The scale of modern civil engineering projects has been expanding, and traditional management methods have been difficult to meet the dual requirements of efficiency and safety. This study explores the application of automation technology in civil engineering project management and constructs an unsafe behavior identification model based on deep learning. Through the integration of on-site monitoring system and schedule control technology, it realizes accurate quality management and safety risk warning of engineering projects. The study designs an unsafe behavior recognition model based on CNN-LSTM, which combines the Inception-v3 framework to extract spatial features, and captures temporal dynamic features through a two-layer LSTM network to realize the intelligent recognition of workers’ unsafe behaviors. The experiments use UCF-101 public dataset and self-constructed construction site dataset for model training and validation, and the results show that the model has a recognition accuracy of 94.52% on the UCF-101 dataset, with a computational complexity of 8.28G, and a parameter count of only 6.16M, which is a 5%-11% increase in the average accuracy value compared with that of the traditional methods. In actual engineering applications, the system collected more than 1,200 pieces of information on personnel’s “three violations” in one year, effectively reducing the incidence of unsafe behaviors on site. The study proves that automation technology combined with deep learning model can effectively improve the safety management efficiency of civil engineering projects and provide technical support for intelligent supervision of engineering.
