A Dual Attention-Based Task Offloading Approach in Computing Power Networks for Object Detection

Intelligent object detection has enabled automatic defect detection at the edges, facilitating smart power transmission inspection and target recognition. However, edge devices often suffer from limited computational resources, resulting in slow model recognition and high energy consumption, making it challenging to meet daily inspection requirements. Computing Power Networks (CPNs) provide a secure and efficient distributed computing solution, enabling the effective offloading of object detection tasks to resource pools within CPNs. Despite numerous studies focusing on optimizing task offloading, some issues persist, including coarse granularity of computing tasks, dispersed computing resources, and weak decision adaptability in complex environments. To address these challenges, we propose a dual attention-based deep reinforcement learning (Dat-DRL) approach, utilizing a custom sequence-to-sequence (seq2seq) neural network to learn effective task offloading strategies. Monitoring tasks at the edge of the power transmission line are modeled as a directed hypergraph, and the dual attention mechanism captures inter-task dependencies and diverse service requirements more effectively. To evaluate the proposed algorithm, we construct a simulation environment to model task offloading scenarios, and extensive experiments demonstrate the efficacy of Dat-DRL in reducing latency and energy consumption across different environments.