Intelligent traffic signal edge computing resource dynamic allocation method and system

Abstract

The invention relates to the technical field of intelligent traffic, in particular to an intelligent traffic signal edge computing resource dynamic allocation method and system.The method comprises the steps that firstly, real-time and historical traffic data of multiple intersections are obtained, a traffic prediction model is constructed, and future flow changes are pre-judged; thirdly, forming an edge computing resource demand matrix according to a prediction result and computing resource demands, and dynamically adjusting resource allocation to generate a resource slicing scheme; then, the flow gradient between intersections is analyzed, an edge node cooperative calculation model is established, and calculation task distribution and migration are optimized; and finally, in combination with system operation and traffic control effects, forming closed-loop feedback, continuously optimizing resource allocation and a cooperative strategy, and through traffic flow prediction driven resource demand modeling, realizing active pre-allocation of computing resources, converting resource allocation from passive response to active pre-adaptation, and improving the system resource utilization rate by 40%-60%.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to an intelligent traffic signal edge computing resource dynamic allocation method and system thereof, which is applied to the field of urban traffic management and control, and solves the problem of dynamic resource allocation of traffic signal control systems in an edge computing environment. Background Art

[0002] With the acceleration of urbanization, urban traffic volume continues to grow. Traditional fixed or simple adaptive traffic signal control systems can no longer meet the complex and changing traffic needs. In recent years, intelligent traffic signal control systems based on edge computing have received widespread attention. By deploying edge computing devices at traffic intersections, real-time processing of traffic data and intelligent control of traffic lights can be achieved, effectively alleviating traffic congestion.

[0003] However, existing edge computing traffic signal control systems generally have problems such as unreasonable resource allocation and untimely system response. Taking the edge computing traffic signal control system and control method disclosed in Chinese patent CN109697866B as an example, the patent constructs a three-layer architecture including information acquisition module, information processing module and information optimization module, collects traffic data through radar, and uses Q learning algorithm to optimize signals. Although the system has improved the intelligent level of traffic signal control to a certain extent, it still has the following shortcomings in practical applications: 1. Edge computing resources are statically allocated and cannot be flexibly adjusted according to the dynamic changes in traffic flow, resulting in insufficient computing resources during peak traffic periods and waste of resources during off-peak periods; 2. There is a lack of direct collaboration mechanism between edge nodes. All data needs to be transferred through the cloud center, which increases the system response delay and cannot achieve rapid collaborative control of adjacent intersections. 3. Only a fixed Q-learning algorithm is used, which is not adaptable enough to cope with diverse traffic scenarios; 4. The elastic recovery mechanism in case of system failure is not considered, which reduces the system reliability; 5. The task division between edge and cloud is fixed, and the distribution of computing tasks cannot be dynamically adjusted according to network conditions and computing load.

[0004] In fact, in the complex and ever-changing urban traffic environment, traffic flow shows obvious temporal and spatial variation characteristics, and the computing requirements at different intersections and at different times vary significantly. Static resource allocation and fixed algorithm strategies are difficult to meet the requirements of intelligent transportation systems for real-time, reliability and flexibility, which greatly limits the application effect of edge computing in the transportation field.

[0005] Therefore, it is urgent to develop an intelligent traffic signal edge computing resource allocation method and system that can dynamically allocate computing resources according to traffic flow predictions, realize collaborative computing between edge nodes, and have the ability to dynamically schedule tasks, so as to improve the intelligence level of traffic management and the efficiency of system operation. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for dynamic allocation of edge computing resources for intelligent traffic signals, aiming to overcome the above-mentioned shortcomings in the prior art, realize dynamic allocation of edge computing resources based on traffic flow prediction, collaborative computing of edge nodes, and flexible scheduling of computing tasks, so as to improve the response speed, resource utilization and system reliability of the traffic signal control system.

[0007] The present invention proposes a method for dynamically allocating intelligent traffic signal edge computing resources, including: acquiring real-time traffic status data and historical traffic data of multiple traffic intersections; based on the real-time traffic status data and historical traffic data, predicting the traffic flow change trend in the future time window and generating a traffic prediction model; According to the traffic prediction model, a resource demand mapping relationship is calculated to determine an edge computing resource demand matrix; Based on the edge computing resource demand matrix, dynamically adjust the resource allocation strategy of the edge computing resource pool and generate a resource slicing solution; According to the resource slicing scheme, determine the traffic flow gradient relationship between intersections and establish an edge node collaborative computing model; Based on the edge node collaborative computing model, the execution location and migration strategy of computing tasks are dynamically determined to achieve optimal allocation of computing tasks; According to the system operation status and traffic control effect, a closed-loop feedback is formed to dynamically optimize resource allocation parameters and collaborative computing strategies.

[0008] Preferably, the acquiring of real-time traffic status data and historical traffic data of multiple traffic intersections specifically includes: Collecting real-time traffic status data through multimodal sensors deployed at traffic intersections, wherein the multimodal sensors include radar sensors, video sensors, and acoustic sensors; Preprocessing the data collected by the multimodal sensor, including data cleaning, outlier detection and data alignment; Extracting historical traffic data of the traffic intersection from a traffic management database; The preprocessed real-time traffic status data is fused with the historical traffic data to generate a standardized traffic status vector.

[0009] Preferably, based on the real-time traffic status data and historical traffic data, predicting the traffic flow change trend in the future time window and generating the traffic prediction model specifically includes: Construct a multi-dimensional traffic state vector, including parameters such as vehicle speed, vehicle density and queue length; Introduce a time series feature extraction matrix, which includes time feature factors such as weekdays / weekends, weather conditions, and special events; Use deep learning models to analyze the temporal changes of traffic state vectors and predict traffic conditions in the next 5 to 30 minutes; The traffic status change rate is calculated based on the prediction results to determine the future traffic flow trend.

[0010] Preferably, according to the traffic prediction model, calculating the resource demand mapping relationship, and determining the edge computing resource demand matrix specifically includes: Establish a mapping function between traffic status and computing resource requirements, and convert traffic status parameters into CPU, memory, network bandwidth and storage resource requirements; Determine the resource demand elasticity coefficient according to the complexity of different traffic scenarios; Calculate the resource demand change curve of each intersection in the future time window; Generate a resource demand prediction matrix including the time dimension to describe the resource demand distribution of each intersection node in the future period.

[0011] Preferably, based on the edge computing resource demand matrix, dynamically adjusting the resource allocation strategy of the edge computing resource pool and generating a resource slicing scheme specifically include: Build an edge computing resource pool virtualization layer to abstract physical resources into a dynamically allocated resource pool; Design a resource pool description structure, including node identification, total resources, and available resources; Define resource allocation utility function and evaluate the efficiency of different resource allocation schemes; Based on the principle of maximizing resource utility, calculate the optimal resource slicing strategy; Generate resource scheduling instruction set and perform dynamic resource slice allocation.

[0012] Preferably, according to the resource slicing scheme, determining the traffic flow gradient relationship between intersections and establishing an edge node collaborative computing model specifically includes: Calculate the traffic flow change rate between adjacent intersections and generate a traffic flow gradient matrix; Based on the traffic flow gradient matrix, identify intersection groups with high coordination needs; Construct edge node affinity matrix to quantify the collaboration efficiency between nodes; Design collaborative computing task description structures, including task priorities, resource requirements, and deadlines; Establish a dynamic collaborative computing graph and determine the set of collaborative edge nodes and their communication relationships.

[0013] Preferably, based on the edge node collaborative computing model, dynamically determining the execution location and migration strategy of computing tasks to achieve optimal allocation of computing tasks specifically includes: Define a task complexity evaluation vector to quantify the computational intensity, memory requirements, and timeliness of the task; Monitor the load status of edge nodes in real time and generate a load status matrix; Build a task migration decision model to determine the optimal execution location of the task based on task complexity, node load and coordination relationship; Establish low-latency point-to-point communication channels to support data transmission during task migration; Execute task offloading instructions to dynamically allocate computing tasks to local execution, neighboring nodes, or the cloud.

[0014] Preferably, the closed-loop feedback is formed according to the system operation status and traffic control effect, and the resource allocation parameters and collaborative computing strategies are dynamically optimized, specifically including: Design system performance indicator structure, including response time, resource utilization, energy consumption and signal control efficiency; Construct the objective function of global and local balance to evaluate the overall effectiveness of the system; Introduce adaptive weight vector to dynamically adjust the weight of each target according to traffic conditions; Based on the performance evaluation results, optimize resource allocation parameters, coordination scope and task priority; Generate a new round of system tuning instructions to achieve a closed-loop feedback of prediction - allocation - coordination - scheduling - evaluation - optimization.

[0015] Preferably, the method further comprises: Monitor the health status of edge nodes and build a distributed health monitoring network; Predict potential failure risks based on node performance abnormality indicators; When a node failure or performance anomaly is detected, the function migration mechanism is automatically triggered to migrate critical tasks to healthy nodes; Implement flexible downgrade operation strategies based on available resources to ensure that core functions are executed first; Record fault information and recovery process to form a system robustness knowledge base.

[0016] The intelligent traffic signal edge computing resource dynamic allocation system for executing the method includes: Multimodal data acquisition module, used to obtain real-time traffic status data and historical traffic data of multiple traffic intersections; A traffic flow prediction module, used to predict the traffic flow change trend in the future time window based on the real-time traffic status data and historical traffic data, and generate a traffic prediction model; A resource demand mapping module, used to calculate the resource demand mapping relationship and determine the edge computing resource demand matrix according to the traffic prediction model; A resource dynamic allocation module, used to dynamically adjust the resource allocation strategy of the edge computing resource pool based on the edge computing resource demand matrix and generate a resource slicing solution; An edge collaborative computing module, used to determine the traffic flow gradient relationship between intersections according to the resource slicing scheme and establish an edge node collaborative computing model; A task scheduling execution module is used to dynamically determine the execution location and migration strategy of computing tasks based on the edge node collaborative computing model to achieve optimal allocation of computing tasks; The closed-loop feedback optimization module is used to form a closed-loop feedback according to the system operation status and traffic control effect, and dynamically optimize resource allocation parameters and collaborative computing strategies.

[0017] The beneficial effects of the present invention include: 1. Through the resource demand modeling driven by traffic flow prediction, active pre-allocation of computing resources is realized, transforming resource allocation from passive response to active pre-adaptation, and improving system resource utilization by 40% to 60%; 2. Constructed a multi-granularity resource pool dynamic slicing technology, which enables computing resources to flow flexibly according to traffic demand, significantly improving the resource allocation efficiency of the system; 3. The edge collaborative calculation mechanism based on traffic gradient realizes direct collaboration between adjacent intersections, reducing system response time by 50% to 70%, especially in the case of edge node collaboration; 4. The dynamic offloading and migration decision engine of computing tasks enables the system to flexibly adjust the task execution location according to the load conditions, improving computing efficiency by 25% to 35%; 5. The edge-cloud collaborative closed-loop feedback optimization system achieves continuous optimization of system parameters, improving overall traffic efficiency by 15% to 25%; 6. The introduction of fault prediction and automatic recovery mechanism has shortened the system fault recovery time by more than 80%, greatly improving system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is an overall architecture diagram of the intelligent traffic signal edge computing resource dynamic allocation system provided by an embodiment of the present invention; Figure 2It is a flow chart of a method for dynamically allocating edge computing resources for intelligent traffic signals provided by an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a traffic flow prediction module provided by an embodiment of the present invention; Figure 4 is a workflow diagram of a resource dynamic allocation module provided by an embodiment of the present invention; Figure 5 It is a performance comparison chart between the embodiment of the present invention and the prior art. DETAILED DESCRIPTION

[0019] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict. For the sake of clarity, the following content omits the detailed description of the known technology.

[0020] Reference Figure 1 The intelligent traffic signal edge computing resource dynamic allocation system provided by the present invention includes: a multimodal data acquisition module 11, a traffic flow prediction module 12, a resource demand mapping module 13, a resource dynamic allocation module 14, an edge collaborative computing module 15, a task scheduling execution module 16 and a closed-loop feedback optimization module 17. These modules work together to form a complete intelligent traffic signal control system.

[0021] Reference Figure 2 The method for dynamic allocation of intelligent traffic signal edge computing resources provided by the present invention includes the following steps: obtaining real-time traffic status data and historical traffic data of multiple traffic intersections; based on the real-time traffic status data and historical traffic data, predicting the traffic flow change trend in the future time window, and generating a traffic prediction model; according to the traffic prediction model, calculating the resource demand mapping relationship, and determining the edge computing resource demand matrix; based on the edge computing resource demand matrix, dynamically adjusting the resource allocation strategy of the edge computing resource pool, and generating a resource slicing scheme; according to the resource slicing scheme, determining the traffic flow gradient relationship between intersections, and establishing an edge node collaborative computing model; based on the edge node collaborative computing model, dynamically determining the execution location and migration strategy of the computing task, and realizing the optimal allocation of the computing task; according to the system operation status and traffic control effect, forming a closed-loop feedback, and dynamically optimizing the resource allocation parameters and collaborative computing strategy.

[0022] Embodiment 1 First, the implementation method of obtaining real-time traffic status data and historical traffic data of multiple traffic intersections is explained in detail.

[0023] In a preferred embodiment of the present invention, the multimodal data acquisition module 11 collects real-time traffic status data through multimodal sensors deployed at traffic intersections, and the multimodal sensors include radar sensors, video sensors and acoustic sensors. Preferably, the radar sensor operates in the 24GHz frequency band, has a detection range of 150 meters, and an accuracy of ±0.5 meters, and can accurately obtain vehicle position, speed and quantity information; the video sensor uses a 4K resolution camera, supports edge image processing, and can identify vehicle models, license plates and traffic events; the acoustic sensor is used to detect the sound characteristics of emergency vehicles and assist in judging special traffic conditions.

[0024] The data acquisition module 11 performs preprocessing on the collected multimodal data, including data cleaning, outlier detection and data alignment. Specifically, the data cleaning process uses a sliding window median filter algorithm with a window size of 5 time units (usually 5 seconds), which can effectively filter sensor noise; outlier detection uses a statistical method to identify data as outliers when they deviate from the mean by more than 3 times the standard deviation, and replace them with adjacent valid values; data alignment uses a timestamp matching method to ensure the timing consistency of different sensor data, and the alignment accuracy is better than 50 milliseconds.

[0025] At the same time, the system extracts the historical traffic data of the intersection from the traffic management database, which usually includes the traffic flow records of the past 6 months and is stored at a granularity of 15 minutes. The historical data is classified by weekdays / weekends / holidays, and marked with influencing factors such as weather conditions and special events.

[0026] In addition, the data acquisition module 11 fuses the preprocessed real-time traffic status data with the historical traffic data to generate a standardized traffic status vector. In addition, the data acquisition module 11 fuses the preprocessed real-time traffic status data with the historical traffic data to generate a standardized traffic status vector. Preferably, the traffic status vector is defined as: , in, Indicates The average speed of each lane (unit: km / h), Indicates The vehicle density of each lane (unit: vehicles / 100m), Indicates The queue length of each lane (unit: vehicle). In actual applications, a four-way intersection usually contains 12 lanes, so A typical value is 12.

[0027] Embodiment 2 Next, the implementation method of predicting the traffic flow change trend in the future time window and generating a traffic prediction model based on real-time traffic status data and historical traffic data is explained in detail.

[0028] In one embodiment of the present invention, the traffic flow prediction module 12 constructs a multi-dimensional traffic state vector, including parameters such as vehicle speed, vehicle density, and queue length. At the same time, a time series feature extraction matrix F is introduced, including time feature factors such as weekdays / weekends, weather conditions, and special events: , in, Indicates the date type (weekday=1, weekend=2, holiday=3), Indicates time period (morning peak=1, off-peak=2, evening peak=3, night=4), Indicates weather conditions (sunny=1, cloudy=2, rainy=3, snowy=4 to Indicates other time characteristic factors, such as special events, etc. In practical applications, The typical value of is 8, which covers the main time characteristics that affect traffic flow.

[0029] Preferably, the traffic flow prediction module 12 uses a deep learning model to analyze the temporal variation of the traffic state vector and predict the traffic state in the next 5 to 30 minutes. Specifically, a prediction model combining a long short-term memory network (LSTM) and an attention mechanism is used, and its structure is as follows: Figure 3 As shown in Figure 2, the model consists of an input layer, two LSTM hidden layers (each layer contains 128 neurons), an attention layer, and an output layer. The model input is the traffic state vector sequence of the past 60 minutes (5 minutes interval, a total of 12 time points), and the output is the predicted traffic state of the next 6 time points (5 minutes, 10 minutes, 15 minutes, 20 minutes, 25 minutes, and 30 minutes).

[0030] The calculation formula of the attention mechanism is: , , , in, is the hidden state of the encoder, is the previous state of the decoder, , and is a learnable parameter is the bias vector, is the attention weight, is the context vector. Based on the prediction results, the traffic flow prediction module 12 calculates the traffic state change rate and determines the future traffic flow trend. Traffic state change rate Defined as: , in, represents the traffic state vector at time t, represents the predicted traffic state vector at time t+Δt, is the prediction time interval.

[0031] In actual applications, the prediction model is trained using historical data sets, using the first 90% of the data for training and the last 10% for validation. The preferred training parameters include: batch size of 64, learning rate of 0.001, training rounds of 200, and early stopping strategy of stopping training if the validation loss does not decrease for 10 consecutive rounds. In this way, the model achieved performance indicators of 5-minute prediction average error of less than 7%, 15-minute prediction average error of less than 12%, and 30-minute prediction average error of less than 18% in actual tests.

[0032] Embodiment 3 The following describes in detail how to implement the edge computing resource demand matrix based on the traffic prediction model and the computing resource demand mapping relationship.

[0033] In a preferred embodiment of the present invention, the resource demand mapping module 13 establishes a mapping function between traffic status and computing resource requirements, and converts traffic status parameters into CPU, memory, network bandwidth and storage resource requirements. Specifically, the resource demand mapping function R (ST, F) is defined as: , Among them, CPU represents the processor requirement (unit: number of cores), MEM represents the memory requirement (unit: GB), NET represents the network bandwidth requirement (unit: Mbps), and STOR represents the storage requirement (unit: GB).

[0034] Preferably, the CPU resource requirement calculation formula is: , in, is the CPU resource coefficient, with a typical value of 0.02 cores / (km / h·vehicle / 100m), It is an adjustment factor based on time characteristics, and its value range is [0.8,1.5]. For example, during the morning and evening peak hours A value of 1.5 indicates that more CPU resources are needed to handle complex traffic conditions. Similarly, the formula for calculating memory resource requirements is: , in, is the memory resource coefficient, with a typical value of 0.05GB / vehicle. It is an adjustment factor based on time characteristics, and its value range is [0.9,1.3].

[0035] The network bandwidth requirement calculation formula is: , in, is the network bandwidth coefficient, the typical value is 0.2Mbps / (vehicle / 100m), It is an adjustment factor based on time characteristics, and its value range is [0.7,1.4].

[0036] The storage requirement calculation formula is: , in, is the storage coefficient, with a typical value of 0.01GB / (vehicle / 100m·hour), It is an adjustment factor based on time characteristics, with a value range of [0.9, 1.2], It is the data retention period (unit: hours), the typical value is 24 hours.

[0037] According to the complexity of different traffic scenarios, the resource demand mapping module 13 determines the resource demand elasticity coefficient Preferably, the elasticity coefficient is defined according to the traffic flow change rate as: , in, is the fluctuation sensitivity coefficient, with a typical value of 0.5. Indicates the relative rate of change of traffic conditions. When traffic conditions change dramatically, As the value increases, the system reserves more resources to cope with traffic fluctuations.

[0038] The resource demand mapping module 13 calculates the resource demand change curve of each intersection in the future time window. , the resource requirement is calculated as: , Finally, the resource demand mapping module 13 generates a resource demand prediction matrix including the time dimension , describing the resource demand distribution of each intersection node in the future period: , in, Represents the prediction time point, with typical values ​​of 5 minutes, 10 minutes, 15 minutes, 20 minutes, 25 minutes and 30 minutes, that is .

[0039] In actual applications, the system updates the resource demand forecast matrix every 3 minutes to ensure that the resource allocation strategy can respond to changes in traffic conditions in a timely manner. When the predicted resource demand change exceeds 20% of the current allocated resources, the resource reallocation process is triggered.

[0040] Embodiment 4 The following describes in detail how to dynamically adjust the resource allocation strategy of the edge computing resource pool based on the edge computing resource demand matrix and generate a resource slicing solution.

[0041] In one embodiment of the present invention, the resource dynamic allocation module 14 first constructs an edge computing resource pool virtualization layer to abstract physical resources into a dynamically allocated resource pool. Figure 4 As shown in the figure, the resource pool virtualization layer abstracts the physical computing resources (CPU, memory, network, and storage) of each edge node to form a unified virtual resource pool, which facilitates flexible allocation of the system.

[0042] Preferably, the resource dynamic allocation module 14 designs a resource pool description structure RP, which includes a node identifier, a total resource amount, and an available resource amount: , Among them, NodeID is the unique identifier of the edge node, CPUTotal and CPUAvail represent the total number of CPU cores and the number of available CPU cores respectively, MEMTotal and MEMAvail represent the total memory and the available memory (unit: GB) respectively, NETTotal and NETAvail represent the total network bandwidth and the available network bandwidth (unit: Mbps) respectively, STORTotal and STORAvail represent the total storage space and the available storage space (unit: GB) respectively.

[0043] The resource dynamic allocation module 14 defines a resource allocation utility function U(RP, S, RM) to evaluate the efficiency of different resource allocation schemes: , in, is the resource slicing strategy matrix, represents the resource utilization utility, represents the resource matching utility, Represents the load balancing utility, , and is the weight coefficient, and .

[0044] In practice, the typical weights are set to Preferably, the resource utilization utility The calculation formula is: , in, , , and Represents the slicing strategy The amount of CPU, memory, network, and storage resources allocated.

[0045] Resource matching utility The calculation formula is: , in, Indicates the amount of resources allocated by the slice strategy. Represents the predicted resource demand in the resource demand matrix. Load balancing utility The calculation formula is: , in, is the number of edge nodes, Is a node The load ratio (the proportion of allocated resources to total resources), is the average load rate of all nodes. Based on the principle of maximizing resource utility, the resource dynamic allocation module 14 solves the optimal resource slicing strategy through genetic algorithm. , The key parameter settings of the genetic algorithm include: population size is 100, chromosome encoding uses real number encoding, selection operation uses tournament selection, crossover probability is 0.8, mutation probability is 0.1, and the number of iterations is 100.

[0046] Finally, the resource dynamic allocation module 14 generates a resource scheduling instruction set and executes dynamic resource slice allocation.

[0047] Preferably, resource scheduling adopts a two-phase commit protocol to ensure the atomicity and consistency of resource allocation. In the first phase, the system sends a resource reservation request to all relevant nodes; in the second phase, after confirming that all nodes can meet the request, a commit instruction is sent to formally execute resource allocation.

[0048] In addition, to prevent system instability caused by frequent resource adjustments, a resource allocation cooldown period (typically 5 minutes) and a resource change threshold (reallocation is triggered only when resource demand changes by more than 20%) are set. In actual tests, the dynamic resource allocation mechanism increased system resource utilization from 43% to 87%, significantly improving the system's resource utilization efficiency.

[0049] Embodiment 5 The following describes in detail how to determine the traffic flow gradient relationship between intersections and establish an edge node collaborative computing model based on the resource slicing solution.

[0050] In a preferred embodiment of the present invention, the edge collaborative computing module 15 first calculates the traffic flow change rate between adjacent intersections to generate a traffic flow gradient matrix G: , in, Indicates The traffic flow gradient between an intersection and its adjacent intersections is calculated as: , in, Indicates from the intersection Outgoing traffic volume (unit: vehicles / hour), Indicates the inflow intersection Traffic flow (unit: vehicles / hour), Represents the distance between adjacent intersections (unit: km). Based on the traffic flow gradient matrix , the edge collaborative computing module 15 identifies the intersection group with high collaborative demand. Preferably, when the absolute value of the traffic flow gradient between two adjacent intersections is Exceeding the threshold (Typical value is 100 vehicles / hour·km), it is determined that the two intersections have coordination demand. The system builds a coordination demand map , where the vertex set Represents the intersection edge set The edge collaborative computing module 15 constructs the edge node affinity matrix , quantify the coordination efficiency between nodes: , in, represents the traffic correlation, Indicates physical distance affinity, represents resource complementarity, , and is the weight coefficient, and In practice, the typical weight setting is .

[0051] Traffic relevance The calculation formula is: , in, Is a node and nodes The physical distance between and are the minimum and maximum collaboration distance thresholds, and their typical values ​​are and 2km.

[0052] Resource complementarity The calculation formula is: , in, and Respectively represent nodes Available and total resources.

[0053] Edge collaborative computing module 15 designs collaborative computing task description structure , Among them, TaskID is the unique identifier of the task, Priority is the task priority (1-10, 10 is the highest), ResourceReq is the resource requirement description, Deadline is the task deadline, and Dependencies is the task dependency list.

[0054] Preferably, collaborative computing tasks are divided into the following types: Traffic status sharing task: priority 8, low resource requirements, strict deadline (<100ms); Signal coordination computing task: priority 9, medium resource requirements, medium deadline (<500ms); Traffic prediction task: priority 7, high resource requirements, loose deadline (<2s); Resource coordination tasks: priority 6, low resource requirements, medium deadline (<1s); Finally, the edge collaborative computing module 15 establishes a dynamic collaborative computing graph to determine the collaborative edge node set and its communication relationship. The collaborative computing graph is a weighted directed graph, where nodes represent edge computing nodes, edges represent collaborative relationships, and edge weights represent collaborative strength (from affinity matrix A). The system establishes a point-to-point (P2P) low-latency communication channel based on the collaborative computing graph to support direct data exchange between edge nodes.

[0055] In actual implementation, the P2P communication channel is implemented based on the UDP protocol, and forward error correction (FEC) technology is used to improve reliability, and the communication delay is controlled within 10 milliseconds. To optimize communication efficiency, the system uses incremental update and data compression technology to reduce the amount of transmitted data. The typical bandwidth requirement is 2-5Mbps for each P2P link.

[0056] The coordination range in the system is dynamically adjusted according to traffic conditions. In congested conditions, the coordination range is expanded to cover up to 10 intersections within 2 kilometers; in unblocked conditions, the coordination range is reduced, usually only including 2 to 3 adjacent intersections. This dynamic coordination mechanism enables the system to reduce unnecessary communication overhead while maintaining control effects.

[0057] Embodiment 6 The following describes in detail how to dynamically determine the execution location and migration strategy of computing tasks based on the edge node collaborative computing model to achieve optimal allocation of computing tasks.

[0058] In a preferred embodiment of the present invention, the task scheduling execution module 16 first defines a task complexity evaluation vector C to quantify the computational intensity, memory requirement, and timeliness of the task: , in, Indicates the computational complexity (1-10, 10 is the highest), Indicates the memory requirement (unit: MB), Indicates the network transmission volume (unit: KB), Indicates the timeliness requirement (unit: ms).

[0059] Preferably, the system predefines complexity templates for different types of tasks, for example: Traffic data processing tasks: ; Signal optimization calculation tasks: ; Traffic prediction tasks: ; Resource scheduling tasks: ; The task scheduling execution module 16 monitors the load status of the edge nodes in real time and generates a load status matrix The load status of each node Contains the following indicators: , in, Indicates CPU utilization (0% to 100%). Indicates memory utilization (0% to 100%). Indicates network bandwidth utilization (0% to 100%). Indicates the length of the task queue. The system updates the load status matrix once a second to ensure that task scheduling decisions are based on the latest load conditions.

[0060] Task scheduling execution module 16 builds a task migration decision model , determine the optimal execution location of the task based on task complexity, node load and coordination relationship: , Among them, 0 means local execution, 1 means offloading to a neighboring node, and 2 means offloading to the cloud. The decision process takes into account multiple factors such as task characteristics, resource availability, node affinity, and network conditions.

[0061] Preferably, the task migration decision adopts a hierarchical strategy, first determining whether the task is suitable for local execution. (Typical value is 70%) and the task complexity is not high When , choose local execution; otherwise, go to the next level of decision-making.

[0062] In the second level decision, the system evaluates the feasibility of offloading tasks to neighboring nodes. Calculate the offloading utility function : , in, is the local node, is the target node, , and is the weight coefficient. When the maximum unloading utility Exceeding the threshold (typical value is 0.6), choose to offload the task to the corresponding adjacent node; otherwise, enter the third level decision.

[0063] In the third level of decision making, the system evaluates the benefits of offloading tasks to the cloud. Or when both the local and neighboring nodes are highly loaded (load rates are both over 80%), choose to offload the task to the cloud for execution.

[0064] The task scheduling execution module 16 establishes a low-latency point-to-point communication channel to support data transmission during task migration. The communication channel is implemented based on the QUIC protocol, which has the characteristics of low latency and high reliability and is suitable for edge computing environments. To improve transmission efficiency, the system compresses and differentially transmits the migration data to reduce network overhead.

[0065] Finally, the task scheduling execution module 16 executes the task unloading instruction and dynamically allocates the computing task to the optimal execution position.

[0066] In practical applications, the dynamic task scheduling mechanism significantly improves the computing efficiency of the system. Test results show that compared with fixed task allocation, dynamic scheduling reduces the average system response time by 68% and increases computing resource utilization by 52%, especially during peak traffic hours.

[0067] Embodiment 7 The following describes in detail how to form a closed-loop feedback loop and dynamically optimize resource allocation parameters and collaborative computing strategies based on the system operating status and traffic control effects.

[0068] In a preferred embodiment of the present invention, the closed-loop feedback optimization module 17 designs a system performance index structure P, including response time, resource utilization, energy consumption and signal control efficiency: , Among them, ResponseTime represents the system response time (unit: ms), and ResourceUtil represents the resource utilization , PowerConsumption represents energy consumption (unit: W), SignalEfficiency represents signal control efficiency, which is evaluated by traffic flow improvement rate. Closed-loop feedback optimization module 17 constructs the objective function of global and local balance , evaluate the overall system performance: , in, represents the global objective function, represents the local objective function, and is the weight coefficient, and .

[0069] Global objective function Defined as: , in, is the maximum acceptable response time (typical value is 1000ms), is the optimal resource utilization (typical value is 85%), is the maximum energy consumption benchmark value, and is the weight coefficient, and . Local objective function Defined as: , Among them, LocalResponseTime represents the local response time performance, LocalResourceBalance represents the local resource balance, and LocalTaskCompletion represents the local task completion rate. , and is the weight coefficient, and .

[0070] Closed-loop feedback optimization module 17 introduces adaptive weight vector , dynamically adjust the weight of each target according to the traffic conditions. Preferably, during peak traffic hours, the system increases the weight of global targets and signal efficiency; during low traffic hours, the system increases the weight of energy consumption and local response time. The weight adjustment follows the following rules: When traffic volume exceeds 80% of capacity, set (Global priority); When the traffic flow is between 50% and 80% of capacity, set (balance global and local); when traffic flow is less than 50% of capacity, set (local priority); The closed-loop feedback optimization module 17 optimizes resource allocation parameters, coordination scope and task priority based on the performance evaluation results. The optimization process uses the Bayesian optimization algorithm to strike a balance between exploration and utilization and quickly converge to the optimal parameter settings.

[0071] The key steps of the Bayesian optimization algorithm include: building a Gaussian process model of the relationship between system performance and parameters; selecting the next set of parameters to evaluate based on the acquisition function (such as expected improvement); updating the model and iterating until convergence or reaching the maximum number of iterations; The optimization parameters include: Resource allocation parameters: resource demand elasticity coefficient , resource allocation utility function weight , , ; Collaborative calculation parameters: Collaborative threshold , affinity matrix weights , , ;Task scheduling parameters: local execution threshold , uninstall utility threshold ; Finally, the closed-loop feedback optimization module 17 generates a new round of system tuning instructions to achieve a closed-loop feedback of prediction-allocation-coordination-dispatching-evaluation-optimization. The tuning instructions are sent to each functional module through the message queue to update the system parameters. The system performs global optimization every 30 minutes and retains multiple sets of parameter configurations to cope with different traffic scenarios.

[0072] In actual tests, the closed-loop feedback optimization mechanism significantly improved the overall performance of the system. As the operating time increased, the system's adaptability to traffic conditions continued to increase. After a three-month test period, compared with the initial configuration, the system response time was reduced by 24%, resource utilization increased by 18%, energy consumption decreased by 15%, and signal control efficiency increased by 22%.

[0073] Embodiment 8 The implementation of the system fault detection and recovery mechanism is described in detail below.

[0074] In one embodiment of the present invention, the system monitors the health status of edge nodes and builds a distributed health monitoring network. Each edge node 21, 22, ..., 2n sends a heartbeat signal regularly (typically once every 5 seconds), including basic health indicators: CPU utilization, memory usage, storage space, network connection status and temperature. At the same time, a mutual monitoring relationship is formed between nodes, and each node is monitored by at least two other nodes to ensure the reliability of fault detection.

[0075] The system predicts potential failure risks based on abnormal node performance indicators. Anomaly detection uses a combination of statistics and machine learning methods, mainly considering the following indicators: CPU utilization exceeds 95% for more than 2 minutes; memory usage suddenly increases by more than 30%; heartbeat signal delay exceeds 3 times the expected value; storage I / O error rate is higher than the threshold (typical value is 0.1%); system temperature exceeds the warning value (typical value is 75°C); When any indicator is abnormal, the system calculates the fault risk score R: , in, is the number of indicators, is the current indicator value, is the normal threshold, is the critical threshold, is the indicator weight. When the risk score When the warning threshold (typical value is 0.7) is exceeded, the system starts preventive maintenance; when the emergency threshold (typical value is 0.9) is exceeded, the emergency migration process is triggered.

[0076] When a node failure or performance anomaly is detected, the system automatically triggers the function migration mechanism to migrate key tasks to healthy nodes. The migration process is based on task priority, first migrating key tasks such as traffic signal control, followed by data processing and prediction tasks. The migration uses a state preservation and recovery mechanism to ensure task continuity.

[0077] Preferably, the function migration process includes the following steps: identifying a set of key tasks to be migrated; selecting the best target node according to the current collaborative computing graph; creating a task state snapshot, including the operating environment, data context and execution status; transmitting the state snapshot to the target node through a high priority channel; rebuilding the task operating environment at the target node and resuming execution; In actual implementation, the entire migration process is completed within 3 seconds to ensure that traffic control functions are not interrupted. The system maintains a task dependency graph to ensure that related tasks are migrated in the correct order to avoid functional loss.

[0078] According to the available resources, the system implements a flexible downgrade operation strategy to ensure that core functions are executed first. The downgrade strategy includes: reducing the priority of non-critical tasks or suspending their execution; reducing data sampling rate and processing accuracy; reducing the scope of collaborative control and focusing on local optimization; using alternative algorithms with lower computational complexity; When operating at a downgraded level, the system still guarantees basic traffic signal control functions and maintains traffic order, but may reduce control accuracy and optimization effects. When system resources are restored, it will automatically return to normal operation.

[0079] The system records fault information and recovery process to form a system robustness knowledge base. Each fault event record includes: fault type, detection method, impact range, recovery strategy, recovery time and effect evaluation. The knowledge base is used to continuously improve fault detection and recovery strategies and improve the overall robustness of the system.

[0080] In actual tests, by simulating different types of node failures (hardware failure, network interruption, resource exhaustion, etc.), the system demonstrated good resilience. The average fault detection time was 1.2 seconds, the function migration completion time was 2.8 seconds, and the overall service interruption time was controlled within 4 seconds, meeting the reliability requirements of the traffic control system.

[0081] Embodiment 9 The following describes in detail the hardware and software architecture implementation of the intelligent traffic signal edge computing resource dynamic allocation system.

[0082] The system of the present invention includes multiple distributed edge nodes 31, 32, ..., 3n and a cloud center node 40, forming a multi-level computing architecture. The edge nodes are deployed at traffic intersections, directly connected to traffic lights and multimodal sensors; the cloud center node provides global resource scheduling and optimization functions.

[0083] In terms of hardware, each edge node contains the following components: the computing unit adopts an ARMCortex-A76 architecture multi-core processor with a main frequency of 2.2GHz and supports dynamic frequency adjustment; storage system: configured with a hierarchical storage architecture, including cache (4GB) and main storage (16GB); network interface: integrated 5G / WiFi6 dual-mode communication module, supporting low-latency P2P communication; power management: adopts an intelligent power management system to support load-based dynamic power consumption control; sensor interface: provides a standard interface to connect radar, camera and acoustic sensors; the cloud center node adopts a high-performance server cluster, configured with Intel Xeon processor, 128GB memory, 10Gbps network interface, providing powerful computing power to support global optimization tasks.

[0084] In terms of software architecture, the system adopts a microservice architecture, which mainly includes the following layers: operating system layer: based on a customized real-time operating system, it supports resource isolation and dynamic allocation; middleware layer: includes a resource virtualization manager, a distributed computing framework, and a message queue system; application layer: implements core functions such as traffic prediction, resource scheduling, and collaborative computing; the system uses container technology to implement task encapsulation and migration, and each functional module is encapsulated as an independent container for flexible deployment and migration. The size of the container image is controlled within 100MB to ensure fast migration.

[0085] In terms of communication, the system builds a three-layer communication network: Intra-edge communication: based on shared memory and local message queues, with a delay of <1ms; Communication between edge nodes: P2P direct communication based on the QUIC protocol, with a delay of <10ms; Edge-cloud communication: based on secure VPN tunnels, supporting dynamic bandwidth allocation; The system adopts a unified data format and interface standard to ensure seamless integration between modules. JSON format is used for data exchange, and RESTAPI and gRPC are used to implement service calls.

[0086] In terms of security, the system implements multi-level security protection: identity authentication: a two-way authentication mechanism based on PKI; communication encryption: using TLS1.3 protocol to ensure data transmission security; access control: fine-grained role-based access control; intrusion detection: real-time monitoring of network traffic and detection of abnormal behavior; In terms of data management, the system adopts a tiered storage strategy: real-time data: stored in the memory database, supporting millisecond-level access; recent data: stored in the local SSD, retained for 24 hours; historical data: compressed and uploaded to the cloud for long-term analysis In actual deployment, edge nodes are installed in traffic light control cabinets or roadside boxes, connecting traffic light controllers and multimodal sensors through standard interfaces. The typical deployment density is 10-15 nodes per square kilometer, covering major road intersections.

[0087] Embodiment 10 The application effects and performance advantages of the present invention in actual traffic scenarios are described in detail below.

[0088] The intelligent traffic signal edge computing resource dynamic allocation system of the present invention was tested and verified in typical traffic sections of multiple cities. Taking a certain city center area as an example, the test area includes 25 traffic intersections with an average daily traffic volume of about 120,000 vehicles. The test lasted for 3 months and covered different traffic scenarios on weekdays, weekends and holidays.

[0089] In terms of resource utilization efficiency, Figure 5As shown in part a of FIG, the computing resource utilization rate of the system of the present invention reaches 87%, which is 102% higher than the static resource allocation method in Comparative Document 1 (CN109697866B). Through dynamic resource allocation driven by traffic forecasting, the system can perceive changes in traffic demand in advance, actively adjust resource allocation, and greatly improve resource utilization efficiency.

[0090] In terms of system response time, as shown in the attached Figure 5 As shown in part b, the average response time of the system of the present invention is 43ms, which is 77% lower than that of the comparative document 1 (CN109697866B). Especially during peak traffic hours, the improvement in response time is more obvious, ensuring that the system can respond to changes in traffic conditions in real time and adjust the signal timing strategy in a timely manner.

[0091] In terms of edge collaboration efficiency, as shown in the attached Figure 5 As shown in part c of the figure, the edge node coordination delay of the system of the present invention is only 28ms, which is 92% lower than that of the comparative document 1 (CN109697866B). The edge coordination calculation mechanism based on the flow gradient enables adjacent intersections to directly exchange information and coordinate control, greatly improving the efficiency of coordinated control of multiple intersections.

[0092] In terms of task scheduling flexibility, as shown in the attached figure Figure 5 As shown in part d, the task scheduling flexibility score of the system of the present invention reaches 92 points (out of 100 points), which is 100% higher than that of comparative document 1 (CN109697866B). Through the dynamic unloading and migration mechanism of computing tasks, the system can flexibly adjust the task execution location according to the load situation and make full use of distributed computing resources.

[0093] In terms of energy efficiency, Figure 5 As shown in part e, the energy efficiency of the present invention reaches 1.35 MIPS / W, which is 221% higher than that of reference document 1 (CN109697866B). Through intelligent resource management and dynamic power consumption control, the system can significantly reduce energy consumption while ensuring performance.

[0094] In terms of traffic efficiency, as shown in the attached Figure 5 As shown in part f, the system of the present invention reduces the average vehicle travel time in the test area by 38%, which is 111% higher than that of reference document 1 (CN109697866B). Through optimized signal timing strategy and multi-intersection coordinated control, the system significantly improves traffic flow efficiency and reduces congestion.

[0095] In terms of system reliability, during the three-month test period, the system of the present invention successfully handled 37 edge node abnormal events, with an average detection time of 1.2 seconds, a function migration completion time of 2.8 seconds, and a service interruption time controlled within 4 seconds, meeting the high reliability requirements of the traffic control system.

[0096] In terms of system scalability, the system of the present invention supports the rapid access of new nodes, and the access time is shortened from the traditional 4 hours / node to 3 minutes / node, which improves the flexible expansion capability of the system. Through standardized interfaces and automated configuration synchronization mechanisms, the system can be easily expanded to a larger-scale transportation network.

[0097] In summary, the intelligent traffic signal edge computing resource dynamic allocation method and system of the present invention significantly improves the performance and efficiency of the traffic signal control system through innovative traffic prediction-driven resource allocation, multi-granularity resource slicing, edge collaborative computing, task dynamic scheduling and closed-loop feedback optimization technologies, providing strong support for intelligent traffic management.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for dynamically allocating edge computing resources for intelligent traffic signals, characterized in that: include: Obtain real-time traffic status data and historical traffic data for multiple traffic intersections; Based on the real-time traffic status data and historical traffic data, predict the traffic flow change trend in the future time window and generate a traffic prediction model; According to the traffic prediction model, a resource demand mapping relationship is calculated to determine an edge computing resource demand matrix; Based on the edge computing resource demand matrix, dynamically adjust the resource allocation strategy of the edge computing resource pool and generate a resource slicing solution; According to the resource slicing scheme, determine the traffic flow gradient relationship between intersections and establish an edge node collaborative computing model; Based on the edge node collaborative computing model, the execution location and migration strategy of computing tasks are dynamically determined to achieve optimal allocation of computing tasks; According to the system operation status and traffic control effect, a closed-loop feedback is formed to dynamically optimize resource allocation parameters and collaborative computing strategies.

2. The method according to claim 1, characterized in that The acquisition of real-time traffic status data and historical traffic data of multiple traffic intersections specifically includes: Collecting real-time traffic status data through multimodal sensors deployed at traffic intersections, wherein the multimodal sensors include radar sensors, video sensors, and acoustic sensors; Preprocessing the data collected by the multimodal sensor, including data cleaning, outlier detection and data alignment; Extracting historical traffic data of the traffic intersection from a traffic management database; The preprocessed real-time traffic status data is fused with the historical traffic data to generate a standardized traffic status vector.

3. The method according to claim 1, characterized in that Based on the real-time traffic status data and historical traffic data, predicting the traffic flow change trend in the future time window and generating a traffic prediction model specifically includes: Construct a multi-dimensional traffic state vector, including parameters such as vehicle speed, vehicle density and queue length; Introduce a time series feature extraction matrix, which includes time feature factors such as weekdays / weekends, weather conditions, and special events; Use deep learning models to analyze the temporal changes of traffic state vectors and predict traffic conditions in the next 5 to 30 minutes; The traffic status change rate is calculated based on the prediction results to determine the future traffic flow trend.

4. The method according to claim 1, characterized in that According to the traffic prediction model, the resource demand mapping relationship is calculated, and the edge computing resource demand matrix is ​​determined, which specifically includes: Establish a mapping function between traffic status and computing resource requirements, and convert traffic status parameters into CPU, memory, network bandwidth and storage resource requirements; Determine the resource demand elasticity coefficient according to the complexity of different traffic scenarios; Calculate the resource demand change curve of each intersection in the future time window; Generate a resource demand prediction matrix including the time dimension to describe the resource demand distribution of each intersection node in the future period.

5. The method according to claim 1, characterized in that Based on the edge computing resource demand matrix, the resource allocation strategy of the edge computing resource pool is dynamically adjusted to generate a resource slicing solution, specifically including: Build an edge computing resource pool virtualization layer to abstract physical resources into a dynamically allocated resource pool; Design a resource pool description structure, including node identification, total resources, and available resources; Define resource allocation utility function and evaluate the efficiency of different resource allocation schemes; Based on the principle of maximizing resource utility, calculate the optimal resource slicing strategy; Generate resource scheduling instruction set and perform dynamic resource slice allocation.

6. The method according to claim 1, characterized in that According to the resource slicing scheme, the traffic flow gradient relationship between intersections is determined, and the edge node collaborative computing model is established, which specifically includes: Calculate the traffic flow change rate between adjacent intersections and generate a traffic flow gradient matrix; Based on the traffic flow gradient matrix, identify intersection groups with high coordination needs; Construct edge node affinity matrix to quantify the collaboration efficiency between nodes; Design collaborative computing task description structures, including task priorities, resource requirements, and deadlines; Establish a dynamic collaborative computing graph and determine the set of collaborative edge nodes and their communication relationships.

7. The method according to claim 1, characterized in that Based on the edge node collaborative computing model, the execution location and migration strategy of computing tasks are dynamically determined to achieve optimal allocation of computing tasks, specifically including: Define a task complexity evaluation vector to quantify the computational intensity, memory requirements, and timeliness of the task; Monitor the load status of edge nodes in real time and generate a load status matrix; Build a task migration decision model to determine the optimal execution location of the task based on task complexity, node load and coordination relationship; Establish low-latency point-to-point communication channels to support data transmission during task migration; Execute task offloading instructions to dynamically allocate computing tasks to local execution, neighboring nodes, or the cloud.

8. The method according to claim 1, characterized in that The closed-loop feedback formed according to the system operation status and traffic control effect, and the dynamic optimization of resource allocation parameters and collaborative computing strategies specifically include: Design system performance indicator structure, including response time, resource utilization, energy consumption and signal control efficiency; Construct the objective function of global and local balance to evaluate the overall effectiveness of the system; Introduce adaptive weight vector to dynamically adjust the weight of each target according to traffic conditions; Based on the performance evaluation results, optimize resource allocation parameters, coordination scope and task priority; Generate a new round of system tuning instructions to achieve a closed-loop feedback of prediction - allocation - coordination - scheduling - evaluation - optimization.

9. The method according to claim 1, characterized in that: The method further comprises: Monitor the health status of edge nodes and build a distributed health monitoring network; Predict potential failure risks based on node performance abnormality indicators; When a node failure or performance anomaly is detected, the function migration mechanism is automatically triggered to migrate critical tasks to healthy nodes; Implement flexible downgrade operation strategies based on available resources to ensure that core functions are executed first; Record fault information and recovery process to form a system robustness knowledge base.

10. An intelligent traffic signal edge computing resource dynamic allocation system that executes the method according to any one of claims 1 to 9, characterized in that: include: Multimodal data acquisition module, used to obtain real-time traffic status data and historical traffic data of multiple traffic intersections; A traffic flow prediction module, used to predict the traffic flow change trend in the future time window based on the real-time traffic status data and historical traffic data, and generate a traffic prediction model; A resource demand mapping module, used to calculate the resource demand mapping relationship and determine the edge computing resource demand matrix according to the traffic prediction model; A resource dynamic allocation module, used to dynamically adjust the resource allocation strategy of the edge computing resource pool based on the edge computing resource demand matrix and generate a resource slicing solution; An edge collaborative computing module, used to determine the traffic flow gradient relationship between intersections according to the resource slicing scheme and establish an edge node collaborative computing model; A task scheduling execution module is used to dynamically determine the execution location and migration strategy of computing tasks based on the edge node collaborative computing model to achieve optimal allocation of computing tasks; The closed-loop feedback optimization module is used to form a closed-loop feedback according to the system operation status and traffic control effect, and dynamically optimize resource allocation parameters and collaborative computing strategies.

Images