The accuracy and effectiveness of fracturing risk identification in oil and gas wells are of vital importance for optimizing resource development, increasing oil and gas production, and planning scientific fracturing construction strategies. In order to effectively prevent downhole fracturing accidents, a discrimination model for fracturing accidents has been established by analyzing on-site fracturing construction data. In this project, the research on sand plug prediction method based on deep learning is carried out to address the problems in fracturing construction, with the goal of improving the safety of fracturing construction and reducing the cost of extraction. First, based on analyzing the morphology of fracturing construction curves, multiple threshold judging methods are used in order to identify the construction stages where sand plugging and pressure tampering accidents may occur. In order to identify fracturing sand plugging accidents, a sand sensitivity index system is established, and the thresholds are optimized using a particle swarm optimization algorithm. Through the method of multi-threshold fusion, accurate identification of sand plugging accidents during fracturing was realized. By analyzing the fracturing construction parameter data and data mining, this paper also reveals the possible hidden information in the time series, and establishes a joint CNN-LSTM network model to predict the time series data of fracturing construction curves 30 s forward. The experimental results show that the joint CNN-LSTM network model can predict the trend more accurately during the fracturing construction process, and the prediction accuracy is greatly improved compared with the traditional LSTM network model.