Intelligent traffic collaborative management method and system based on large model and federated learning

Abstract

The application relates to the technical field of intelligent traffic, and provides an intelligent traffic cooperative management method and system, which comprises the following steps: regarding regional road side units as federal learning nodes, completing local traffic data processing, uploading encrypted model gradients to a central coordinator for aggregation; injecting Laplace noise into user GPS track data, and ensuring that individual information of the user cannot be identified through differential privacy technology; predefining cross-region data sharing rules, automatically performing data access permission verification and data flow circulation whole-process auditing; performing shunt processing on real-time traffic data collected by a road side camera, and completing local real-time analysis by an edge node; when the node load is greater than a preset load threshold, fragmenting and unloading LLM inference tasks to adjacent idle nodes. Through technical means such as hierarchical intelligent agent architecture design, privacy compliance enhancement and edge-cloud collaborative calculation, the application can realize intelligentization, high efficiency, safety and compliance of traffic cooperative management.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically, to an intelligent transportation collaborative management method and system based on large models and federated learning. Background Technology

[0002] With the acceleration of urbanization and the rapid development of Intelligent Traffic Systems (ITS), traffic management faces increasingly complex challenges, such as traffic congestion, frequent accidents, and energy waste. Traditional traffic control methods rely on single sensor data, making it difficult to handle the fusion requirements of multimodal data (such as time series, audio, and video), and they consume high computational resources, making real-time deployment on edge devices difficult. Although Large Language Models (LLM) and Reinforcement Learning (RL) have shown potential in intelligent transportation, existing technologies still have significant shortcomings in multimodal data fusion, computational efficiency, generalization ability, and multi-agent collaboration.

[0003] Therefore, it is necessary to develop a smart transportation collaborative management method to solve the real-time decision-making and resource allocation problems in complex traffic scenarios through efficient multimodal fusion, dynamic federated learning and resource optimization technologies, and promote the intelligent and efficient development of smart transportation systems. Summary of the Invention

[0004] To improve the intelligence and efficiency of intelligent transportation systems, this invention provides a collaborative management method and system for intelligent transportation based on large models and federated learning, the specific technical solution of which is as follows:

[0005] The intelligent transportation collaborative management method based on large models and federated learning includes the following steps: Build a hierarchical intelligent agent system and deploy a large model responsible for natural language interaction and decision generation; Each regional roadside unit is used as a federated learning node to complete local traffic data processing and upload encrypted model gradients to the central coordinator for aggregation. The traffic simulation and accident prediction modules are used as task agents, which are dynamically accessed and assigned to the corresponding traffic demand nodes through API interfaces. Injecting Laplace noise into user GPS trajectory data ensures that individual user information is unidentifiable through differential privacy technology; Deploy blockchain smart contracts, predefine cross-regional data sharing rules, and automatically execute data access permission verification and full-process audit of data circulation; Real-time traffic data collected by roadside cameras is processed in a diversion manner, and edge nodes perform local real-time analysis, extract key events, and upload event summaries to the cloud. Monitor the computing load of each node, and when the node load exceeds the preset load threshold, shard the LLM inference task and offload it to a nearby idle node.

[0006] The intelligent traffic collaborative management method based on large models and federated learning uses roadside units in each region as federated learning nodes to complete local traffic data processing, upload encrypted model gradients to the central coordinator for aggregation, and use specialized modules for traffic simulation and accident prediction as task agents. These modules are dynamically accessed and allocated to corresponding traffic demand nodes through API interfaces. This optimizes resource allocation, solves the load imbalance problem caused by multi-agent task competition and the processing delay problem caused by cloud dependence in existing technologies, and meets the real-time decision-making needs in complex traffic scenarios.

[0007] By injecting Laplace noise into user GPS trajectory data, differential privacy technology ensures that individual user information is unidentifiable, and blockchain smart contracts are deployed to predefine cross-regional data sharing rules. This automatically executes data access permission verification and full-process auditing of data circulation, which helps to solve the privacy leakage risks brought about by centralized storage, realizes the secure processing and compliant sharing of traffic data, and thus builds a comprehensive privacy protection system.

[0008] By diverting and processing real-time traffic data collected by roadside cameras, edge nodes perform local real-time analysis, extract key events, upload event summaries to the cloud, and monitor the computing load of each node. When the node load exceeds a preset load threshold, the LLM inference task is fragmented and unloaded to a nearby idle node. This breaks the restrictions on traffic data sharing imposed by data sovereignty regulations, allowing the scope of traffic optimization to be expanded to multi-city road networks. It achieves global optimization of cross-regional traffic collaborative management and solves the technical problem of limited cross-regional collaboration in existing systems.

[0009] In summary, this invention addresses various technical shortcomings of existing intelligent transportation systems through layered intelligent agent architecture design, enhanced privacy compliance, and edge-cloud collaborative computing, thereby achieving intelligent, efficient, safe, and compliant traffic collaborative management.

[0010] Preferably, the intelligent transportation collaborative management method further includes the following steps: By visualizing the decision-making basis of intelligent agents through attention heatmaps, the influencing factors and weight allocation of traffic decisions can be obtained; Construct a blockchain audit log chain to record the decision paths and data source hash values ​​of all intelligent agents.

[0011] Preferably, the intelligent transportation collaborative management method further includes the following steps: Traffic metadata is transmitted using the MQTT protocol to optimize communication bandwidth. Traffic metadata includes a congestion index.

[0012] Preferably, the specific method for uploading encrypted model gradients to the central coordinator for aggregation includes the following steps: Obtain the local training dataset, and based on the local training dataset, the federated learning local nodes complete the model parameter adjustment and gradient calculation locally; Before uploading their local gradients, local nodes in federated learning perform double encryption on their local gradient matrices and upload the encrypted gradient data to the central coordinator via the MQTT protocol. The central coordinator verifies the encrypted gradient data received from each local node of the federated learning system one by one. The gradient data that passes the verification is classified, stored, and marked with node identifiers and timestamps. The central coordinator serves as the sole gradient aggregation node, completing the decryption and global aggregation of all valid gradient data.

[0013] Preferably, the specific method for fragmenting and offloading the LLM inference task to a nearby idle node when the node load exceeds a preset load threshold includes the following steps: Establish a two-layer load monitoring architecture consisting of local node self-monitoring and central coordinator unified monitoring to obtain node load in real time; Node load levels are classified based on both computing power usage and task processing efficiency, and preset load thresholds are obtained. Based on the global node load status table, the optimal nearest idle node is selected as the target node; A hybrid sharding approach combining model-level sharding and task-level sharding is adopted, and the sharding process is completed on the local node. Encrypted transmission of fragmented tasks is achieved based on the MQTT lightweight communication protocol. The offloading process is monitored and coordinated by the central coordinator. After receiving the task, the target node completes the verification and computing power scheduling. The original load node retains overall control over the LLM inference task. After the target node completes the fragmented inference, it sends back the encrypted result fragment. After the original load node completes the result aggregation and integrity verification, it generates a complete traffic decision.

[0014] A smart transportation collaborative management system based on large models and federated learning is used to implement the aforementioned smart transportation collaborative management method, comprising: The federated collaborative architecture module is used to build a hierarchical intelligent agent system, deploy a large model responsible for natural language interaction and decision generation, use roadside units in each region as federated learning nodes to complete local traffic data processing, upload encrypted model gradients to the central coordinator for aggregation, and use traffic simulation and accident prediction specialized modules as task intelligent agents, which are dynamically accessed and allocated to the corresponding traffic demand nodes through API interfaces. The privacy compliance enhancement module is used to inject Laplace noise into user GPS trajectory data, ensure the unidentifiable nature of individual user information through differential privacy technology, and deploy blockchain smart contracts to predefine cross-regional data sharing rules, automatically execute data access permission verification and full-process audit of data circulation. The edge cloud computing module is used to offload and process real-time traffic data collected by roadside cameras. Edge nodes perform local real-time analysis, extract key events, upload event summaries to the cloud, and monitor the computing load of each node. When the node load exceeds a preset load threshold, the LLM inference task is fragmented and offloaded to a nearby idle node.

[0015] Preferably, the intelligent transportation collaborative management system based on large models and federated learning further includes: The interpretability enhancement module is used to visualize the decision-making basis of intelligent agents through attention heatmaps, obtain the influencing factors and weight allocation of traffic decisions, and build a blockchain audit log chain to record the decision paths and data source hash values ​​of all intelligent agents.

[0016] Preferably, the federated collaborative architecture module includes: The training data acquisition unit is used to acquire the local training dataset and, based on the local training dataset, to perform model parameter adjustment and gradient calculation locally. The local nodes of federated learning are used to perform double encryption on the local gradient matrix before uploading the local gradient, and upload the encrypted gradient data to the central coordinator via the MQTT protocol. The central coordinator is used to verify the encrypted gradient data received from each local node of the federated learning, classify and store the gradient data that passes the verification, and mark the node with a timestamp. As the sole gradient aggregation node, it completes the decryption and global aggregation of all valid gradient data.

[0017] Preferably, the edge cloud computing module includes: The monitoring architecture building unit is used to build a two-layer load monitoring architecture of local node self-monitoring and central coordinator unified monitoring to obtain node load in real time. The load threshold acquisition unit is used to classify node load levels based on both computing power usage and task processing efficiency, and to obtain preset load thresholds. The target node acquisition unit is used to select the best nearby idle node as the target node based on the global node load status table. The inference task sharding unit is used to complete the sharding process on the local node by adopting a hybrid sharding method that combines model-level sharding and task-level sharding. The data encryption transmission unit is used to implement encrypted transmission of fragmented tasks based on the MQTT lightweight communication protocol. The traffic decision acquisition unit is used to transmit encrypted result fragments after the target node completes the fragmented inference. After the original load node completes the result aggregation and integrity verification, it generates a complete traffic decision. Attached Figure Description

[0018] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0019] Figure 1 This is a schematic diagram of the overall process of a smart transportation collaborative management method based on large models and federated learning in one embodiment of the present invention. Figure 2 This is a flowchart illustrating a specific method for uploading encrypted model gradients to a central coordinator for aggregation in one embodiment of the present invention. Figure 3 This is a schematic diagram of the process structure of a specific method for fragmenting and unloading an LLM inference task to a nearby idle node in one embodiment of the present invention. Figure 4 This is a schematic diagram of the overall process of a smart transportation collaborative management method based on large models and federated learning in another embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0021] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] In this invention, "first" and "second" do not represent a specific quantity or order, but are merely used to distinguish names.

[0024] Before describing the embodiments of the present invention, a brief introduction to the prior art will be given.

[0025] With the continued acceleration of global urbanization, urban traffic flow is experiencing explosive growth. Intelligent Transportation Systems (ITS), as a core means to solve urban traffic congestion, frequent accidents, and excessive energy consumption, have been widely researched and applied. Traditional traffic control methods mostly rely on traffic data collected by single sensors for decision-making. The data dimensions are limited, and the processing methods are rigid, making it difficult to meet the current demand for the fusion and processing of multimodal traffic data (time-series traffic flow data, audio monitoring data, video surveillance stream data, etc.). At the same time, traditional methods consume large amounts of computational resources, and the algorithm models are too complex to be deployed and run in real time on edge devices, resulting in insufficient real-time response capabilities for traffic management.

[0026] To address the aforementioned issues, existing technologies have begun to introduce Large Model Learning (LLM) and Reinforcement Learning (RL) into the field of intelligent transportation, demonstrating certain application potential in scenarios such as traffic flow prediction, route optimization, and accident early warning. However, these technical solutions still suffer from many key technical deficiencies in actual implementation, failing to achieve global optimization and efficient collaboration of intelligent transportation systems.

[0027] First, the multi-agent task scheduling mechanism is imperfect. Task competition among agents leads to an imbalance in computing load, significant delays in real-time system response, and uneven resource allocation, which seriously restricts the real-time performance of traffic management.

[0028] Secondly, traffic data processing still uses a centralized storage model. Sensitive data such as user GPS trajectories and driving behavior are vulnerable to malicious attacks or data misuse, posing a very high risk of privacy leaks.

[0029] Third, due to restrictions imposed by national and regional data sovereignty regulations (such as the EU GDPR), existing systems struggle to achieve compliant sharing and collaborative processing of cross-regional traffic data, hindering cross-regional traffic collaborative management and preventing the completion of global optimization and scheduling of urban road networks.

[0030] Fourth, the decision-making process of large models has a "black box" characteristic. The traffic decisions generated by them, such as route optimization and accident warning, lack interpretability and are difficult to gain the trust of users and traffic management departments, thus restricting the actual implementation of the technology. In addition, the existing system relies too much on the cloud for processing massive traffic data. Real-time data collected by roadside cameras and radars must be uploaded to the cloud for analysis, resulting in excessively high data processing delays and failing to meet the real-time requirements of traffic management.

[0031] The aforementioned technical challenges are mutually restrictive, making it difficult for existing intelligent transportation systems to meet the management needs of complex urban traffic scenarios in terms of intelligence level, operational efficiency, and safety compliance. There is an urgent need to propose a smart transportation collaborative management solution that balances efficiency, safety, compliance, and explainability to promote the technological upgrading and practical application of intelligent transportation systems.

[0032] One of the objectives of this invention is to improve the intelligence and efficiency of intelligent transportation systems. To this end, such as... Figure 1 As shown, one embodiment of the present invention provides a smart traffic collaborative management method based on large models and federated learning, comprising the following steps: S1, build a hierarchical intelligent agent system and deploy a large model responsible for natural language interaction and decision generation.

[0033] Specifically, a hierarchical intelligent agent system is built, and the lightweight large model TinyBERT with compressed parameters through knowledge distillation is deployed as the large language model intelligent agent (LLM Agent) to be responsible for natural language interaction and decision generation.

[0034] S2 uses roadside units in each region as federated learning nodes to complete local traffic data processing and upload encrypted model gradients to the central coordinator for aggregation.

[0035] Specifically, the federated learning network setup and basic initialization are performed first. The city is divided into subnets according to traffic control areas. One or more Road Side Units (RSUs) in each subnet serve as a single local node for federated learning. All RSU nodes and the central coordinator deployed in the cloud / regional traffic control center form a star-shaped federated learning network. An encrypted communication link based on TLS / SSL is established between the RSU nodes and the central coordinator. The lightweight MQTT protocol is used as the gradient data transmission carrier, which is only used to transmit lightweight data such as encrypted gradients and model parameters to reduce bandwidth consumption.

[0036] The central coordinator symmetrically encrypts the basic parameters of the pre-trained lightweight large model Tiny BERT, such as those for adapting to intelligent transportation tasks like traffic flow prediction, accident warning, and congestion scheduling, and then distributes them all at once to all regional RSU federated nodes. This unifies the initial state of the model on all local nodes and clarifies the feature dimensions for model gradient calculation (such as traffic flow, vehicle speed, road occupancy, traffic light status, weather, and other dimensions related to traffic decisions).

[0037] For node parameters, each RSU federated node is assigned a unique node identifier, encryption key, and initial value of weight coefficient. The weight coefficient is pre-set based on the traffic data volume, data quality, road importance (such as core business districts / highways) and RSU computing power of the node's region, providing a basis for subsequent gradient weighted aggregation and supporting dynamic adjustment.

[0038] The RSU nodes then complete the entire process of traffic data collection, preprocessing, and dataset construction locally, without transmitting any raw traffic data externally, thus mitigating privacy risks from the source. Each RSU node collects multimodal traffic data for its assigned area, including camera video streams, radar traffic flow data, anonymized vehicle GPS trajectory data, traffic light status data, road construction / accident reporting data, etc. All data is stored in the RSU's local edge storage module and is not synchronized to the cloud or other nodes.

[0039] Local data preprocessing includes, but is not limited to, data cleaning, feature extraction, privacy enhancement, and data standardization. Data cleaning specifically includes removing outliers, missing values, and duplicates from the collected data (such as invalid vehicle speed data from radar or blurry frames from cameras). Feature extraction specifically involves extracting key features (such as accident features, traffic density features, and vehicle violation features) from unstructured data such as video and audio streams using lightweight algorithms, transforming them into structured feature vectors. Privacy enhancement specifically involves injecting Laplacian noise into sensitive data such as user GPS trajectories and driving behavior to achieve differential privacy protection, ensuring that individual information is unidentifiable. Data standardization specifically involves unifying the dimensions and value ranges of all feature data to form a local training dataset that meets the input requirements of the Tiny BERT model.

[0040] The preprocessed dataset is divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). Only the training set is used for local model gradient calculation, and the validation set is used for local model accuracy verification.

[0041] As a preferred technical solution, such as Figure 2 As shown, the specific method for uploading encrypted model gradients to the central coordinator for aggregation includes the following steps: S21: Obtain the local training dataset, and based on the local training dataset, the local nodes of the federated learning system complete the model parameter adjustment and gradient calculation locally.

[0042] Here, RSU nodes utilize local computing resources, using the Tiny BERT base model issued by the central coordinator as a benchmark, and employ mini-batch gradient descent to fine-tune the model locally. The training tasks are matched with the traffic management needs of the node's region (e.g., RSUs in urban core areas focus on congestion scheduling model training, while RSUs on highways focus on accident warning model training). During training, the gradient values ​​of the parameters of each layer of the Tiny BERT model are calculated based on the model loss function (e.g., cross-entropy loss), resulting in the local model gradient matrix; only the gradient calculation results are retained, and the complete model parameters after local training are not saved. If an RSU node is under high load (e.g., a surge in data volume during morning and evening rush hours), according to the edge-cloud collaboration requirements, some gradient calculation tasks are distributed to nearby idle RSU nodes. After completing the distributed gradient calculation, the local gradient matrix is ​​aggregated at the main RSU node to avoid single-point overload.

[0043] S22, in federated learning, local nodes perform double-layer encryption on their local gradient matrix before uploading the local gradient, and then upload the encrypted gradient data to the central coordinator via the MQTT protocol.

[0044] To prevent gradient data from being stolen or tampered with during transmission, RSU nodes perform double-layer encryption on their local gradient matrices before uploading, ensuring the security of the gradient data. The specific encryption steps include: normalizing the local gradient matrix to eliminate potential sensitive feature correlations in the gradient data; and simultaneously compressing the gradient data using Huffman coding to reduce data volume and adapt to edge-side communication bandwidth limitations. The compressed gradient data is encrypted using the AES-256 symmetric encryption algorithm. The encryption key is a unique key pre-assigned between the central coordinator and the RSU node, and is distributed offline using the RSA-2048 asymmetric encryption algorithm, not transmitted online. A node identifier, timestamp, and data hash value are added to the encrypted gradient data. The hash value is used by the central coordinator to verify the integrity of the gradient data, while the node identifier and timestamp are used for gradient traceability and management.

[0045] S23, the central coordinator verifies the encrypted gradient data received from each local node of the federated learning network one by one. The gradient data that passes the verification is classified, stored, and marked with node identifiers and timestamps.

[0046] RSU nodes upload encrypted gradient data (including verification information) to the central coordinator via the MQTT protocol. During transmission, only gradient data is sent; no original traffic data or model parameters are transmitted. A time-slot transmission strategy is employed to avoid network congestion caused by simultaneous uploads from multiple nodes. The central coordinator verifies the encrypted gradient data received from each RSU node one by one, using data hash values ​​to check for tampering or missing data. Gradient data that fails verification is discarded, and a retransmission command is sent to the corresponding RSU node. Gradient data that passes verification is categorized, stored, and tagged with node identifiers and timestamps.

[0047] S24 uses the central coordinator as the sole gradient aggregation node to complete the decryption and global aggregation of all valid gradient data.

[0048] The central coordinator, acting as the sole gradient aggregation node, decrypts and globally aggregates all valid gradient data. Its core employs a federated averaging algorithm to achieve global gradient optimization. Specifically, the central coordinator uses a pre-allocated AES-256 key to decrypt the valid encrypted gradient data of each RSU node, reconstructing the local model gradient matrix of each node. Based on the pre-defined weighting coefficients of each RSU node, a weighted average is calculated across all decrypted local gradient matrices to obtain the global model gradient matrix.

[0049] The central coordinator uses the aggregated global model gradient matrix and the learning rate to update the parameters of the initial Tiny BERT base model, generating the latest global federated model. The central coordinator then encrypts the updated global federated model parameters using AES-256 and distributes them to all normal RSU federated nodes via the MQTT protocol. Each RSU node receives the encrypted global model parameters, decrypts them, and updates the base parameters of its local Tiny BERT model, completing one federated learning iteration.

[0050] The process of local data processing → gradient calculation → encrypted upload → central aggregation → model distribution is repeated multiple times for federated learning iterations. After each iteration, each RSU node uses its local validation set to verify the task accuracy of the updated model. After the iteration terminates, each RSU node uses the final global federated model to perform real-time processing of local traffic data and decision generation (such as congestion scheduling suggestions and accident warnings), realizing local real-time decision-making for intelligent transportation.

[0051] S3 uses specialized modules for traffic simulation and accident prediction as task agents, which are dynamically accessed and assigned to corresponding traffic demand nodes through API interfaces.

[0052] The MQTT protocol can be used to transmit traffic metadata such as congestion index in order to optimize communication bandwidth.

[0053] S4 injects Laplace noise into the user's GPS trajectory data and uses differential privacy technology to ensure that the user's individual information is unidentifiable.

[0054] Since ensuring the unidentifiable nature of individual user information through injecting Laplace noise and differential privacy technology are conventional techniques in this field, they will not be elaborated upon here.

[0055] S5 deploys blockchain smart contracts, predefines cross-regional data sharing rules, and automatically executes data access permission verification and full-process auditing of data circulation.

[0056] Specifically, smart contracts are first deployed in the blockchain network to predefine data sharing rules that comply with regulations such as GDPR / CCPA. These rules include core content such as data types (anonymous traffic / sensitive GPS), access subject permissions, cross-regional circulation compliance requirements, and audit record field specifications. All federated learning nodes / access subjects complete on-chain identity registration and rights confirmation.

[0057] The data access subject then initiates an access request, recording information such as identity identifier, data type, purpose, and access region on the blockchain. The smart contract automatically matches pre-configured rules to complete multi-dimensional automatic verification of the subject's permissions, compliance, and legality of the purpose. If the verification passes, the smart contract automatically generates a unique access credential and records it on the blockchain, allowing the authorized subject to access the data based on the credential; if the verification fails, the request is directly rejected, and the reason for rejection (exceeding permissions / cross-regional violation, etc.) is recorded on the blockchain in real time.

[0058] Information such as the operating entity, timestamp, data hash value, and operation type is recorded on the blockchain in real time to ensure traceability and immutability. The smart contract verifies the hash value of the data in real time during circulation. If the data is tampered with, circulation is immediately terminated and an on-chain alert is triggered.

[0059] Based on on-chain data from all stages, the blockchain automatically generates standardized audit logs. These logs are linked to corresponding data operations, enabling full-process traceability and querying. Meanwhile, smart contracts automatically flag and issue compliance warnings for behaviors such as unauthorized access and illegal cross-regional circulation.

[0060] S6 processes real-time traffic data collected by roadside cameras, performs local real-time analysis by edge nodes, extracts key events, and uploads event summaries to the cloud.

[0061] Specifically, the RSU classifies the collected data through a built-in diversion module, imports the core data (video stream, real-time traffic data) that needs to be analyzed in real time into the local analysis link, and keeps the auxiliary data in local edge storage for backup only without participating in real-time analysis. It also uses lightweight algorithms (object detection, traffic density calculation, etc.) to perform local real-time analysis on the core data to identify key traffic events such as accidents, congestion, and road anomalies.

[0062] The identified key events are structured by extracting crucial information such as event type, occurrence time, geographical location, event level, and core features to generate a concise event summary, while redundant data such as the original video stream is removed. The structured event summary is then encrypted and uploaded to the cloud via the MQTT lightweight communication protocol, while the original data and analysis records are stored locally.

[0063] S7 monitors the computing load of each node. When the node load exceeds the preset load threshold, the LLM inference task is fragmented and unloaded to a nearby idle node.

[0064] As a preferred technical solution, such as Figure 3 As shown, the specific method for fragmenting and offloading LLM inference tasks to nearby idle nodes when the node load exceeds a preset load threshold includes the following steps: S71 establishes a two-layer load monitoring architecture consisting of local node self-monitoring and central coordinator-managed monitoring, enabling real-time acquisition of node load.

[0065] A lightweight load monitoring submodule is deployed on each RSU federated learning node to collect core metrics of node computational load in real time. The sampling frequency is set to 100ms / time (which can be dynamically adjusted according to traffic scenarios). The monitoring metrics are aligned with the computational consumption characteristics of LLM inference tasks, specifically including: CPU utilization, GPU computing power utilization (on nodes with deployed computing cards), memory / video memory utilization, LLM inference task queue length, single LLM inference request processing latency, and priority of currently pending traffic tasks. All monitoring data is first stored in the node's local edge storage unit and is not directly uploaded to the cloud.

[0066] The central coordinator sends load status heartbeat requests to all RSU nodes via the MQTT lightweight communication protocol (every 500ms). Each node only uploads normalized statistical values ​​of load metrics (such as CPU utilization of 85% and inference latency of 40ms), rather than raw monitoring data. The central coordinator builds a global node load status table and updates the load data, node geographical coordinates, communication link quality, and remaining computing power of each node in real time, providing global data support for subsequent task unloading.

[0067] Both local nodes and the central coordinator are configured with data filtering rules to remove abnormal load data caused by sensor failures or network jitter (such as instantaneous CPU utilization of 100% but no actual tasks being executed); the load index of three consecutive samples is averaged to ensure the accuracy of load perception and avoid misjudgment.

[0068] S72 classifies node load levels based on both computing power usage and task processing efficiency, and obtains preset load thresholds.

[0069] In light of the real-time requirements of LLM inference tasks (traffic flow prediction, accident warning, route optimization, etc.) in intelligent transportation scenarios, the node load levels are divided according to two dimensions: computing power occupancy and task processing efficiency. High load judgment thresholds are dynamically set to adapt to RSU nodes with different computing power, such as different thresholds for high computing power RSU in core urban areas and low computing power RSU in suburban areas, and the triggering conditions for task unloading are clearly defined.

[0070] Node states are categorized into five levels (idle, lightly loaded, moderately loaded, heavily loaded, and overloaded). The core performance thresholds for each level can be pre-configured and dynamically adjusted by the central coordinator based on the node's computing power. Different levels correspond to different CPU utilization, LLM inference latency, and task queue lengths. When a node's load status reaches the high-load / overload level for three consecutive samplings, and the currently executing LLM inference task is not a single atomic task, the local node automatically triggers an offload request, reporting the type of task to be offloaded, task priority, and required computing power to the central coordinator, and requesting resources from nearby idle nodes.

[0071] The central coordinator periodically updates the node load judgment threshold based on the traffic characteristics of the node's region (such as core business districts, highways, and suburban roads) and historical load data (morning and evening peak hours, holidays). For example, the high load threshold is lowered to 60% during morning peak hours, triggering the unloading mechanism in advance to prevent the node from entering an overload state.

[0072] S73 selects the optimal nearby idle node as the target node based on the global node load status table.

[0073] The central coordinator uses a global node load status table, combined with three dimensions: geospatial location, communication link quality, and remaining computing power matching degree, to quickly select the optimal nearby idle nodes for high-load nodes, ensuring low latency and high reliability of task unloading.

[0074] The central coordinator pre-constructs a regional traffic edge node topology map based on the geographical coordinates of each RSU node, divides the network into subnets according to the traffic control area, and records the actual communication distance, link bandwidth, and packet loss rate between nodes; communication distance ≤5km, link packet loss rate ≤1%, and bandwidth ≥100Mbps are used as the basic screening criteria for neighboring nodes.

[0075] Among the nodes that meet the proximity criteria, nodes with idle / light load levels are selected, their remaining computing power values ​​are extracted, and they are matched with the computing power required for the LLM inference sharding task of the high-load node. The candidate unload node list is generated by sorting the computing power matching degree from high to low, and 3-5 candidate nodes are matched for each high-load node.

[0076] The central coordinator sends instantaneous verification requests for computing power occupancy to candidate nodes to confirm that the nodes have not experienced sudden load changes in a short period of time. The optimal node that passes the verification is designated as the primary offloading target, and the rest are designated as backup targets. At the same time, the offloading target information is pushed to high-load nodes.

[0077] S74 employs a hybrid sharding approach that combines model-level sharding with task-level sharding, completing the sharding process on local nodes while balancing sharding efficiency with the completeness of inference results.

[0078] The sharding granularity is adjusted based on the remaining computing power and communication link bandwidth of the target node to be unloaded. High computing power + high bandwidth corresponds to large granular sharding (reducing the number of shards and communication overhead), while medium computing power + low bandwidth corresponds to small granular sharding (reducing the computing power pressure on a single node). At the same time, the minimum number of shards per task is limited to 2 and the maximum number of shards is limited to 8 to avoid the increase in result aggregation complexity caused by excessive sharding.

[0079] S75 uses the MQTT lightweight communication protocol to implement encrypted transmission of sharded tasks. The offloading process is monitored and coordinated by the central coordinator. After receiving the task, the target node completes the verification and computing power scheduling.

[0080] The weightlifting system still utilizes the MQTT lightweight communication protocol to achieve encrypted transmission of fragmented tasks. The offloading process is monitored and coordinated by the central coordinator. After receiving the task, the target node quickly completes verification and computing power scheduling to ensure seamless connection of inference tasks.

[0081] S76: The original load node retains overall control over the LLM inference task. After the target node completes the fragmented inference, it sends back the encrypted result fragment. After the original load node completes the result aggregation and integrity verification, it generates a complete traffic decision.

[0082] Specifically, the original high-load node retains overall control over the LLM inference task, while neighboring nodes only send back encrypted result fragments after completing the fragmented inference. The original node generates a complete traffic decision after completing result aggregation and integrity verification, ensuring the consistency of the inference results.

[0083] After the sharding is unloaded, the original high-load nodes are continuously monitored for load. When the node load returns to a normal level, the task migration mechanism is triggered to restore the node's local full LLM inference capability and release the computing resources of neighboring nodes to ensure global node load balance.

[0084] In summary, the intelligent traffic collaborative management method based on large models and federated learning uses roadside units in each region as federated learning nodes to complete local traffic data processing, upload encrypted model gradients to the central coordinator for aggregation, and use specialized modules for traffic simulation and accident prediction as task agents. These modules are dynamically accessed and allocated to corresponding traffic demand nodes through API interfaces. This optimizes resource allocation, solves the load imbalance problem caused by multi-agent task competition and the processing delay problem caused by cloud dependence in existing technologies, and meets the real-time decision-making needs in complex traffic scenarios.

[0085] By injecting Laplace noise into user GPS trajectory data, differential privacy technology ensures that individual user information is unidentifiable, and blockchain smart contracts are deployed to predefine cross-regional data sharing rules. This automatically executes data access permission verification and full-process auditing of data circulation, which helps to solve the privacy leakage risks brought about by centralized storage, realizes the secure processing and compliant sharing of traffic data, and thus builds a comprehensive privacy protection system.

[0086] By diverting and processing real-time traffic data collected by roadside cameras, edge nodes perform local real-time analysis, extract key events, upload event summaries to the cloud, and monitor the computing load of each node. When the node load exceeds a preset load threshold, the LLM inference task is fragmented and unloaded to a nearby idle node. This breaks the restrictions on traffic data sharing imposed by data sovereignty regulations, allowing the scope of traffic optimization to be expanded to multi-city road networks. It achieves global optimization of cross-regional traffic collaborative management and solves the technical problem of limited cross-regional collaboration in existing systems.

[0087] In one embodiment, such as Figure 4 As shown, the intelligent transportation collaborative management method further includes the following steps: S8 visualizes the decision-making basis of intelligent agents through attention heatmaps, and obtains the influencing factors and weight allocation of traffic decisions.

[0088] S9 constructs a blockchain audit log chain to record the decision paths and data source hash values ​​of all intelligent agents.

[0089] Specifically, the Tiny BERT model deployed in the LLM Agent is lightweighted and fine-tuned, retaining the core architecture of a 12-layer encoder + 6-head attention layer (adapting to the computing power of RSU edge nodes), and the model's attention layer is specifically modified: a traffic feature-specific embedding layer is added to the model input layer, transforming traffic decision-related structured features (such as traffic flow, vehicle speed, and road occupancy), unstructured features (such as weather, road construction, and accident information), and environmental features (such as time period, holidays, and region type) into feature vectors that the model can recognize; at the same time, the model's general semantic encoding layer is frozen, and only the attention layer parameters related to traffic features are trained, ensuring that the model's attention weight calculation in traffic scenarios can accurately focus on actual traffic influencing factors. The core influencing factors of different decision-making tasks in intelligent transportation are sorted out, and a standardized feature library is constructed according to first-level features and second-level features. A one-to-one mapping relationship between the model's hidden layer feature dimensions and actual traffic influencing factors is established, ensuring that the subsequently extracted attention weights can directly correspond to specific traffic factors, avoiding the disconnect between weights and actual factors.

[0090] To address the multi-attention head and multi-attention layer architecture of the Tiny BERT model, raw attention weights are extracted from the inference process and then fused at the head and layer levels to obtain global attention weights that reflect the model's decision-making logic, ensuring the accuracy and representativeness of the weights. Specifically, this involves several steps: acquiring the raw attention weight matrix, fusing weights from multiple attention heads, fusing weights from multiple attention layers, and filtering out invalid weights.

[0091] Next is the quantification and normalization of attention weights. By converting the fused global attention weights from model tensor form into a quantifiable and comparable numerical form and then normalizing them, the weight ratio of each traffic influencing factor can be intuitively interpreted, providing standardized data support for heatmap generation.

[0092] Based on the actual application scenarios of intelligent transportation decision-making, three types of attention heatmaps are designed: feature dimension, spatial dimension, and temporal dimension. Different rendering rules and display formats are adopted to adapt to different visualization carriers such as RSU local edge terminals and large screens in urban traffic control centers. The colors, scales, and labels of the heatmaps all follow the principles of simplicity, intuitiveness, professionalism, and ease of understanding, making it easy for traffic management personnel to quickly interpret the basis of model decision-making.

[0093] Based on the generated attention heatmap, the core influencing factors, factor correlation logic, and decision rules of traffic decision-making are systematically analyzed. The implicit decision-making logic of the LLM Agent is transformed into an explicit natural language description, ensuring that traffic management personnel can quickly understand the basis of the model's decision-making, rather than just looking at the heatmap values.

[0094] This embodiment also provides a smart transportation collaborative management system based on large models and federated learning to implement the aforementioned smart transportation collaborative management method. It includes a federated collaborative architecture module, a privacy compliance enhancement module, and an edge cloud computing module.

[0095] The federated collaborative architecture module is used to build a hierarchical intelligent agent system, deploy a large model responsible for natural language interaction and decision generation, use roadside units in each region as federated learning nodes to complete local traffic data processing, upload encrypted model gradients to the central coordinator for aggregation, and use traffic simulation and accident prediction specialized modules as task intelligent agents, which are dynamically accessed and assigned to corresponding traffic demand nodes through API interfaces.

[0096] Specifically, preferably, the federated collaborative architecture module includes a training data acquisition unit, federated learning local nodes, and a central coordinator.

[0097] The training data acquisition unit is used to acquire the local training dataset and, based on the local training dataset, to complete the model parameter adjustment and gradient calculation locally; the federated learning local node is used to perform double-layer encryption on the local gradient matrix before uploading the local gradient, and upload the encrypted gradient data to the central coordinator via the MQTT protocol.

[0098] The central coordinator is used to verify the encrypted gradient data received from each local node of the federated learning, classify and store the gradient data that passes the verification, and mark the node with a timestamp. As the only gradient aggregation node, it completes the decryption and global aggregation of all valid gradient data.

[0099] The privacy compliance enhancement module injects Laplace noise into user GPS trajectory data, ensures the unidentifiable nature of individual user information through differential privacy technology, and deploys blockchain smart contracts to predefine cross-regional data sharing rules, automatically executing data access permission verification and full-process auditing of data circulation.

[0100] The edge cloud computing module is used to offload and process real-time traffic data collected by roadside cameras. Edge nodes perform local real-time analysis, extract key events, upload event summaries to the cloud, and monitor the computing load of each node. When the node load exceeds a preset load threshold, the LLM inference task is fragmented and offloaded to a nearby idle node.

[0101] The edge cloud computing module includes a monitoring architecture building unit, a load threshold acquisition unit, a target node acquisition unit, an inference task sharding unit, a data encryption transmission unit, and a traffic decision acquisition unit.

[0102] The monitoring architecture building unit is used to build a two-layer load monitoring architecture of local node self-monitoring and central coordinator unified monitoring to obtain node load in real time; the load threshold acquisition unit is used to classify node load levels according to two dimensions of computing power usage and task processing efficiency, and obtain preset load thresholds; the target node acquisition unit is used to select the best nearby idle node as the target node based on the global node load status table.

[0103] The inference task sharding unit is used to complete the sharding process on the local node by using a hybrid sharding method that combines model-level sharding and task-level sharding; the data encryption transmission unit is used to implement encrypted transmission of sharded tasks based on the MQTT lightweight communication protocol; the traffic decision acquisition unit is used to send back encrypted result fragments after the target node completes sharded inference, and after the original load node completes result aggregation and integrity verification, it generates a complete traffic decision.

[0104] As a preferred technical solution, the specific method for monitoring the computing load of each node and sharding and offloading the LLM inference task to a nearby idle node when the node load exceeds a preset load threshold includes the following steps: The first step involves real-time collection of key resource metrics for each node, ensuring lightweight data and privacy compatibility. Nodes sample resource data every second, including key metrics such as CPU utilization (Ct, current CPU usage percentage), memory usage (Mt, current memory usage percentage), network latency (Nt, data transmission latency), metadata calculations based on the MQTT protocol, and task queue length (Qt, number of pending tasks, such as LLM inference shards). These key resource metrics are then smoothed to reduce noise.

[0105] A multi-dimensional sliding sampling mechanism is introduced, using time-series windows to reduce noise, and encrypted gradient aggregation among federated learning nodes is used to calibrate metric biases, avoiding direct sharing of raw data. For example, RSU nodes collect CPU utilization, memory usage, network latency, and task queue length, but only upload summaries to the central coordinator for encrypted aggregation.

[0106] Gradient aggregation information from federated learning is used to adjust local metrics and prevent inter-node bias. Let the gradient aggregation interval be T, then the calibration factor... . .in, This represents the average change in the encryption gradient, where α is the decay coefficient (default 0.1). This calibration factor is used to calibrate critical resource metrics, i.e., calibrated critical resource metrics = uncalibrated critical resource metrics × FCF, to ensure that load metrics remain consistent under privacy protection and avoid misjudgments due to data deviations.

[0107] The second step involves integrating multi-dimensional key resource indicators for efficient judgment. Weights can be dynamically allocated based on information entropy (high entropy indicates greater indicator volatility and thus higher weight), and traffic event impact factors (such as the output of the accident prediction module) can be introduced to make the index more aligned with intelligent transportation scenarios.

[0108] First, the multi-dimensional key resource indicators are normalized, and then the dynamic weighted entropy index is obtained. .in, These are the edge load balancing coefficient and the traffic event sensitivity coefficient, respectively, with default values ​​of 0.7 and 0.3. The edge load balancing coefficient controls the weight of basic load indicators; a larger value indicates a more significant impact of multi-dimensional resource status on the load. The traffic event sensitivity coefficient is mainly used to adjust the contribution of real-time traffic events to the load.

[0109] This represents dynamic weights, calculated based on the entropy values ​​of key resource indicators. Entropy It can be determined by calculating the probability distribution over the past 30 seconds. This represents the traffic event impact factor, which is equal to the product of the event probability output by the accident prediction module and the congestion index. Represents the key resource indicator after normalization of the i-th time. Here, the dynamic weighted entropy exponential function assigns higher weights to highly volatile indicators (such as sudden increases in network latency), and combined with real-time traffic data, it can achieve coupling between load and traffic status.

[0110] The third step involves combining exponential smoothing to predict future load trends and using a global view from federated learning to set dynamic thresholds to avoid overload or resource waste caused by fixed thresholds.

[0111] First, load trend prediction is performed using double exponential smoothing to predict the weighted entropy exponent for the next period. Then, the current and predicted loads are combined to obtain the load stress score. .in, The weighted entropy index represents the predicted weight for the next period. This represents the current weight entropy index. λ represents the risk coefficient and the standard deviation of DWEI over the past 10 seconds, respectively, with λ defaulting to 0.5.

[0112] Step 4: Preset load threshold .in, These represent the adjustment coefficient and the standard deviation of the weighted entropy exponent, respectively. The adjustment coefficient is set to 1.5 by default.

[0113] Fifth, if the load pressure score exceeds the preset load threshold for more than 3 seconds, the node is marked as high load. Otherwise, the status is normal.

[0114] In this way, predictive mechanisms can be used to provide early warnings (e.g., before sudden traffic peaks) and thresholds can be adapted through federal global data, thereby better triggering task offloading.

[0115] The intelligent transportation collaborative management system based on large models and federated learning also includes an interpretability enhancement module.

[0116] The interpretability enhancement module is used to visualize the decision-making basis of agents through attention heatmaps, obtain the influencing factors and weight allocation of traffic decisions, and build a blockchain audit log chain to record the decision paths and data source hash values ​​of all agents.

[0117] By visualizing the decision-making basis of large models through attention heatmaps, the influencing factors and weight allocation of traffic decisions can be clearly displayed, solving the problem of the black box nature of large model decision-making. At the same time, relying on blockchain to build an audit log chain, the decision-making paths and data source hash values ​​of all agents are fully recorded, realizing full-process traceability of traffic decisions.

[0118] In summary, this invention, through hierarchical intelligent agent architecture design, enhanced privacy compliance, edge-cloud collaborative computing, and optimized decision interpretability, can specifically address various technical deficiencies in existing intelligent transportation systems, thereby achieving intelligent, efficient, safe, and compliant traffic collaborative management.

[0119] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A smart transportation collaborative management method based on large models and federated learning, characterized in that, The intelligent transportation collaborative management method includes the following steps: Build a hierarchical intelligent agent system and deploy a large model responsible for natural language interaction and decision generation; Each regional roadside unit is used as a federated learning node to complete local traffic data processing and upload encrypted model gradients to the central coordinator for aggregation. The traffic simulation and accident prediction modules are used as task agents, which are dynamically accessed and assigned to the corresponding traffic demand nodes through API interfaces. Injecting Laplace noise into user GPS trajectory data ensures that individual user information is unidentifiable through differential privacy technology; Deploy blockchain smart contracts, predefine cross-regional data sharing rules, and automatically execute data access permission verification and full-process audit of data circulation; Real-time traffic data collected by roadside cameras is processed in a diversion manner, and edge nodes perform local real-time analysis, extract key events, and upload event summaries to the cloud. Monitor the computing load of each node, and when the node load exceeds the preset load threshold, shard the LLM inference task and offload it to a nearby idle node.

2. The intelligent transportation collaborative management method based on large models and federated learning as described in claim 1, characterized in that, The intelligent transportation collaborative management method also includes the following steps: By visualizing the decision-making basis of intelligent agents through attention heatmaps, the influencing factors and weight allocation of traffic decisions can be obtained; Construct a blockchain audit log chain to record the decision paths and data source hash values ​​of all intelligent agents.

3. The intelligent transportation collaborative management method based on large models and federated learning as described in claim 2, characterized in that, The intelligent transportation collaborative management method also includes the following steps: Traffic metadata is transmitted using the MQTT protocol to optimize communication bandwidth. Traffic metadata includes a congestion index.

4. The intelligent transportation collaborative management method based on large models and federated learning as described in claim 3, characterized in that, The specific method for uploading encrypted model gradients to the central coordinator for aggregation includes the following steps: Obtain the local training dataset, and based on the local training dataset, the federated learning local nodes complete the model parameter adjustment and gradient calculation locally; Before uploading their local gradients, local nodes in federated learning perform double-layer encryption on their local gradient matrices and upload the encrypted gradient data to the central coordinator via the MQTT protocol. The central coordinator verifies the encrypted gradient data received from each local node of the federated learning system one by one. The gradient data that passes the verification is classified, stored, and marked with node identifiers and timestamps. The central coordinator serves as the sole gradient aggregation node, completing the decryption and global aggregation of all valid gradient data.

5. The intelligent transportation collaborative management method based on large models and federated learning as described in claim 4, characterized in that, The specific method for fragmenting and offloading LLM inference tasks to nearby idle nodes when the node load exceeds a preset load threshold includes the following steps: Establish a two-layer load monitoring architecture consisting of local node self-monitoring and central coordinator unified monitoring to obtain node load in real time; Node load levels are classified based on both computing power usage and task processing efficiency, and preset load thresholds are obtained. Based on the global node load status table, the optimal nearest idle node is selected as the target node; A hybrid sharding approach combining model-level sharding and task-level sharding is adopted, and the sharding process is completed on the local node. Encrypted transmission of fragmented tasks is achieved based on the MQTT lightweight communication protocol. The offloading process is monitored and coordinated by the central coordinator. After receiving the task, the target node completes the verification and computing power scheduling. The original load node retains overall control over the LLM inference task. After the target node completes the fragmented inference, it sends back the encrypted result fragment. After the original load node completes the result aggregation and integrity verification, it generates a complete traffic decision.

6. A smart transportation collaborative management system based on large models and federated learning, used to implement the smart transportation collaborative management method as described in any one of claims 1-5, characterized in that, The intelligent transportation collaborative management system includes: The federated collaborative architecture module is used to build a hierarchical intelligent agent system, deploy a large model responsible for natural language interaction and decision generation, use roadside units in each region as federated learning nodes to complete local traffic data processing, upload encrypted model gradients to the central coordinator for aggregation, and use traffic simulation and accident prediction specialized modules as task intelligent agents, which are dynamically accessed and allocated to the corresponding traffic demand nodes through API interfaces. The privacy compliance enhancement module is used to inject Laplace noise into user GPS trajectory data, ensure the unidentifiable nature of individual user information through differential privacy technology, and deploy blockchain smart contracts to predefine cross-regional data sharing rules, automatically execute data access permission verification and full-process audit of data circulation. The edge cloud computing module is used to offload and process real-time traffic data collected by roadside cameras. Edge nodes perform local real-time analysis, extract key events, upload event summaries to the cloud, and monitor the computing load of each node. When the node load exceeds a preset load threshold, the LLM inference task is fragmented and offloaded to a nearby idle node.

7. The intelligent transportation collaborative management system based on large models and federated learning as described in claim 6, characterized in that, The intelligent transportation collaborative management system based on large models and federated learning also includes: The interpretability enhancement module is used to visualize the decision-making basis of intelligent agents through attention heatmaps, obtain the influencing factors and weight allocation of traffic decisions, and build a blockchain audit log chain to record the decision paths and data source hash values ​​of all intelligent agents.

8. The intelligent transportation collaborative management system based on large models and federated learning as described in claim 7, characterized in that, The federal collaborative architecture module includes: The training data acquisition unit is used to acquire the local training dataset and, based on the local training dataset, to perform model parameter adjustment and gradient calculation locally. The local nodes of the federated learning are used to perform double encryption on the local gradient matrix before uploading the local gradient, and upload the encrypted gradient data to the central coordinator via the MQTT protocol. The central coordinator is used to verify the encrypted gradient data received from each local node of the federated learning, classify and store the gradient data that passes the verification, and mark the node with a timestamp. As the sole gradient aggregation node, it completes the decryption and global aggregation of all valid gradient data.

9. The intelligent transportation collaborative management system based on large models and federated learning as described in claim 8, characterized in that, The edge cloud computing module includes: The monitoring architecture building unit is used to build a two-layer load monitoring architecture that combines local node self-monitoring and central coordinator-managed monitoring to obtain node load in real time. The load threshold acquisition unit is used to classify node load levels based on both computing power usage and task processing efficiency, and to obtain preset load thresholds. The target node acquisition unit is used to select the best nearby idle node as the target node based on the global node load status table. The inference task sharding unit is used to complete the sharding process on the local node by adopting a hybrid sharding method that combines model-level sharding and task-level sharding. The data encryption transmission unit is used to implement encrypted transmission of fragmented tasks based on the MQTT lightweight communication protocol. The traffic decision acquisition unit is used to transmit encrypted result fragments after the target node completes the fragmented inference. After the original load node completes the result aggregation and integrity verification, it generates a complete traffic decision.

Images