For identification of human malicious behaviors in nuclear security, a novel framework is proposed in this research, including computer vision module (CVM) and reasoning module (RM). In CVM, multiple CV-based deep learning models are combined to extract multiple human actions. Then in RM, we developed three types of novel reasoning methods which could comprehensively analyze the combination and sequential features of extracted human actions in an explainble way.Furthermore, to address the insufficient datasets due to the confidentiality of nuclear engineering, a self-collected dataset including shot-videos and game-engined data is utilized. The results show that our proposed framework is promising.