Aiming at the problems of low accuracy and missing information when UAVs utilize a single sensor for obstacle avoidance, this paper designs and proposes an autonomous UAV obstacle avoidance method based on multi-sensor fusion. The improved Bayesian fusion algorithm contributes to the multi-sensor fusion, considers the use of multiple UAVs to perform power system inspection tasks collaboratively, and utilizes deep reinforcement learning for multi-UAV inspection path optimization. On the basis of the AnoGAN anomaly detection algorithm, the performance enhancement optimization of the anomaly detection technology is carried out, and a SE-f-AnoGAN model for anomaly detection of UAV power inspection images is designed. The model draws on the idea of attention mechanism, and introduces a compressed activation network based on channel attention into the encoder of fAnoGAN, which captures the information of each channel from the global field of view category, so as to improve the accuracy of anomaly detection. Deep reinforcement learning multi-drone optimization path and multi-drone inspection image anomaly detection techniques are performed for model training and performance analysis, respectively.The DQN algorithm is designed to enable mobile drones to complete collision-free inspection path planning, and can continuously shorten the inspection path through training and learning to save inspection time.The SE-f-AnoGAN model has a high accuracy and precision rate in different dataset categories.
