In order to achieve the civil aviation safety management goal of‘safety first，prevention first and com? prehensive management’，a deep learning model is established to learn from reports and assess the risk level. Based on the 10-year incident reports available in the Aviation Safety Reporting System，we first establish quantitative indi? cators of incident consequences and classify all incidents into 5 levels according to their severity：high，moderately high，moderate，moderately low and low risk，which helps to eliminate the impact of unbalanced and intricate event consequences. Then，the relationship between the unstructured incident synopsis and the risk level is explored by convolutional neural network（CNN），and the events are classified by the model to determine the risk level. The clas? sification model proves its superiority by comparing it with different quantitative indicators and methods，with an ac? curacy of 96%，which is better than the compared models. Finally，the 2020’s incident reports are predicted by this model，which enables rapid risk assessment of the synopsis of the incident，with an accuracy rate of 80%. The CNNbased civil aviation risk assessment model can fully mine the text-formatted incident synopsis，and quickly assess and actively perceive the risk level，which helps support the early warning of civil aviation safety.