A smart vehicle decision-making model based on deep reinforcement learning and risk assessment is proposed to solve the problem of safe driving decisions for vehicles in highway environments. Firstly, a Bayesian based position uncertainty quantification method is proposed for modeling and quantifying driving risks; Then, a self attention mechanism is introduced into the decision model to help vehicles perceive potential dangers in complex scenes and avoid dangerous decision execution; Finally, a simulation environment was constructed on the Highway env simulation platform, and the model was trained and tested using simulation experiments. Multiple experimental comparisons were also designed. The results show that the proposed RA-PPO-Mul model achieves 98% safety rate and higher driving efficiency, which is superior to the traditional reinforcement learning model and the model that only introduces a single module.