Research on Task Allocation for Continuous Transmission Line Inspection Based on Reinforcement Learning

This paper systematically explores the task allocation challenge within the realm of power inspection, particularly addressing the issue that power inspections often cannot be completed by a single drone due to limitations in power endurance, UAV hardware, and task configurations. Consequently, multiple drones must perform the inspection in a continuous manner. Leveraging the sophisticated framework of reinforcement learning, the study commences by delineating a comprehensive modeling methodology, meticulously elucidating the pivotal components of task allocation, namely the state space, action space, and the reward function design. These foundational elements establish the theoretical underpinnings for the subsequent deployment of reinforcement learning algorithms. Subsequently, the Deep Q-Network (DQN) algorithm is adopted, wherein we initially train individual task allocation models tailored for single and dual unmanned aerial vehicle (UAV) scenarios. Subsequent to this, these models are integrated to construct a hybrid reinforcement learning framework. The empirical findings reveal that both the single UAV task allocation model and the dual UAV task allocation model exhibit distinct limitations: the former demonstrating shortcomings in safety, while the latter lacks efficiency. When comparing the composite reinforcement learning model with the traditional reinforcement learning model, it is found that the former has higher accuracy when deciding the number of inspection drones needed. Moreover, the composite reinforcement learning model performs excellently in power inspection tasks, significantly improving the efficiency of task completion.