Ramp traffic signal adaptive control method and system based on cooperation of unmanned aerial vehicles

Abstract

The application provides a ramp traffic signal adaptive control method and system based on unmanned aerial vehicle cooperation, belongs to the technical field of intelligent traffic control, and comprises the following steps: collecting traffic data in real time based on an unmanned aerial vehicle cluster dynamically arranged above a ramp, constructing a global traffic state matrix, and performing preliminary processing through edge computing; generating a local optimal signal timing suggestion in real time through edge computing of the edge node on the preliminarily processed data; integrating and uniformly training the local optimal signal timing suggestions uploaded by multiple unmanned aerial vehicles on the cloud to generate a global optimal signal timing strategy; and adjusting the signal lamp parameters in real time based on the global optimal signal timing strategy, thereby enhancing the direct intervention ability on the on-site traffic through the air guidance function and providing a high-reliability solution for a smart traffic system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic control technology, and in particular relates to an adaptive control method and system for ramp traffic signals based on unmanned aerial vehicle (UAV) collaboration. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In urban expressways, such as ring roads and elevated highways, ramps are crucial nodes connecting main roads with auxiliary / local roads. If not properly controlled, they can easily become bottlenecks leading to congestion. Especially during peak hours, multi-lane ramps need to coordinate vehicles merging from ramps onto the main road or exiting from the main road onto ramps to ensure a smooth connection between main road and ramp traffic.

[0004] Existing ramp signal control systems primarily rely on fixed-point monitoring devices such as ground-based induction coils and cameras to acquire vehicle flow and queue length information. They then employ preset control strategies or model-based algorithms for signal timing, such as feedback control and fuzzy control. However, these methods have significant limitations: First, ground sensors have limited coverage and cannot fully perceive the complex multi-lane traffic conditions at ramp entrances. Detection and response are often insufficient in the event of sudden surges in traffic flow or abnormal events such as accidents or vehicle malfunctions. Second, fixed or semi-fixed control strategies lack adaptability and struggle to adapt to dynamic changes in traffic flow. Especially in scenarios with highly variable traffic flow, such as ring road entrances or urban elevated highway ramps, traditional control can easily lead to oversaturation of one lane while wasting capacity in the other. Third, existing technologies often lack coordination between main road and ramp signals. Ramp signal control fails to adequately consider main road traffic conditions, resulting in poor traffic flow between the main road and ramps.

[0005] Due to their advantages such as wide field of view and flexible deployment, drones hold promise for traffic data collection and auxiliary control in ramp areas. Currently, there are related solutions using drone technology to achieve highway traffic flow guidance, highway ramp pavement condition detection, etc. However, these solutions cannot achieve real-time and comprehensive perception of ramp traffic conditions, nor can they coordinate the signal timing between the mainline and the ramps. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides an adaptive traffic signal control method for ramps based on UAV collaboration, which can improve ramp traffic efficiency, reduce vehicle delay rate, and intelligently coordinate signal timing between main roads and ramps.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] Firstly, a method for adaptive control of ramp traffic signals based on UAV collaboration is disclosed, including:

[0009] Based on the real-time collection of traffic data by a swarm of drones dynamically deployed above the ramp entrance, a global traffic state matrix is ​​constructed and preliminarily processed through edge computing.

[0010] By utilizing edge nodes, edge computing is performed on the pre-processed data to generate local optimal signal timing suggestions in real time.

[0011] The system integrates and trains locally optimal signal timing suggestions uploaded by multiple drones in the cloud to generate a globally optimal signal timing strategy.

[0012] The traffic light parameters are adjusted in real time based on the globally optimal signal timing strategy.

[0013] As a further technical solution, a drone swarm is dynamically deployed over the ramp entrance. Specifically, the drone swarm is deployed in a grid-like topology structure, covering the ramp entrance, exit and adjacent main road areas. The drones are equipped with multispectral cameras, lidar and millimeter-wave radar to collect video streams and 3D point cloud data at set frequencies.

[0014] As a further technical solution, preliminary processing via edge computing includes:

[0015] Vehicle detection and parameter calculation are performed based on real-time traffic data.

[0016] Based on vehicle detection and parameter calculation, identify sudden events at ramp entrances and generate corresponding emergency signs.

[0017] As a further technical solution, the method of identifying sudden events at ramp entrances and generating corresponding emergency signs based on vehicle detection and parameter calculation includes the following parameter calculation:

[0018] Traffic flow change detection: Calculate the rate of change of traffic flow in adjacent time windows and trigger an emergency indicator;

[0019] Speed ​​variance detection: Calculate the variance of vehicle speeds within a lane to detect abnormal congestion or accidents;

[0020] By combining sudden changes in traffic flow and speed variance, a comprehensive emergency sign is generated.

[0021] As a further technical solution, edge nodes are used to perform edge computing on the pre-processed data to generate local optimal signal timing suggestions in real time, specifically including:

[0022] Based on real-time traffic data, a bidirectional LSTM network is used to process the time series data of traffic parameters and extract traffic state features.

[0023] A lightweight DDPG model is constructed to perform local reinforcement learning on the input traffic state features and generate locally optimal signal timing suggestions.

[0024] As a further technical solution, the locally optimal signal timing suggestions uploaded by multiple drones are integrated and uniformly trained in the cloud to generate a globally optimal signal timing strategy, specifically including:

[0025] Each drone maintains a local experience replay cache pool to store local interaction samples and corresponding TD errors;

[0026] The cloud periodically samples and merges samples from each drone experience pool according to priority. The priority can be designed to be proportional to the absolute value of the TD error.

[0027] Slowly synchronize the Critic target network parameters to ensure that training information can be smoothly transferred between multiple agents;

[0028] Construct a digital twin model that closely matches the actual ramp entrance, minimize conflict time by making adjustments, compare the target flow or target queue length of each lane with the actual observed value in real time, and dynamically adjust the green light duration.

[0029] Simulate different traffic flow scenarios, compare and optimize the average ramp delay, average mainline speed, and conflict time indicators before and after optimization, and obtain the globally optimal signal timing strategy in the cloud through multiple iterations of training and simulation verification.

[0030] As a further technical solution, the traffic light parameters are adjusted in real time based on a globally optimal signal timing strategy, specifically including:

[0031] Dynamic control of traffic lights based on green light duration adjustment formula;

[0032] The priority release policy is triggered based on the generated emergency sign or emergency vehicle request.

[0033] By using drones as V2X communication relay nodes, signal status and speed suggestions are broadcast to ground vehicles, and guidance information is projected from the air using lasers to collaboratively optimize traffic flow.

[0034] As a further technical solution, it also includes:

[0035] Real-time monitoring and diagnosis of the failure status of each drone;

[0036] If a drone is detected i If a failure occurs, dynamic task reorganization and coverage area reconstruction will be performed.

[0037] Once the failure status of a drone is confirmed, if a hot standby drone exists, its status will be changed. The activated hot standby drone will autonomously fly to the location of the failed drone or take over its mission after the deployment delay, based on the optimized new location.

[0038] In extreme cases where multiple drones fail simultaneously and no backups are available, limited sensing resources are reallocated based on key indicators for each road segment.

[0039] Secondly, a drone-based adaptive traffic signal control system for ramp intersections was disclosed, including:

[0040] The real-time traffic data acquisition module is configured to: collect traffic data in real time based on a cluster of drones dynamically deployed above the ramp entrance, construct a global traffic state matrix, and perform preliminary processing through edge computing;

[0041] The local optimal signal timing suggestion generation module is configured to: use edge nodes to perform edge calculations on the pre-processed data to generate local optimal signal timing suggestions in real time.

[0042] The global optimal signal timing strategy generation module is configured to integrate and uniformly train the local optimal signal timing suggestions uploaded by multiple drones in the cloud to generate a global optimal signal timing strategy.

[0043] The real-time signal light parameter adjustment module is configured to adjust the signal light parameters in real time based on the globally optimal signal timing strategy.

[0044] The above one or more technical solutions have the following beneficial effects:

[0045] This invention utilizes the wide-area perception capabilities of a multi-UAV swarm, combined with low-latency decision-making through edge computing and dynamic optimization through multi-agent reinforcement learning, to achieve comprehensive, real-time perception and intelligent, efficient, and adaptive control of traffic flow at ramp entrances. This enables the system to automatically adjust traffic light timings based on real-time traffic flow, queue length, and speed information, significantly improving ramp efficiency and reducing vehicle delays.

[0046] Furthermore, when some drone nodes fail, the system can automatically trigger task reorganization, dynamic replacement, and even "graceful degradation" operation, ensuring the continuity and stability of core control functions. It can also quickly respond to emergencies such as traffic accidents and emergency vehicles, prioritizing emergency passage and improving safety and efficiency in special circumstances. In addition, drones, as communication relay nodes, expand V2X communication coverage, promote vehicle-road cooperation, and enhance the ability to directly intervene in on-site traffic through aerial guidance, providing a highly reliable solution for intelligent transportation systems.

[0047] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0048] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0049] Figure 1 This is a system architecture diagram of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the data acquisition and fusion process of the present invention;

[0051] Figure 3 This is a schematic diagram of the artificial intelligence processing module of the present invention;

[0052] Figure 4 This is a flowchart of the multi-UAV self-healing mechanism of the present invention;

[0053] Figure 5 This is a schematic diagram of the overall hardware connection of the system of the present invention. Detailed Implementation

[0054] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0055] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0056] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0057] Based on existing ramp signal control methods, such as fixed-cycle control or simple inductive control, there is a lack of flexible and mobile three-dimensional monitoring means. It is often difficult to adjust in a timely manner according to instantaneous traffic conditions, resulting in long waiting times for vehicles on ramps or queues of vehicles on the main line, which reduces traffic efficiency and increases delays.

[0058] Terminology Explanation:

[0059] Multi-UAV Collaboration: A technology in which multiple UAVs share real-time observation data through distributed communication networks (such as V2X, 5G) and jointly optimize traffic control objectives based on collaborative decision-making algorithms (such as multi-agent reinforcement learning). In this invention, the UAV swarm is deployed in a grid-like topology, covering multi-lane areas at ramp entrances, and achieves dynamic generation of global traffic state perception and control commands through data fusion and task allocation.

[0060] Edge computing is a technology that pushes data processing and decision-making functions closer to the data source (such as onboard computing units on drones or roadside nodes) to reduce system latency. In this system, the drone carries a lightweight reinforcement learning model to process traffic data in real time and generate local control commands, ensuring that the response latency for signal timing adjustments is ≤200ms.

[0061] Multi-Agent Reinforcement Learning (MARL) is a distributed artificial intelligence algorithm that treats multiple agents, such as drones and signal controllers, as independent decision-making units. By sharing state information and collaborative reward functions, they jointly optimize a global objective, such as minimizing total delay. This invention employs the Deep Deterministic Policy Gradient (DDPG) algorithm, trains an Actor-Critic network in the cloud, and deploys a lightweight model at the edge layer to achieve collaborative optimization of ramp and mainline signals.

[0062] V2X (Vehicle-to-Everything Communication): A core component of vehicle-to-everything (V2X) technology, supporting real-time information exchange between vehicles and infrastructure (V2I), between vehicles and other vehicles (V2V), and between vehicles and cloud platforms (V2N). In this system, the drone acts as a communication relay node, broadcasting traffic light status, countdown information, and speed guidance instructions via DSRC or 5G NR-U protocols, while simultaneously receiving priority passage requests from vehicles.

[0063] Cloud-Edge-Air-Ground Collaborative Architecture: A layered system architecture that integrates global optimization in the cloud, real-time decision-making at the edge, UAV swarm perception, and execution functions by the ground controller. The cloud is responsible for model training and policy distribution, the edge layer processes local data, UAVs provide wide-area monitoring and communication relay, and the ground controller executes signal timing commands to achieve closed-loop optimization of traffic control.

[0064] Heartbeat Mechanism: This refers to a communication mechanism used to monitor whether nodes in a system or network are functioning properly. In this implementation example, a node refers to a drone.

[0065] Overall Concept: This embodiment's sub-solution addresses the issues of limited perception and lack of global coordination in existing technologies. Existing technologies typically use a single drone or a loosely distributed group of drones to monitor specific road sections, failing to provide real-time, comprehensive, and complete monitoring of the overall traffic status of the area connecting ramps and main roads. This sub-solution, by deploying a grid-like topology drone swarm and combining it with multi-spectral cameras, LiDAR, and millimeter-wave radar—multimodal sensors—constructs an integrated cloud-edge-air-ground collaborative architecture. This achieves comprehensive, real-time perception of traffic flow in ramp areas, resolving the issue of limited perception. It also addresses the problems of delayed response and lack of real-time decision-making and control.

[0066] Existing drone-based solutions often require data collection and subsequent offline analysis in the cloud or manual intervention, failing to provide real-time, automatic signal timing adjustments for rapidly changing traffic conditions such as sudden accidents and traffic surges. This invention introduces edge computing, enabling direct data preprocessing and local decision-making on drones or roadside units. Utilizing bidirectional LSTM and a lightweight DDPG model, it can generate locally optimal signal timing suggestions in real time, ensuring low decision latency (≤200ms) and meeting the new real-time requirements of ramp control. It also solves the problems of isolated control and lack of intelligent coordination between the main line and ramps.

[0067] Existing technologies have failed to effectively address the coordination and optimization problem between ramp signals and main road signals, often leading to contradictions such as smooth ramps impacting the main road or smooth main road traffic causing ramp congestion. This embodiment employs Multi-Agent Reinforcement Learning (MARL) in the cloud, integrating local suggestions uploaded by various edge nodes and generating a globally optimal signal timing strategy through training and simulation verification in a digital twin environment. This achieves dynamic and intelligent coordination of ramp and main road traffic flows, solving the problem of isolated control. It also addresses the issues of poor system robustness and lack of fault tolerance.

[0068] Existing drone application solutions rarely consider the impact of drone failures, such as battery depletion, communication interruptions, or hardware malfunctions, on the overall traffic management system. This embodiment proposes a complete system self-healing mechanism. When a drone failure is detected, the system can automatically perform dynamic task reorganization, activate a hot standby drone for replacement, and even implement degraded operation in extreme cases. This meets the new demands of high-reliability and high-stability traffic control systems and is a crucial element generally lacking in existing technologies.

[0069] Example 1

[0070] See appendix Figure 2 , 3 As shown in Figure 4, this embodiment discloses an adaptive control method for ramp traffic signals based on UAV cooperation, including:

[0071] Step 1: Multi-UAV Collaborative Perception and Data Acquisition: A cluster of UAVs is dynamically deployed over the ramp entrance, collecting traffic data in real time through multimodal sensors, constructing a global traffic state matrix, and performing preliminary processing through edge computing;

[0072] Step 2: Edge computing and distributed reinforcement learning decision-making: Generate local optimal signal timing suggestions in real time at edge nodes to ensure low latency response;

[0073] Step 3: Cloud-based global optimization and strategy fusion: Generate the globally optimal signal timing strategy through multi-agent collaborative training and securely distribute it to the execution layer;

[0074] Step 4: Dynamic Execution of Signals and Coordinated Traffic Flow Guidance: Adjust traffic light parameters in real time and guide vehicles through V2X communication to achieve global traffic flow optimization;

[0075] Step 5: System response and self-healing mechanism for partial drone failure: Ensure that when a single or a small number of drones unexpectedly leave service due to reasons such as power depletion, communication interruption or hardware failure, the entire ramp traffic signal adaptive control system can still maintain its core functions, achieve graceful degradation and restore full functionality as soon as possible.

[0076] In one implementation example, step 1 is used for traffic data collection, as follows:

[0077] S1-1: The DJI Matrice 300 RTK drone is used, equipped with an H20T multispectral camera, LiDAR and millimeter-wave radar. It collects high-definition video streams and 3D point cloud data at a frequency of 10Hz, and collects traffic status data such as vehicle flow, vehicle speed and queue length in real time above the ramp, and constructs a global traffic status matrix.

[0078] Global Traffic State Matrix Construction Method: The global traffic state matrix is ​​the result of data fusion. Local vehicle location, speed, and flow data collected by various drones are mapped onto a unified geographic coordinate system (GIS) grid. Each element of the matrix corresponds to a specific area at a ramp entrance, such as lane ID, and traffic parameters for a specific location segment at a specific time. These traffic parameters include flow rate f, speed v, and queue length q.

[0079] This matrix, as input data, is transmitted to the edge computing nodes and used as input to the bidirectional LSTM network in step 2 to extract temporal features. It is also used to construct the state space of the reinforcement learning DDPG, serving as the basis for agent decision-making.

[0080] S1-2: The drones are deployed in a grid-like topology to cover the ramp entrances, exits, and adjacent main road areas, ensuring comprehensive monitoring without blind spots. Each drone is spaced 150m apart and flies at a height of 30m, ensuring a sensor field of view (FOV) overlap rate of ≥20%, eliminating blind spots, and improving the breadth, real-time performance, and accuracy of data acquisition.

[0081] S1-3: Based on the onboard sensors and the distribution of the UAV, vehicle detection and parameter calculation are performed. An attention mechanism, namely the CBAM module, is introduced on the basis of YOLOv11 to improve the detection accuracy of small target vehicles.

[0082] Vehicle detection is achieved by using a camera mounted on a drone to acquire video streams, which are then input into a YOLOv11+CBAM object detection model deployed on an edge computing unit. The model outputs the bounding box and category of the vehicle in each frame of the image.

[0083] Parameter calculation implementation: Flow: The number of vehicle bounding boxes passing through the virtual detection line per unit time. Speed: The instantaneous speed is calculated by the displacement of the same vehicle bounding box between adjacent frames and the frame rate of 10Hz. Queue length: The physical length of a continuous, closely spaced queue of vehicles within a lane whose speed is below a threshold, such as 5km / h.

[0084] YOLOv11 and CBAM are the core algorithms in the aforementioned vehicle detection steps. CBAM (Convolutional Block Attention Module) is inserted into the YOLO network as an attention module to weight key features such as vehicles and suppress background noise such as road surfaces and trees, directly determining the detection accuracy and thus affecting the accuracy of subsequent traffic flow and speed calculations. Execution hardware includes high-performance computing modules onboard the UAV, such as NVIDIA Jetson series, and roadside edge computing nodes with communication connections.

[0085] S1-4: Based on vehicle detection and parameter calculation, it is necessary to quickly identify emergencies such as accidents and severe congestion, and generate corresponding emergency flags so that the system can trigger priority control strategies. Therefore, calculations are performed on traffic flow mutations, speed variance, and multi-indicator joint decision-making.

[0086] The calculations are performed on hardware: edge computing nodes, located on the UAV's onboard or roadside end. These calculations are used to generate the Emergency Flag. If the calculated value exceeds a threshold, the system will trigger an emergency priority control strategy in step 4 (S4-2), forcibly switching the traffic light phase, such as all red or a green light in a specific direction, and initiating UAV airborne guidance.

[0087] Flow mutation detection: Calculate the adjacent time window Δt win A flow rate change rate of 10 seconds triggers the emergency indicator.

[0088]

[0089] in, Indicates the length of the time window, in seconds, for example, the time interval used to calculate the rate of change in flow. Indicates the current time window Traffic flow within the area This indicates the vehicle flow rate within the previous time window. For the current time window The difference between the traffic flow in the previous time window and the traffic flow in the previous time window.

[0090] The flow rate change calculated by the above formula It needs to be compared with a preset mutation threshold, such as a change rate > 50%. When the calculated result is greater than this threshold, an emergency flag is triggered, and the Flag is set to 1.

[0091] Speed ​​Variance Detection: Statistically calculates the variance of vehicle speeds within a lane to detect abnormal congestion or accidents.

[0092]

[0093] in, The variance of vehicle speed within a lane measures the degree of speed dispersion and reflects the likelihood of congestion or accidents. This represents the total number of vehicles detected in the lane during the statistical period. Indicates the first i The instantaneous speed of the vehicle, This represents the average speed of all vehicles in the lane.

[0094] Multi-indicator joint decision-making: Combining flow mutation and velocity variance to generate a comprehensive emergency indicator.

[0095]

[0096] in, For comprehensive emergency indicator values; , These are the weighting coefficients for traffic mutation and speed variance, respectively, which can be optimized offline based on historical data. Traffic flow indicates the number of vehicles passing through a specific detection section on a ramp / mainline per unit time, used to represent vehicle arrival intensity. It represents the speed variance, reflecting speed fluctuations and dispersion. It tends to increase during accidents, bottlenecks, shock waves, and lane-changing / merging interference.

[0097] In this implementation example, It is a weighted value, and the system will set a comprehensive judgment threshold. In the signal control logic of step 4, if the real-time calculation , This is a threshold threshold, also known as a trigger threshold, used to determine whether the current situation is an abnormal congestion or an accident. At this point, the signal control algorithm exits the conventional PID / reinforcement learning mode and enters the S4-2 emergency priority control mode, executing specific emergency timing schemes, such as extending the mainline red light.

[0098] The overall advantages of the above data collection steps are as follows: Compared with traditional ground sensors, such as coils and geomagnetic sensors, this step utilizes a drone swarm to achieve wide-area, blind-spot-free perception, dynamically adapting to the complex geometry of ramp entrances, and solving the problems of difficult maintenance, limited coverage, and inability to perceive queue length and traffic flow trajectories from a macroscopic perspective.

[0099] The advantages of each of the above steps are as follows: S1-1 Multimodal Acquisition: Combining visible light, LiDAR, and millimeter-wave radar, this achieves all-weather, high-precision perception. LiDAR supplements depth information, and millimeter-wave radar ensures detection stability in adverse weather conditions. S1-2 Mesh Topology: Setting the FOV overlap rate to ≥20% eliminates blind spots, ensures continuity of vehicle tracking across cameras, and achieves true global matrix construction. In S1-3, YOLO+CBAM improves the detection accuracy of small targets such as distant vehicles. The attention mechanism suppresses interference from complex backgrounds and improves the accuracy of traffic parameter calculations. S1-4 Emergency Recognition: This achieves millisecond-level anomaly event detection, providing a trigger source for subsequent emergency control without manual intervention.

[0100] Specifically, the construction used for edge computing in step 2 is as follows:

[0101] The computing units integrated on the drone or nearby edge computing nodes perform low-latency preprocessing and local optimization of the collected traffic data. Specifically, this includes traffic state feature extraction, processing time-series traffic parameter data using a bidirectional LSTM network, and constructing a lightweight DDPG model for local reinforcement learning policy generation. The Actor network outputs actions such as signal period, phase difference, and green light duration, while the Critic network evaluates the value of these actions and stores local interaction data for experience replay, ensuring low-latency decision-making response.

[0102] In this implementation example, the computational tasks of step 1 (data acquisition and preprocessing), step 2 (local optimization suggestions), and sudden event detection are all undertaken by the UAV's onboard computing unit or roadside edge computing nodes. This demonstrates the low latency characteristic of edge computing.

[0103] During traffic state feature extraction: the raw parameters collected in step 1 are serialized and input into a bidirectional LSTM (Long Short-Term Memory) network. The bidirectional LSTM can simultaneously utilize historical and future contextual information, mapping time-series data into high-dimensional feature vectors. Specific characteristics include, but are not limited to, the historical traffic flow trends of each lane, the average speed fluctuation characteristics, the queue length growth rate, and the spatial and temporal occupancy rate.

[0104] The lightweight DDPG model described above is constructed using the DDPG deep deterministic policy gradient, which is an algorithm based on an Actor-Critic architecture.

[0105] The lightweight DDPG model includes an input layer that receives state features extracted by LSTM. Network Structure: To achieve lightweight design and adaptability to edge devices, a neural network structure with fewer layers (e.g., 2-3 fully connected layers) and a moderate number of nodes (e.g., 64 / 32 nodes) is adopted. Output Layer: The Actor network directly outputs continuous action values, i.e., timing parameters.

[0106] Traffic parameter time series data is a sequence of traffic parameters collected and calculated in step 1, such as flow rate, speed, and queue length at times t, t-1, t-2, etc. These constitute time series data.

[0107] S2-1: Based on the data collected in step 1, traffic state features are extracted; a bidirectional LSTM network is used to process the time series data of traffic parameters. The network structure is: 64 nodes in the input layer, 128 nodes in the bidirectional LSTM layer, and 32 nodes in the fully connected layer.

[0108] State vector generation formula It can be represented as:

[0109]

[0110] in, For bidirectional LSTM at time Hidden state splicing, These represent the weights and biases of the fully connected layer. This is the activation function.

[0111] S2-2: Construct a lightweight DDPG model to generate local reinforcement learning policies and achieve low-latency local signal timing decisions.

[0112] Actor network: Input Output action , For signal period, For phase difference, Green light duration;

[0113]

[0114] in, Indicates the Actor network at time... t The output action vector, This represents the Actor network policy function. Indicates exploratory noise. This represents the deterministic policy function of the Actor network, i.e., the mapping relationship within the network itself. The vector representing the state feature at time t is the input data. These represent the weight and bias parameters in the Actor neural network, which are variables that need to be updated during training.

[0115] The Actor policy network is responsible for making decisions. Based on the current traffic conditions, it directly calculates and outputs specific signal control action vectors, namely: signal cycle length, phase difference, and green light duration.

[0116] Critic network: used to evaluate the value of actions. Guide the Actor network update; input state vector With action Output the Q value corresponding to the action. Used to evaluate in state Take action below The value of.

[0117]

[0118]

[0119] in, The immediate reward obtained at the current moment can be designed as a comprehensive indicator such as reduced ramp delays and reduced mainline queue length; Discount factor; and These are the network parameters for the target Critic and the target Actor, obtained through soft updates. Indicates at time Under this decision, actions ( The estimated value of the action that should be achieved; It is the loss function of the Critic, i.e., Q-network, used to train / update the parameters of the Critic. Let Critic predict Try to get as close as possible to the TD target value , It represents the expectation of the empirical sample; and These are the parameters of the online Actor / Critic network.

[0120] Meanwhile, to generate preliminary optimization suggestions for local ramp signals in real time and ensure a decision latency of less than 200 ms, a local playback pool is maintained at the edge node to store interaction samples. Periodic sampling is used for network updates.

[0121] Specifically, in step 3, the cloud-based global optimization and multi-agent collaborative training method involves the cloud control platform integrating and uniformly training the local experiences uploaded by multiple UAVs and then distributing the globally optimal strategy. This mainly consists of two parts: experience fusion and simulation verification.

[0122] The cloud-based control platform is responsible for the unified management and scheduling of multiple drones, and utilizes powerful computing resources to perform global optimization and policy fusion of multi-agent reinforcement learning models. Specifically, this includes multi-agent experience fusion, which uses a distributed experience pool and a soft update mechanism to slowly synchronize the Critic target network parameters; and the construction of a digital twin model in the SUMO traffic simulation platform for simulation verification, optimizing conflict time minimization and green light duration adjustment to evaluate the feasibility and effectiveness of the strategies.

[0123] S3-1: Multi-agent experience fusion: Each UAV and its corresponding edge signal control node maintain a local experience replay cache pool to store local interaction samples and corresponding TD errors (Temporal Difference Error). The cloud periodically samples and merges samples from each drone's experience pool according to priority. The priority can be designed to be proportional to the absolute value of the TD error.

[0124]

[0125] in, Indicates the first i The probability that a sample of experience is selected from the cloud for training. This is a priority index factor used to control the sampling probability of high-error samples. Indicates drone i The Temporal Difference Error of empirical samples is used to measure the importance of the sample for updating the Critic network.

[0126] Then, a soft update is performed to slowly synchronize the Critic target network parameters with S2-2 to ensure that training information can be smoothly transferred between multiple agents.

[0127] ( )

[0128] This represents the parameters of the Target Critic Network, including weights and biases. This represents the soft update coefficient. This represents the parameters of the current Main Critic Network.

[0129] S3-2: Simulation Verification and Strategy Evaluation: In the SUMO (Simulation of Urban MObility) traffic simulation environment, construct a digital twin model that is highly consistent with the real ramp, including: multi-lane ramps, mainline traffic flow patterns, intersection signal phase logic, etc.

[0130] First, conflict time must be minimized; conflict time refers to the total time that vehicles on ramps and mainlines wait for each other due to improper signal phases. This can be achieved by adjusting the phase difference. This will reduce conflicts during peak periods.

[0131]

[0132] This indicates that this is an optimization problem, and the goal is to find the optimal phase difference. This ensures that the subsequent calculation result, i.e., the conflict time index, reaches its minimum value. Indicates time step Accumulate the sum. That is, calculate the sum of all moments over a period of time. It is a logic judgment switch.

[0133] At the current time During peak periods, the function value is 1, and the phase difference is included in the total cost.

[0134] At the current time During off-peak periods, the function value is 0, and the effect of phase difference is ignored.

[0135] Here, 1 is an indicator function, taking the value 1 when the peak period is in effect and 0 otherwise. The key is to reduce collision time through phase difference optimization.

[0136] Secondly, the green light duration is adjusted. In the simulation, the target flow rate or target queue length of each lane is compared with the actual observed value in real time, and the green light duration is dynamically adjusted to achieve local balance. Based on PID feedback control, the flow rate of each lane is dynamically balanced.

[0137]

[0138] in, This represents the difference between the target queue length and the actual queue length. These are the proportional, integral, and derivative coefficients of the PID controller, respectively. The time interval between each system update / sampling control is called the sampling time interval. This indicates the length of the time window, which is adjusted using the PID control described above. Feedback is sent to the simulation model to update the signal timing. Indicates at time / the Each control cycle represents the change in the green light duration, i.e., the adjustment amount.

[0139] Different traffic scenarios, including peak / non-peak and accident disturbances, were repeatedly simulated in SUMO, and the average ramp delay, average mainline speed, and collision time were compared before and after optimization. Through multiple iterative training and simulation verification, the globally optimal signal timing strategy in the cloud was obtained, including the optimal phase difference. Adjustment scheme for cycle length and green light duration.

[0140] In step 4, the global strategy optimized in the cloud is distributed to the edge execution layer, namely the ground signal controller and the UAV, and real-time execution and coordination are achieved through V2X communication and UAV air guidance.

[0141] Ground traffic signal controllers act as execution terminals, receiving and executing optimized signal timing instructions from the cloud, with drones assisting in traffic flow guidance. This includes dynamic signal light control, adjusting the cycle, phase, and green light duration of traffic lights at ramp entrances in real time; emergency priority control, automatically triggering emergency control mode and forcibly switching signal phases when an accident or approaching emergency vehicle is detected; and V2X collaborative guidance, traffic flow guidance, and information broadcasting. Drones act as communication relay nodes, broadcasting signal light status, countdowns, and speed guidance information via DSRC or 5G NR-U protocols, and using onboard speakers or visible light signaling devices to directly broadcast this information to vehicles in the ramp area from the air.

[0142] The details of the signal dynamic execution and traffic flow coordinated guidance method are as follows:

[0143] S4-1: Dynamic control of traffic lights.

[0144] PID parameter tuning includes proportional coefficient Kp=0.8, integral time Ti=10s, and derivative time Td=2s;

[0145] The formula for adjusting green light duration is:

[0146]

[0147] in, Indicates the first The green light duration for the current release phase, i.e., the ramp phase, in each control cycle; Indicates at time / the Each control cycle represents the amount of change in the green light duration, i.e., the adjustment amount. This indicates the new green light duration to be executed in the next control cycle after adding the above adjustment amount to the current green light duration.

[0148] In traffic signal control, the signal cycle is usually fixed or varies slowly. Red light duration is typically a derived variable. For a two-phase signal (ramp vs. mainline), if the calculated green light duration for the ramp phase is... The red light duration for the main phase is usually... Yellow light time Indicates the duration of the red light in the main phase. The duration of the green light for the ramp phase. Indicates the yellow light (clear / transition) time during phase switching.

[0149] The PID formula calculates the adjustment amount for the green light duration of the current traffic phase, such as a ramp. The light duration for other phases is based on the total signal cycle length. The result is automatically calculated by subtracting the current green light time. , This indicates the total green light duration that can be allocated to other phases or directions besides the current phase / current permitted direction; The green light duration currently allocated to the phases or directions that are being prioritized / controlled. This indicates the transition time taken for the yellow light and the necessary full red light clear, used for safe phase switching.

[0150] The specific execution process is as follows: The ground signal controller receives the current cycle initialization green light duration from the edge node or the cloud. Then calculate the error in real time. The PID formula in S3-2 is used to obtain... Finally, update and execute the new green light duration, and continue the cycle in the next period.

[0151] S4-2: Emergency Priority Control: Emergency flags generated by S1-4 for emergency events. Or an emergency vehicle request triggers a priority release policy.

[0152] First, the calculation of the upper limit of emergency vehicle passage time can be expressed as:

[0153]

[0154] in, This represents the path distance from the current ramp entrance to the destination for emergency vehicles. The preset safe driving speed; To allow for an additional buffer time, such as 5 seconds, to deal with unexpected situations. This indicates the maximum time limit allowed for completion or effectiveness under the current task / emergency scenario.

[0155] When an emergency vehicle requests arrival, if the main lane light is currently green and the new cycle will switch to a red light, the red light duration will be forcibly extended until... Priority will be given to allowing access to the emergency lanes on the ramps.

[0156] like If a V2X communication request is received from an emergency vehicle within a set time period, the ground signal controller immediately switches the signal phase, turning the ramp lights green or extending the mainline red lights; the drone hovers above the emergency vehicle's current path, using visible lights or loudspeakers to alert other vehicles to slow down and give way. After the emergency vehicle has passed, the system returns to PID dynamic control mode.

[0157] S4-3: V2X Collaborative Bootstrapping.

[0158] By using drones as V2X communication relay nodes, signal status and speed suggestions are broadcast to ground vehicles, and guidance information is projected from the air using lasers to collaboratively optimize traffic flow.

[0159] The recommended broadcast formula is as follows:

[0160]

[0161] in, Recommended speed (km / h); The current queue length (m); The estimated remaining time (s) for the next green light phase; Speed ​​limits are imposed on roads.

[0162] If the target speed is to pass through the current queue at a constant speed according to the remaining green light time, then the target speed should satisfy the following condition: If the speed exceeds the speed limit, then the speed limit will be used. .

[0163] Meanwhile, in response to sudden emergency scenarios, drones use lasers to project green navigation arrows or lane guidance lines onto the road surface, guiding vehicles to actively avoid emergency lanes and improving passage efficiency.

[0164] Step 5: System fault tolerance and self-healing mechanism guidance information.

[0165] The cloud-based control platform monitors the health status of the drone swarm in real time. When a failure of some drones is diagnosed through a heartbeat mechanism or fault codes, the system immediately activates a self-healing mechanism. This mechanism includes: first, dynamically reconstructing the positions and tasks of the remaining healthy drones through optimized algorithms to prioritize coverage of key areas and minimize perception blind spots; second, the multi-agent reinforcement learning model automatically adapts to changes in the number of agents to ensure decision-making continuity; and finally, if hot standby drones are deployed, the system can automatically activate and dispatch them to replace the failed units, quickly restoring the system's full functionality, thereby ensuring high stability and reliability of the entire control system in the face of single-point failures.

[0166] S5-1: Real-time Failure Monitoring and Diagnosis: The cloud-based control platform continuously monitors the status of each drone in the cluster via a heartbeat mechanism. (Drone) i Failure status flag (0 indicates normal, 1 indicates failure) The following logic applies:

[0167]

[0168] in, To determine whether a drone was received at time t i The heartbeat signal is represented by 1 for received and 0 for not received. For heart rate monitoring cycle, The preset timeout threshold is used for judgment. It is a drone i The reported fault code, once The system immediately activated its contingency plan. Indicates the first The error / fault codes currently reported by each node; This represents the set of fatal error codes. Once an error code belongs to this set, it is directly determined to be a fault.

[0169] S5-2: Dynamic Task Reorganization and Coverage Area Reconstruction: Once a drone is detected... i Failure =1, the cloud control platform will aim to minimize perception blind spots and reconstruction costs for the remaining healthy drone swarm. Perform dynamic reconstruction. This process can be formalized as an optimization problem:

[0170]

[0171] in, It is a new set of target location coordinates for health drones. This refers to the uncovered monitoring area resulting from the new location deployment, calculated using the following formula: ,in It is the total area under monitoring. It is a drone j new location The effective coverage area. It is a drone j From the original position Move to a new location The square of the distance represents the cost of movement. and These are preset weighting coefficients, used to balance the importance of coverage and move costs.

[0172] Meanwhile, in the multi-agent reinforcement learning (MARL) model, when agent i fails, the agent set A will be dynamically updated to... Global state observation vector S t and joint action vector A t The dimensions are reduced accordingly, and the model continues to make collaborative decisions based on the new set of agents A′.

[0173] S5-3: Redundancy Backup and Rapid Replacement Mechanism: To achieve rapid self-healing, the system can activate a hot standby drone. When the drone... i Failure status =1 is confirmed if a hot standby drone exists. s Its state Ss This will trigger a transformation:

[0174]

[0175] Once activated, the hot standby drone will autonomously fly to the location of the disabled drone. p i Alternatively, based on the optimized new location of S5-2, it can take over its tasks after the deployment delay, restoring the system to full functionality. This indicates the deployment / takeover delay time, representing the time required from triggering activation to the standby unit actually entering the Active state. Indicates the standby unit at time The state variable. No. Each active unit at time Fault identification indicators, for in-service units such as a drone performing a mission.

[0176] S5-4: System Degradation Mode: In extreme cases where multiple drones fail simultaneously and no backups are available, the system will enter degradation mode. The system will adjust its operation based on the "critical index" of each road segment. Reallocate limited sensory resources.

[0177] Specifically, based on the critical index, ) Perform a greedy allocation.

[0178] First, the system calculates the key indices for all road segments to be monitored. The index is determined by queue length, velocity variance, and whether it is a confluence point.

[0179] Secondly, the remaining available health drones will be prioritized for assignment to The highest value Location of each road segment.

[0180] then, Non-critical areas with lower values ​​will be temporarily de-monitored or will only utilize existing ground sensors to achieve a "graceful degradation".

[0181] The key index is defined as:

[0182]

[0183] in, k Represents a section of road within the ramp area. It is a section of road k Historical average queue length It is the historical vehicle speed variance, reflecting the stability of traffic flow. ∈{0,1} are topological key point identifiers, such as sink points and branch points. =1, , , It is the weighting coefficient.

[0184] In degraded mode, the optimization objective in S5-2 will be transformed into maximizing the sum of the key indices of the covered road segments, i.e. This ensures that the most critical traffic bottlenecks are continuously monitored. "Covered" indicates the set of objects that have been determined to be covered.

[0185] Through the implementation of the above steps, based on wide-area perception through multi-UAV collaboration, combined with low-latency decision-making through edge computing and global optimization through cloud-based multi-agent reinforcement learning, this invention effectively solves the problems of limited perception, delayed response, and insufficient ability to cope with complex and dynamic traffic conditions in ramp traffic signal control. Furthermore, through a built-in dynamic reconstruction and self-healing mechanism, it significantly improves the system's robustness in the event of partial hardware failure. The system can accurately acquire multi-lane traffic flow data and adaptively adjust signal timing strategies, significantly improving ramp traffic efficiency, reducing vehicle delays, and enhancing the ability to respond quickly to emergencies. Therefore, it provides a closed-loop solution for intelligent traffic management of complex urban ramps that combines high efficiency and high reliability.

[0186] Its core lies in constructing an integrated "cloud-edge-air-ground" collaborative architecture to achieve global optimization of traffic flow and emergency response. The system's methodology specifically includes: collaborative perception and data collection through a multi-UAV swarm, utilizing UAVs equipped with multimodal sensors to acquire traffic data in real time and perform comprehensive monitoring and emergency event detection; distributed reinforcement learning decision-making in edge computing units, extracting traffic state features through a bidirectional LSTM network and constructing a lightweight DDPG model to generate local signal timing strategies; global optimization and strategy fusion on a cloud control platform, training and distributing globally optimal signal timing strategies through a multi-agent experience fusion mechanism and simulation verification; dynamic signal execution and traffic flow collaborative guidance on the ground traffic signal controller, achieving real-time adjustment of traffic lights, emergency priority control, and V2X collaborative guidance; furthermore, the method includes a complete self-healing mechanism, which automatically triggers task reorganization, dynamic position compensation, and model adaptive adjustment when some nodes fail, and can quickly restore functionality by activating hot standby UAVs to ensure the continuity and stability of system operation. The present invention aims to solve the problems of limited perception, delayed response, and insufficient coordination in traditional ramp control through the above-mentioned technical solutions, significantly improve the traffic efficiency of ramp entrances, reduce vehicle delays, enhance the ability to respond quickly to sudden traffic situations, and ensure the stable operation of the entire system in the face of partial hardware failures. Ultimately, it provides a comprehensive solution for intelligent transportation systems that combines high efficiency, intelligence, and high reliability.

[0187] Example 2

[0188] The purpose of this embodiment is to provide a computer-readable storage medium.

[0189] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0190] Example 3

[0191] The purpose of this embodiment is to provide an adaptive control system for ramp traffic signals based on UAV collaboration, including:

[0192] The real-time traffic data acquisition module is configured to: collect traffic data in real time based on a cluster of drones dynamically deployed above the ramp entrance, construct a global traffic state matrix, and perform preliminary processing through edge computing;

[0193] The local optimal signal timing suggestion generation module is configured to: use edge nodes to perform edge calculations on the pre-processed data to generate local optimal signal timing suggestions in real time.

[0194] The global optimal signal timing strategy generation module is configured to integrate and uniformly train the local optimal signal timing suggestions uploaded by multiple drones in the cloud to generate a global optimal signal timing strategy.

[0195] The real-time signal light parameter adjustment module is configured to adjust the signal light parameters in real time based on the globally optimal signal timing strategy.

[0196] See appendix Figure 1 , 5 As shown, the hardware architecture of the aforementioned ramp traffic signal adaptive control system consists of a multi-layered architecture encompassing cloud, edge, and ground layers. This includes multiple drones, edge computing units, communication relay modules, ground traffic signal controllers, and a cloud control platform. Multiple drones work collaboratively to perform various functions in the ramp and main road areas, including data acquisition, computational decision-making, communication and guidance, and a self-healing mechanism in case of malfunctions.

[0197] Aerial traffic data acquisition: Each drone is equipped with a high-definition camera or other sensors to perform real-time video monitoring and image acquisition of traffic conditions such as vehicle flow, speed, and queue length at ramp entrances / exits and adjacent main roads. The drone's aerial perspective can cover multiple lanes and a large area of ​​the road network, providing more comprehensive traffic information compared to ground sensors.

[0198] Edge computing decision-making: Computing units integrated on the drone or nearby roadside edge computing nodes process the collected traffic data in real time, extracting traffic parameters using pre-deployed intelligent algorithms, such as deep learning-based object detection and tracking algorithms, and using reinforcement learning-based control decision models to make preliminary judgments on the current traffic state and generate signal timing adjustment suggestions. The deployment of edge computing ensures low latency in decision-making, enabling the system to respond quickly to changes in ramp and mainline traffic conditions.

[0199] Communication Relay and V2X Support: UAVs are equipped with wireless communication modules, acting as relay nodes for vehicle-to-infrastructure (V2I) communication. On one hand, wireless networks are formed between UAVs and between UAVs and ground signal controllers and cloud platforms, enabling high-speed transmission of collected data and decision-making instructions. On the other hand, UAVs interact with nearby connected vehicles via V2X communication, broadcasting current traffic light status, countdown times, or speed guidance information, and receiving status data or priority passage requests from vehicles. UAV relay expands communication coverage and overcomes ground communication blind spots, ensuring timely and reliable information sharing among all participants at ramp entrances.

[0200] Traffic guidance and information broadcasting: Drones equipped with loudspeakers or visible light signaling devices can directly broadcast guidance information to vehicles in the ramp area from the air. For example, when vehicles on the ramp need to accelerate to merge into the main road, the drone can guide vehicles to pass in an orderly manner through voice prompts or light signals; in the event of an accident or when emergency vehicles request priority passage, the drone can hover over the relevant road section, broadcast warning information, and guide other vehicles to give way. This local guidance function enhances the system's ability to directly intervene in on-site traffic, improving traffic safety and efficiency in special circumstances.

[0201] Ground traffic signal control module: This specifically refers to the traffic signal hardware installed at intersections and its execution program. It has a network communication interface, enabling it to receive timing parameters from edge nodes or the cloud, including cycle time and green light ratio, and physically drive the switching of traffic lights. This module is the system's execution terminal, responsible for receiving and executing optimized signal timing instructions, dynamically adjusting the cycle time, phase, and green light duration of traffic lights at ramp intersections, and collaborating with drone swarms, edge computing layers, and cloud platforms to achieve global traffic flow optimization and emergency response.

[0202] The self-healing mechanism module is primarily deployed in the cloud control platform as part of the system management software. Leveraging its cloud-based global perspective, it receives heartbeat packets from all drones, making it suitable for global fault diagnosis and task rescheduling, such as reorganization and hot standby deployment.

[0203] This module is the core of ensuring the system's high reliability and operational continuity, responsible for real-time monitoring of the health status of each unit in the drone swarm. Once a drone failure is diagnosed through heartbeat detection or fault codes, this module will automatically trigger contingency plans. Its functions include: dynamic task reconfiguration, instructing nearby healthy drones to adjust their positions to quickly fill perception blind spots; activating and dispatching "hot standby" drones to autonomously fly to predetermined positions to replace the failed unit; and in extreme cases of multi-point failures, implementing a "graceful degradation" strategy to concentrate remaining resources on ensuring traffic control in the most critical areas. This mechanism enables the system to have self-healing capabilities, ensuring the continuity and stability of control tasks in complex environments.

[0204] The aforementioned drones work collaboratively to optimize the adaptive control of ramp signals. A cloud-based control platform manages and schedules multiple drones, utilizing the powerful computing resources of the cloud to train and update reinforcement learning models, and distributing optimized control strategies to edge drone nodes via a wireless network. The control algorithm employs a multi-agent reinforcement learning architecture, treating each drone-signal control node as an agent. These agents share observation information and collaborate on decision-making to achieve dynamic optimization of the ramp signal light cycle and phase. This algorithm enables the signal timing to automatically adjust based on real-time traffic flow, self-organizing changes to parameters such as green light duration and cycle length to adapt to traffic characteristics at different times, achieving global optimization of the mainline and ramp traffic efficiency.

[0205] Compared with existing technologies, the multi-UAV, multi-role fusion control method of this embodiment breaks through the dependence of traditional ramp control on fixed sensors, realizing the perception of the overall traffic situation from the air and flexible intervention in ground signals. The introduction of reinforcement learning enables the system to learn itself through continuous interaction with the traffic environment, gradually optimizing the control strategy. It can dynamically adjust signal timing in both normal peak hours and abnormal emergencies, thereby improving ramp traffic efficiency and significantly reducing vehicle delays. In addition, the system has a rapid response capability for abnormal events: when an accident is detected or an emergency vehicle approaches, the algorithm will automatically trigger an emergency control mode, prioritizing the provision of traffic signals, such as temporarily extending the mainline red light to allow the ramp rescue vehicle to pass. The UAV also simultaneously guides other vehicles to give way, winning valuable time for emergency handling. In summary, this invention provides an intelligent, efficient, and flexible adaptive traffic signal control system for ramps through integrated cloud-edge-air-ground collaboration, with significant practical value and innovation.

[0206] An adaptive traffic signal control system for ramp entrances is based on multi-UAV collaborative perception, edge computing, and multi-agent reinforcement learning. Utilizing the wide-area dynamic monitoring capabilities provided by a multi-UAV swarm, combined with the BeiDou Navigation Satellite System and Vehicle-to-Everything (V2X) technology, real-time traffic flow data from multiple lanes at ramp entrances is collected. The data is processed with low latency by an edge computing unit, and the Multi-Agent Reinforcement Learning (MARL) algorithm is applied to dynamically optimize traffic light timing strategies.

[0207] This invention integrates aerial perspectives from UAVs, ground sensor data, and cloud-based global optimization capabilities to construct a cloud-edge-air-ground collaborative architecture that combines high-efficiency collaboration with fault tolerance. This addresses the challenges of limited perception, delayed response, and insufficient collaboration in traditional ramp control. Furthermore, the system introduces a dynamic priority control mechanism for rapid response to emergencies such as accidents and emergency vehicles. More importantly, this invention includes a system self-healing mechanism to ensure stable operation of core functions even when some UAV nodes fail. The aim is to comprehensively improve the traffic efficiency of complex urban ramps, reduce vehicle delays, and provide a truly reliable and robust solution for intelligent transportation systems.

[0208] An adaptive traffic signal control system for ramp entrances is based on multi-UAV collaborative perception, edge computing, and multi-agent reinforcement learning. It utilizes the wide-area dynamic monitoring capabilities provided by an Unmanned Aerial Vehicle Swarm (UAV Swarm), combined with the BeiDou Navigation Satellite System (BDS) and Vehicle-to-Everything (V2X) technology, to collect real-time traffic flow data from multiple lanes at the ramp entrances. The data is processed with low latency by an edge computing unit, and a multi-agent reinforcement learning (MARL) algorithm is applied to dynamically optimize traffic light timing strategies.

[0209] This invention integrates aerial perspectives from UAVs, ground sensor data, and cloud-based global optimization capabilities to construct a "cloud-edge-air-ground" collaborative architecture that combines efficient collaboration with fault tolerance, addressing the challenges of limited perception, delayed response, and insufficient coordination in traditional ramp control. Furthermore, the system introduces a dynamic priority control mechanism for rapid response to emergencies such as accidents and emergency vehicles. More importantly, this invention includes a system self-healing mechanism to ensure stable operation of core functions even when some UAV nodes fail. The aim is to comprehensively improve traffic efficiency at complex urban ramps, reduce vehicle delays, and provide a truly reliable and robust solution for intelligent transportation systems.

[0210] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0211] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0212] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An adaptive traffic signal control method for ramp intersections based on UAV collaboration, characterized by: include: Based on the real-time collection of traffic data by a swarm of drones dynamically deployed above the ramp entrance, a global traffic state matrix is ​​constructed and preliminarily processed through edge computing. By utilizing edge nodes, edge computing is performed on the pre-processed data to generate local optimal signal timing suggestions in real time. The system integrates and trains locally optimal signal timing suggestions uploaded by multiple drones in the cloud to generate a globally optimal signal timing strategy. The traffic light parameters are adjusted in real time based on the globally optimal signal timing strategy. System fault tolerance and self-healing mechanism guidance information, including: Real-time monitoring and diagnosis of the failure status of each drone; If a drone is detected i If a failure occurs, dynamic task reorganization and coverage area reconstruction will be performed. Once the failure status of a drone is confirmed, if a hot standby drone exists, its status will be changed. The activated hot standby drone will autonomously fly to the location of the failed drone or take over its mission after the deployment delay, based on the optimized new location. In the extreme scenario where multiple drones fail simultaneously and no backups are available, limited sensing resources are reallocated based on key indicators for each road segment; specifically: S5-1: Real-time Failure Status Monitoring and Diagnosis: The cloud control platform continuously monitors the status of each drone in the cluster through a heartbeat mechanism; drones i Failure status flag 0 indicates normal operation, and 1 indicates failure, determined by the following logic: in, To determine whether a drone was received at time t i The heartbeat signal is represented by 1 for received and 0 for not received. For heart rate monitoring cycle, The preset timeout threshold is used for judgment. It is a drone i The reported fault code, once If the value is 1, the system will immediately activate its contingency plan. Indicates the first The error / fault codes currently reported by each node; This represents the set of fatal error codes. Once an error code belongs to this set, it is directly determined to be a fault. S5-2: Dynamic Task Reorganization and Coverage Area Reconstruction: Once a drone is detected... i Failure =1, the cloud control platform will aim to minimize perception blind spots and reconstruction costs for the remaining healthy drone swarm. Perform dynamic refactoring; this process can be formalized as an optimization problem: in, It is a new set of target location coordinates for health drones; This refers to the uncovered monitoring area resulting from the new location deployment, calculated using the following formula: ,in It is the total area under monitoring. It is a drone j new location The effective coverage area; It is a drone j From the original position Move to a new location The square of the distance represents the cost of movement; and These are preset weighting coefficients, used to balance the importance of coverage and move costs; Meanwhile, in the multi-agent reinforcement learning (MARL) model, when agent i fails, the agent set A will be dynamically updated to... Global state observation vector S t and joint action vector A t The dimensions are reduced accordingly, and the model continues to make collaborative decisions based on the new set of agents A′. S5-3: Redundancy Backup and Rapid Replacement Mechanism: To achieve rapid self-healing, the system can activate hot standby drones; when the drone... i Failure status =1 is confirmed if a hot standby drone exists. s Its state Ss This will trigger a transformation: Once activated, the hot standby drone will autonomously fly to the location of the disabled drone. p i Alternatively, based on the optimized new location of S5-2, it can take over its tasks after the deployment delay, restoring the system to full functionality. This indicates the deployment / takeover delay time, representing the time required from triggering activation to the standby unit actually entering the Active state. Indicates the standby unit at time State variables; No. Each active unit at time The fault determination indicator indicates that the in-service unit is a certain drone that is currently performing a mission; S5-4: System Degradation Operation Mode: In extreme cases where multiple drones fail simultaneously and no backups are available, the system will enter degradation mode; the system will adjust its operation based on the "critical indices" of each road segment. Reallocate limited sensory resources; Specifically, based on key indices Perform a greedy allocation; First, the system calculates the key indices for all road segments to be monitored. This index is determined by queue length, velocity variance, and whether it is a confluence point; Secondly, the remaining available health drones will be prioritized for assignment to The highest value Location of each road segment; then, Non-critical areas with lower values ​​will be temporarily de-monitored or will only utilize existing ground sensors to achieve a "graceful degradation". The key index is defined as: in, k Represents a section of road within the ramp area. It is a section of road k Historical average queue length It is the historical vehicle speed variance, reflecting the stability of traffic flow; ∈{0,1} is a topological key point identifier, used for confluence points and branching points. =1, , , These are weighting coefficients; In degraded mode, the optimization objective in S5-2 will be transformed into maximizing the sum of the key indices of the covered road segments, i.e. This ensures that the most critical traffic bottlenecks are continuously monitored; Covered indicates the set of objects that have been determined to be covered.

2. The adaptive traffic signal control method for ramp intersections based on UAV collaboration as described in claim 1, characterized in that, A drone swarm is dynamically deployed over the ramp entrance. Specifically, the drone swarm is deployed in a grid-like topology, covering the ramp entrance, exit, and adjacent main road areas. The drones are equipped with multispectral cameras, lidar, and millimeter-wave radar to collect video streams and 3D point cloud data at set frequencies.

3. The adaptive traffic signal control method for ramp intersections based on UAV collaboration as described in claim 1, characterized in that, Preliminary processing via edge computing includes: Vehicle detection and parameter calculation are performed based on real-time traffic data. Based on vehicle detection and parameter calculation, identify sudden events at ramp entrances and generate corresponding emergency signs.

4. The adaptive traffic signal control method for ramp intersections based on UAV cooperation as described in claim 3, characterized in that, The method of identifying sudden events at ramp entrances and generating corresponding emergency signs based on vehicle detection and parameter calculation includes: Traffic flow change detection: Calculate the rate of change of traffic flow in adjacent time windows and trigger an emergency indicator; Speed ​​variance detection: Calculate the variance of vehicle speeds within a lane to detect abnormal congestion or accidents; A comprehensive emergency indicator is generated by combining flow mutation and velocity variance.

5. The adaptive traffic signal control method for ramp intersections based on UAV collaboration as described in claim 1, characterized in that, Edge computing is performed on the pre-processed data using edge nodes to generate real-time local optimal signal timing suggestions, specifically including: Based on real-time traffic data, a bidirectional LSTM network is used to process the time series data of traffic parameters and extract traffic state features. A lightweight DDPG model is constructed to perform local reinforcement learning on the input traffic state features and generate locally optimal signal timing suggestions.

6. The adaptive traffic signal control method for ramp intersections based on UAV cooperation as described in claim 1, characterized in that, The cloud integrates and uniformly trains locally optimal signal timing suggestions uploaded by multiple drones to generate a globally optimal signal timing strategy, specifically including: Each drone maintains a local experience replay cache pool to store local interaction samples and corresponding TD errors; The cloud periodically samples and merges samples from each drone experience pool according to priority. The priority can be designed to be proportional to the absolute value of the TD error. Slowly synchronize the Critic target network parameters to ensure that training information can be smoothly transferred between multiple agents; Construct a digital twin model that closely matches the actual ramp entrance, minimize conflict time by making adjustments, compare the target flow or target queue length of each lane with the actual observed value in real time, and dynamically adjust the green light duration. Simulate different traffic flow scenarios, compare and optimize the average ramp delay, average mainline speed, and conflict time indicators before and after optimization, and obtain the globally optimal signal timing strategy in the cloud through multiple iterations of training and simulation verification.

7. The adaptive traffic signal control method for ramp intersections based on UAV cooperation as described in claim 1, characterized in that, The traffic light parameters are adjusted in real time based on the globally optimal signal timing strategy, specifically including: Dynamic control of traffic lights based on green light duration adjustment formula; The priority release policy is triggered based on the generated emergency sign or emergency vehicle request. By using drones as V2X communication relay nodes, signal status and speed suggestions are broadcast to ground vehicles, and guidance information is projected from the air using lasers to collaboratively optimize traffic flow.

8. An adaptive traffic signal control system for ramp intersections based on UAV collaboration, characterized in that: include: The real-time traffic data acquisition module is configured to: collect traffic data in real time based on a cluster of drones dynamically deployed above the ramp entrance, construct a global traffic state matrix, and perform preliminary processing through edge computing; The local optimal signal timing suggestion generation module is configured to: use edge nodes to perform edge calculations on the pre-processed data to generate local optimal signal timing suggestions in real time. The global optimal signal timing strategy generation module is configured to integrate and uniformly train the local optimal signal timing suggestions uploaded by multiple drones in the cloud to generate a global optimal signal timing strategy. The real-time signal light parameter adjustment module is configured to adjust the signal light parameters in real time based on the globally optimal signal timing strategy. System fault tolerance and self-healing mechanism guidance information, including: Real-time monitoring and diagnosis of the failure status of each drone; If a drone is detected i If a failure occurs, dynamic task reorganization and coverage area reconstruction will be performed. Once the failure status of a drone is confirmed, if a hot standby drone exists, its status will be changed. The activated hot standby drone will autonomously fly to the location of the failed drone or take over its mission after the deployment delay, based on the optimized new location. In the extreme scenario where multiple drones fail simultaneously and no backups are available, limited sensing resources are reallocated based on key indicators for each road segment; specifically: S5-1: Real-time Failure Status Monitoring and Diagnosis: The cloud control platform continuously monitors the status of each drone in the cluster through a heartbeat mechanism; drones i Failure status flag 0 indicates normal operation, and 1 indicates failure, determined by the following logic: in, To determine whether a drone was received at time t i The heartbeat signal is represented by 1 for received and 0 for not received. For heart rate monitoring cycle, The preset timeout threshold is used for judgment. It is a drone i The reported fault code, once If the value is 1, the system will immediately activate its contingency plan. Indicates the first The error / fault codes currently reported by each node; This represents the set of fatal error codes. Once an error code belongs to this set, it is directly determined to be a fault. S5-2: Dynamic Task Reorganization and Coverage Area Reconstruction: Once a drone is detected... i Failure =1, the cloud control platform will aim to minimize perception blind spots and reconstruction costs for the remaining healthy drone swarm. Perform dynamic refactoring; this process can be formalized as an optimization problem: in, It is a new set of target location coordinates for health drones; This refers to the uncovered monitoring area resulting from the new location deployment, calculated using the following formula: ,in It is the total area under monitoring. It is a drone j new location The effective coverage area; It is a drone j From the original position Move to a new location The square of the distance represents the cost of movement; and These are preset weighting coefficients, used to balance the importance of coverage and move costs; Meanwhile, in the multi-agent reinforcement learning (MARL) model, when agent i fails, the agent set A will be dynamically updated to... Global state observation vector S t and joint action vector A t The dimensions are reduced accordingly, and the model continues to make collaborative decisions based on the new set of agents A′. S5-3: Redundancy Backup and Rapid Replacement Mechanism: To achieve rapid self-healing, the system can activate hot standby drones; when the drone... i Failure status =1 is confirmed if a hot standby drone exists. s Its state Ss This will trigger a transformation: Once activated, the hot standby drone will autonomously fly to the location of the disabled drone. p i Alternatively, based on the optimized new location of S5-2, it can take over its tasks after the deployment delay, restoring the system to full functionality. This indicates the deployment / takeover delay time, representing the time required from triggering activation to the standby unit actually entering the Active state. Indicates the standby unit at time State variables; No. Each active unit at time The fault determination indicator indicates that the in-service unit is a certain drone that is currently performing a mission; S5-4: System Degradation Operation Mode: In extreme cases where multiple drones fail simultaneously and no backups are available, the system will enter degradation mode; the system will adjust its operation based on the "critical indices" of each road segment. Reallocate limited sensory resources; Specifically, based on key indices Perform a greedy allocation; First, the system calculates the key indices for all road segments to be monitored. This index is determined by queue length, velocity variance, and whether it is a confluence point; Secondly, the remaining available health drones will be prioritized for assignment to The highest value Location of each road segment; then, Non-critical areas with lower values ​​will be temporarily de-monitored or will only utilize existing ground sensors to achieve a "graceful degradation". The key index is defined as: in, k Represents a section of road within the ramp area. It is a section of road k Historical average queue length It is the historical vehicle speed variance, reflecting the stability of traffic flow; ∈{0,1} is a topological key point identifier, used for confluence points and branching points. =1, , , These are weighting coefficients; In degraded mode, the optimization objective in S5-2 will be transformed into maximizing the sum of the key indices of the covered road segments, i.e. This ensures that the most critical traffic bottlenecks are continuously monitored; Covered indicates the set of objects that have been determined to be covered.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-7.

Images