Personnel violations in power construction sites pose significant safety risks and may lead to serious accidents. This paper proposes a method based on multi-sensor fusion and deep neural networks to recognize staff violations and fully understand on-site personnel behavior. By integrating data from multiple sensors, including visual and distance sensors, this method allows for precise personnel positioning and detailed behavior analysis. Fine-grained behavior analysis and recognition can be achieved by incorporating 3D point cloud data to obtain three-dimensional spatial information of personnel, and combining texture features of visible light images. The method first performs pre-processing and feature extraction on multimodal data, followed by integrating information from both data sources using a fusion strategy, and constructing a behavior recognition model using deep neural networks. Experimental results in complex real-world scenarios show that this method significantly outperforms single data source approaches in terms of recognition accuracy and robustness, effectively enhancing personnel behavior recognition in complex ground environments.