Abstract: Human behavior recognition can be used in power inspections to ensure the safety of personnel in complex and harsh environments. In order to solve the diversity and uncertainty of human behavior activities, this paper defined a probability framework based on atomic activities, used Allen time series interval relations to express local time dependence, and established a complex activity recognition model based on probabilistic time series intervals. We applied it to power patrol inspection to finally realize the functions of personnel's dangerous action detection and alarm. Experimental evaluation results show that the proposed method is significantly better than other complex activity recognition methods, and can effectively recognize complex power inspection behaviors.