To address congestion and safety challenges in expressway weaving areas, a CAV lane-changing decision-making model integrating a risk field is proposed. The model formulates the CAV decision-making problem as a Markov Decision Process (MDP), designs a 12-dimensional state space containing dynamic risk assessment values, and constructs a risk quantification index based on composite risk field theory for inclusion in a multi-objective reward function. Simulation results indicate that the proposed model reduces the average risk value to near zero and significantly increases the final reward from +180 (conventional model) to +400, effectively balancing safety and efficiency. The Dueling DQN algorithm performs best, demonstrating good adaptability across low, medium, and high traffic flow densities; notably, under high-density congestion conditions, it maintains an 81.6% lane-changing success rate. Furthermore, in multi-scenario testing under extreme conditions such as sudden collisions and dropped obstacles, the model adopts a "defensive driving" strategy, increasing the lane-changing success rate by over 10% compared to the baseline, thereby validating its strong robustness and generalization capability in complex dynamic environments.