Vehicle and road cloud data processing method and system based on edge computing center

Abstract

The invention relates to the technical field of intelligent traffic, and discloses a vehicle and road cloud data processing method and system based on an edge computing center, and the method comprises the steps: deploying a sensing module, collecting vehicle, road and environment data, and carrying out the standardization and alignment; constructing a three-dimensional tensor, and calculating a local gradient to reflect data aggregation; evaluating a node collaboration demand, and constructing a graph structure and a collaboration weight based on a gradient; constructing an optimization objective function, and balancing delay, resources and task importance scheduling; when the gradient exceeds a threshold value, nodes are selected, and state saving and task migration are executed; when communication is abnormal, autonomous reasoning processing is completed based on a local model and features; the system comprises a sensing module, a tensor modeling module, a resource collaborative evaluation module, a task scheduling module, a task migration module and an autonomous reasoning module. According to the method, the three-dimensional tensor structure is introduced to express the multi-dimensional information of the vehicle, the roadside equipment and the environment sensing terminal, the multi-dimensional information is fused and modeled, and the purpose of automatically extracting the area aggregation situation is achieved.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, specifically to a vehicle-road-cloud data processing method and system based on an edge computing center. Background Art

[0002] Intelligent transportation systems are gradually emerging from the laboratory and beginning to be deployed on a large scale in diverse scenarios, including urban arterial roads, highways, and within industrial parks. Driven by the widespread coverage of 5G communication technology, V2X (Vehicle-to-Everything) is becoming a core supportive tool for vehicle-road collaboration. Through information exchange between onboard units (OVUs), roadside equipment, and the remote cloud, vehicle-road-cloud systems are expected to achieve comprehensive awareness of road conditions, traffic flows, and abnormal events, enabling intelligent responses.

[0003] To improve overall system computing power and data processing depth, existing vehicle-road-cloud data processing technologies often utilize cloud computing centers for centralized processing of collected vehicle-road data. The basic logic is to upload data from vehicles, RSUs, and environmental sensors to a cloud server, where it is computed using a unified model before returning decision-making instructions. This centralized architecture offers the advantages of concentrated computing power and ease of management, making it suitable for scenarios such as complex model training and large-scale data archiving. Consequently, it was widely used in early intelligent transportation pilot projects.

[0004] However, existing vehicle-road-cloud data processing technology and its centralized processing architecture have poor delay control capabilities when faced with high-density, real-time traffic scenarios. Secondly, data needs to be uploaded to the cloud across multiple hops and then returned to the edge for decision execution. The entire process is prone to significant network congestion and response lags. In addition, there is a lack of effective collaborative expression mechanisms between nodes. Each computing unit often operates as an independent entity, making it difficult to dynamically share loads and resulting in inefficient resource scheduling. Therefore, the present invention provides a vehicle-road-cloud data processing method and system based on an edge computing center to address the shortcomings of the existing technology. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of this application is to provide a vehicle-road-cloud data processing method and system based on an edge computing center, which solves the problems of high data transmission delay, slow computing response and lack of local autonomous processing capabilities in the cloud computing architecture of the existing vehicle-road-cloud data processing technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A vehicle-road-cloud data processing method based on an edge computing center includes the following steps: S1. Deploy a perception module in the edge computing center to collect real-time data from vehicles, roadside units, and environmental sensing equipment. Standardize and time-align the collected real-time data to obtain structured pre-processed data. S2. Based on the structured preprocessed data, construct a three-dimensional tensor model with spatial position, time state, and data type dimensions, and calculate the local gradient of the three-dimensional tensor model to characterize the data aggregation trend; S3. Evaluate the resource collaboration requirements between edge computing nodes based on the local gradient of the tensor model. Build a graph of edge nodes based on the node features and gradient information in the tensor model and calculate the collaboration weights between nodes. S4. Based on the collaborative weights and tensor gradient output results, a task scheduling optimization objective function is constructed to calculate latency, resource usage status, and task importance, and allocate resources to tasks within the current edge node. S5. During the scheduling process, when the tensor gradient amplitude of a node exceeds a preset threshold, a target node that meets the migration conditions is selected based on the collaborative weight, and task state preservation, model deployment, and data migration operations are performed; S6. In the event of communication anomalies or node failures, the edge computing center calls the pre-deployed model to complete autonomous reasoning processing based on the locally constructed tensor model and graph feature information.

[0007] Preferably, in step S1, the standardization and time alignment processing of the collected real-time data includes the following steps: The perception module collects data from vehicles, roadside units, and environmental sensing devices in real time and transmits this raw data to the edge computing center; Based on the characteristics of data collected by each device, data from different sources are normalized through standardization algorithms; The collected data is time-aligned based on the timestamp, and the collected data from different devices are time-aligned by synchronizing the clocks.

[0008] Preferably, in step S2, the local gradient of the three-dimensional tensor model is calculated using a partial derivative formula, and the partial derivative formula is: Where, T i,j,k is an element of the tensor model, f(T) is the loss function of the model, is the gradient value of the element, is the symbol of partial derivative.

[0009] Preferably, in step S2, constructing a tensor model with spatial position, temporal state, and data type dimensions comprises the following steps: Based on the preprocessed data, the three data types from vehicles, roadside units, and environmental sensing devices are mapped into different dimensions of a three-dimensional tensor; By building a tensor model, the processed data is filled into the corresponding elements of the tensor, and the number of dimensions of the tensor is selected according to the characteristics of the data; Through the tensor decomposition method, the tensor is reduced in dimension and decomposed to generate a set of low-rank tensor factors.

[0010] Preferably, in step S3, evaluating resource collaboration requirements between edge computing nodes includes the following steps: Based on the constructed tensor model, the gradient descent method is used to derive each element of the tensor to obtain the local gradient, which represents the computational requirements and resource consumption of each node; Based on the calculated local gradient information, the resource consumption and load of each node are evaluated; The collaborative weights between nodes are calculated based on the local gradient values ​​and resource requirements between each node.

[0011] Preferably, the collaborative weight between the computing nodes adopts a weighted calculation formula based on gradient similarity, and the weighted calculation formula based on gradient similarity is: Where W ij is the collaborative weight between edge node i and node j, and are the gradient values ​​of node i and node j respectively.

[0012] Preferably, in step S4, constructing the task scheduling optimization objective function includes the following steps: Based on the collaborative weights and tensor gradient outputs calculated in the task, combined with the latency, resource usage status, and importance of each task, the optimization objective function for task scheduling is constructed: Objective=λ1×Delay+λ2×Resource Usage+λ3×TaskImportance; Where λ1, λ2, and λ3 represent the impact of delay, resource usage, and task importance on the optimization objective, respectively. Delay represents the expected waiting time or deadline deviation of the task on the current node. Resource Usage represents the total amount of resources expected to be occupied by the task on the specified node. TaskImportance is the task priority indicator set by the system. Objective is the original optimization objective function value. By optimizing the objective function, the execution delay and resource consumption of each task on different nodes are calculated to evaluate the importance and priority of the task; Based on the results of the optimization objective function, the task scheduling order and resource allocation strategy are dynamically adjusted.

[0013] Preferably, in step S5, the execution of task state preservation, model deployment and data migration operations includes the following steps: When the tensor gradient amplitude of a node exceeds the preset threshold, the collaborative weight and adaptability between nodes are calculated and task migration is performed; Perform task status preservation operations to ensure that tasks remain consistent during the migration process, and save and transfer task status, data, and model information to the target node; Migrate the computing tasks and data of the task to the target node and update the computing model of the target node.

[0014] Preferably, in step S6, calling the pre-deployed model to complete the autonomous reasoning process includes the following steps: When communication anomalies or node failures occur, the autonomous reasoning mechanism of the edge computing center is activated by using the locally built tensor model and graph feature information; The edge computing center analyzes the node's status data based on the pre-deployed inference model and automatically infers potential failures or data anomalies; Based on the inference results, the edge computing center adjusts the computing tasks and resource allocation of the nodes.

[0015] It also provides a vehicle-road-cloud data processing system based on the edge computing center, including: The perception module collects real-time data from vehicles, roadside units, and environmental sensing devices, and standardizes and time-aligns the collected data based on the device data characteristics to generate structured pre-processed data. The tensor modeling module is used to map spatial location information, temporal state information, and data type information from different sources to corresponding dimensions of a three-dimensional tensor model based on structured preprocessed data, thereby constructing a complete tensor data structure to extract data aggregation trend characteristics. The resource collaboration evaluation module is used to evaluate the computing requirements and resource consumption of each edge node based on the local gradient information output by the tensor modeling module, build a graph structure model of the edge nodes, and calculate the collaboration weight between nodes; The task scheduling module is used to dynamically construct the task scheduling optimization objective function based on the calculated coordination weight, combined with the task delay requirements, current resource usage status and task importance information; The task migration module is used to select a suitable target node based on the collaborative weight analysis results when the computing load of a node exceeds the preset threshold during the execution of the task scheduling module, and perform task status preservation, data transmission and model deployment operations; The autonomous reasoning module is used to call the pre-deployed autonomous reasoning model in the edge computing center based on the model and graph structure established by the tensor modeling module and the resource collaborative evaluation module in scenarios of communication anomalies or node failure anomalies, and automatically infer the node failure type and status anomaly.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention introduces a three-dimensional tensor structure to represent the multidimensional information of vehicles, roadside equipment, and environmental perception terminals. This model integrates spatial location, temporal state, and data type into a unified semantic representation of perception data and automatically extracts regional clustering patterns. Compared to existing approaches that rely on isolated processing of distributed perception data, this approach avoids the fragmentation of heterogeneous data and dimensionality inconsistencies, significantly improving the computability of the data structure and the scalability of subsequent analysis.

[0017] 2. This invention uses a collaborative weighting mechanism based on gradient similarity to measure the intensity of task coupling between edge nodes. This dynamically reflects the correlation of computational loads between nodes and guides task allocation, thereby achieving precise collaboration during task scheduling. Traditional technologies, which are often based on static topology or manual weight configuration, are difficult to adapt to resource regulation in time-varying network states. This makes the scheduling mechanism more closely aligned with the needs of real-time edge computing scenarios.

[0018] 3. This invention constructs an optimization function model that combines task latency, resource consumption, and business importance, quantifying the multi-factor task scheduling problem into a computable optimal solution, eliminating reliance on manual rule configuration. This approach is more flexible than conventional scheduling strategies based on fixed priorities or round-robin mechanisms, avoiding the common problems of uneven resource utilization or long waiting times for low-priority tasks, and providing more controllable and robust scheduling results.

[0019] 4. In the event of a communication anomaly or failure at a node, the edge computing center, without relying on the cloud, leverages locally constructed tensor and graph models to initiate autonomous reasoning, promptly identifying abnormal conditions and reconstructing the task execution path. This mechanism eliminates reliance on central servers and addresses the technical limitations of existing solutions, which often prevent them from maintaining system continuity in the event of an edge network outage. This advantage is particularly evident when deploying safety-critical tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method steps of this application; Figure 2 This is the system architecture diagram of this application. DETAILED DESCRIPTION

[0021] The following is combined with Figure 1 -Attached Figure 2 , further details of this application are given.

[0022] Please see the attached Figure 1 , an embodiment of the present invention provides a vehicle-road-cloud data processing method based on an edge computing center, comprising the following steps: S1. Deploy a perception module in the edge computing center to collect real-time data from vehicles, roadside units, and environmental sensing equipment. Standardize and time-align the collected real-time data to obtain structured pre-processed data. S2. Based on the structured preprocessed data, construct a three-dimensional tensor model with spatial position, time state, and data type dimensions, and calculate the local gradient of the three-dimensional tensor model to characterize the data aggregation trend; S3. Evaluate the resource collaboration requirements between edge computing nodes based on the local gradient of the tensor model. Build a graph of edge nodes based on the node features and gradient information in the tensor model and calculate the collaboration weights between nodes. S4. Based on the collaborative weights and tensor gradient output results, a task scheduling optimization objective function is constructed to calculate latency, resource usage status, and task importance, and allocate resources to tasks within the current edge node. S5. During the scheduling process, when the tensor gradient amplitude of a node exceeds a preset threshold, a target node that meets the migration conditions is selected based on the collaborative weight, and task state preservation, model deployment, and data migration operations are performed; S6. In the event of communication anomalies or node failures, the edge computing center calls the pre-deployed model to complete autonomous reasoning processing based on the locally constructed tensor model and graph feature information.

[0023] Regarding step S1, in this embodiment, when performing vehicle-road collaborative data processing in the edge computing center, in order to ensure the accuracy and consistency of subsequent modeling and analysis, the raw data collected by the perception module must first be standardized and time-aligned. This step serves as the starting stage of data access processing in the vehicle-road-cloud-edge collaborative system and constitutes the basis of the entire processing flow. The effective execution of standardization and time alignment operations directly affects the quality of tensor modeling and the rationality of subsequent resource evaluation and task scheduling strategies.

[0024] Edge computing centers deploy perception modules, which can include but are not limited to cameras, millimeter-wave radars, lidars, V2X communication terminals, environmental sensing devices, and other types of terminals. These modules collect time-stamped data from vehicles, roadside units (RSUs), and the surrounding environment. Generally, these sensors have different data sampling periods, encoding formats, and physical units, necessitating standardization of the collected data.

[0025] Normalization includes unit conversion, data amplitude normalization, and format unification. As an option, linear normalization can be used to map data from different sources to the [0, 1] or [-1, 1] intervals. The following normalization formula is defined: Where, X is the original collected data; X ′ is the normalized result; X min 、X max are the minimum and maximum observation values ​​corresponding to the data source, respectively. The above method can significantly reduce the interference of sampling scale differences between different devices on the accuracy of tensor model construction.

[0026] After the standardization process is completed, time alignment is further performed on the multi-source data. Since heterogeneous sensing devices have asynchronous sampling characteristics, direct parallel use will lead to offset problems in the time dimension. Therefore, the present invention introduces a synchronous clock mechanism to perform timestamp-based alignment.

[0027] Specifically, in one possible implementation, the system has a built-in unified coordinated clock source (such as a GPS clock or a local NTP server), resamples the sampled data of various devices according to a preset time granularity (such as 100ms or 1s), discards some data that exceeds the window range, and performs linear interpolation or sliding average filling operations on the missing data. This process can be expressed as: Where, D is the output data of the i-th device at time point t after time alignment; i (t) is the original sampling value of this type of device; ∈ i (t) is the interpolation error term; N is the total number of acquisition devices; Indicates that the following condition holds for every element in the set; {1,2,…,N} is a set of integers. Alternatively, if the missing rate within a time window exceeds a set threshold, the data for that period will be marked as unavailable to avoid introducing unstable samples.

[0028] To further enhance system stability and flexibility, some implementations can incorporate device tagging and data quality identification mechanisms. Each data packet is assigned a source ID, confidence score, and latency information, and these are processed simultaneously during the normalization and alignment processes. This multi-dimensional tagging mechanism can be embedded as additional features within the tensor structure during subsequent tensor modeling, enhancing the model's expressiveness.

[0029] It's important to note that data standardization and time alignment not only prepare data for modeling but also provide a solid foundation for subsequent tensor local gradient derivation and task scheduling. It's through the standardized execution of this step that we ensure structural consistency across the three dimensions of space, time, and type for data from different sources, laying the foundation for the construction of high-quality three-dimensional tensor models.

[0030] In step S2, in this embodiment, after standardizing and temporally aligning the multi-source sensor data, a tensor representation model with spatial, temporal, and type semantics is further developed for the preprocessed data to construct a data representation structure suitable for edge computing scenarios. This modeling process not only characterizes the data's organization but also provides structured input support for subsequent local trend analysis, collaborative computing, and resource optimization.

[0031] For the standardized and time-aligned structured data, the system projects the perception data onto three different dimensions of a three-dimensional tensor based on the category and physical location of the data source.

[0032] Specifically, in one possible implementation, the first dimension is set as the spatial location index of the data source, corresponding to the vehicle position or the deployment coordinates of the roadside equipment; the second dimension is set as the time state sequence, representing the time process of data collection or the moment number in the sliding window; the third dimension is the data type dimension, which identifies the type information corresponding to the data source, such as speed, temperature, communication quality, image features, etc.

[0033] For example, in some embodiments, the speed of vehicle A at position x at time t is V, and this data will be filled into the corresponding position T in tensor T. ij,k , where i represents the vehicle's index in spatial coordinates, j is the current sampling time index, and k is the data type index corresponding to "speed." After filling the three-dimensional tensor, to compress the model structure and extract significant features, tensor decomposition methods are generally used to reduce the dimensionality of the original high-dimensional tensor. Alternatively, methods such as Tucker decomposition, CP decomposition, or Tensor Principal Component Analysis (Tensor PCA) can be used to map the original tensor into a set of low-rank tensor factors, thereby preserving the dominant characteristics of the data structure and removing noise and redundancy.

[0034] After the tensor model is constructed, the local gradient of each element in the tensor needs to be evaluated to determine the data change trend and regional sensitivity.

[0035] In a preferred implementation, the present invention derives the loss function of the tensor model through a partial derivative formula to quantify the impact of each tensor unit on the overall model structure. The calculation formula is as follows: Where, T i,j,kis an element of the tensor model; f(T) is the loss function of the model; is the gradient value of the element; is the symbol of partial derivative.

[0036] Generally speaking, tensor elements with larger gradient amplitudes often correspond to data mutation points or highly variable regions, and are suitable for identifying edge nodes where resource demand surges.

[0037] As an option, we can combine the correlation constraints between the dimensions within the tensor and introduce a regularization term to control the stability of the gradient calculation process and prevent the abnormal diffusion of local gradients.

[0038] Furthermore, some embodiments can set a gradient threshold to trigger subsequent resource evaluation and migration strategies. For example, if the gradient magnitude of a node exceeds a threshold θ over consecutive time slices, the system can mark the node as a "high-load node to be migrated" for further decision-making by the resource scheduling module.

[0039] Regarding step S3, in this embodiment, after completing the tensor model construction and its local gradient solution, the system further identifies and evaluates the resource consumption and load status between edge computing nodes to achieve coordinated scheduling and optimal resource allocation among multiple nodes. Based on the gradient information of the tensor model, the resource coupling strength between each node is quantitatively calculated and the node graph structure is constructed accordingly.

[0040] First, a gradient descent analysis is performed on the three-dimensional tensor model obtained in the previous stage to obtain the local gradient amplitude of each tensor element. This gradient not only describes the trend of data aggregation or mutation, but also indirectly reflects the resource sensitivity of each node in data processing.

[0041] Specifically, the system maps the sub-block corresponding to each node in the tensor into a computational state description of the edge node, and calculates the overall gradient norm of the node as an indicator of its resource requirements and computational intensity.

[0042] In one possible implementation, the tensor gradient of node i can be expressed as follows based on the tensor differentiation method: Where, Represents the gradient value of the tensor at the i-th spatial index, the j-th time index, and the k-th data type dimension; It represents the overall gradient vector of the area corresponding to the node.

[0043] The system then aggregates the local gradients of all edge nodes and, based on the current resource consumption of each node (including CPU occupancy, memory usage, communication bandwidth, etc.), constructs a load vector for each node: Where R i is the resource state vector of edge node i; It represents the real-time usage level of the lth type of resource at node i.

[0044] Generally, the system performs a weighted combination of the node's gradient norm and resource usage to obtain a node load intensity index to support collaborative scheduling.

[0045] On this basis, the system needs to further evaluate the degree of collaboration between edge nodes. To achieve quantitative representation, this paper proposes a collaborative weight calculation method based on gradient similarity, which is mathematically expressed as follows: Where W ij is the collaborative weight between edge node i and node j; and are the gradient values ​​of node i and node j, respectively. This formula essentially measures the similarity between nodes in data change trends. High similarity usually indicates a strong correlation between tasks between nodes, making it suitable for computing collaboration.

[0046] In some embodiments, in order to improve robustness, a threshold control strategy can be introduced: if the collaborative weight W ij If it is lower than the set threshold δ, it is considered that there is no significant collaborative demand between the nodes and no edge connection is established.

[0047] Alternatively, a time-varying adjustment factor can be introduced to correct the time window sensitivity in the gradient vector to maintain model stability in highly dynamic environments.

[0048] In addition to collaboration weights, the system can also introduce resource load similarity metrics to supplement scenarios where gradient similarity is insufficient, enabling the collaborative graph structure to have multi-dimensional dynamic adjustment capabilities. Ultimately, the system inputs the collaboration weight matrix W and the node load matrix into the graph construction module, outputting an edge node graph structure that forms the basis for generating scheduling functions and migration strategies.

[0049] In step S4, in this embodiment, after calculating the coordination weights between edge nodes and outputting tensor gradients, a task scheduling objective function with quantitative optimization metrics is constructed to achieve dynamic and efficient allocation of various tasks in a multi-node environment. This function, as the core foundation for executing the scheduling strategy, enables task priority assessment and rational resource utilization control based on multi-factor trade-offs.

[0050] In this embodiment, the scheduling module first receives input information from the upstream module, including: the collaborative weight matrix W between each edge node ij , node local gradient information output by the tensor model As well as attribute label data of the task itself, such as deadline, resource consumption estimation and functional level.

[0051] On this basis, the scheduling module integrates three indicators, namely task delay sensitivity, resource consumption, and task importance, into a unified scheduling trade-off model by constructing a linear weighted optimization function. The objective function is defined as follows: Objective = λ1×Delay+λ2×Resource Usage+λ3×TaskImportance; Where λ1, λ2, and λ3 are the weighted coefficients corresponding to the three scheduling target dimensions respectively; Delay represents the expected waiting time or deadline deviation of the task on the current node; Resource Usage represents the total amount of resources (such as CPU cycles, memory, bandwidth, etc.) expected to be occupied by the task on the specified node; TaskImportance is the task priority indicator set by the system, which can be set based on the task source, service level, or real-time requirements; Objective is the original optimization objective function value.

[0052] Generally, the three weighting coefficients need to be initialized and configured according to the scenario, and the system is allowed to adaptively adjust based on performance feedback during operation.

[0053] In one possible implementation, the task scheduling module substitutes the execution status of each candidate task on each available node into the above objective function to form a two-dimensional scheduling score matrix: S mn =λ1D mn +λ2R mn +λ3I m ; Where S mn is the optimization score of task m on node n; D mn is the estimated delay of task m on node n; R mn Valuing resource use; I m is the importance level label of task m; λ1, λ2, and λ3 correspond to the influence weights of delay, resource usage, and task importance in the optimization score, respectively.

[0054] By comparing S mn The system can identify the optimal task-node matching pair and perform the corresponding scheduling operation based on the minimum value in the score matrix.

[0055] In some embodiments, in order to cope with the demand for rapid queue-jumping of sudden tasks or high-priority tasks, a penalty term may be added to the above objective function, for example: Objective'=Objective+θ·I MissDeadline ; Where, Objective' is the optimization target value after adding the penalty term; Objective is the original optimization target function value; θ is the penalty factor; I MissDeadline It is an indicator function that takes 1 if the task is expected to be overdue and 0 otherwise, and is used to suppress scheduling plans that do not meet the time requirements.

[0056] After the scheduling scores are generated, the scheduling module determines the scheduling order in a greedy manner or using heuristic rules (such as minimum value first, adjacency weighted scheduling based on collaborative graphs, etc.).

[0057] Based on the results of the above optimization process, the system can dynamically adjust the task scheduling order and resource allocation strategy to achieve multi-dimensional consideration of resources, timeliness and business goals in complex task scenarios, and effectively support the stability and intelligence of task distribution in the vehicle-road cooperative system.

[0058] Regarding step S5, in this embodiment, after completing the construction of the task scheduling optimization objective function, the system will adjust the task execution location in real time based on the scheduling results. When the computing pressure of an edge node exceeds a preset threshold, or resource consumption exceeds the safety limit, the task migration process must be immediately triggered to ensure the overall system performance and the continued effectiveness of the scheduling strategy. At this time, operations such as task state preservation, model deployment, and data migration must be performed to ensure the consistency of the computing state and task continuity.

[0059] In this embodiment, the edge computing center first monitors the computational load of each node based on the local gradient amplitude of the tensor model. When the tensor gradient norm of a node exceeds a set threshold, the node is considered to have entered a high-load state. At this point, the system automatically calls the constructed collaborative weight graph structure to find the target node with the highest collaborative weight and sufficient available resources.

[0060] After selecting the target node, the system first saves the task state. Specifically, task state data typically includes the task context, intermediate execution variables, progress markers, and pointers to related resource references. This state information must be snapshotted in a structured format and stored in a temporary task buffer in the edge computing center.

[0061] As an option, to ensure task consistency, a consistency check mechanism can be used during state preservation, such as using a task summary check value (Hash) or a concurrent write protection strategy based on atomic operations to prevent write overwrites or data corruption during state transfer.

[0062] Specifically, when task migration is triggered, the system constructs the following task migration data structure: TaskState i → j ={Ci ,D i ,M i}; In the formula, TaskState i → j Represents the complete state data structure when the task is migrated from node i to node j; C i It is the task control state; D i is real-time data related to the task; M i The model weight parameter that the task currently depends on.

[0063] After the above data structure is transmitted to the target node j, the target node will complete the following operations in sequence: Loading mission control status C i , restore the task scheduling context; Initialize dataset D i , ensuring that computational semantics continue correctly; Deployment model structure M i , ensuring the consistency of task operation logic.

[0064] In one possible implementation, task migration can be optimized by combining cold and hot migration mechanisms. If the task is still in the initialization phase, only the task definition and schedule table can be migrated (cold migration). If the task has entered the runtime phase, the intermediate state and model parameters must be fully migrated (hot migration).

[0065] In some embodiments, to improve migration efficiency, a migration-while-execution strategy may be adopted, that is, during the batch transfer of task data, key control parameters are first loaded for preheating and starting, and non-critical auxiliary data are subsequently added.

[0066] After the migration is complete, the target node will synchronously update its local computational graph model and take over the subsequent scheduling and execution of the task, replacing the original node to continue the task process. At the same time, the original node will release occupied resources and enter a cooling waiting state to avoid resource waste.

[0067] Regarding step S6, in this embodiment, after completing task migration and resource reallocation, to enhance the edge computing system's ability to recover from emergencies, it is necessary to establish an autonomous reasoning mechanism independent of external dependencies. This mechanism enables automatic state determination and resource adjustment in abnormal situations such as communication link interruptions or edge node failures. By invoking pre-deployed reasoning models on the edge, lightweight, low-latency fault diagnosis and scheduling optimization are performed locally, ensuring the system's continued resilience and intelligence.

[0068] When the system detects communication anomalies, node disconnection, signal loss, or node heartbeat timeouts, it first triggers the autonomous perception module of the edge computing center. Typically, anomaly triggering conditions are based on link quality metrics at the communication layer (such as Packet Loss Rate and Round Trip Delay) or node activity at the system layer (such as CPU heartbeat interval and state synchronization timeout).

[0069] In a possible implementation, the communication module detects that the target node has not received any return data for a continuous period of Δt, or finds that the link quality indicator η is lower than the set threshold η min , the system will determine that the target node has entered an unstable state and then start the autonomous reasoning process.

[0070] The autonomous reasoning module will call the tensor model and graph feature module previously pre-deployed on the edge node to analyze the structural relationship and data status of the current node set. Specifically, the reasoning model uses the spatial-temporal-type feature information T of the currently available nodes in the tensor model. i,j,k , combined with the edge connection weight W in the node graph structure ij , construct the inference function g(·) to perform node status prediction and anomaly attribution. The inference expression is as follows: Where, The node / current status label inferred by the system (such as normal / abnormal / to be migrated); T i is the tensor fragment data corresponding to the node; N i is the graph structure neighbor set of node i; W ij is the edge connection weight between node i and its neighbor node j.

[0071] In some embodiments, the reasoning process can be implemented based on edge-runnable algorithm architectures such as graph neural networks (GNNs), lightweight decision trees, or multi-scale time series models, and the trade-off configuration between model complexity and reasoning accuracy can be flexibly selected based on the system deployment capabilities.

[0072] After completing the inference analysis, the system will further execute compensation scheduling decisions based on the current node inference state. For example, if a node's inference state is "task interrupted," the scheduling module will automatically initiate a backup node wake-up operation and call the stored task state and tensor model based on the node redundancy mapping table M:i→i, thereby quickly restoring the original task operating environment locally and maintaining system continuity.

[0073] As an option, the system redundancy level R can be set, which represents the number of backup nodes preset for each core task. When the inference confirms that a node fails and the main node cannot be restored, the system can automatically select {i1,i2,…,iR}Select the most suitable node to complete the takeover.

[0074] Finally, the system state label set output by reasoning The system completes node-level scheduling reconstruction and resource reallocation, and dynamically adjusts the task proportion, data routing, and resource application limit of the remaining available nodes.

[0075] The vehicle-road-cloud data processing system based on the edge computing center described below and the vehicle-road-cloud data processing method based on the edge computing center described above can be referenced to each other.

[0076] Please see the attached Figure 2 The present invention also provides a vehicle-road-cloud data processing system based on an edge computing center, including: The perception module collects real-time data from vehicles, roadside units, and environmental sensing devices, and standardizes and time-aligns the collected data based on the device data characteristics to generate structured pre-processed data. The tensor modeling module is used to map spatial location information, temporal state information, and data type information from different sources to corresponding dimensions of a three-dimensional tensor model based on structured preprocessed data, thereby constructing a complete tensor data structure to extract data aggregation trend characteristics. The resource collaboration evaluation module is used to evaluate the computing requirements and resource consumption of each edge node based on the local gradient information output by the tensor modeling module, build a graph structure model of the edge nodes, and calculate the collaboration weight between nodes; The task scheduling module is used to dynamically construct the task scheduling optimization objective function based on the calculated coordination weight, combined with the task delay requirements, current resource usage status and task importance information; The task migration module is used to select a suitable target node based on the collaborative weight analysis results when the computing load of a node exceeds the preset threshold during the execution of the task scheduling module, and perform task status preservation, data transmission and model deployment operations; The autonomous reasoning module is used to call the pre-deployed autonomous reasoning model in the edge computing center based on the model and graph structure established by the tensor modeling module and the resource collaborative evaluation module in scenarios of communication anomalies or node failure anomalies, and automatically infer the node failure type and status anomaly.

[0077] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0078] The examples of this specific embodiment are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, any equivalent changes made based on the structure, shape, and principle of this application should be included in the scope of protection of this application.

Claims

1. A vehicle-road-cloud data processing method based on an edge computing center is characterized in that: The following steps are involved: S1. Deploy a perception module in the edge computing center to collect real-time data from vehicles, roadside units, and environmental sensing equipment. Standardize and time-align the collected real-time data to obtain structured pre-processed data. S2. Based on the structured preprocessed data, construct a three-dimensional tensor model with spatial position, time state, and data type dimensions, and calculate the local gradient of the three-dimensional tensor model to characterize the data aggregation trend; S3. Evaluate the resource collaboration requirements between edge computing nodes based on the local gradient of the tensor model. Build a graph of edge nodes based on the node features and gradient information in the tensor model and calculate the collaboration weights between nodes. S4. Based on the collaborative weights and tensor gradient output results, a task scheduling optimization objective function is constructed to calculate latency, resource usage status, and task importance, and allocate resources to tasks within the current edge node. S5. During the scheduling process, when the tensor gradient amplitude of a node exceeds a preset threshold, a target node that meets the migration conditions is selected based on the collaborative weight, and task state preservation, model deployment, and data migration operations are performed; S6. In the event of communication anomalies or node failures, the edge computing center calls the pre-deployed model to complete autonomous reasoning processing based on the locally built tensor model and graph feature information.

2. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S1, the standardization and time alignment of the collected real-time data includes the following steps: The perception module collects data from vehicles, roadside units, and environmental sensing devices in real time and transmits this raw data to the edge computing center; Based on the characteristics of data collected by each device, data from different sources are normalized through standardization algorithms; The collected data is time-aligned based on the timestamp, and the collected data from different devices are time-aligned by synchronizing the clocks.

3. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S2, the local gradient of the three-dimensional tensor model is calculated using a partial derivative formula, which is: Where, T i,j,k is an element of the tensor model, f(T) is the loss function of the model, is the gradient value of the element, is the symbol of partial derivative.

4. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S2, constructing a tensor model with spatial position, temporal state, and data type dimensions includes the following steps: Based on the preprocessed data, the three data types from vehicles, roadside units, and environmental sensing devices are mapped into different dimensions of a three-dimensional tensor; By building a tensor model, the processed data is filled into the corresponding elements of the tensor, and the number of dimensions of the tensor is selected according to the characteristics of the data; Through the tensor decomposition method, the tensor is reduced in dimension and decomposed to generate a set of low-rank tensor factors.

5. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S3, evaluating the resource coordination requirements between edge computing nodes includes the following steps: Based on the constructed tensor model, the gradient descent method is used to derive each element of the tensor to obtain the local gradient, which represents the computational requirements and resource consumption of each node; Based on the calculated local gradient information, the resource consumption and load of each node are evaluated; The collaborative weights between nodes are calculated based on the local gradient values ​​and resource requirements between each node.

6. The vehicle-road-cloud data processing method based on the edge computing center according to claim 5 is characterized in that: The collaborative weight between the computing nodes adopts a weighted calculation formula based on gradient similarity, and the weighted calculation formula based on gradient similarity is: Where W ij is the collaborative weight between edge node i and node j, and are the gradient values ​​of node i and node j respectively.

7. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S4, constructing the task scheduling optimization objective function includes the following steps: Based on the collaborative weights and tensor gradient outputs calculated in the task, combined with the latency, resource usage status, and importance of each task, the optimization objective function for task scheduling is constructed: Objective=λ1×Delay+λ2×Resource Usage+λ3×TaskImportance; Where λ1, λ2, and λ3 represent the impact of delay, resource usage, and task importance on the optimization objective, respectively. Delay represents the expected waiting time or deadline deviation of the task on the current node. Resource Usage represents the total amount of resources expected to be occupied by the task on the specified node. TaskImportance is the task priority indicator set by the system. Objective is the original optimization objective function value. By optimizing the objective function, the execution delay and resource consumption of each task on different nodes are calculated to evaluate the importance and priority of the task; Based on the results of the optimization objective function, the task scheduling order and resource allocation strategy are dynamically adjusted.

8. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S5, the execution of task status preservation, model deployment and data migration operations includes the following steps: When the tensor gradient amplitude of a node exceeds the preset threshold, the collaborative weight and adaptability between nodes are calculated and task migration is performed; Perform task status preservation operations to ensure that tasks remain consistent during the migration process, and save and transfer task status, data, and model information to the target node; Migrate the computing tasks and data of the task to the target node and update the computing model of the target node.

9. The vehicle-road-cloud data processing method based on the edge computing center according to claim 1 is characterized in that: In step S6, calling the pre-deployed model to complete the autonomous reasoning process includes the following steps: When communication anomalies or node failures occur, the autonomous reasoning mechanism of the edge computing center is activated by using the locally built tensor model and graph feature information; The edge computing center analyzes the node's status data based on the pre-deployed inference model and automatically infers potential failures or data anomalies; Based on the inference results, the edge computing center adjusts the computing tasks and resource allocation of the nodes.

10. A vehicle-road-cloud data processing system based on an edge computing center, applied to the vehicle-road-cloud data processing method based on an edge computing center according to any one of claims 1 to 9, characterized in that: include: The perception module collects real-time data from vehicles, roadside units, and environmental sensing devices, and standardizes and time-aligns the collected data based on the device data characteristics to generate structured pre-processed data. The tensor modeling module is used to map spatial location information, temporal state information, and data type information from different sources to corresponding dimensions of a three-dimensional tensor model based on structured preprocessed data, thereby constructing a complete tensor data structure to extract data aggregation trend characteristics. The resource collaboration evaluation module is used to evaluate the computing requirements and resource consumption of each edge node based on the local gradient information output by the tensor modeling module, build a graph structure model of the edge nodes, and calculate the collaboration weight between nodes; The task scheduling module is used to dynamically construct the task scheduling optimization objective function based on the calculated coordination weight, combined with the task delay requirements, current resource usage status and task importance information; The task migration module is used to select a suitable target node based on the collaborative weight analysis results when the computing load of a node exceeds the preset threshold during the execution of the task scheduling module, and perform task status preservation, data transmission and model deployment operations; The autonomous reasoning module is used to call the pre-deployed autonomous reasoning model in the edge computing center based on the model and graph structure established by the tensor modeling module and the resource collaborative evaluation module in scenarios of communication anomalies or node failure anomalies, and automatically infer the node failure type and status anomaly.

Images