Unmanned traffic node adaptive control method and system based on edge computing

Abstract

The application belongs to the technical field of intelligent traffic control, and particularly relates to an unmanned traffic node adaptive control method and system based on edge computing, which comprises the following steps: extracting target density, speed dispersion, conflict trajectory number, queue length gradient and link jitter amplitude to construct a state vector, and calculating a traffic conflict risk index and its conflict evolution characteristics. Both are mapped into a task priority matrix and an execution sequence is generated. In combination with temperature rise constraints and link congestion thresholds, resource scaling steps, container migration amounts and core number distribution amounts are calculated. A migration disturbance factor is used to adjust task deadline and data age constraints. When the traffic conflict risk index exceeds the threshold, the signal release sequence and the unmanned vehicle passing window are optimized and rolled, and the optimization period is adjusted based on the resource scaling step and the migration disturbance factor. The application improves the availability and passing safety guarantee level of the edge roadside management and control system under complex mixed traffic flow.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology. More specifically, this invention relates to an adaptive control method and system for unmanned traffic nodes based on edge computing. Background Technology

[0002] In modern autonomous transportation scenarios, intersections, as key traffic nodes, are typically equipped with multimodal sensors such as cameras and millimeter-wave radar, and are accompanied by dense traffic flow of both autonomous vehicles and human-driven vehicles. To achieve efficient traffic control and safety warnings, the system needs to process massive amounts of sensor data in real time and rapidly issue traffic signal control commands. Edge computing technology, by decentralizing computing power to roadside devices and performing data fusion and computation locally, has become a core technological means to solve the problems of high communication latency and limited bandwidth in traditional centralized cloud computing architectures.

[0003] However, urban traffic flow is highly random and time-varying. During rush hours or in case of emergencies, intersections are prone to frequent vehicle congestion, queue sprawl, and speed dispersion. This rapidly changing traffic situation not only places extremely high real-time demands on capturing instantaneous traffic conflict risks but also brings a huge instantaneous computing power impact to roadside edge computing nodes. If edge devices trigger high-temperature frequency throttling due to full-load operation in harsh outdoor environments, or if link jitter occurs in the underlying communication network, the data processing time of edge nodes will fluctuate drastically. How to ensure efficient and stable coordination between underlying computing resources and top-level traffic signal control under changing road conditions and stringent edge hardware constraints such as computing power, temperature rise, and jitter is a critical pain point that urgently needs to be addressed in the field of intelligent transportation.

[0004] To reduce system latency, existing roadside control solutions typically incorporate edge computing architectures to replace cloud-based control. For example, Chinese patent application CN112466116A discloses a distributed control method for intersection groups based on edge computing. This method integrates edge computing nodes directly within the intersection signal controllers, directly integrating and processing traffic flow data uploaded by local detectors, and generating signal control schemes at the edge nodes before issuing and executing them. This technology, by decentralizing computing power to physical locations, avoids to some extent the significant transmission latency associated with cloud computing.

[0005] While the aforementioned solutions shorten the physical distance for data transmission, they still have significant technical shortcomings when dealing with sudden high-density mixed traffic flows. Existing technologies mainly employ a static, coarse-grained computing power processing mode. Edge nodes, acting as a black box, perform data perception, prediction, and control generation. They neither decouple the internal algorithm chain through task fragmentation based on prior dependencies and transmission overhead, nor incorporate the real-time computational load, temperature constraints, and network link jitter of the edge hardware into the traffic control cycle. This rigid processing mechanism ignores the tightness of coupling between tasks and abrupt changes in the physical environment in practical applications. When faced with bursts of high-dimensional, multimodal sensor data, the edge system cannot perform fine-grained distributed coordination and resource scaling. Once hardware approaches its temperature limit or network congestion occurs, it is forced to perform disordered container migration, directly generating huge computational and timing delays at the underlying level. Due to the lack of a feedback compensation mechanism for underlying computing power disturbances to upper-layer services, once such delays occur, the system lag at the lower level will directly impact the traffic management operations at the upper level. The upper-level system still issues control commands based on fixed optimization cycles or outdated data ages, resulting in a severe time misalignment between traffic light timings and the actual traffic congestion at intersections. Ultimately, this control method, which separates computing power from business operations, not only easily leads to the depletion of edge resources and system crashes, but also prevents the system from accurately planning conflict-free continuous passage windows for autonomous vehicles. This can easily trigger secondary traffic congestion and safety accidents, and simply cannot meet the stringent requirements of high-frequency control stability and safety for autonomous traffic nodes. Summary of the Invention

[0006] To address the technical problem that existing edge-side control schemes suffer from severe computational latency and instruction misalignment under conditions of hardware temperature rise / fall and network link jitter when dealing with sudden high-density mixed traffic flows due to the lack of a feedback compensation mechanism for upper-layer services caused by the low-level computing power disturbance, which leads to system resource exhaustion and failure of safety management of unmanned traffic nodes, this invention provides an adaptive control method and system for unmanned traffic nodes based on edge computing.

[0007] In a first aspect, the present invention provides an adaptive control method for unmanned traffic nodes based on edge computing, comprising: collecting video streams, radar traces, communication latency, and edge computing power occupancy information of the unmanned traffic node; extracting target density, speed dispersion, number of conflict trajectories, queue length gradient, and link jitter amplitude; constructing a state vector; calculating traffic conflict risk indicators based on the state vector; and generating conflict evolution characteristics by combining the rate of change and fluctuation intensity of the traffic conflict risk indicators within a time window; mapping the computational load from the traffic conflict risk indicators, conflict evolution characteristics, and edge computing power occupancy information into a task priority matrix, and performing edge-side perception and control tasks. The system performs segmented processing, calculates coupling values ​​based on prior dependencies, result reuse, and transmission overhead, and generates a distributed execution sequence. It calculates resource scaling steps based on historical computing load and queue length, and determines container migration and core allocation based on temperature rise constraints and link congestion thresholds. It generates migration disturbance factors based on link jitter amplitude and container migration volume, adjusting task deadlines and data age constraints. When traffic conflict risk indicators exceed preset thresholds, it uses rolling optimization to calculate signal release order, duration, and unmanned vehicle passage window, and adjusts the optimization cycle based on resource scaling steps, migration disturbance factors, and execution deviations of adjacent nodes before issuing execution orders to unmanned traffic nodes.

[0008] By adopting the above technical solution, this invention deeply integrates traffic conflict risk indicators with underlying hardware temperature rise constraints, link congestion thresholds, and computational load, breaking the limitations of traditional static computing power processing modes and realizing microservice dynamic sharding and adaptive resource scheduling based on task coupling values. Furthermore, by dynamically tightening task deadlines and adjusting optimization cycles using migration disturbance factors, it effectively avoids computational latency caused by disordered container migration, ensuring that even in harsh outdoor environments and under conditions of limited computing power, it can still accurately and without timeout issue conflict-free passage windows and signal release sequences for unmanned vehicles, thus improving the stability and security of high-frequency traffic management.

[0009] Preferably, the construction of the state vector includes: dividing the total number of vehicles within the detection range of the video stream and radar points by the actual physical area of ​​coverage to obtain the target density; extracting the current instantaneous speed of the vehicles and calculating the absolute deviation of the instantaneous speed from the average speed of the vehicles in the historical sampling period to obtain the speed dispersion; extending each vehicle along its current driving direction to obtain the trial running trajectory, and counting the cumulative number of intersection collision points of the trajectories in the two-dimensional space to obtain the number of conflict trajectories; calculating the difference in the total length of the parking queue in two adjacent sampling periods and dividing it by the sampling time interval to obtain the queue length gradient; extracting the sending and receiving timestamps of the underlying communication data packets and calculating the range of the transmission delay of adjacent data packets to obtain the link jitter amplitude; and concatenating the target density, speed dispersion, number of conflict trajectories, queue length gradient, and link jitter amplitude into a column vector according to a preset dimension to generate the state vector.

[0010] By adopting the above technical solution, this invention assesses the macroscopic traffic flow situation and the health status of the underlying physical network, eliminates the accuracy loss error caused by the blind spot of single sensor detection and data drift, and builds a solid and multidimensional data foundation for subsequent computing power scheduling and risk assessment.

[0011] Preferably, the step of calculating the traffic conflict risk index based on the state vector and generating conflict evolution features by combining the rate of change and fluctuation intensity of the traffic conflict risk index within a time window includes: normalizing each data element in the state vector and assigning a corresponding weight coefficient to each normalized data element; weighting and summing the normalized state vector and the weight coefficients to obtain the traffic conflict risk index; calculating the absolute value of the difference between the traffic conflict risk index in the current time window and the previous time window to obtain the rate of change; extracting the traffic conflict risk index from a preset number of consecutive historical time windows before the current time window and calculating the standard deviation to obtain the fluctuation intensity; and linearly weighting and combining the traffic conflict risk index, the rate of change, and the fluctuation intensity according to a preset ratio to generate the conflict evolution features in scalar form.

[0012] By adopting the above technical solution, the present invention overcomes the limitations of traditional solutions that rely solely on the absolute value of the current risk inventory for assessment. It can keenly capture the instantaneous acceleration of the rapid spread of risk and the intensity of internal oscillation disturbances, thereby improving the accuracy of traffic node safety perception and forward-looking early warning capabilities.

[0013] Preferably, the step of performing segmented processing on the edge-side perception and control tasks, and calculating the coupling value based on prior dependencies, result reuse, and transmission overhead, includes: decoupling the edge-side algorithms according to a directed acyclic graph structure, dividing the perception, prediction, timing, and control tasks into independent execution segments; counting the number of call interfaces of each segment to the output data of the preceding segment, and calculating the prior dependencies; counting the proportion of globally shared variables in the total data stream of each segment, and calculating the result reuse; counting the number of network bandwidth bits occupied by inter-segment interaction instructions and data packet headers, and calculating the transmission overhead; and using a preset weighted formula to perform a weighted summation of the prior dependencies, the result reuse, and the transmission overhead to obtain the coupling value representing the degree of correlation between the task segments.

[0014] By adopting the above technical solution, the present invention can identify the real physical resistance of algorithm nodes at the memory exchange and interface call level, avoid the disruption of the smoothness of the original high-efficiency computing link due to forced disordered cutting and scheduling, and reduce the unnecessary physical bandwidth consumption of cross-chassis communication nodes.

[0015] Preferably, the step of determining the container migration amount and core allocation amount by combining temperature rise constraints and link congestion thresholds includes: calculating the average processor utilization rate within a historical time window as the historical computing load; inputting the historical computing load and queue length into a resource mapping function to calculate the cross-cycle resource pool expansion / contraction capacity as the resource scaling step size; obtaining the processor's highest safe operating temperature as a temperature rise constraint and calculating the upper limit of the currently available CPU cores; obtaining the maximum allowed transmission bandwidth of the upstream switch and deducting the bandwidth occupied by basic communication, extracting the remaining available bandwidth as the link congestion threshold; allocating the resource scaling step size to available physical cores based on the upper limit of the available CPU cores, and calculating the core allocation amount for this node; when the required computing resources exceed the core allocation amount, converting the excess computing tasks into task loads that need to be offloaded according to the link congestion threshold, and obtaining the container migration amount.

[0016] By adopting the above technical solution, the present invention can schedule physical cores on demand according to the core allocation amount, and safely offload excess load into container migration when the computing power limit is exceeded, thus eliminating the fatal defects of the past, which were easily caused by simply relying on processor lag utilization feedback, resulting in device thermal runaway, lag, disconnection and crash.

[0017] Preferably, the step of generating a migration disturbance factor based on the link jitter amplitude and container migration amount, and adjusting the task deadline and data age constraint, includes: multiplying the link jitter amplitude and the container migration amount by their respective dimension conversion coefficients and then adding them together, and scaling them using a preset normalization coefficient to generate the migration disturbance factor with time dimensions for time compensation; subtracting a time compensation amount proportional to the migration disturbance factor from the preset maximum tolerance deadline to update the task deadline of the control task; and adding a transmission attenuation time loss value corresponding to the migration disturbance factor to the predicted initial data age of the task to update the data age constraint.

[0018] Preferably, the step of using rolling optimization to calculate the signal release order, duration, and autonomous vehicle passage window includes: within a preset simulation test domain, optimizing the release sequence of different traffic flows as decision variables with the objective function of maximizing intersection throughput, to obtain the optimal signal release order; constructing a traffic flow dissipation model using vehicle arrival rate and queue dissipation rate, calculating the shortest time required to clear the queue of vehicles in the current phase, and taking the smaller value between the shortest time and a preset maximum allowable duration threshold as the duration; calculating the predicted time for the autonomous vehicle to reach the intersection stop line, and calculating a conflict-free continuous passage period based on the queue dissipation time and safe following distance of the preceding human-driven vehicles, as the autonomous vehicle passage window.

[0019] Preferably, the step of adjusting the optimization cycle based on the resource scaling step size, migration disturbance factor, and adjacent node execution deviation includes: receiving the expected arrival traffic flow and the actual arrival traffic flow from adjacent nodes, and calculating the absolute value of the difference between the two as the adjacent node execution deviation; using a preset initial rolling optimization cycle as a benchmark value; multiplying the resource scaling step size, the migration disturbance factor, and the adjacent node execution deviation by their respective corresponding adjustment coefficients with dimensional conversion attributes, and superimposing them to obtain a time adjustment amount with time dimensions; adding the benchmark value and the time adjustment amount to generate the optimization cycle required for the next time window.

[0020] Preferably, the step of mapping the computational load from traffic conflict risk indicators, conflict evolution characteristics, and edge computing power occupancy information into a task priority matrix includes: extracting the real-time CPU utilization rate within the edge computing power occupancy information as the initial computational load; performing scaling mapping processing on the traffic conflict risk indicators, conflict evolution characteristics, and initial computational load using the max-min normalization algorithm to construct the current environment state vector; calculating the inner product dot product value between the preset feature weight vector and the current environment state vector to deduce the dynamic priority score of each independent task segment; and arranging and combining the network identifiers, dynamic priority scores, and estimated computing resource requirements of all task segments in descending order of dynamic priority scores to construct a two-dimensional task priority matrix.

[0021] Secondly, the present invention provides an edge computing-based adaptive control system for unmanned traffic nodes, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned edge computing-based adaptive control method for unmanned traffic nodes is implemented.

[0022] By adopting the above technical solution, the above-mentioned edge computing-based adaptive control method for unmanned traffic nodes is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of terminal devices based on the memory and processor, making them convenient to use.

[0023] The technical solution of the present invention has the following beneficial technical effects: This invention improves the accuracy of traffic safety perception by comprehensively collecting multi-source perception data and edge computing power occupancy information from unmanned traffic nodes, deeply extracting various traffic flow characteristics, and accurately assessing traffic conflict risks and their evolution characteristics. By deeply integrating traffic conflict risk indicators with computational load, a task priority matrix is ​​constructed to achieve task fragmentation and distributed collaborative processing of perception and control tasks. Combined with temperature rise constraints and link congestion thresholds, the resource scaling step size and core allocation are accurately calculated, improving the utilization rate and processing timeliness of edge computing resources. By dynamically correcting task deadlines and data age constraints using migration perturbation factors, a rolling optimization mechanism is adopted when traffic conflict risk indicators exceed preset thresholds. This mechanism flexibly calculates the signal release order, duration, and unmanned vehicle passage window, and precisely adjusts the optimization cycle by combining resource scaling step size, migration perturbation factors, and execution deviations of adjacent nodes. This achieves deep collaboration between edge computing power scheduling and traffic node management, reducing system communication and computation latency, and enhancing the overall operational efficiency, resource coordination capabilities, and traffic safety assurance level of the unmanned transportation system. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the adaptive control method for unmanned traffic nodes based on edge computing in this invention. Figure 2 This is a diagram illustrating the degree of correlation between task fragments; Figure 3 This is a schematic diagram of the historical computing load sequence; Figure 4 This is a schematic diagram of the time sequence distribution of continuous passage of unmanned vehicles. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses an adaptive control method for unmanned traffic nodes based on edge computing, referring to... Figure 1 This includes steps S1-S3: S1: Construct state vectors and generate conflict evolution features The system collects video streams, radar traces, communication latency, and edge computing power occupancy information of unmanned traffic nodes, extracts target density, speed dispersion, number of conflict trajectories, queue length gradient, and link jitter amplitude, and constructs state vectors. Based on the state vectors, it calculates traffic conflict risk indicators and generates conflict evolution characteristics by combining the rate of change and fluctuation intensity of traffic conflict risk indicators within a time window.

[0027] In an optional embodiment, during the specific perception data acquisition and processing stage, video stream data is first continuously acquired by cameras deployed in the intersection area. The FFmpeg multimedia processing library is then used to parse the video stream data into a sequence of RGB image frames. Simultaneously, the YOLOv8 target detection algorithm is used to identify all vehicle targets corresponding to the image frame sequence. For the underlying hardware parameter settings, the preferred fixed setting for the video stream resolution is 1920×1080, and the video recording frame rate is forcibly constrained to 30 frames per second. The total number of vehicles within the detection range of the video stream and radar points is then calculated. Divide by the actual physical area covered Thus, the target density is extracted. For example, when 141 cars are detected within the detection area of ​​an intersection, the target density is calculated accordingly. The specific figure is 0.002 vehicles per square meter.

[0028] Specifically, the target density satisfies the following relationship:

[0029] in, This indicates the target density, expressed in vehicles per square meter. Represents the total number of vehicles within the detection range of the video stream and radar points; a dimensionless quantity. This indicates the actual physical area covered by the intersection, in square meters.

[0030] The density calculation formula mentioned above is based on the common sense of physical calculation of traffic flow density in the field of traffic engineering. This invention uses multimodal visual perception detection technology to accurately locate the total number of vehicles and use it as the divisor, replacing the traditional solution that relies solely on inductive loops, thereby reducing the accuracy loss error caused by detection blind spots.

[0031] For vehicle speed feature extraction, real-time 3D spatial point cloud data is acquired using millimeter-wave radar equipment. Then, the DBSCAN clustering algorithm is used to continuously extract and track the current radar points. The physical coverage range of the millimeter-wave radar point detection is limited to a radius of 150 meters, and the corresponding physical detection area is calculated to be 70,650 square meters. Subsequently, the Kalman filter algorithm is used to calculate the instantaneous speed of each vehicle. In practical implementation, the absolute time length of the hardware sampling period is preferably fixed at 100 milliseconds, while simultaneously extracting the overall average speed of all vehicles within the past 10 consecutive sampling periods. Calculate the instantaneous speed of each car. Relative to the above overall average speed The absolute deviation is calculated by accumulating all absolute deviations and taking the average value to generate the velocity dispersion. For example, suppose the current instantaneous speed of a target car is... The monitored value was 15.2 meters per second, and the calculated value of the corresponding absolute deviation of the car was 2.7 meters per second.

[0032] Specifically, the velocity dispersion satisfies the following relationship:

[0033] in, This indicates the velocity dispersion, measured in meters per second. This represents the total number of vehicles calculated within the current detection coverage area; it is a dimensionless quantity. Indicates the first The instantaneous speed of the target vehicle is currently measured in meters per second; This represents the average speed of vehicles during the historical sampling period, expressed in meters per second.

[0034] The above-mentioned absolute deviation calculation model is based on the known calculation method of obtaining the average absolute deviation in classical mathematical statistics. The algorithm of this invention combines Kalman front-end filtering to smooth out the original speed data with severe spikes, overcoming the error drift caused by the inherent noise of the transient speed measurement of the underlying millimeter-wave radar.

[0035] In the traffic conflict trajectory hazard assessment stage, the DeepSORT multi-target tracking algorithm is used to deeply fuse and match continuous image frames and radar spot features to generate the relative continuous motion trajectories of all detected target vehicles. Based on the current physical speed and current heading angle of the vehicles, each vehicle's forward time-series extension generates a corresponding trial trajectory. The preset value of the corresponding control extension extension physical running time parameter is preferably limited to 3 seconds. A time collision algorithm is used to statistically retrieve the geometric and physical intersection points generated by the intersection of all trial running trajectory segments within the preset two-dimensional coordinate system plane. Subsequently, the estimated time difference for each pair of intersecting running trajectories to reach the aforementioned single physical intersection point is calculated. When determining the expected time difference If the value is less than the preset safety threshold, the trajectory contact judgment is directly defined as a real cross-collision danger event. The total number of real cross-collision points that occur within the entire calculation and detection cycle is accumulated and summarized to extract the number of generated conflict trajectories. For example, within the currently specified data collection timeframe, the system statistics show that the total number of actual intersection collision points is exactly 12.

[0036] For assessing vehicle queuing congestion, the system automatically identifies the spatial coordinates of the tail end of the parking queue behind the stop line of each approach lane at intersections by analyzing roadside video surveillance streams. Based on these collected coordinates, the total length of the parking queue is extracted and identified during the current detection cycle. And the total length of the parking queue extracted from the previous detection sampling cycle. A specific difference algorithm is used to calculate the physical difference in the total queue length within adjacent periods. Finally, this distance difference result is directly divided by the corresponding specified sampling time interval. The gradient of queue length is calculated. For example, if the total queue length in the previous sampling period was 45 meters and it has increased to 58 meters in the current detection period, then the sampling time interval will be determined based on the current settings. With a fixed physical condition of 5 seconds, the gradient of the queue length is calculated by division. The accurate value is 2.6 meters per second.

[0037] Specifically, queue length gradient Calculate the expression that satisfies the following relation:

[0038] in, This represents the queue length gradient, in meters per second. This indicates the total length of the parking queue identified during the current detection cycle, in meters. This indicates the total length of the parking queue extracted in the previous sampling period, in meters. This indicates the specific time interval spanned between two consecutive sampling and measurement processes, in seconds.

[0039] The gradient calculation formula is based on the general and common knowledge model of instantaneous change rate in classical kinematics. The present invention uses the first-order instantaneous difference motion of the physical length of the queue in space to intuitively present the trend of the spread and congestion of the convoy ahead, overcoming the natural shortcoming effect caused by the simple counting method being easily affected by the obstruction of different large vehicle chassis.

[0040] At the underlying communication system network health status detection level, the Internet Control Message Protocol (ICP) is invoked to calculate and obtain the specific send and receive timestamp records of continuous network data packets in real time. A single data measurement monitoring window is defined as 1 second in length, and a total of 50 UDP data packets are randomly and continuously captured during this observation period to construct and generate an overall network transmission latency statistical sequence. Furthermore, the maximum communication latency caused by a single data transmission process is retrieved individually from the aforementioned latency sequence. and minimum communication latency Next, the maximum communication latency will be extracted directly. With minimum communication delay The range between the two values ​​is calculated and used as the link jitter amplitude. In this verification example execution, the network program calculates the aforementioned range subtraction to obtain the corresponding link jitter amplitude. The value is 33 milliseconds.

[0041] Specifically, link jitter amplitude Calculate the expression that satisfies the following relation:

[0042] in, This indicates the current underlying link jitter amplitude, in milliseconds. This indicates the maximum network data communication latency within the set monitoring window, measured in milliseconds. This indicates the minimum communication latency of network data within the set monitoring window, measured in milliseconds.

[0043] The range calculation subtraction formula is directly generated and constructed based on the common statistical knowledge and common sense of evaluating port congestion delay peak difference in the field of computer system network. The algorithm can accurately capture transient network congestion events that may cause control command issuance timeouts.

[0044] In addition, the underlying process library psutil is invoked, and the CPU utilization and memory usage percentage of the edge computing nodes are obtained in real time directly through the interface of calling this program, serving as edge computing power utilization information. Extracted edge computing power occupancy information The output will be used in subsequent steps for computing power allocation decisions.

[0045] After completing the basic physical data extraction and detection actions described above, the parameters reflecting external traffic flow and network operation status are sequentially concatenated and merged according to the preset extraction dimensional requirements. The target density obtained from the detection is then calculated. Overall speed dispersion of the vehicle Number of conflict trajectories at intersections Gradient of parking queue length and the link jitter amplitude of the aforementioned underlying network. The five variables are subjected to an ordered unidirectional permutation and combination. The five parameters are concatenated in a preset order to generate a 5×1 state vector. Based on the specific sets of measurement results for different dimensions provided by the aforementioned measurement module, the corresponding current state vector is extracted by mapping and assembling. .

[0046] After successfully constructing and obtaining the complete set of data parameters for the current feature space state, the preset data analysis engine is invoked to extract and collect the state vector of the above system using the principal component analysis method. It directly performs high-dimensional feature extraction, clustering, and dimensionality reduction. This dimensionality reduction process removes redundant variables from environmental data and extracts the main features representing traffic flow trends. Subsequently, the extracted column vectors of these dimensionality-reduced main features are used as the original input matrix features of the underlying algorithm data source. These features are directly input into the pre-trained, offline-deployed multilayer perceptron model network in the system memory to initiate deep inference and fitting calculations. This allows for the direct output of a target intersection assessment and the system's predicted traffic conflict risk index for the current moment. .

[0047] In the calculation step of the supplementary embodiment for obtaining the predicted risk using another mathematical algebraic path, the Min-Max normalization method is used to normalize the above 5-dimensional state vector. Internally, all independent environmental feature data elements are individually processed using linear scaling and numerical extreme value normalization mapping. This numerical limit mapping method maps each data element to the [0,1] interval. The system generates constraint reference theoretical limit boundary values ​​for each input data element, pre-set based on the historical statistical extreme values ​​of the intersection. For example, the theoretical maximum boundary value for the system's monitored target density is set to 0.01 vehicles per square meter, the maximum tolerance for the speed dispersion of the traffic flow is limited to 10 meters per second, and the theoretical maximum number of conflict trajectories that can occur at the intersection is limited to 50. Simultaneously, the maximum allowable measurement value for the queue length gradient calculation in the system's analysis and processing is fixed at 5 meters per second, and the theoretical maximum absolute measurement value for the communication link jitter amplitude measured by the judgment extraction device is 100 milliseconds. After this step, the influence of dimensions is eliminated, resulting in the normalized state vector. The above-mentioned parameter settings are used to calculate and generate a normalized state vector. .

[0048] Preload a set of validated and optimized proportional allocation combination weight coefficient vectors. For the aforementioned five items, the extracted feature parameters, after shrinkage and normalization, are assigned specific proportional weight coefficients according to their extraction positions in the queue, with each coefficient defined as 0.25, 0.15, 0.35, 0.15, and 0.10 respectively. The state vector, with all dimensions and unit restrictions removed, is then directly normalized. Combined with the above weight coefficient vector Perform matrix inner product operations and subsequent weighted summation and accumulation operations. The normalized state vector is then summed with the weight coefficients to obtain a scalar form of the traffic conflict risk index. By performing the parallel matrix inner product calculations in the background of the system, based on the numerical examples of the specifications presented in the previous embodiments, the calculated value of the indicator corresponding to the current time of the intersection can be obtained, which is equal to 0.2855.

[0049] Specifically, traffic conflict risk indicators under mathematical paths Calculate the expression that satisfies the following relation:

[0050] in, This indicates that the overall road segment traffic conflict risk index is derived using algebraic weighted fusion logic; it is a dimensionless quantity. This indicates that the preset allocation loads a feature column vector specifying the weight coefficients for each individual feature. This represents the dimensionless normalized state vector after each environmental data element has undergone linear scaling and same-scale translation mapping operations.

[0051] The above inner product algebra evaluation formula is strictly based on the common weighted numerical evaluation of multi-attribute composite condition decision-making schemes within the conventional operations research system. It is a general and well-known common sense model for construction and application. The modified algorithm of this invention directly establishes that the characteristics of different independent operating parameters have completely different contribution roles and proportions to the overall security risks of the current monitoring system's underlying control. This avoids the arithmetic imbalance and deviation caused by the extreme imbalance of absolute numerical measurements on the surface when physical properties cross heterogeneous data sources.

[0052] It is understandable that the combined weight coefficient vector The specific weight coefficients for each component are obtained by using principal component analysis combined with information entropy weighting or hierarchical analysis based on historical big data statistics of the target intersection over a long period, such as the past year. The system offline analyzes the variance and information content of various characteristic parameters before real collisions occurred in the past, objectively assessing the contribution of individual environmental elements such as target density, speed dispersion, and the number of conflict trajectories to the overall safety hazard, and then solidifying these into a unique weight coefficient feature vector for this intersection. Those skilled in the art should understand that the above specific values ​​are only preferred examples of this embodiment; for different road structure scenarios, the weight vectors can be recalculated and dynamically loaded.

[0053] After calculating and generating the current traffic conflict risk index, the most recently generated traffic conflict risk indices are continuously recorded and stored to construct a time series. A fixed constant parameter, preferably set to 5 seconds, is used to define the physical objective calculation time span limit for the corresponding sliding calculation window for retrieval, analysis, extraction, and comparison. A first-order forward difference algorithm is employed to calculate the absolute value of the difference between the traffic conflict risk index of the current time window and the previous time window. The calculated absolute value of the difference is used as the rate of change for the current stage. For example, suppose the traffic conflict risk index extracted in the previous capture window was 0.2512, and the index value captured in the current system window is 0.2855. The control background processing network, after performing the above subtraction and taking the absolute value (without the negative sign), calculates the final rate of change corresponding to the current stage. It equals 0.0343.

[0054] Furthermore, a sequence of traffic conflict risk indicators for a preset number of consecutive historical time windows preceding the current time window is extracted. The standard deviation of this sequence is calculated to reflect the degree of fluctuation of the indicator within the historical period. In this embodiment, the unmanned traffic node can specifically be a crossroads. The fluctuation intensity of the unmanned traffic node within a large period range is obtained by measurement. If, through calculation and prior estimation, the statistical standard deviation of the aforementioned long-term sequence is calculated and determined to be 0.015, this result will be used as the fluctuation intensity of the current unmanned traffic node. .

[0055] Then, the traffic conflict risk indicators intercepted at the moment are extracted simultaneously. The time dimension subtraction difference operation yields the stage risk change rate. And the fluctuation intensity extracted, summarized, and judged by the above statistical methods. Three key parallel variables are used. Based on the pre-built model, corresponding linear weighting coefficients are pre-set for these three risk assessment parameters. These coefficients strictly follow the order of appearance of the three variables: 0.6, 0.3, and 0.1. After linearly weighting and combining these parameters with their corresponding coefficients, the conflict evolution characteristics are ultimately generated through this process. If all the sample representative values ​​obtained from the above calculation process, including the three parallel analysis examples, are substituted into the above-mentioned set constants for weighted summation, the final aggregate model can be accurately and smoothly calculated to obtain the comprehensive scalar characteristic determination of the evolution direction of the macro-comprehensive risk environment. The final specific calculation result is equal to 0.18309.

[0056] Specifically, the characteristics of conflict evolution satisfy the following relation:

[0057] in, It represents the characteristics of conflict evolution and is a dimensionless quantity. , , These represent the pre-set weighting coefficients for different security feature data, respectively. This indicates the extraction of current traffic conflict risk indicators; This indicates the rate of change of traffic conflict risk indicators within a time window, used to characterize the overall change in risk. It represents the intensity of volatility, used to characterize the degree of volatility in the long-term risk of a system.

[0058] The above-mentioned scalar feature extraction output formula is based entirely on the common knowledge and general engineering application of traditional general time-series signals undergoing nonlinear state drastic evolution, which is used in the comprehensive weighted cross-evaluation of system safety. The technical advantages of this calculation process are not only focused on the current risk inventory and inherent absolute danger level of the system control system, but also include the rate of change of risk and the internal disturbance frequency into the comprehensive safety management and control evaluation system.

[0059] Understandable , , The specific data acquisition and setting are based on the following method: During the offline simulation phase, the system calculates the data using a combination of multiple linear regression analysis and historical traffic accident causal retrospective analysis for high-dimensional time series data. By fitting historical hazard evolution slice data, the system objectively assesses the proportion of the current static hazard level, the instantaneous acceleration slope, and the frequency of environmental oscillations and disturbances in catalyzing the final traffic accident. This objective data-driven weight allocation mechanism ensures the rigor and reliability of the system's assessment of the macro-level comprehensive risk evolution direction.

[0060] Thus, this step improves the accuracy of traffic node safety perception, overcomes the shortcomings of traditional solutions such as single perception dimension and lack of underlying physical computing power feedback, and lays a solid data foundation for subsequent edge task sharding and adaptive control scheduling.

[0061] S2: Generating distributed execution sequences and determining computing power allocation The computational load in traffic conflict risk indicators, conflict evolution characteristics, and edge computing power occupancy information is mapped into a task priority matrix. Edge-side perception and control tasks are processed in segments. Coupling values ​​are calculated based on pre-order dependencies, result reuse, and transmission overhead to generate a distributed execution sequence. Resource scaling step size is calculated based on historical computing power load and queue length. Container migration amount and core allocation amount are determined by combining temperature rise constraints and link congestion thresholds.

[0062] In an optional embodiment, during the task priority matrix evaluation and construction computation phase, the real-time CPU utilization rate within the edge computing power occupancy information is first extracted. This processor utilization rate value is then directly used as the initial computational load. Subsequently, the min-max normalization algorithm was used to analyze the traffic conflict risk index, scalar conflict evolution characteristics, and the aforementioned initial computational load. The scaling mapping process is performed independently. This numerical scaling operation smoothly compresses the physical variables of each dimension to the standard measurement value range of 0 to 1. The normalized feature parameters are extracted separately and used to construct the current environmental state vector. The overall intersection processing algorithm is decoupled and divided into four independently running segments using a microservice architecture based on a directed acyclic graph (DAG): a video and radar target-level fusion perception segment, a Kalman filter-based trajectory prediction segment, a fuzzy logic timing segment, and a final signal controller control segment. Pre-defined feature weight vectors for each independently running segment are derived and set in advance using the analytic hierarchy process (AHP). The feature weight vector The parameters internally encompass risk sensitivity, evolution sensitivity, and system load tolerance characteristics. The computation engine calculates this preset feature weight vector. With the current environment state vector Based on the corresponding inner product dot product values, the dynamic priority score of each independent task fragment can be derived. For example, when the underlying perceived risk indicators show a sudden increase, the system assigns a very high dynamic score to the signal control segment. Conversely, when the system's computational load approaches the hardware's physical limit, the system forcibly reduces the corresponding score value of the perceived segment proportionally. Ultimately, the system identifies the specific network identifiers of all task segments and calculates the dynamic priority score. The estimated computational resource requirements are then arranged and combined in descending order of their scores. Based on this descending sequence array, a corresponding two-dimensional task priority matrix is ​​constructed. .

[0063] Specifically, the dynamic priority score of task fragmentation satisfies the following relationship:

[0064] in, This indicates that a specific task segment obtains a dynamic priority score, which is a dimensionless quantity. This indicates that the analytic hierarchy process (AHP) is used to set the preset feature weight vectors for each execution segment; This means integrating current multi-dimensional risk indicator parameters to construct a vector corresponding to the current environmental state.

[0065] This dynamic score inner product evaluation formula is built upon a common-sense model of dynamic weight adaptive allocation decision-making in general operations research. This solution integrates the instantaneous hazard evolution indicators of upper-level traffic with the transient load characteristics of the underlying cabinet chips through a dot product. This enables the system to spontaneously block inefficient edge sensing processes when high-risk road accidents occur or network computing power is on the verge of collapse, thereby ensuring the priority execution of critical intervention commands such as signal control.

[0066] It should be noted that the specific method for obtaining the preset feature weight vector is as follows: During the hardware and software integration phase before the edge gateway is deployed at the factory, a two-dimensional stress test matrix of traffic risk and edge computing power is constructed. Preset traffic conflict risk flows of different levels are injected into the system, and extreme heating and full computing power load conditions are simultaneously superimposed. The analytic hierarchy process (AHP) is used to record and extract the optimal decision weight solution set that ensures the system does not crash and that critical collision avoidance commands have zero timeouts, ultimately generating the preset feature weight vector. This method avoids the randomness of resource allocation and ensures the rationality of priority scoring under special operating conditions.

[0067] During the task inter-relationship assessment and global execution scheduling phase, the dependencies of each independent subtask node on its preceding data requirements are analyzed. The total number of API calls initiated by each execution shard to the output data of its preceding shards is counted. This number of API calls is directly designated as the preceding dependency. For example, trajectory prediction segmentation requires cross-process calls to three data interfaces: target coordinates, target velocity, and a specific heading angle, output by the perception segment. Simultaneously, fuzzy logic timing segmentation requires calls to two data interfaces: traffic information and queue length information. Continuous monitoring and analysis of the overall data flow in the system's internal global memory pool is performed. The percentage of physical traffic volume in the total input / output data flow of each segment, determined by globally shared variables, is statistically analyzed. This traffic percentage is then extracted and reused as a result. Assume that the overall input / output data stream size for trajectory prediction segments during a specific time period is 50 megabytes per second, and the shared vehicle status table variables occupy 30 megabytes per second. The system calculates and reuses the corresponding results through division. The specific value is 0.6. During cross-shard virtual communication interactions, the probe captures the interaction commands and underlying network packet headers. The actual network bandwidth bits used by the standard JSON format interaction commands appended to the packet headers are calculated. Based on this communication scale, the node transmission overhead can be extrapolated. The system measures that there are 1000 interactions per second between the sensing segment and the trajectory prediction segment. Each data interaction, including the header and the instruction itself, has a total length of 200 bytes. The system calculates the corresponding transmission overhead. The value is 1.6 megabits per second. The three measured parameters are uniformly normalized to eliminate dimensions. The normalization value for prior dependencies is set to 0.6, and the normalization value for result reuse is also set to 0.6. Simultaneously, the normalization value for the 1.6 megabits per second transmission overhead is set to 0.32. A preset weighted formula is used to perform linear combination calculations on the dimensionless parameters. The weights within the preset weighted formula are set to 0.4, 0.35, and 0.25 respectively. Using the Pearson correlation coefficient algorithm, the prior dependencies, reuse parameters, and transmission overhead are superimposed into an independent task coupling value. The coupling value is calculated based on the above series of input parameters. The result is equal to 0.53. Finally, a greedy algorithm is used to optimize the search, aiming to minimize the coupling value and maximize the priority weight. This search generates a secure, collaborative, distributed execution sequence across nodes. .

[0068] Specifically, the coupling values ​​of edge task nodes satisfy the following relationship:

[0069] in, A dimensionless coupling value representing the degree of correlation between task slices; This indicates that the prior dependency of the dimensionless proportional shrinkage normalization has been eliminated. This indicates that the normalization results after removing flow metric features are reused. This indicates that the normalized transmission overhead is obtained through transformation and extraction. , , These represent the three network parameters that are individually pre-assigned with corresponding proportional weights based on an empirical model.

[0070] This network aggregation formula is built upon general engineering design principles for evaluating the interrelationship of sub-services within a computer microservice architecture. This design approach precisely calculates the intricate resistance data between algorithm nodes at the memory exchange and interface retrieval levels, specifically avoiding disruption of the original efficient computing links due to disordered forced scheduling and segmentation. This significantly reduces the enormous scale of unnecessary physical communication bandwidth consumption across chassis communication nodes.

[0071] It should be noted that, , , The specific settings are based on benchmark test results of the underlying communication overhead of the microservice architecture. By controlling variables in the experimental environment, the number of preceding dependency interface calls between each subtask node, the number of bytes reused in the memory pool, and the additional bandwidth for cross-process communication were independently amplified, and the resulting bus latency growth rate was measured. Based on the actual physical impact of these three indicators on overall communication congestion, normalized calculations were performed to extract corresponding weighted values, thereby accurately guiding the subsequent segmentation and generation of cross-node secure collaborative distributed execution sequences.

[0072] In the underlying hardware computing power supply and demand matching and scheduling decision-making stage, historical computing power load and queue length data are input into a long short-term memory neural network model for trend prediction. This AI prediction model directly outputs the predicted future load increment of the system through inference calculations. The predicted load increment is extracted and divided by the nominal processing capacity of a single CPU physical core to obtain the corresponding theoretical quotient, which is used as the base number of cores for scaling. Simultaneously, the long short-term memory neural network model synchronously outputs the predicted queue length gradient value corresponding to the future evolution analysis cycle. Based on this predicted queue length gradient, a mapping is dynamically assigned to match the specific physical scaling factor of the corresponding system. Finally, the system's basic scaling capacity or basic computing power increment is multiplied by the scaling factor to calculate the resource scaling step size. To address the issue of computing power fluctuations across cycles.

[0073] As an alternative implementation path for multidimensional nonlinear data analysis and computing power extrapolation, a resource monitoring probe is used to cyclically read the CPU utilization rate sequence within a past observation window at a fixed scanning period of 1 second. For example, the probe component measures a total of 60 sampling points, with the corresponding core utilization rates stably distributed within a specific load range of 45% to 75%. The arithmetic mean of the above parameter set is calculated to obtain the historical computing power load representing the normal stable load of the system. This set of statistical data calculates the historical computing load. The average value is 62%. The maximum queue length at the intersection during the current traffic congestion outbreak is simultaneously extracted and measured. The length of the queue was determined visually by radar sensors. The spatial extension is 85 meters. This historical computing load... The baseline value of 62% and the queue length value of 85 are used as the dual-path input parameter set. These joint feature variables are input into the system's pre-set polynomial resource mapping function for fitting operations. Alternatively, the system can input the corresponding extracted parameters into a support vector regression model for inference operations. The inference and determination model ultimately outputs the cross-cycle resource pool expansion / shrinkage capacity for addressing traffic congestion evolution directly to the system. This cross-cycle resource pool expansion / shrinkage capacity is directly used as the resource scaling step size. For example, if calculations determine that an additional 0.4 TFLOPS is needed, the system will directly lock this 0.4 TFLOPS requirement as the current resource scaling step. .

[0074] Specifically, the resource scaling step size satisfies the following relationship:

[0075] in, This indicates that the final output of the computational model is the predicted resource scaling step size for the cross-cycle resource pool, expressed in TFLOPS. This indicates the extraction of polynomial operation mapping functions or regression support inference operators that have been trained offline using a large amount of pre-trained load test data. This indicates that the monitoring probe reads and obtains the historical computing load of the corresponding current edge computing device; This indicates that the current maximum queue length at the target intersection is obtained through a road surface detection system.

[0076] This expansion / contraction expectation resource calculation formula is generated based on the well-known principle of using classic regression analysis algorithms in machine learning to compensate for dynamic and variable high-voltage load disturbances in control systems. This innovative design directly utilizes the characteristic data of large-scale traffic congestion in the upper-level physical queues as a prerequisite feature variable to predict the peak computing power demand of the underlying cabinet control chips. This eliminates and compensates for the shortcomings of traditional methods that rely solely on lagging processing unit utilization feedback, which can easily lead to irreversible crashes and disconnections.

[0077] During the cross-chassis transfer of algorithm containers and the corresponding core allocation calculation stage, the nominal maximum safe operating temperature limit of the core motherboard microprocessor is read as the system scheduling temperature rise constraint. If the core motherboard thermodynamic temperature sensor reads 78 degrees Celsius, the system automatically invokes an empirical model that incorporates the chip's own physical silicon wafer thermodynamic decay law and internal temperature rise and power consumption to perform calculations. This model then infers the maximum number of usable CPU cores under the current stringent thermal conditions. The constraint is locked at 6 cores. The maximum allowed physical bandwidth of the network backbone fiber is queried from the upstream aggregation layer switch. The bandwidth reserved for basic communication for monitoring video streaming is forcibly and exclusively deducted from the maximum total bandwidth. The remaining available bandwidth is extracted and used directly as the assessment link congestion threshold. For example, if the underlying calculation yields a remaining available communication bandwidth of 400 megabits per second, this measured limit is directly used as the link congestion threshold. Based on the temperature constraints mentioned above, a maximum of 6 usable physical cores is determined. The target resource scaling step size will then be calculated. The corresponding 0.4 TFLOPS is directly allocated on demand to the currently healthy and available physical cores. Currently, due to high temperature and frequency reduction, each core can only nominally provide 0.2 TFLOPS. The core allocation for this node is calculated by performing mutual division. This equates to 2 physical cores. If the calculated peak computational load for the current collision avoidance safety task requires a forced concurrent allocation of a total of 4 computing cores, this task's burst demand has exceeded the currently available and supplementary core allocation for this hardware node. This means a resource limit of 2 physical cores, which triggers a container migration mechanism to package computing tasks exceeding the computing power limit (equivalent to 2 cores) into containers. Based on the aforementioned measurement of a 400 Mbps link congestion threshold... Strictly limit the physical size of the generated container image file. Calculations and deductions determine the specific container migration volume that must be forcibly relocated and unloaded to other idle security edge nodes in the vicinity. .

[0078] Specifically, the core allocation satisfies the following relationship:

[0079] in, This indicates the number of additional cores required to be allocated, calculated and verified by the system's operations, in units of physical cores. This represents the calculated resource scaling step size of the system in response to increased load demand, expressed in TFLOPS. This represents the current stable level of computing power provided by a single usable physical core under the influence of extreme high temperature, power consumption, heat generation, and frequency reduction physical control, in units of TFLOPS.

[0080] The above relationship is generated based on the common logic of physical deduction and division, which is a general instruction set of the underlying von Neumann computer architecture. This design can accurately map and transform the load demand scale of the large and complex congestion prediction abstract algorithm at the top level to the physical boundary limit provided by the underlying microelectronics heat-generating physical computing power hardware. This achieves a closed-loop binding protection effect between the macro-level software control code operation and the heat limit of the rack silicon chip.

[0081] Thus, this step solves the technical problem of control command delay caused by the overload of underlying computing power under high-concurrency mixed traffic, and builds a solid hardware resource guarantee for the global timing deadline warning and control cycle issuance.

[0082] S3: Adjust task constraints and continuously optimize command issuance. Based on the link jitter amplitude and container migration amount, a migration disturbance factor is generated, and the task deadline and data age constraints are adjusted. When the traffic conflict risk index exceeds the preset threshold, the signal release order, duration and unmanned vehicle passage window are calculated using rolling optimization. The optimization cycle is adjusted according to the resource scaling step size, migration disturbance factor and execution deviation of adjacent nodes, and the execution is sent to the unmanned traffic nodes.

[0083] In an optional embodiment, in the underlying communication interference quantization and time limit adjustment module, the link jitter amplitude is first calculated by extracting the current environment. With the required cross-node transport container migration amount To eliminate the objective physical dimension difference between the two heterogeneous parameters, each variable is independently multiplied by a matching corresponding dimension conversion coefficient. The link jitter amplitude is then set. The conversion factor is fixed at 1 to preserve the original time attributes. The container migration amount is also set. The corresponding conversion factor is 0.2 milliseconds per megabyte to represent the time equivalent consumed by data migration per unit size. After conversion, arithmetic addition is performed to obtain an intermediate value of 63 milliseconds. Subsequently, the built-in preset normalization scaling factor is used to scale this intermediate sum, and the final value is 49.5 milliseconds. This result of 49.5 milliseconds is defined as the migration perturbation factor. This factor has a clear time dimension and objectively reflects the combined delay caused by network instability and computing power relocation on the front-end system response.

[0084] Specifically, the migration disturbance factor satisfies the following relationship:

[0085] in, The migration disturbance factor represents the latency caused by the combined effects of network instability and relocation, expressed in milliseconds. This indicates the currently measured link jitter amplitude, in milliseconds; This indicates the amount of container migration required for cross-node flow transfer in the system, in megabytes. This represents the preset time-dimension conversion coefficient for link jitter; This indicates the time-equivalent conversion factor configured for data per unit volume; This represents the predefined normalization scaling factor of the internal system.

[0086] This relation is constructed and applied based on a common knowledge system of unified dimensional scaling synthesis of multiple interference factors commonly found in automation control engineering. This invention addresses the numerical estimation of the physical transmission resistance and link jitter amplification effect caused by the relocation process itself during software and hardware transfer, avoiding blind migration that could directly drag down or even exceed the originally expected optimization time margin.

[0087] After completing the above interference factor synthesis operation, retrieve the initial allocation and preset initial task deadlines for various control tasks within the edge computing service stack. And the initial data age constraint for the allocation of forecasting business. The front-end hard real-time control task system itself is configured with the most stringent time boundary bottom line and the maximum tolerable deadline. Since container migration scheduling actions at this time must allow for a safety margin of time for additional cross-chassis communication, the maximum tolerable deadline corresponding to the extracted value of 200 milliseconds is used directly. Subtracting the deduction and migration disturbance factor The time compensation amount is directly proportional to the original time limit. Setting the deduction ratio to a fixed 1.5, the calculated required time reduction compensation is 74.25 milliseconds. After compressing and updating the original time limit and removing this compensation amount, the final updated task deadline is obtained. The compression was forced to 125.75 milliseconds. This, in turn, forced a constraint that the control result must be output within a shorter available time window after compression. On the other hand, there are constraints on the initial data age for the trajectory prediction task. Superimposed supplementary and migration perturbation factors This directly corresponds to the relevant network transmission attenuation time loss value. If the system selects a communication loss coefficient equal to 1, a tolerance index of 49.5 milliseconds will be added to the original constraint. The latest available data age constraint is then directly updated and reset. The tolerance threshold has been extended to 549.5 milliseconds.

[0088] Specifically, the updated task deadline satisfies the following relationship:

[0089] in, This indicates the updated task deadline, in milliseconds. This indicates the maximum tolerance cutoff time for native hard real-time control, in milliseconds. This represents the integrated network and relocation disturbance factor, expressed in milliseconds. This represents the control task time minus the compensation ratio coefficient, which is proportional to the disturbance factor.

[0090] In real-time control systems, instructions must be issued within the maximum tolerable time to ensure security. This invention, through logical derivation, forcibly deducts the time compensation amount caused by edge computing power migration and network jitter—that is, the product of the migration disturbance factor and the time deduction compensation coefficient—based on the maximum tolerable deadline, constructing a dynamically tightened task time limit calculation model. This provides sufficient safety margin for cross-chassis communication and task flow of underlying containers, eliminating the risk of traffic control execution windows being blocked due to the time-consuming nature of computing power scheduling actions themselves.

[0091] Specifically, the data age constraint satisfies the following relation:

[0092] in, This indicates the updated data age constraint, in milliseconds. This indicates the initial data age constraint extracted from the original forecasting business, in milliseconds. This represents the loss factor for the relaxation of transmission attenuation time tolerance; This represents the integrated network and relocation disturbance factor, expressed in milliseconds.

[0093] This formula is constructed based on the theoretical derivation and construction principles of the predictive data lifecycle. The effectiveness of traffic prediction data decays over time. This invention constructs a data age elastic expansion derivation formula, which, based on the original data age constraint, superimposes the transmission attenuation time loss value caused by container migration and network jitter. Through a mathematical compensation mechanism, the availability time of data under extreme fluctuation conditions is appropriately relaxed, ensuring that road network collaborative prediction data can still be smoothly accessed under high load conditions at edge nodes, preventing the system from falling into deadlock and recalculation due to frequent data expiration.

[0094] In the traffic operation control strategy rolling execution calculation module, corresponding safety warning boundary parameters are set as preset thresholds for judgment. Traffic conflict risk indicators are compared in real time. The aforementioned safety threshold is predefined. When the monitoring detects that the objective value of this indicator is greater than and directly exceeds the predefined threshold red line, an emergency traffic control intervention procedure is directly triggered. The internal model predictive control algorithm is actively invoked to formally initiate the global rolling optimization decision cycle. A lightweight microscopic simulation test domain environment is constructed, pre-limiting the planned future global prediction time domain coverage to a total of three complete signal cycle time axes. For the model predictive control algorithm, the prediction time domain is calculated, and the corresponding road signal control state transition mathematical equation is directly constructed based on the Markov decision process model specific to traffic operations research. The core of the operation uses maximizing the total traffic flow of the entire intersection system and minimizing the overall average vehicle backlog delay as the joint objective function with two-way constraints. The eight parallel independent traffic flow release sequences, including the four different approach lanes carrying straight and left turns, are directly extracted as control variables in the discrete decision space. An integrated genetic algorithm program is invoked to quickly and concurrently search for the optimal global solution within the constructed state space. The final applicable optimal signal release order is directly output through this step of inference and calculation. For example, the optimal signal release order is presented by optimizing the output, which includes a four-segment arrangement of north-south straight, north-south left turn, east-west straight, and east-west left turn. At the same time, for the above sequence, the duration of the corresponding independent traffic light needs to be matched for the current round for each release order is calculated.

[0095] During the calculation of the specific time-phase duration and the dedicated passage lane for autonomous vehicles, two key features—the arrival rate of input vehicles and the saturation queue dissipation rate—are acquired using real-time perception. An independent traffic flow dissipation model is then constructed within the test domain based on these inputs. Assuming that the number of vehicles queuing ahead of the green signal release phase detection is 15, the shortest time required to clear and dissipate the total number of vehicles queuing in the current phase is calculated. This is equal to 50 seconds. Simultaneously, a threshold is set for the maximum allowed duration of resource occupation corresponding to this one-way independent release phase. The duration is 60 seconds. The two metrics are compared, and the smaller extreme value of 50 seconds is selected as the duration for which the current physical phase will be allocated. In scenarios involving complex traffic flow where human-machine interaction is intertwined, the system utilizes microwave radio frequency communication between vehicles and roads to remotely receive and intercept reports from unmanned vehicles (UVs) actively traveling 200 meters from the stop line at intersections. Simultaneously, the estimated time required for the designated UV to physically reach the stop line at the intersection is calculated. =Equal to 20 seconds. Combining multimodal forward-looking visual intelligent perception, it is directly determined that there are 3 stationary human-driven vehicles queuing at the forefront of the lane where the unmanned vehicle is being tested. Based on environmental measurements, the overall average queuing dissipation rate of humans is fixed at 0.5 vehicles per second. Based on this, it is calculated that it will take 6 seconds for the 3 human-driven vehicles blocking the queuing to completely start and dissipate. At the same time, combined with the pre-set safety following buffer distance according to safety regulations, a safety following time margin is generated. Combining the above two dissipation and following buffer margins, it is objectively deduced that the unmanned vehicle must be forced to be at the time node 8 seconds after the current green light officially turns on in order to obtain the condition of unobstructed passage. If the current green light phase of the test will end at the 15th second of the cycle according to the traffic signal timetable, the internal spatiotemporal conflict resolution algorithm is called to perform intersection judgment inference, and a conflict-free continuous passage period covering the current signal cycle idle time is directly calculated and issued for the target unmanned vehicle. The start and end point of this dedicated channel is defined from the 8th second after the current green light turns on to the 15th second. The above-mentioned time tolerance interval with a total duration of 7 seconds is extracted and used as the autonomous vehicle passage window. This directly guides and controls the unmanned vehicle to perform optimal acceleration and deceleration trajectory calculation and planning based on this window.

[0096] Specifically, the duration relationship is as follows:

[0097] in, This indicates the duration for selecting the optimal phase as the final evaluation phase, expressed in seconds. This indicates the shortest time required to clear all vehicles in the queue using flow rate calculations, expressed in seconds. This indicates the maximum permissible duration threshold for one-way traffic lights as defined by traffic management rules, expressed in seconds.

[0098] This formula is based on traffic light timing safety constraints in traffic engineering. The conventional approach in this field is to ensure sufficient clearance of queuing vehicles at intersections while preventing excessively long one-way green light times that could lead to a global deadlock and paralysis of the intersection. Therefore, the smaller of the minimum clearance time and the maximum allowable time is selected. This invention improves upon this by applying it to a business scenario, using it as a single-step execution constraint for a Markov decision model in a microscopic simulation domain. This ensures that the duration of each phase output by the system through the genetic algorithm maximizes current local traffic efficiency while strictly adhering to the global safe time tolerance limit of the road network.

[0099] Specifically, the start time of the autonomous vehicle passage window satisfies the following relationship:

[0100] in, Indicates the start time of the autonomous vehicle passage window, in seconds; This represents the number of human-driven vehicles queuing ahead; it is a dimensionless quantity. This indicates the queue dissipation rate, measured in vehicles per second. This indicates the safe following time margin, in seconds.

[0101] After calculating the start time of the autonomous vehicle passage window, and combining it with the remaining duration of the current green light phase, a conflict-free continuous passage period is extracted and used as the autonomous vehicle passage window.

[0102] This formula is based on the theoretical derivation of a vehicle kinematics dissipation model. To calculate the absolutely safe passage time for an autonomous vehicle in mixed traffic flow, this invention uses theoretical derivation to divide the total number of manually driven vehicles locked by forward visual perception by the average queue dissipation rate to determine the time required for the preceding convoy to fully start, and adds a safety following time margin to avoid collisions. This formula is reasonable and rigorous, eliminating the physical obstruction caused by human drivers' starting delays and slow reactions in mixed traffic flow, and accurately deriving a conflict-free starting point for autonomous vehicles.

[0103] It should be noted that the queue dissipation rate is obtained by calculating the average speed at which vehicles pass the stop line one after another after the green light turns on for all permitted phases, based on historical monitoring video streams of the target intersection during pre-set time periods, such as the past week's morning and evening rush hours. The safe following time margin is obtained by calculating and deriving it based on the legal speed limit of the target road segment and the standard braking response time of the autonomous vehicle's drive-by-wire chassis system, combined with the classic kinematic safe braking distance formula, to ensure an absolute physical collision safety boundary in mixed traffic flow.

[0104] During the final command issuance and control phase of the network-wide collaborative adaptive feedback and new cycle adjustment, at the end of each optimization cycle, a collaborative data request is proactively sent to adjacent upstream and downstream traffic network nodes via dedicated network cables. Subsequently, the system successfully receives the traffic flow summary data from the previous cycle from the upstream adjacent nodes. If the adjacent node's message feedback indicates that its original expected system output to this receiving node was 45 vehicles per cycle, while this control node actually verifies the actual arrival traffic flow as 38 vehicles per cycle through the underlying visual checkpoint sensing equipment, a subtraction calculation is performed to determine the absolute difference between the two traffic flow data, resulting in a difference of 7 vehicles per cycle. The absolute value of this calculated difference is extracted and used as the execution deviation of the adjacent node. Among them, the execution deviation of adjacent nodes This is used to characterize the degree of distortion in upstream and downstream traffic flow prediction coordination. The hardware resource scaling step size is extracted and calculated from the aforementioned stages. Time-constrained control of migration disturbance factor And the aforementioned acquisition of spatial error and adjacent node execution deviation As independent input parameter variables, to fuse the three heterogeneous attribute variables, each is multiplied by its corresponding bound adjustment coefficient parameter with corresponding dimension transformation attribute. The adjustment coefficient corresponding to the set scaling step size is extracted and confirmed to be 10 seconds per TFLOPS; the disturbance factor matching correlation adjustment coefficient is directly set to 0.02 seconds per millisecond; and the execution deviation spatial distortion adjustment coefficient is set to 0.5 seconds per vehicle. The time adjustment amount is obtained by multiplying the above three parameters by their respective adjustment coefficients and then summing them. The time is 8.49 seconds. During the initial startup phase of operation or during periods of stable and safe off-peak traffic flow, a fixed, preset initial rolling optimization cycle parameter is extracted and used as the decision-making benchmark value for the entire system. The extracted baseline value This represents the standard safety time interval for the intersection collision avoidance control algorithm to perform a global data recalculation. When the target monitored intersection is experiencing high temperatures due to computing power overheating or in a harsh environment, a preset baseline value will be retrieved. The extracted time adjustment amount is calculated by superimposing the value of 30 seconds with the aforementioned values. For a time interval of 8.49 seconds, the two values ​​are directly added together for cumulative compensation. The resulting cumulative value equals 38.49 seconds, which is then used as the new actual optimization cycle required for the next time window. The system proactively extends its decision-making and computational buffer cycle by actively applying a stretching or appropriate compression optimization cycle adjustment mechanism within the aforementioned control architecture. The globally optimal signal release order, phase duration, planned unmanned vehicle passage window data, and the newly calculated optimization cycle are jointly packaged and structured into a lightweight network data packet in a specified JSON format. This decision data packet is then published and pushed to a dedicated message buffer queue on the edge nodes via an integrated industrial MQTT IoT communication high-speed protocol. Finally, after receiving, subscribing, capturing, and reverse-parseting the commands, the system transmits them to the road relays through a high-voltage interface to execute the traffic signal color conversion and vehicle right-of-way scheduling commands.

[0105] Specifically, the execution deviation between adjacent nodes satisfies the following relationship:

[0106] in, This indicates the degree of distortion in the collaborative prediction between this intersection and external detection, and the deviation of adjacent nodes is determined by the number of vehicles per cycle. This represents the expected arrival traffic volume reported by adjacent upstream nodes, in units of vehicles per cycle; This indicates the actual number of arriving vehicles, expressed in units of vehicles per cycle.

[0107] The formula is based on the calculation of absolute error in mathematical statistics. The conventional approach in this field is to intuitively assess the distortion of the prediction model by calculating the absolute value of the difference between the predicted value and the actual physical observation. This invention improves upon this conventional approach, specifically using it to measure the collaborative control reliability of upstream and downstream edge computing nodes. This transforms the degree of collaborative distortion in the road network space into a specific vehicle difference scale, which serves as a key independent input variable to trigger subsequent adaptive adjustments in the rolling optimization cycle.

[0108] Specifically, the time adjustment amount and the optimization period satisfy the following relationships:

[0109]

[0110] in, This represents the time adjustment amount with time dimensions obtained from the superposition calculation, with the unit being seconds; This indicates that a new optimization cycle, measured in seconds, is generated after comprehensive environmental and hardware compensation calculations to adapt to the next time window. This indicates that the system's preset initial rolling optimization cycle baseline value is fixed when there is no heat generation interference, and the unit is seconds; Indicates the resource scaling step size calculated by the hardware; This indicates the migration disturbance factor extracted due to network interference. Indicates the execution deviation between adjacent nodes; , , These represent the respective adjustment coefficients that have dimension conversion properties.

[0111] This formula is based on the theoretical derivation principle of multivariable closed-loop feedback compensation for complex systems. Existing traffic signal control typically uses a fixed calculation cycle, which is prone to stalling when computing power is insufficient. This invention constructs a theoretical calculation formula for a flexible optimization cycle. Based on the system's inherent initial optimization cycle, it uses resource scaling step size (computing power dimension), migration disturbance factor (network dimension), and adjacent node execution deviation (business dimension) as three independent feedback variables. Each variable is assigned a corresponding dimension transformation attribute adjustment coefficient and then linearly superimposed for compensation. The rationality of this formula lies in its rigorous mathematical combination, which enables adaptive stretching and compression of the control cycle under extreme computing power and road conditions, ensuring that edge nodes do not crash.

[0112] In this way, this step combines the underlying network disturbances and prediction deviations to dynamically adjust the task time limit and optimization cycle, overcoming the computation timeout defects that are easily caused by fixed control time windows, and ensuring the timely and stable issuance of collision avoidance control signals and unmanned vehicle passage window instructions under fluctuating system conditions.

[0113] To verify the practical technical effectiveness of the aforementioned edge computing-based adaptive control method and system for unmanned traffic nodes, an edge computing gateway equipped with an eight-core processor and an AI accelerator card was used, along with a two-megapixel camera and millimeter-wave radar deployed at the intersection. A microservice architecture system was deployed in the software environment, and a road network model containing four connected intersections was built in microscopic traffic simulation software. Traffic flow input conditions were set to a high-saturation state during the morning peak, with vehicle arrival rates fluctuating between 0.2 and 0.5 vehicles per second. Simultaneously, network transmission latency differences of 50ms to 150ms were randomly injected into the underlying communication links to construct link jitter. The unmanned vehicle penetration rate was set at 15%, and the experiment duration was set at 2 hours. Control groups included: traditional fixed-sequence control and static resource allocation methods, conventional control methods without container migration and periodic adjustments, and the complete multi-source fusion and computational power collaborative scheduling method proposed in this invention.

[0114] Under a 2-hour morning rush hour traffic flow stress test, the verification data shows that the traditional fixed-time method has an average intersection throughput of 2100 vehicles / hour, an average queue length of 95 meters, and edge node CPU utilization consistently above 85%, with three instances of task timeouts due to system overheating and frequency throttling. The conventional control method increases the average throughput to 2450 vehicles / hour and reduces the average queue length to 68 meters, but when facing high-concurrency prediction tasks, the control instruction latency reaches a maximum of 400ms, exceeding the maximum tolerable deadline. In contrast, the complete method of this invention, through container migration and rolling optimization cycle adjustment, achieves an average throughput of 2830 vehicles / hour and reduces the average queue length to 42 meters. Regarding computing power scheduling, this method precisely triggers 5 adjacent node container migration operations, keeping the local node processor utilization stably within a reasonable range of around 60%, and strictly limiting the maximum instruction response latency to within 125ms.

[0115] The proposed solution achieves a maximum increase in traffic volume of 34.7% and a maximum reduction in queue length of 55.7%, effectively eliminating the risks of thermal runaway and system crashes caused by computational overload. This technological improvement directly benefits from the deep coupling of edge-side computing resource scaling with traffic signal evolution characteristics, particularly by utilizing migration disturbance factors and optimization cycle adjustment mechanisms to mitigate the impact of network link jitter and sudden surges in traffic on system computing power. The precise decoupling of task microservice sharding and the cross-node allocation of cores not only ensure the safe and stable issuance of dedicated passage windows for autonomous vehicles in mixed traffic flows but also guarantee the high availability and low-latency control capabilities of the overall road network system in multi-intersection collaborative scheduling scenarios.

[0116] Figure 2 This diagram illustrates the degree of correlation between task fragments. The two sets of parallel bars represent the actual normalized index values ​​and the preset weights assigned by the system for each core dimension. The three segments on the horizontal axis represent the three correlation evaluation dimensions: prior dependencies, result reuse, and transmission overhead. The scale on the vertical axis shows the specific dimensionless numerical magnitudes corresponding to each parameter.

[0117] The graph shows that the bars representing the actual indicators of each dimension are not evenly distributed. Instead, the bars representing prior dependencies and result reuse are significantly higher than those representing transmission overhead. This proves that the system evaluates the real physical resistance between algorithm nodes at the memory exchange and interface call levels, and does not incorrectly apply a mechanically equal distribution to all network loss dimensions. Observing the bars representing preset weights, it can be seen that the system assigns the highest priority to prior dependencies. As the evaluation dimension shifts towards transmission overhead, the weight bars show a decreasing stepwise trend. This corresponds to the physical characteristics described in the specific implementation method, which utilizes the Pearson correlation coefficient algorithm to accurately aggregate feature parameters to generate task coupling values.

[0118] Figure 3 This is a schematic diagram of the historical computing load sequence. The undulating broken lines and their nodes in the diagram represent the actual fluctuation trajectory of the processor utilization of edge computing devices under continuous historical sampling points; the horizontal dashed line that runs through the entire coordinate system represents the historical average computing load benchmark obtained through arithmetic deduction.

[0119] The graph shows that the continuous utilization rate curve did not remain stable at a fixed baseline, but rather exhibited violent high-frequency jumps and oscillations around a core level. This proves that the hardware resource monitoring probe successfully captured the transient disturbances in the underlying physical computing power under the impact of complex traffic flow, and did not incorrectly use a single occasional extreme value to represent the overall system load. Observing the horizontal dashed line running through the entire graph, it can be found that it is precisely anchored at the center of the oscillation range of the curve. This corresponds to the technical feature described in the specific implementation method of smoothing out transient heating spikes and extrapolating a normalized stable load benchmark through long-term window statistics.

[0120] Figure 4 This is a schematic diagram of the continuous passage time distribution of autonomous vehicles. The three independent rectangular blocks in the diagram represent, from bottom to top, the dissipation of human vehicle queues, the safety following time margin, and the dedicated passage window for autonomous vehicles. The horizontal span of the rectangular blocks represents the physical duration of each stage on the overall passage time axis. The scale on the horizontal axis shows the second-level timing nodes of the specific control actions.

[0121] The image shows that the rectangular blocks representing each traffic phase do not overlap or intersect on the timeline, but rather exhibit a strict, step-like sequential arrangement. This proves that the spatiotemporal conflict resolution algorithm successfully eliminates physical collisions and cross-interference caused by mixed traffic flows, and does not incorrectly include the start-up time of human vehicles into the target autonomous vehicle's traffic margin. Observing the topmost autonomous vehicle-specific traffic window block, it can be seen that its starting node is closely connected to the moment the following buffer margin ends and continues to extend to the end of the current signal cycle. This corresponds to the technical feature described in the specific implementation method of accurately mining clean and empty traffic intervals from the gaps in congested mixed traffic flows to alleviate the stagnation of autonomous vehicles.

[0122] This invention also discloses an edge computing-based adaptive control system for unmanned traffic nodes, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the edge computing-based adaptive control method for unmanned traffic nodes according to this invention.

[0123] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0124] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An adaptive control method for unmanned traffic nodes based on edge computing, characterized in that, include: Collect video streams, radar traces, communication latency and edge computing power occupancy information of unmanned traffic nodes, extract target density, speed dispersion, number of conflict trajectories, queue length gradient and link jitter amplitude, and construct state vectors; Traffic conflict risk indicators are calculated based on state vectors, and conflict evolution characteristics are generated by combining the rate of change and fluctuation intensity of traffic conflict risk indicators within a time window. The computational load from traffic conflict risk indicators, conflict evolution characteristics, and edge computing power occupancy information is mapped into a task priority matrix. Edge-side perception and control tasks are segmented for processing. Coupling values ​​are calculated based on pre-order dependencies, result reuse, and transmission overhead to generate a distributed execution sequence. Resource scaling step size is calculated based on historical computing power load and queue length. Container migration amount and core allocation amount are determined by combining temperature rise constraints and link congestion thresholds. Migration disturbance factors are generated based on link jitter amplitude and container migration amount to adjust task deadlines and data age constraints. When traffic conflict risk indicators exceed preset thresholds, rolling optimization is used to calculate the signal release order, duration, and unmanned vehicle passage window. The optimization cycle is adjusted based on resource scaling step size, migration disturbance factors, and execution deviations of adjacent nodes before execution is issued to unmanned traffic nodes.

2. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The construction of the state vector includes: dividing the total number of vehicles within the detection range of the video stream and radar points by the actual physical area of ​​coverage to obtain the target density; extracting the current instantaneous speed of the vehicles and calculating the absolute deviation of the instantaneous speed from the average speed of the vehicles in the historical sampling period to obtain the speed dispersion; extending each vehicle along its current driving direction to obtain the trial running trajectory, and counting the cumulative number of intersection collision points of the trajectories in the two-dimensional space to obtain the number of conflict trajectories; calculating the difference in the total length of the parking queue in two adjacent sampling periods and dividing it by the sampling time interval to obtain the queue length gradient; extracting the sending and receiving timestamps of the underlying communication data packets and calculating the range of the transmission delay of adjacent data packets to obtain the link jitter amplitude; and concatenating the target density, speed dispersion, number of conflict trajectories, queue length gradient, and link jitter amplitude into a column vector according to a preset dimension to generate the state vector.

3. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The method of calculating traffic conflict risk indicators based on state vectors and generating conflict evolution features by combining the rate of change and fluctuation intensity of traffic conflict risk indicators within a time window includes: normalizing each data element in the state vector and assigning corresponding weight coefficients to each normalized data element; weighting and summing the normalized state vector and the weight coefficients to obtain the traffic conflict risk indicator; calculating the absolute value of the difference between the traffic conflict risk indicator in the current time window and the previous time window to obtain the rate of change; extracting the traffic conflict risk indicator from a preset number of consecutive historical time windows before the current time window and calculating the standard deviation to obtain the fluctuation intensity; and linearly weighting and combining the traffic conflict risk indicator, the rate of change, and the fluctuation intensity according to a preset ratio to generate the conflict evolution features in scalar form.

4. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The step of performing segmented processing on edge-side perception and control tasks, and calculating coupling values ​​based on prior dependencies, result reuse, and transmission overhead, includes: decoupling the edge-side algorithms according to a directed acyclic graph structure to divide the perception, prediction, timing, and control tasks into independent execution segments; counting the number of call interfaces of each segment to the output data of the preceding segment to calculate the prior dependencies; counting the proportion of globally shared variables in the total data stream of each segment to calculate the result reuse; counting the number of network bandwidth bits occupied by inter-segment interaction instructions and data packet headers to calculate the transmission overhead; and using a preset weighted formula to perform a weighted summation of the prior dependencies, the result reuse, and the transmission overhead to obtain the coupling value representing the degree of correlation between the task segments.

5. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The method of determining container migration and core allocation by combining temperature rise constraints and link congestion thresholds includes: calculating the average processor utilization rate within a historical time window as the historical computing load; inputting the historical computing load and queue length into a resource mapping function to calculate the cross-cycle resource pool expansion / contraction capacity as the resource scaling step size; obtaining the processor's highest safe operating temperature as a temperature rise constraint and calculating the current upper limit of available CPU cores; obtaining the maximum allowed transmission bandwidth of the upstream switch and deducting the bandwidth occupied by basic communication, extracting the remaining available bandwidth as the link congestion threshold; allocating the resource scaling step size to available physical cores based on the upper limit of available CPU cores to calculate the core allocation for this node; when the required computing resources exceed the core allocation, converting the excess computing tasks into offloadable task loads according to the link congestion threshold to obtain the container migration amount.

6. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The step of generating a migration disturbance factor based on the link jitter amplitude and container migration amount, and adjusting the task deadline and data age constraints, includes: multiplying the link jitter amplitude and the container migration amount by their respective dimension conversion coefficients and then adding them together, and scaling them using a preset normalization coefficient to generate the migration disturbance factor with time dimensions for time compensation; subtracting a time compensation amount proportional to the migration disturbance factor from the preset maximum tolerance deadline to update the task deadline of the control task; and adding a transmission attenuation time loss value corresponding to the migration disturbance factor to the predicted initial data age of the task to update the data age constraints.

7. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The method of using rolling optimization to calculate the signal release order, duration, and autonomous vehicle passage window includes: within a preset simulation test domain, optimizing the release sequence of different traffic flows as decision variables with the objective function of maximizing intersection throughput, to obtain the optimal signal release order; constructing a traffic flow dissipation model using vehicle arrival rate and queue dissipation rate, calculating the shortest time required to clear the queue of vehicles in the current phase, and taking the smaller value between the shortest time and a preset maximum allowable duration threshold as the duration; calculating the predicted time for the autonomous vehicle to reach the intersection stop line, and calculating a conflict-free continuous passage period based on the queue dissipation time and safe following distance of the preceding human-driven vehicles, as the autonomous vehicle passage window.

8. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The step of adjusting the optimization cycle based on the resource scaling step size, migration disturbance factor, and adjacent node execution deviation includes: receiving the expected arrival traffic flow and the actual arrival traffic flow from adjacent nodes, calculating the absolute value of the difference between the two as the adjacent node execution deviation; using a preset initial rolling optimization cycle as a baseline value; multiplying the resource scaling step size, the migration disturbance factor, and the adjacent node execution deviation by their respective corresponding adjustment coefficients with dimensional conversion attributes, and superimposing them to obtain a time adjustment amount with time dimensions; adding the baseline value and the time adjustment amount to generate the optimization cycle required for the next time window.

9. The adaptive control method for unmanned traffic nodes based on edge computing according to claim 1, characterized in that, The process of mapping the computational load from traffic conflict risk indicators, conflict evolution characteristics, and edge computing power occupancy information into a task priority matrix includes: extracting the real-time CPU utilization rate within the edge computing power occupancy information as the initial computational load; performing scaling mapping processing on the traffic conflict risk indicators, conflict evolution characteristics, and initial computational load using the max-min normalization algorithm to construct the current environment state vector; calculating the inner product dot product value corresponding to the preset feature weight vector and the current environment state vector to deduce the dynamic priority score of each independent task segment; and arranging and combining the network identifiers, dynamic priority scores, and estimated computing resource requirements of all task segments in descending order of dynamic priority scores to construct a two-dimensional task priority matrix.

10. An adaptive control system for unmanned traffic nodes based on edge computing, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the edge computing-based adaptive control method for unmanned traffic nodes according to any one of claims 1-9.

Images