<p>The severe safety situation of construction engineering urgently requires research on active risk prevention and control. Based on the multimodal data fusion theory, this paper constructs a construction engineering operating state monitoring and risk prediction model. It solves the multimodal distribution solution for missing data in construction engineering flow. By using adversarial generative training to determine the multimodal distribution of missing data in construction engineering flow, the model develops a multi-filling strategy based on data fusion and suggests a new time-series generative adversarial network model to realize the filling of missing data in construction engineering flow. A deep feature clustering layer is intended to iteratively optimize the clustering effect and feature extraction process. The self-encoding network is utilized to extract the risk feature information from the construction engineering parameters. Finally, the missing data imputation performance of the proposed method is verified by using the actual construction engineering flow parameter data. The experimental results show that the model selects detection rate (false detection rate = 5% and 10%) and false detection rate (detection rate = 90% and 95%). The performance of risk prediction and detection of different algorithms and the risk prediction level of each construction project state are quantified, which significantly improves the effectiveness of the proposed method for risk state classification.</p>