Research On Collaborative Technology For Cooperative Transportation Network

With the advancement of massive data processing and artificial intelligence technologies,traffic networks are now receiving comprehensive intelligent assistance.Network nodes’ self-perception and self-decision-making abilities are continuously improving,becoming a vital technological foundation for intelligent transportation systems.One of the essential infrastructures of intelligent transportation systems is the cooperative traffic network,which enables the efficient operation of smart travel and urban management by coordinating people,vehicles,sensing devices,and road network data while also supporting emerging industries such as vehicle networking.The main characteristics of cooperative traffic network coordination technologies are decentralization and shared collaboration.Decentralized architecture offers advantages such as low cost,fast service establishment,and user privacy protection,while shared collaboration mechanisms enable efficient communication and information exchange between nodes.This mechanism breaks the isolation of individuals,achieving collaboration and support for each other,thus promoting more efficient and excellent network operation.Cooperative traffic network technologies encompass various aspects;this thesis focuses on four core areas: path planning,behavior understanding,traffic prediction,and cooperation framework.Cooperative path planning realizes optimal route selection by sharing real-time traffic conditions and information among participants to avoid congestion.Behavior understanding helps nodes better perceive the surrounding traffic environment and other participants’ actions,enhancing decision-making efficiency.Traffic prediction combines multi-source data aggregation and analysis,providing more accurate flow forecasts and congestion warnings for participants.The cooperation framework significantly promotes data sharing and algorithm collaboration,assisting parties in making optimal coordinated decisions,optimizing resources,and improving efficiency.In summary,path planning provides safe driving solutions based on behavior intentions and traffic condition predictions.Behavior understanding analyzes traffic participants’ intentions and actions,providing fundamental information for path planning and the cooperation framework.Traffic prediction offers precise information for collaborative work.The cooperation framework integrates various technologies,enabling node coordination and enhancing traffic system performance.Despite extensive research in these fields,many challenges remain.To address the traffic flow balance problem in cooperative path planning,existing multi-agent learning-based path planning algorithms have limited capability in describing path conflicts and cannot fully meet the planning demands under traffic flow balance constraints.This thesis proposes a multi-agent personalized path planning method for traffic network flow balance based on reinforcement learning.The method adopts a centralized training-decentralized execution architecture,using a centralized value network during training to fuse and compute the global traffic flow state of the network,guiding the distributed policy networks to learn point-of-interest navigation and flow balance path planning strategies based on local observations.Additionally,the method utilizes imitation learning based on real taxi and pedestrian trajectories to further optimize the policy networks,achieving personalization and traffic flow balance in multi-user path planning.To address the behavior understanding problem in cooperative path planning,existing methods have insufficient capabilities in handling unexpected travel times of other intelligent agents.This thesis proposes a hierarchical multi-agent path planning method for unexpected behavior understanding.The method utilizes semi-Markov processes to model unpredictable travel time patterns of network nodes as decisionmaking behaviors based on non-uniform time intervals.The upper-level agents are responsible for understanding and evaluating the unexpected behaviors of nodes,while the lower-level agents plan according to the time recommendations and path conflict constraints provided by the upper-level agents.Furthermore,the upper-level agents employ graph attention networks to fuse the path planning reward signals shared by lower-level agents with local traffic flow balance reward signals,guiding the distributed policy networks to learn to recognize and bypass potentially congested road segments.This method provides a new approach for multi-user,multi-objective cooperative path planning with unexpected behavior understanding.To address the instability of traffic feature construction for traffic prediction based on sparse data,existing graph neural network-based methods mainly rely on the road network to construct the basic graph structure.However,under sparse traffic data with a high proportion of missing values,it is difficult for the model to capture the flow transfer,leading to a significant decline in prediction performance.This thesis proposes a traffic flow prediction method based on dynamic graph generation.The method generates a logical road network graph structure over several forecasting periods to reorganize the transfer relationship of traffic flows.Focusing on scenic area traffic networks,the method is driven by external events to predict the model,combining attraction relationships,tour time relationships,and event elements to reconstruct the flow transfer graph.During prediction,several historical flow transfer graph snapshots are stacked,with graph attention networks capturing the correlations between graph nodes,and sequence selfattention networks capturing the temporal features of traffic flows in the historical flow graph snapshots.In validation experiments conducted on real scenic area traffic flow datasets,the traffic flow prediction based on dynamic graph generation can better utilize the impact of external events on traffic flow,showing a significant advantage in medium and long-term predictions compared to baseline algorithms.To address the trust-building problem in collaborative frameworks,establishing trustworthy cooperative relationships based on reputation requires timely updates of each node’s reputation score.However,existing reputation update mechanisms struggle to cope with the rapid changes in traffic network topology.This thesis proposes a fully decentralized shared collaborative framework based on Directed Acyclic Graph(DAG)distributed ledgers.The framework models the cooperative behavior of nodes in the traffic network as node reputation and utilizes the distributed ledger for management.A partition-based consensus method is then introduced to further reduce the management cost of the distributed ledger.Lastly,a hybrid measure of node reputation is designed to comprehensively assess cooperation among nodes and ledger maintenance work,encouraging nodes to actively participate in collaborative sharing and ledger maintenance.Experiments conducted on real traffic trajectory datasets and open-source ledger datasets demonstrate that the convergence and scalability of the ledger have significant advantages.