set V corresponds to each intersection in the traffic road network, and the edge set E represents the physical connection relationship between intersections. The historical traffic mutual information is used to calculate the edge weight a ij , and the formula for the edge weight between intersections i and j is = historical traffic mutual information I(qi, qj) / maximum mutual information value of the entire road network. In this calculation, I(qi, qj) represents the mutual information of the historical traffic sequence of intersections i and j, which quantifies the dependence between nodes and highlights key correlation paths. Meanwhile, a 3rd-order Chebyshev graph convolution operation is introduced, which realizes spatial feature extraction through the formula: kth layer spatial feature = activation function σ acting on the weighted sum of the standardized Laplacian matrix (Chebyshev polynomial T m × Laplacian matrix × input feature X s （k-1） ). This effectively reduces the impact of data heterogeneity; T m is the Chebyshev polynomial, which optimizes the complexity of graph convolution through recursive calculation; X s （k-1） is the input feature of the previous layer, which can efficiently capture the traffic propagation rules across regions in the road network, and has improved adaptability to complex road network structures by 23% compared to the traditional adjacency matrix method.

[0061] Temporal sequence layer: To address the dynamic time-varying characteristics of traffic flow, a time sequence processing module based on gated recurrent unit (GRU) and attention mechanism is designed. The GRU unit effectively handles the long-term dependence problem of traffic flow through the synergistic effect of reset gate and update gate. The self-attention mechanism is introduced to adaptively allocate weights according to the importance of traffic characteristics at different times, focusing on key scenarios such as peak hours and sudden accidents. Through actual measurement verification, the response delay of this module to sudden congestion can be controlled within 200ms. Compared with the traditional STGCN model, the average absolute error is reduced by 15% and the prediction accuracy is improved by 12% in the short-term traffic prediction task, significantly enhancing the system's perception and response ability to real-time traffic state changes. In the spatio-temporal graph neural network optimization model, to further improve the model's generalization ability and training efficiency, the multi-modal data collected at the vehicle end and roadside are normalized and used as the input of the model.

[0062] By constructing a large-scale road network simulation dataset, M-STGCN is pre-trained, and in actual scenarios, a transfer learning strategy is adopted to quickly adapt to different regional traffic characteristics using a small amount of real-time data, reducing the training time of the model in new scenarios and enhancing the model's adaptability to complex and variable traffic environments.

[0063] 3. Federated learning enhancement mechanism A hierarchical federated learning framework is designed to balance data privacy and model generalization: Edge training: Each regional mobile edge computing (MEC) node serves as the core of distributed training, using a spatio-temporal graph neural network (STGNN) model to process local traffic flow data. In the gradient calculation stage, Laplace noise that meets differential privacy (privacy budget ε=0.5) is injected to effectively prevent the leakage of original data. The training process uses a dynamic batch size strategy, adjusting the batch size to 16-64 according to real-time data traffic, ensuring that regional traffic dynamic characteristics are quickly captured under limited local computing resources.

[0064] Cloud aggregation: The central server serves as the global model hub, innovatively adopting a dynamic intersection weight distribution mechanism. In the model aggregation stage, the actual number of intersections n i is used as a weighted factor to calculate the global model parameters θ i = weighted average of regional model parameters θ i .

[0065] The global model parameters θ global are calculated. This strategy makes the model updates of traffic flow dense areas more significantly influence the global model, improving the model's adaptability to complex road networks. A parameter verification mechanism is also introduced to filter abnormal parameters in the aggregation process according to the 3σ principle, ensuring the stability of the model.

[0066] Adaptive adjustment: To enhance the model's adaptability to special scenarios such as scenic spots and industrial parks, a parameter difference evaluation system based on Euclidean distance is constructed. By calculating the regional model difference = regional parameter θ i The Euclidean distance of the global parameter θ global . Quantify the parameter difference between regional model and global model, set dynamic threshold (initial value 0.8) as adjustment trigger condition. When D i >τ, automatically increase the training round of the corresponding region by 20%, and simultaneously enable the transfer learning strategy, use the historical training parameters of similar scenarios as the initialization weight, accelerate the model convergence speed in special scenarios, and finally realize the dynamic optimization of global traffic signal control strategy.

[0067] The core algorithm of traffic signal control includes: 1. Green wave line dynamic generation algorithm Based on spatio-temporal correlation mining to realize intelligent construction and optimization of green wave corridor: High delay section identification: Road delay index is the core index of traffic congestion quantification, its calculation formula is delay index = (actual travel time average t actual - free flow time t free ) / free flow time × 100%. The difference between actual travel efficiency and theoretical free flow state is considered. Among them, t actual Take the past 1 hour rolling average passing time, which is updated dynamically through real-time traffic flow data; t free =L seg / v speed_limit Calculate the theoretical free flow time based on the legal speed limit and length of the road section. L seg : represents the length of the road section (unit: meter or kilometer, determined according to the actual scene unit conversion); V speed_limit : represents the speed limit of the road section (unit: km / h or m / s, need to match the length unit); t free : refers to the theoretical travel time of a vehicle traveling at the legal speed limit on the road section, which is used to compare and calculate the delay index. When the road delay index exceeds 30% threshold for 15 minutes in a row, the system automatically triggers the green wave corridor creation process, and combines with the real-time traffic data of Gaode / Baidu map for secondary verification to ensure the accuracy of the identification result.

[0068] Multi-objective optimization model: To achieve the global optimization of traffic efficiency, a multi-dimensional objective function is constructed, with the first term being the total delay time equal to the sum of (the actual travel time of all vehicles in the green wave zone minus the free flow travel time), where the free flow travel time = the length of the road segment divided by the legal speed limit value; the second term being the average number of stops equal to the total number of stops of all vehicles in the green wave zone divided by the total number of vehicles, with a stop event defined as a speed less than or equal to 5 km / h and a duration greater than or equal to 3 seconds; the third term being the speed consistency deviation equal to the absolute value of the actual average speed of each road segment in the green wave zone minus the target green wave speed; the final optimization objective being to minimize (the first term multiplied by the delay weight coefficient + the second term multiplied by the stop weight coefficient + the third term multiplied by the speed weight coefficient), with the weight coefficients dynamically adjusted according to the traffic state (the delay weight is increased to 0.6 during the morning peak).

[0069] The model includes three core optimization dimensions: Total delay minimization: D total The cumulative delay time of all vehicles in the aggregated road segment group directly reflects the improvement effect of the green wave scheme on travel efficiency; Stop number control: S stop The average number of stops of vehicles in the green wave zone is counted, effectively reducing idling emissions and fuel consumption; Speed consistency optimization: ∑ N i=1 |V-v i | measures the deviation of the actual speed of each road segment from the green wave speed, ensuring smooth traffic flow; N represents the total number of road segments included in the green wave corridor; V represents the target green wave speed (i.e., the ideal vehicle travel speed designed, unit: km / h); vi represents the actual average speed of the ith road segment (unit: km / h); The model constraint condition adopts a phase difference coordination mechanism: the phase difference between adjacent intersections is calculated according to the following rules: the reference value is calculated as (the physical distance from the upstream intersection to the downstream intersection divided by the target green wave speed) multiplied by 3600; the period constraint is handled when the reference value is greater than the signal period: the final phase difference = the reference value minus the signal period multiplied by N, where N is the maximum integer that satisfies (the final phase difference is less than the signal period); a positive or negative three-second floating compensation amount is added based on real-time traffic data. By accurately calculating the mathematical relationship between intersection spacing, green wave speed, and signal period, seamless connection of green light phases is ensured.

[0070] The weight coefficients a, b, and g are dynamically adjusted based on the entropy weight method. This algorithm analyzes the information entropy of historical traffic data and adaptively allocates the optimization focus of different time periods: the morning peak focuses on reducing delay (a weight increases), the flat peak focuses on speed stability (g weight increases), and a fine-grained control strategy is formed for different time periods. The model solution uses the Alternating Direction Method of Multipliers (ADMM) to decompose the complex multi-objective optimization problem into sub-problems for iterative solution. In the solving process, auxiliary variables and augmented Lagrange functions are introduced to achieve coordination and balance between different optimization objectives. At the same time, distributed computing architecture is used to accelerate the solving process, so that the algorithm can still converge quickly in large-scale road network scenarios, meeting the real-time control requirements of traffic signals.

[0071] Dynamic iteration mechanism: The system builds a three-level data-driven adaptive optimization system. First, based on the historical traffic flow data of the past half year, the non-parametric kernel density estimation method is used to establish the prior probability distribution model of green wave speed, and the statistical characteristics of the mean value of 45 km / h and the standard deviation of 5 km / h are determined to provide a basic parameter framework for the control strategy. In the real-time running phase, the system continuously collects multi-source data from the signal cycle, including the cross-section flow of roadside sensing devices, the driving trajectory of vehicle terminals, and the cloud traffic situation prediction information. Through the Bayesian optimization algorithm, the real-time data and the prior distribution are dynamically fused, and the phase difference parameters of each intersection are iteratively updated.

[0072] 2. Reinforcement learning dynamic timing algorithm Deep Deterministic Policy Gradient (DDPG) is used to realize continuous optimization of multi-dimensional timing parameters: a deep neural network architecture containing policy network and value network is constructed to fuse real-time traffic flow data perceived by vehicles, road state information collected by roadside devices, and historical traffic rule data stored in the cloud. The data is input into the DDPG algorithm framework. The policy network generates a multi-dimensional timing parameter action sequence in continuous space based on the fused data, such as green light duration and phase switching sequence. The value network evaluates the actions output by the policy network, taking the cumulative discounted reward as the optimization objective, and updates the network parameters using the temporal difference error backpropagation. In the training process, the experience replay mechanism and target network are introduced to effectively alleviate the data correlation and value estimation bias problems in the training process. Through multiple rounds of iterative learning, the system can adaptively output the optimal multi-dimensional timing parameter combination according to the dynamically changing traffic scene, achieving fine-grained and intelligent adjustment of traffic signal control strategies.

[0073] State space (20 dimensions): The state vector s comprehensively represents the intersection traffic state, which is composed of 7 key features, including: P uses One-Hot Encoding to describe 4 basic phase combinations, and realizes the digitalization of discrete state selection through binary vectors [1, 0, 0, 0] to [0, 0, 0, 1]; Q is the real-time flow of 4 import channels (unit: vehicle / minute), and the sliding window algorithm is used to take the average of the past 5 minutes of data to smooth fluctuations; L records the queuing length of each import channel (unit: meters), which is dynamically updated based on vehicle contour recognition technology of video detection equipment; D is the comprehensive delay index, which integrates vehicle stop time, start-stop times and other parameters to construct a composite evaluation index; V avg is the average speed of the road section (unit: km / h), which is collected through the interaction between the OBU device and the roadside unit; H distinguishes between peak and flat peak periods, and dynamically divides them in combination with historical flow data and real-time congestion index; W covers three weather states: sunny, rainy and snowy, and accesses the real-time weather information from the meteorological department API.

[0074] Action space: The continuous action vector a output by the system contains 3 adjustable parameters: a=[△C, △Φ, δ p ], where: △C is the period adjustment amount, with a value range of [-10, 10] seconds, which optimizes the overall traffic efficiency by changing the signal light cycle; △Φ is used to adjust the east-west / north-south direction phase difference, allowing dynamic compensation for the time difference of vehicle flow arrival within the interval [-5, 5] seconds; p is the phase priority coefficient, 0 means straight priority, 1 means left turn priority, and intermediate values support mixed priority strategies for smooth transition.

[0075] Reward function: Based on multi-objective optimization, a weighted reward function is constructed: weighted comprehensive reward = delay reward R D + parking reward R S + efficiency reward R E + stability reward R C , the weights ω i are dynamically adjusted, and the meanings of each sub-item are as follows: The delay reward encourages the system to reduce vehicle stop time by calculating the ratio of the current period to the maximum delay; ​Parking rewards are based on the ratio of the increase in parking times to the historical peak, reducing unnecessary vehicle starts and stops; Efficiency rewards are based on the ratio of lane throughput change rate to the historical maximum value, improving overall traffic capacity; The stability reward adjusts the amplitude by constraining the period to avoid drastic fluctuations in the control strategy.

[0076] The weight parameters are dynamically adjusted using a fuzzy logic controller. For example, in a congested state (average vehicle speed < 15 km / h), 0.4 and 0.3 are automatically allocated to prioritize alleviating delays and parking problems.

[0077] Network architecture and training results: Using the classic Actor-Critic architecture, both networks consist of three fully connected layers, each containing 256 neurons and using the ReLU activation function. The experience replay pool capacity is set to 100,000 samples, and random sampling is used to alleviate data correlation. A soft target update mechanism is introduced to The algorithm gradually synchronizes the target network parameters with the coefficients of the roadside units (RSUs) and on-board units (OBUs). After training with real-world traffic data collected by actual roadside units (RSUs) and on-board units (OBUs), the average delay at typical intersections was reduced by 42%, and traffic efficiency increased by 35%, effectively verifying the algorithm's engineering practicality.

[0078] 3. Phase sequence intelligent matching algorithm Construct a fusion model of fuzzy rules and neural networks to achieve dynamic optimization of phase solutions: Multi-scenario phase library: Based on traffic flow characteristics and road topology, 8 standard phase combinations are predefined to cover common and complex scenarios in urban traffic: Four-phase solution: Suitable for traffic balancing scenarios. It adopts the classic phase sequence (east-west straight → east-west left turn → north-south straight → north-south left turn) and ensures traffic efficiency in all directions through fixed-cycle allocation. It is often used at intersections on main roads.

[0079] Three-phase solution: When the proportion of east-west left-turn traffic is ≤30%, the east-west straight-through and left-turn phases are dynamically merged to reduce phase switching losses and improve the overall green-to-signal ratio utilization rate. This solution is typically used on sections with tidal traffic characteristics.

[0080] Variable phase solution: For high-frequency pedestrian crossing scenarios (such as around schools and shopping districts), the system can insert 1-2 independent pedestrian phases in real time to form a dynamic combination of 5-6 phases, triggered by pedestrian detection sensors, to ensure the safety of both pedestrians and vehicles.

[0081] Emergency priority phase: Reserve special phase channel, when emergency vehicles such as ambulances and fire engines trigger RFID / video identification signals, give priority to release and dynamically adjust the surrounding intersection phase to build an emergency green wave zone.

[0082] Bus priority phase: Combined with bus lane and vehicle positioning system, extend the green light duration or release early before the arrival of the bus, reduce the bus delay rate.

[0083] Ramp coordination phase: For the expressway entrance and exit, dynamically adjust the phase of the merging area according to the main road and ramp traffic data to relieve the interlaced conflict.

[0084] Tidal lane phase: Cooperate with lane indicators, switch the corresponding phase of the tidal lane according to the change of peak traffic direction.

[0085] Adaptive expansion phase: Support custom phase combination to meet the temporary control needs of special scenes such as construction sections and large events.

[0086] Fuzzy adaptation degree calculation: Real-time collection of traffic flow feature vectors f through multi-source sensing devices (magnetic, radar, video), including left turn, straight, right turn traffic density, and pedestrian crossing flow. Adopt Gaussian membership function to build fuzzy matching model: the adaptation degree of scheme i = the product of the Gaussian probability of each flow feature and its ideal value. i f k c i,k . In the formula, c i,k is the ideal adaptation center value of phase scheme i to flow type k, obtained by historical data clustering analysis; the variance parameter reflects the tolerance range of different schemes to traffic fluctuations. This model deeply integrates expert experience (predefined phase rules) and real-time sensing data, quantifying the adaptation degree of each scheme under the current traffic state.

[0087] Dynamic decision mechanism: Use Gated Recurrent Unit (GRU) to extract features from the past 30 minutes of traffic time series data, capture the periodicity and sudden change trend of traffic flow, and output the switching probability vector p of each phase scheme. Use the fusion decision formula to select the phase scheme with the maximum switching probability p i and fuzzy adaptation degree μ i weighted sum (weight 0.6). Comprehensive evaluation of scheme priority, where 0.6 is the experience weight coefficient, balancing the timing prediction and real-time adaptation degree. The system decision delay is controlled within 50ms, and the phase matching accuracy rate of the system in complex intersections (including tidal flow, pedestrian-intensive, emergency response, etc.) reaches 92.7%, which is more than 25% more efficient than traditional schemes.​​

[0088] The functional module design is specifically: 1. Green wave line diagnosis module Realize full-dimensional monitoring and intelligent diagnosis of green wave band operation state: Real-time visualization: dynamically draw the time-distance graph, the horizontal axis is distance (meters), the vertical axis is time (seconds), mark the green light period with green color, superimpose vehicle passing time scatter points, and display the green wave passing rate (target ≥ 85%) in real time; Deviation analysis: calculate the speed deviation = (actual vehicle speed V actual - Target green wave vehicle speed V target ) / target green wave vehicle speed, phase difference deviation = the absolute error sum of actual phase difference of each intersection and planned value, delay deviation ΔD = D real -D pred , the threshold values are ±10%, 5 seconds, and 15% respectively; D real : represents the actual delay time (unit: seconds) of a road section or intersection, which is collected in real time through multi-source data such as vehicle end track and roadside perception; D pred : represents the delay time (unit: seconds) predicted based on the traffic signal control large model; ΔD: the difference between actual delay and predicted delay, used to measure the prediction accuracy of the model and the effectiveness of the control strategy. Intelligent diagnosis: based on Bayesian network to locate faults (such as radar fault of a certain intersection leading to abnormal phase difference), generate a structured report containing problem causes, impact range, and optimization suggestions, support PDF export and email push, and the fault location accuracy is ≥95%.

[0089] Several traffic intersections in Binjiang District of Hangzhou City are selected as pilots, and the lane vehicle flow per hour reaches 2000 during morning and evening peak hours, and the traffic congestion problem is prominent. This pilot aims to realize significant improvement of regional traffic conditions by using the traffic signal control large model SaaS system and method based on vehicle-road cloud multi-modal data fusion proposed in the invention, and the specific goals are to reduce the regional average delay by more than 30%, and to increase the green wave band coverage rate to 70%, effectively alleviate traffic congestion, and improve regional traffic efficiency. The measured data shows that compared with the traditional genetic algorithm, the parameter convergence speed of this mechanism is improved by 40%, and in the typical two-way green wave scenario, the average number of vehicle stops in the green wave band is reduced from 2.3 to 1.5, with a reduction of 35%, which significantly improves the main line traffic efficiency.

[0090] At the data collection level, 2000 connected vehicles were successfully accessed, accounting for 70% of the total number of vehicles in the area. These connected vehicles upload real-time and stable key data such as vehicle trajectory, speed, and parking data to the system through 5G slice networks with a latency of ≤50ms. Meanwhile, 12 mobile edge computing (MEC) devices were deployed in the area, which can accurately and real-time output intersection flow data through advanced radar and camera fusion technology, with an accuracy of 95% and an error control of ≤3 meters in queue length, achieving precise monitoring of the micro traffic situation at the intersection.

[0091] At the cloud processing level, the system integrates real-time traffic data from Gaode, which updates the congestion index every 10 seconds. By deeply integrating this internet traffic data with data collected from vehicles and roadside, a comprehensive three-dimensional traffic situation map is constructed, providing strong data support for the cloud-deployed traffic signal control large model and enabling it to make fast and accurate real-time decisions.

[0092] The control strategy implementation includes: Green wave corridor construction: The system automatically identifies 3 high-delay road segments (delay index 40%-50%), generates an initial green wave speed of 45km / h based on historical commuting data, and adjusts the phase difference through Bayesian optimization to reduce the number of stops from an average of 5 times per kilometer to 3.2 times per kilometer, and the green wave pass rate is increased to 88%.

[0093] Reinforcement learning timing: During the morning peak, the algorithm detects a 150% surge in left-turn traffic on the main road and dynamically extends the left-turn phase by 15 seconds and shortens the straight-ahead phase on the branch road by 10 seconds, reducing the average delay at the intersection from 80 seconds to 55 seconds and improving traffic efficiency by 35%; during the mid-peak period, the system automatically switches to a variable phase scheme based on pedestrian crossing requests (≥10 times per minute), reducing pedestrian waiting time by 25%.

[0094] In terms of quantitative indicators, this pilot has achieved remarkable results. The overall delay in the area has decreased by 42%, the number of vehicle stops has decreased by 38%, the vehicle speed during peak hours has increased from 20km / h to 25km / h, with an increase of 25%, and the green wave coverage rate has reached 75%, far exceeding the expected target.

[0095] In terms of qualitative effects, through the SaaS platform of the application, efficient human-computer collaborative operation is realized. Users can conveniently perform drag-and-drop time-distance diagram editing, and the phase difference adjustment accuracy can reach 1 second, greatly improving the flexibility and accuracy of traffic signal control. In the deployment of temporary plans for major activities, the time is greatly shortened from 2 hours in the traditional way to 15 minutes, significantly improving the efficiency of emergency response. At the same time, the operation and maintenance cost is reduced by 60% compared with the traditional scheme, effectively reducing the economic burden of urban traffic management, fully demonstrating the great advantages and value of the application in practical application.

[0096] The application breaks the data isolation in the traditional traffic system and constructs a deeply collaborative vehicle-road cloud architecture. Under this architecture, vehicles can collect high-precision trajectories with an accuracy of 2.5m; road-side sensing devices can update data at a level of seconds and real-time feedback the traffic state at intersections; the time delay of the cloud decision system is controlled within 100ms, truly realizing the real-time closed loop of vehicle trajectory, road-side state and cloud decision. Compared with the traditional system, the data utilization rate is improved by 300%, the decision response speed is greatly shortened from minutes to seconds, and an efficient control chain of "perception - decision - execution" is successfully constructed, laying a solid foundation for precise and real-time control of traffic signals.

[0097] For the problem of heterogeneous data fusion, the application proposes a hybrid model based on spatio-temporal graph neural network (STGNN) and federated learning. This model effectively solves the spatio-temporal alignment problem of multi-modal data, controls the spatio-temporal alignment error within 10 seconds, and significantly improves the cross-domain generalization ability of the model, with a performance decline of ≤5% in off-site deployment. According to actual verification, the prediction accuracy of the model is as high as 92%, which is improved by 25% compared with the single data source scheme, and can effectively deal with complex and variable road network scenarios, providing more accurate and reliable decision basis for traffic signal control.

[0098] The application constructs a three-level optimization model including green wave generation, reinforcement learning timing, and phase matching, forming a perfect dynamic intelligent decision system. Among them, the response time of the green wave generation model is only 5 minutes, which can quickly identify high delay sections and generate optimized green wave schemes; the iteration period of the reinforcement learning timing model is 20 seconds, which can dynamically adjust the signal timing parameters according to the real-time traffic state; the decision delay of the phase matching model is controlled within 50ms, which can quickly and accurately select the optimal phase scheme. This decision system realizes multi-scale control from macroscopic green wave corridor planning to microscopic phase fine adjustment, and the response speed to complex traffic scenarios such as tidal flow and sudden congestion is improved by 5 times, effectively reducing the average delay at intersections by more than 40%, significantly improving the traffic efficiency.

[0099] Through the combination of centralized deployment of cloud large models and lightweight inference of edge nodes, "zero localization deployment" is realized, the initial deployment cost of small and medium-sized cities is reduced to 40% of that of the traditional scheme, and the rapid popularization of intelligent traffic technology is promoted.

[0100] The present application is not limited to the above-mentioned embodiments, and any changes in shape or material composition are allowed, as long as the structure design provided by the present application is adopted.

Claims

1. A traffic signal control method based on vehicle-road-cloud multimodal data fusion, characterized in that: include: Real-time and synchronous collection of multimodal traffic data through vehicle-side sensors, roadside sensing equipment, and cloud-based Internet platforms; Use spatiotemporal alignment algorithm to fuse heterogeneous data and construct a standardized spatiotemporal feature matrix; Traffic flow prediction is performed using a multi-layer spatiotemporal graph neural network trained with a federated learning mechanism, and signal control solutions are generated through reinforcement learning and multi-objective optimization models. The green wave parameters, dynamic timing plan and cross-domain collaborative strategy are sent to the roadside signal machines for execution control through the cloud-edge collaborative architecture.

2. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 1 is characterized in that: The synchronous collection of multimodal traffic data includes: using the extended Kalman filter algorithm through the vehicle terminal to fuse positioning and inertial data to generate vehicle spatiotemporal trajectories with millimeter-level accuracy; obtaining engine start and stop signals and speed sensor data in real time through the vehicle interface, and identifying parking events based on predefined speed thresholds and durations; receiving six-dimensional dynamic information of surrounding vehicles in real time through the vehicle-road communication interface to construct a local traffic flow matrix; and collecting environmental data through temperature and humidity sensors and raindrop sensors.

3. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 1 or 2, characterized in that: The use of a spatiotemporal alignment algorithm to fuse heterogeneous data includes: designing a multi-layer fusion architecture, performing spatiotemporal calibration processing on vehicle-side data, roadside data, and Internet data, unifying the coordinate systems and timestamps of data of different modalities, and minimizing spatiotemporal errors; processing heterogeneous data through a feature extraction method, including vehicle-side features, roadside features, and Internet features, to construct a standardized feature vector covering multiple dimensions; wherein the vehicle-side features include the average speed of a road section and the density of the number of stops, the roadside features include the flow rate and queue length, and the Internet features include the real-time congestion index and historical flow deviation.

4. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 3 is characterized in that: The method of constructing a standardized spatiotemporal feature matrix includes calculating the association weights between road network nodes based on the mutual information of historical traffic flow, and constructing a directed weighted road network association graph; using graph convolution operations to process spatial features, and optimizing the complexity of graph convolution using polynomial recursive calculations; using a time series processing method, extracting the dynamic change characteristics of traffic flow through gated recurrent units and a self-attention mechanism, and constructing a standardized multi-dimensional spatiotemporal feature matrix.

5. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 4 is characterized in that: The multi-layer spatiotemporal graph neural network trained based on the federated learning mechanism performs traffic flow prediction, including: adopting a hierarchical federated learning framework, performing local model training at edge nodes, injecting privacy-preserving noise and dynamically adjusting the training batch size based on real-time data traffic; performing global model aggregation on a cloud server, calculating global parameters using the actual number of intersections in each area as a weighting factor, and introducing a parameter verification mechanism to filter outliers; evaluating model differences through Euclidean distance, triggering adaptive adjustment mechanisms including increasing training rounds and transfer learning.

6. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 4 or 5, characterized in that: The signal control scheme generated through reinforcement learning and multi-objective optimization models includes: identifying high-delay sections based on the section delay index and creating green wave corridors; constructing a multi-objective optimization function that minimizes total delay, number of stops, and speed deviation, and using a phase difference coordination mechanism to constrain adjacent intersections; using a decomposition iterative algorithm and combining historical traffic flow data to establish a prior distribution, and dynamically updating parameters through real-time data fusion.

7. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 6 is characterized in that: Generating a signal control scheme through reinforcement learning and a multi-objective optimization model further includes: constructing a state space based on real-time traffic status, the state space including real-time traffic flow and delay indexes and traffic period types; The action space is defined to include adjustable signal control parameters, including signal cycle adjustment, phase difference adjustment for adjacent directions, and priority coefficients for straight and turn phases. A reward function is designed to evaluate the effectiveness of the output signal control parameters of the action space through a weighted combination of delay reduction rewards, stop reduction rewards, capacity improvement rewards, and strategy stability rewards. A fuzzy logic controller is then used to dynamically adjust the weight coefficients of each reward item. A deep deterministic policy gradient algorithm is adopted to construct a policy network and a value network, and a dynamic timing plan is generated using the state space as input, the action space as output, and the reward function as the optimization target.

8. The traffic signal control method based on vehicle-road-cloud multimodal data fusion according to claim 7 is characterized in that: The control execution sent to roadside signal machines through the cloud-edge collaborative architecture includes: generating green wave band parameters, dynamic timing plans and cross-domain collaborative strategy control instructions through the cloud platform, and processing them using mobile edge computing nodes to achieve coordination between cloud model decision-making and edge real-time execution; control instructions are sent to roadside signal machines through message protocols to execute phase switching and timing adjustments; the green wave operating status is visually monitored and intelligently diagnosed, and faults are located based on deviation analysis and optimization reports are generated.

9. A traffic signal control system based on vehicle-road-cloud multimodal data fusion, adopting a traffic signal control method based on vehicle-road-cloud multimodal data fusion according to any one of claims 1 to 8, characterized in that: It includes a vehicle-side perception module, a roadside intelligent module and a cloud platform module; the vehicle-side perception module is connected to the roadside intelligent module through a communication link, and uploads vehicle spatiotemporal trajectory data and local traffic flow matrix in real time; the roadside intelligent module deploys edge computing nodes, processes multimodal sensor data through spatiotemporal calibration and feature extraction, generates intersection state vectors, and compresses and uploads them to the cloud platform module.

10. The traffic signal control system based on vehicle-road-cloud multimodal data fusion according to claim 9 is characterized in that: The cloud platform module includes a data middle platform and a large model engine. The data middle platform receives and integrates the trajectory data of the vehicle-side perception module, the state vector of the roadside intelligent module, and the Internet traffic data to build a spatiotemporal database; the large model engine processes the fused feature vectors based on the spatiotemporal graph neural network, outputs green wave parameters, dynamic timing plans, and collaborative strategy instructions, and the decision instructions of the large model engine are sent to the roadside intelligent module through a low-latency link to drive the signal control equipment to execute the dynamic timing plan.

Images