ABSTRACT With the rapid development of intelligent scenarios such as intelligent transportation and urban perception, delay‐sensitive and computationally intensive applications continue to grow, especially in tasks such as anomaly detection based on machine learning, which puts higher requirements on the real‐time processing capabilities and resource scheduling efficiency of edge computing systems. Mobile edge computing (MEC), as a key supporting architecture, plays a core role in ensuring quality of service (QoS). On one hand, the occupation and release of server resources during task processing lead to dynamic changes in system resources within edge computing networks. Since system resources are often difficult to effectively replenish, services that are well adapted to the current time point may fail to accommodate new service requests at subsequent time points. On the other hand, due to the heterogeneous nature of tasks, resource consumption varies significantly across different task processing, causing some servers to easily become overloaded and unable to meet the processing demands of new tasks continuously. To tackle the challenges presented by the intensified dynamic changes in edge server resources due to task heterogeneity and the difficulty of processing new task requests under high‐load conditions in edge computing scenarios, we first devise a collaborative scheduling and offloading strategy for heterogeneous tasks across multiple edge servers. A task sorting mechanism and priority algorithm based on time groups and score values are designed. Then, with the optimization objective of minimizing task processing latency, a regional resource optimization algorithm based on Deep‐Q‐Network (DQN) is proposed to enable the effective processing of tasks. Finally, extensive experimental results show that this strategy can effectively achieve edge node load balancing, significantly reduce system processing delay, improve overall resource utilization, and has good heterogeneous task adaptability, which is suitable for multiple intelligent scene requirements including anomaly detection.
Zhang et al. (Wed,) studied this question.