Dynamic traffic signal control method and device based on data fusion and medium

Abstract

The embodiment of the invention discloses a dynamic traffic signal control method and device based on data fusion and a medium, belongs to the technical field of intelligent traffic, and solves the problem that the existing traffic signal control is difficult to make effective response when an accident occurs. Performing data fusion on first traffic data acquired by a road monitoring and shooting device and second traffic data acquired by a positioning device of the floating car through a multi-head attention mechanism at the edge node to obtain multi-source traffic characteristic data; the cloud end inputs the multi-source traffic characteristic data into a preset signal optimization model every first preset duration; the preset signal optimization model is updated and optimized through a PPO algorithm and a federated average method, and a global traffic model is generated; and through knowledge distillation, compressing the global traffic model, and migrating an edge traffic model obtained after compression to a corresponding edge node so as to carry out traffic signal strategy output on a corresponding road section through the edge traffic model.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, and in particular to a dynamic traffic signal control method, device and medium based on data fusion. Background Art

[0002] With the acceleration of urbanization, traffic congestion has become more and more serious, bringing many challenges to people's daily travel, logistics and transportation, and the sustainable development of cities. The core problems of traffic congestion include three points: rigid timing of traffic lights, that is, traditional timing strategies are difficult to adapt to dynamic traffic; data islands, that is, multi-source data such as cameras and GPS are not fully integrated; response delays, that is, the control strategy is slow to update.

[0003] Traditional traffic light optimization methods use a timing control strategy. Based on historical traffic flow statistics, a fixed-length traffic light cycle is pre-set. Vehicle arrival is detected through ground sensors or radars, and the green light duration is dynamically adjusted. Alternatively, traffic police direct on-site or signal parameters are manually modified to temporarily respond to sudden congestion.

[0004] However, the traditional traffic light optimization method relies on a single sensor with limited data coverage, making it difficult to obtain multi-dimensional information such as real-time traffic density and vehicle type distribution of the entire road network. The data collection capability is weak, and the lack of data at some intersections leads to a mismatch between traffic light timing and traffic flow, which in turn aggravates congestion. In addition, due to the rigid control strategy, it is difficult to adapt to the spatiotemporal heterogeneity of traffic flow when encountering sudden accidents, weather changes, etc., making it difficult for traffic signal control to respond effectively. Summary of the invention

[0005] The embodiments of the present application provide a dynamic traffic signal control method, device and medium based on data fusion, which are used to solve the following technical problems: the traditional traffic light optimization method has limited data coverage and rigid control strategies. In the event of an emergency, it is difficult for traffic signal control to respond effectively.

[0006] The present application embodiment adopts the following technical solutions: The embodiment of the present application provides a dynamic traffic signal control method based on data fusion. It includes: at the edge node, weights are assigned to the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle through a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data; after each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model; wherein the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1, and is constructed by integrating meta-learning; the preset signal optimization model is updated and optimized by the PPO algorithm and the federated average method to generate a global traffic model; the global traffic model is compressed by knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0007] The embodiment of the present application integrates the first traffic data of the road monitoring device and the second traffic data of the floating vehicle positioning device through a multi-head attention mechanism, which can fully obtain multi-dimensional information such as real-time traffic density and vehicle model distribution of the entire road network, making up for the defect of insufficient coverage of traditional single sensor data. Secondly, the embodiment of the present application uses the PPO algorithm and the federal average method to update the optimization model, improve traffic fluency, integrate multiple edge node data, make the model training more comprehensive and accurate, avoid the influence of single node data deviation, and generate a global traffic model that is more universal and reliable. The edge traffic model is obtained by compressing the global traffic model through knowledge distillation and migrating it to the corresponding edge node. Without losing too much performance, the model scale is reduced, the computing and storage burden of the edge node is reduced, and the model can run efficiently on the edge device. The edge node can use local data and the edge traffic model to output traffic signal strategies in real time, reduce data transmission delays, and quickly respond to local traffic changes. Improve the timeliness and accuracy of traffic control.

[0008] In one implementation of the present application, at an edge node, weights are assigned to the first traffic data obtained by a road monitoring device and the second traffic data obtained by a positioning device of a floating vehicle through a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data, specifically including: at the edge node, first feature extraction is performed on the first traffic data and the second traffic data based on spatial features; and, second feature extraction is performed on the first traffic data and the second traffic data based on time features; dynamic weight assignment is performed on the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data through a multi-head attention mechanism; data fusion is performed based on the dynamic weights, the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data to obtain multi-source traffic feature data.

[0009] In one implementation of the present application, before inputting the multi-source traffic feature data into the preset signal optimization model, the method further includes: taking minimizing the average delay time at the intersection as the core goal, and taking reducing the number of emergency brakes and giving priority to public transportation vehicles as auxiliary goals, and constructing a reward function: ; in, R is the reward function; w 1 is the first weight coefficient; w 2 is the second weight coefficient; w 3 is the third weight coefficient, ; D is the average delay time at the intersection; B is the number of emergency brakes; P Score the priority of public transportation vehicles; adjust the weight of the reward function based on the traffic scenario corresponding to the edge node, so as to optimize the preset signal optimization model based on the adjusted reward function.

[0010] In one implementation of the present application, the preset signal optimization model is updated and optimized through the PPO algorithm and the federated average method to generate a global traffic model, specifically including: based on the PPO algorithm, maximizing the long-term cumulative reward through the optimization strategy; wherein the objective function of the PPO algorithm is: ; in, represents the probability ratio of the new and old strategies; Optimize strategies for the new; Optimize strategies for legacy; Indicates the signal light control strategy; represents the advantage function; Indicates the clipping threshold; clip is the clipping function; are model parameters; For the agent at time step t According to the status s t the action performed; For the agent at time step t state; use the federated average algorithm to aggregate the model parameters corresponding to multiple edge nodes and generate a global traffic model.

[0011] In one implementation of the present application, a federal averaging algorithm is used to aggregate model parameters corresponding to multiple edge nodes to generate a global traffic model, specifically including: at each edge node, fine-tuning the preset signal optimization model through local data to obtain fine-tuning parameters; uploading the obtained fine-tuning parameters and the local data corresponding to each edge node to the cloud; at the cloud, weighted averaging the data uploaded by multiple edge nodes to generate a global traffic model.

[0012] In one implementation of the present application, weighted averaging is performed on the data uploaded by multiple edge nodes, specifically including: based on the function: ; ; Based on the amount of data corresponding to each edge node, the data uploaded by multiple edge nodes is weighted averaged; is the global traffic model; Indicates the number of edge nodes participating in federated learning; Representation Node The number of local N is the total amount of data; For edge nodes k The corresponding model parameters.

[0013] In one implementation of the present application, the global traffic model is compressed through knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, specifically including: based on the function: ; ; The compressed edge traffic model is migrated to the corresponding edge node; L is the loss function; For the cross entropy loss, align the marginal traffic model prediction data With the true label ; is the divergence loss, aligning the output probability of the edge traffic model and the global traffic model probability ; is the weight coefficient; Logits output of the global traffic model; Logits output for edge traffic model; Represents the temperature parameter.

[0014] In one implementation of the present application, a traffic signal strategy is output for a corresponding road section through an edge traffic model, specifically including: each edge node inputs the acquired traffic information into the corresponding edge traffic model after each second preset time interval, so as to output a traffic signal strategy corresponding to the current road section through the edge traffic model; and, when the average vehicle speed of the current road section is less than a preset vehicle speed threshold, or the vehicle queue length of the current road section is greater than a preset lane capacity threshold, a congestion response mechanism is triggered; the cloud determines the target vehicle speed and phase difference of the main road based on the global traffic data, and when the growth rate of the number of vehicles at the upstream intersection of the main road is greater than the preset growth rate threshold, the traffic lights at the downstream intersection are delayed; and, the cloud determines the vehicles heading towards the main road, and sends recommended filtered vehicle speed information to the vehicles heading towards the main road through V2X communication.

[0015] The embodiment of the present application provides a dynamic traffic signal control device based on data fusion, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: at the edge node, weight the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle through the multi-head attention mechanism, and perform data fusion based on the assigned weights to obtain multi-source traffic feature data; after each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model; wherein the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1, and is constructed by integrating meta-learning; the preset signal optimization model is updated and optimized by the PPO algorithm and the federated average method to generate a global traffic model; the global traffic model is compressed by knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0016] A non-volatile computer storage medium provided in an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: at an edge node, weights are assigned to first traffic data obtained by a road monitoring device and second traffic data obtained by a positioning device of a floating vehicle through a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data; after each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model; wherein the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1, and is constructed by integrating meta-learning; the preset signal optimization model is updated and optimized through the PPO algorithm and the federated averaging method to generate a global traffic model; the global traffic model is compressed through knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0017] At least one of the above technical solutions adopted in the embodiment of the present application can achieve the following beneficial effects: the embodiment of the present application integrates the first traffic data of the road monitoring device and the second traffic data of the floating vehicle positioning device through the multi-head attention mechanism, and can fully obtain multi-dimensional information such as real-time traffic density and vehicle model distribution of the entire road network, making up for the defect of insufficient coverage of traditional single sensor data. Secondly, the embodiment of the present application uses the PPO algorithm and the federal average method to update the optimization model, improve traffic fluency, integrate multiple edge node data, make the model training more comprehensive and accurate, avoid the influence of single node data deviation, and generate a global traffic model that is more universal and reliable. The edge traffic model is obtained by compressing the global traffic model through knowledge distillation and migrating it to the corresponding edge node. Without losing too much performance, the model scale is reduced, the computing and storage burden of the edge node is reduced, and the model can run efficiently on the edge device. The edge node can use local data and the edge traffic model to output traffic signal strategies in real time, reduce data transmission delays, and quickly respond to local traffic changes. Improve the timeliness and accuracy of traffic control. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings: Figure 1 A flow chart of a dynamic traffic signal control method based on data fusion provided in an embodiment of the present application; Figure 2A schematic diagram of the algorithm design principle structure of a DRL signal optimization model provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a dynamic control strategy mechanism provided in an embodiment of the present application; Figure 4 A practical application diagram of a dynamic traffic signal optimization system under a vehicle-road-cloud collaborative framework provided in an embodiment of the present application; Figure 5 A schematic diagram of the structure of a dynamic traffic signal control device based on data fusion provided in an embodiment of the present application.

[0019] Reference numerals: 200: Dynamic traffic signal control device based on data fusion, 201: Processor, 202: Memory. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a dynamic traffic signal control method, device and medium based on data fusion.

[0021] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0022] The technical solution proposed in the embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0023] Figure 1 A flow chart of a dynamic traffic signal control method based on data fusion provided in an embodiment of the present application is as follows: Figure 1 As shown, the process of the dynamic traffic signal control method based on data fusion includes the following steps: Step 101: At the edge node, weights are assigned to the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle through a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data.

[0024] In one implementation of the present application, at the edge node, the first feature extraction is performed on the first traffic data and the second traffic data respectively based on the spatial feature. And, based on the time feature, the second feature extraction is performed on the first traffic data and the second traffic data respectively. Dynamic weights are assigned to the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data through a multi-head attention mechanism. Based on the dynamic weight, the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data, data fusion is performed to obtain multi-source traffic feature data.

[0025] Specifically, multi-source data is first obtained, including the first traffic data, i.e., camera data, and the second traffic data, i.e., GPS data. Specifically: the camera data is the length of the vehicle queue and the distribution of vehicle types, including private cars, buses, bicycles, pedestrians, etc. The GPS data is the vehicle trajectory, speed, and location, including floating vehicles such as taxis, online ride-hailing vehicles, and buses.

[0026] Secondly, the data is preprocessed and features are extracted. On the one hand, spatiotemporal alignment and outlier filtering are required during data preprocessing. In terms of space, the camera and GPS data are unified into the same coordinate system, such as WGS84. In terms of time, ensure that the data timestamps are synchronized, such as sampling every 5 seconds. When filtering outliers, stationary vehicles, such as GPS points in parking lots, are eliminated, and abnormal traffic data caused by camera occlusion or failure is identified and eliminated. On the other hand, feature extraction mainly divides data features in terms of time and space.

[0027] For spatial features, camera data includes lane-level vehicle density, queue length, pedestrian and bicycle flow, and GPS data includes vehicle trajectory and speed distribution. For temporal features, camera data includes traffic flow trends, such as queue length growth rate, and GPS data includes vehicle arrival rate and driving intention, such as left turn and straight ahead.

[0028] Furthermore, attention weights are calculated. The basis for weight distribution is data reliability, coverage, and real-time performance. First, the camera has a high weight when there is sufficient light, and the GPS has a high weight in the tunnel. Second, the camera covers the intersection area, and the GPS covers the entire road network. Third, the GPS data update frequency is high (1 second / time), and the camera update frequency is low (5 seconds / time). The specific attention mechanism is designed to use multi-head attention (MHA) to calculate the weight: ; in, Q represents the query vector, i.e., the current traffic status; K represents the key vector, i.e., the camera and GPS features; V Represents a value vector, i.e., camera and GPS data; Represents the feature dimension. T is the transposed symbol of the matrix K. In the dynamic adjustment of weights, when the camera detects a pedestrian or a bicycle, the camera data weight is increased; when the GPS data has a wide coverage, the GPS data weight is increased.

[0029] Furthermore, the camera data and GPS data are fused. Specifically, the features of the camera and GPS data are weighted and summed: ; in, Represents the camera data characteristics, Indicates the characteristics of GPS data, Both represent dynamic weights. When a data source is missing, such as a camera failure, historical data or another data source is used to interpolate.

[0030] Furthermore, the Kalman filter algorithm is used to smooth the logarithm, eliminate noise, and output the result. The data result is to generate lane-level traffic status, namely, traffic flow, queue length, vehicle speed, etc., which is used for traffic light timing optimization.

[0031] Edge collaborative computing plays a core role in real-time data processing and low-latency decision-making in intelligent transportation systems. By sinking computing power to edge nodes close to data sources, such as intersection signal controllers, cloud transmission delays can be significantly reduced and system response speed can be improved.

[0032] Specifically, an edge-cloud collaborative architecture is built through corresponding hardware devices, and the pre-processed data is uploaded to the cloud for model training and updating.

[0033] Step 102: After each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model.

[0034] In one implementation of the present application, based on the collaboration of multi-source data fusion and edge computing, the embodiment of the present application constructs a DRL signal optimization model through a multimodal Transformer architecture and meta-learning to achieve dynamic optimization of traffic lights. Figure 2 A schematic diagram of the algorithm design principle structure of a DRL signal optimization model provided in an embodiment of the present application. The algorithm design principle of the model is as follows Figure 2 The core goal of this step is to dynamically optimize the timing of traffic lights through the algorithm output control strategy of the DRL signal optimization model, minimize the average delay time at the intersection, and take into account auxiliary goals such as energy consumption and bus priority.

[0035] In one implementation of the present application, the fused multi-source feature data is first input into the DRL model: Core goal: Minimize the average delay time at the intersection (seconds): ; in, Indicates the actual time the vehicle passes, calculated by the edge device through data; Indicates the theoretical travel time at free-flow speed, which is pre-calculated and sent by the cloud; Auxiliary goals: reduce the number of emergency brakes (reduce energy consumption) and give priority to public transportation. Emergency brake events are detected through GPS data, with an acceleration threshold of <-2m / s 2 Calculate the number of emergency brakes per minute. B Priority is given to public transport vehicles, that is, when the bus is less than 200 meters away from the intersection, the priority signal is triggered. P =1, otherwise P =0; Constructing the reward function: ; in, R is the reward function; w 1 is the first weight coefficient; w 2 is the second weight coefficient; w 3 is the third weight coefficient, ; D is the average delay time at the intersection; B is the number of emergency brakes; P Score the transit priority for public transport vehicles; Peak hours (congestion index>7): ; Off-peak hours: ; Based on the traffic scenario corresponding to the edge node, the reward function is weighted and adjusted to optimize the preset signal optimization model based on the adjusted reward function.

[0036] Furthermore, the multimodal Transformer architecture based on DeepSeek-R1 in the embodiment of the present application includes: Input layer: receives multimodal feature data such as lane-level traffic flow and historical delay time; Transformer encoder: uses multi-head attention mechanism to capture spatiotemporal propagation patterns; Output layer: Generates the traffic light timing plan, i.e. the green light duration and phase switching sequence.

[0037] Meta-Learning adapts to different urban road networks: Objective: To make the model quickly adapt to the characteristics of different urban road networks; Task division: Divide the road network data of different cities into multiple tasks; Meta-training: train the model on multiple tasks to generate initial parameters; Meta-testing: Send historical data to the cloud on the new city road network to fine-tune model parameters, so that the model can quickly adapt to the local road network; Continuous Updates: Models are dynamically updated based on real-time data from edge devices.

[0038] Step 103: Update and optimize the preset signal optimization model through the PPO algorithm and the federated average method to generate a global traffic model.

[0039] In one implementation of the present application, the PPO algorithm is centrally trained in the cloud.

[0040] Goal: By optimizing the strategy , the signal light controls the probability distribution of actions to maximize the long-term cumulative reward.

[0041] Core formula: The objective function of the PPO algorithm includes policy gradient and importance sampling, and clipping is used to avoid excessive policy updates: ; in, represents the probability ratio of the new and old strategies; Optimize strategies for the new; Optimize strategies for legacy; Indicates the signal light control strategy; represents the advantage function; It represents the clipping threshold, usually set to 0.2; clip is the clipping function; are model parameters; For the agent at time step t According to the status s t the action performed; For the agent at time step t status.

[0042] Further, The advantage function is calculated using the generalized advantage estimate (GAE): ; ; in, is a discount factor of 0.99, The GAE smoothing coefficient is 0.95, represents the state value estimated by the value network; is the time step t TD residual of ; Represents the probability ratio of the new and old strategies.

[0043] In one implementation of the present application, by maximizing the gradient ascent , update the policy parameters .

[0044] In one implementation of the present application, at each edge node, the preset signal optimization model is fine-tuned using local data to obtain fine-tuning parameters. The obtained fine-tuning parameters and the local data corresponding to each edge node are uploaded to the cloud. At the cloud, the data uploaded by multiple edge nodes are weighted averaged to generate a global traffic model.

[0045] Specifically, based on the function: ; ; Based on the amount of data corresponding to each edge node, the data uploaded by multiple edge nodes is weighted averaged; in, is the global traffic model; Indicates the number of edge nodes participating in federated learning; Representation Node The number of local N is the total amount of data; For edge nodes k The corresponding model parameters.

[0046] That is, edge distributed fine-tuning is achieved through federated averaging. The goal is to aggregate the local model parameters of multiple edge nodes to generate a global model. Each edge node uses local data to fine-tune the model and obtain the parameters. Each node uploads and The data is then sent to the cloud, where it is weighted averaged according to the amount of data to generate a new global model.

[0047] Step 104: compress the global traffic model through knowledge distillation, and migrate the compressed edge traffic model to the corresponding edge node to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0048] In one implementation of the present application, model lightweighting is achieved through knowledge distillation.

[0049] Goal: Migrate knowledge from large models in the cloud to small models at the edge.

[0050] Function-based: ; ; Migrate the compressed edge traffic model to the corresponding edge node; in, L is the loss function; For the cross entropy loss, align the marginal traffic model prediction data With the true label ; is the divergence loss, aligning the output probability of the edge traffic model and the global traffic model probability ; is the weight coefficient, usually 0.1; Logits output of the global traffic model; Logits output for edge traffic model; Represents the temperature parameter.

[0051] Furthermore, the model output is softened: ; in, and are the logits outputs of the global traffic model and the edge traffic model respectively. represents the temperature parameter, which is usually taken as 3, softening the probability distribution to transfer more dark knowledge.

[0052] Specifically, by Compressed into a lightweight edge model to adapt to edge devices.

[0053] In one implementation of the present application, each edge node inputs the acquired traffic information into the corresponding edge traffic model after every second preset time interval, so as to output the traffic signal strategy corresponding to the current road section through the edge traffic model. Also, when the average vehicle speed of the current road section is less than the preset speed threshold, or the vehicle queue length of the current road section is greater than the preset lane capacity threshold, the congestion response mechanism is triggered. The cloud determines the target speed and phase difference of the main road based on the global traffic data, and delays the traffic lights at the downstream intersection when the growth rate of the number of vehicles at the upstream intersection of the main road is greater than the preset growth rate threshold. Also, the cloud determines the vehicles heading to the main road, and sends recommended filtered speed information to the vehicles heading to the main road through V2X communication.

[0054] Specifically, the control strategy is mainly divided into two aspects: hierarchical response and green wave coordination. Among them, the hierarchical response mechanism is divided into normal optimization and sudden congestion response. The edge device integrates lane-level traffic flow, queue length and other data every 5 minutes, and generates a timing plan through a lightweight DRL model. When the real-time detection lane average speed is <15km / h, or the queue length exceeds 80% of the lane capacity, the congestion response mechanism is triggered, and the current green light phase is extended or an emergency phase is inserted in seconds to forcibly evacuate congested traffic.

[0055] Furthermore, the cloud platform calculates the ideal speed and phase difference of the main road based on the global traffic data. When the traffic at the upstream intersection surges, the green light duration at the downstream intersection is extended to avoid traffic accumulation. Through V2X communication, the green wave speed is recommended to approaching vehicles, such as prompting the driver through the navigation APP that "X km / h driving can pass Y green lights in a row."

[0056] Specifically, the edge device obtains the vehicle speed, driving time and license plate information corresponding to the vehicle, generates a speed driving sequence based on the corresponding speeds at different times, and uploads the speed driving sequences corresponding to different vehicles to the cloud.

[0057] The cloud analyzes the acquired vehicle speed sequence and inputs the vehicle speed sequence into the preset vehicle speed prediction model to output the predicted vehicle speed of each vehicle in the future time period, and determines the time when each vehicle arrives at the stop line of the main road according to the position and predicted vehicle speed of each vehicle. The preset vehicle speed prediction model can be a neural network model, and the training process of the model is to take the historical vehicle speed sequence samples as input, take the historical predicted vehicle speed data corresponding to the input samples as output samples, train the preset neural network model, and obtain the preset vehicle speed prediction model.

[0058] According to the number of intersections on the main road, the main road is divided into multiple sub-roads. According to the obtained arrival time of each vehicle at the stop line of the main road, the predicted traffic flow corresponding to the multiple sub-roads in the future time period is determined.

[0059] Furthermore, when the predicted traffic volume of the sub-trunk road is greater than the preset traffic volume threshold, the cloud sends a green light delay instruction to the edge device related to the sub-trunk road.

[0060] And, the vehicle's driving route is matched with each sub-road. When the vehicle's driving route is successfully matched with the sub-road, the location of all traffic lights on the sub-road is determined based on the geographical information of the sub-road. These traffic lights are grouped into a set to determine the set of traffic lights that the vehicle needs to pass through.

[0061] According to the control information of each signal light in the traffic light set and the predicted speed of the vehicle in the future time period, the recommended green wave speed corresponding to the vehicle is determined. Among them, the control information is the cycle length of the signal light, the green light start time, the green light duration and other information. Specifically, the time required for the vehicle to travel from the current position to the signal light can be obtained by dividing the distance between the vehicle and the signal light by the predicted speed of the vehicle. According to the cycle length of the signal light, the green light start time and the green light duration, the state of the signal light when the vehicle arrives is determined. If the signal light is in the green light state when the vehicle arrives, the remaining time of the green light is recorded; if it is in the red light state, the time when the next green light is on is calculated. Within the safe driving speed range of the vehicle, different speeds are traversed, and the total time when the vehicle encounters the green light or the number of green lights passing when passing the traffic light set at each speed is calculated. The speed that makes the number of green lights passing the largest or the total green light time the longest is selected as the recommended green wave speed.

[0062] It should be noted that when determining the recommended green wave speed, some actual conditions need to be considered, such as road speed limit, traffic flow, vehicle performance, etc. If the recommended green wave speed exceeds the road speed limit or the actual drivable speed range of the vehicle, the recommended speed needs to be adjusted to ensure the safety and feasibility of vehicle driving.

[0063] Figure 3 A schematic diagram of a dynamic control strategy mechanism provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, first, the data such as delays and emergency brake times corresponding to the lanes are collected, and the acquired data is uploaded to the cloud to update the reward function weights, and the optimized model is sent to the edge device to perform signal control through the edge device.

[0064] Figure 4 A practical application diagram of a dynamic traffic signal optimization system under a vehicle-road-cloud collaborative framework provided in an embodiment of the present application is shown in FIG. Figure 4 As shown in the figure, multiple edge devices upload the acquired lane data to the cloud respectively. The cloud optimizes the model through the PPO algorithm and the federated average algorithm, and sends the optimized model to the edge devices.

[0065] Furthermore, the embodiment of the present application is also provided with a system anti-interference and robustness design.

[0066] Specifically, when data is missing, the multimodal attention mechanism dynamically adjusts the weights to ensure reliable fusion results. In addition, an adaptive pruning method is used to remove neurons with low contribution, reduce the amount of calculation, achieve model lightweight, and improve anti-interference ability and robustness.

[0067] For example, in order to verify the technical effect of the invention, based on the actual road network parameters of a certain city's intersection, the morning rush hour traffic flow was simulated through the SUMO1.14.1 simulation platform for comparative experiments. Table 1 is a table of intersection timing simulation results provided by an embodiment of the present application, as shown in Table 1 for the experimental group parameters, road network topology: 4-phase signal lights, 3 main roads. Morning rush hour 7:00-9:00 traffic: 1200 vehicles / hour, penetration rate 10%; vehicle types: private cars 70%, buses 10%, trucks 20%; DRL model output dynamic timing delay <200ms. Comparison group parameters, traditional timing control fixed cycle 120 seconds, phase: straight 40s / left turn 30s / right turn 20s; ground sensor coil induction control triggers the green light to extend the maximum 15s. The simulation duration is 2 hours for the morning rush hour, of which the warm-up time is 30 minutes and the effective data collection time is 1.5 hours. During the model training and optimization process, the PPO algorithm trains the global DRL model in the cloud, and the federated learning aggregates 7 days of local data to fine-tune the parameters. The lightweight model is compressed to 9.8 million, and the edge inference delay is 45ms. As shown in Table 1, normal optimization: the timing plan is adjusted every 5 minutes, and the green wave coordinated control covers 3 intersections upstream and downstream. Emergency response: When the vehicle speed is detected to be <15km / h, the green light is extended by 15 seconds or the bus priority phase is inserted.

[0068] Table 1

[0069] Figure 5 The following is a schematic diagram of the structure of a dynamic traffic signal control device based on data fusion provided in an embodiment of the present application. Figure 5As shown, a dynamic traffic signal control device 200 based on data fusion includes: at least one processor 201; and a memory 202 in communication with the at least one processor 201; wherein the memory 202 stores instructions that can be executed by the at least one processor 201, and the instructions are executed by the at least one processor 201, so that the at least one processor 201 can: at the edge node, weight the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle through the multi-head attention mechanism, and perform data fusion based on the assigned weights to obtain multi-source traffic feature data; After each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into the preset signal optimization model; wherein, the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1, and is constructed by integrating meta-learning; the preset signal optimization model is updated and optimized through the PPO algorithm and the federated averaging method to generate a global traffic model; the global traffic model is compressed through knowledge distillation, and the compressed edge traffic model is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0070] A non-volatile computer storage medium provided in an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: at an edge node, weights are assigned to first traffic data obtained by a road monitoring device and second traffic data obtained by a positioning device of a floating vehicle through a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data; after each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model; wherein the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1, and is constructed by integrating meta-learning; the preset signal optimization model is updated and optimized through the PPO algorithm and the federated averaging method to generate a global traffic model; the global traffic model is compressed through knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

[0071] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0072] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various changes and variations. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A dynamic traffic signal control method based on data fusion, characterized in that: The method comprises: At the edge node, the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle are weighted by a multi-head attention mechanism, and data fusion is performed based on the assigned weights to obtain multi-source traffic feature data; After each first preset time interval, the cloud receives the multi-source traffic feature data uploaded by the edge node, and inputs the multi-source traffic feature data into a preset signal optimization model; wherein the preset signal optimization model is a multimodal Transformer architecture based on DeepSeek-R1 and is constructed by integrating meta-learning; The preset signal optimization model is updated and optimized by using the PPO algorithm and the federated average method to generate a global traffic model; Through knowledge distillation, the global traffic model is compressed, and the edge traffic model obtained after compression is migrated to the corresponding edge node, so as to output the traffic signal strategy for the corresponding road section through the edge traffic model.

2. A dynamic traffic signal control method based on data fusion according to claim 1, characterized in that: At the edge node, the first traffic data obtained by the road monitoring device and the second traffic data obtained by the positioning device of the floating vehicle are weighted by a multi-head attention mechanism, and data fusion is performed based on the weights assigned to obtain multi-source traffic feature data, specifically including: At the edge node, extracting first features from the first traffic data and the second traffic data based on spatial features; and, based on the time feature, extracting second features from the first traffic data and the second traffic data respectively; Dynamically weighting the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data, by using a multi-head attention mechanism; Based on the dynamic weight, the first feature and the second feature corresponding to the first traffic data, and the first feature and the second feature corresponding to the second traffic data, data fusion is performed to obtain the multi-source traffic feature data.

3. The method for dynamic traffic signal control based on data fusion according to claim 1, characterized in that: Before inputting the multi-source traffic characteristic data into a preset signal optimization model, the method further includes: Taking minimizing the average delay time at the intersection as the core goal, and reducing the number of emergency brakes and giving priority to public transportation vehicles as auxiliary goals, we construct a reward function: ； in, R is the reward function; w 1 is the first weight coefficient; w 2 is the second weight coefficient; w 3 is the third weight coefficient, ; D is the average delay time at the intersection; B is the number of emergency brakes; P Score the transit priority for public transport vehicles; Based on the traffic scenario corresponding to the edge node, the reward function is weighted to optimize the preset signal optimization model based on the adjusted reward function.

4. The method for dynamic traffic signal control based on data fusion according to claim 1, characterized in that: The preset signal optimization model is updated and optimized by the PPO algorithm and the federated average method to generate a global traffic model, specifically including: Based on the PPO algorithm, maximize the long-term cumulative rewards by optimizing the strategy; Among them, the objective function of the PPO algorithm is: ； in, represents the probability ratio of the new and old strategies; Optimize strategies for the new; Optimize strategies for legacy; Indicates the signal light control strategy; represents the advantage function; Indicates the clipping threshold; clip is the clipping function; are model parameters; For the agent at time step t According to the status s t the action performed; For the agent at time step t Status; The model parameters corresponding to the plurality of edge nodes are aggregated by using a federated averaging algorithm to generate the global traffic model.

5. A dynamic traffic signal control method based on data fusion according to claim 4, characterized in that: The method of using a federated average algorithm to aggregate the model parameters corresponding to the plurality of edge nodes to generate the global traffic model specifically includes: At each of the edge nodes, fine-tuning the preset signal optimization model using local data to obtain fine-tuning parameters; Uploading the obtained fine-tuning parameters and local data corresponding to each edge node to the cloud; At the cloud end, a weighted average is performed on the data uploaded by the plurality of edge nodes to generate the global traffic model.

6. A dynamic traffic signal control method based on data fusion according to claim 5, characterized in that: The weighted average of the data uploaded by the plurality of edge nodes specifically includes: Through the function: ； ； Based on the data volume corresponding to each edge node, weighted average is performed on the data uploaded by the plurality of edge nodes; in, is the global traffic model; Indicates the number of edge nodes participating in federated learning; Representation Node The number of local N is the total amount of data; For edge nodes k The corresponding model parameters.

7. The method for dynamic traffic signal control based on data fusion according to claim 1, characterized in that: The global traffic model is compressed through knowledge distillation, and the edge traffic model obtained after compression is migrated to the corresponding edge node, specifically including: Function-based: ； ； Migrate the compressed edge traffic model to the corresponding edge node; in, L is the loss function; For the cross entropy loss, align the marginal traffic model prediction data With the true label ; is the divergence loss, aligning the output probability of the edge traffic model and the global traffic model probability ; is the weight coefficient; Logits output of the global traffic model; Logits output for edge traffic model; Represents the temperature parameter.

8. The method for dynamic traffic signal control based on data fusion according to claim 1, characterized in that: The outputting of the traffic signal strategy for the corresponding road section by using the edge traffic model specifically includes: Each of the edge nodes inputs the acquired traffic information into the corresponding edge traffic model every second preset time interval, so as to output a traffic signal strategy corresponding to the current road section through the edge traffic model; And, when the average vehicle speed of the current road section is less than a preset vehicle speed threshold, or the length of the vehicle queue of the current road section is greater than a preset lane capacity threshold, a congestion response mechanism is triggered; The cloud determines the target vehicle speed and phase difference of the main road based on the global traffic data, and when the growth rate of the traffic volume at the upstream intersection of the main road is greater than the preset growth rate threshold, the traffic lights at the downstream intersection are delayed; Furthermore, the cloud determines the vehicles heading toward the main road, and sends recommended filtered vehicle speed information to the vehicles heading toward the main road through V2X communication.

9. A dynamic traffic signal control device based on data fusion, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 8.

Images