Modern manufacturing systems are marked by high levels of complexity and uncertainty. To address diverse production goals, it is imperative to develop scheduling methods that can effectively balance solution quality with robustness. This study addresses the flexible job shop scheduling problems (FJSP) involving urgent job insertions and machine breakdowns by proposing a Dueling Deep Q-Network (DDQN) framework based on Deep Reinforcement Learning (DRL), aiming to minimize tardiness while reducing the standard deviation of all machining loads. Firstly, due to the frequent changes in job shop status caused by dynamic events, multiple state features were designed to accurately reflect the job shop environment. Secondly, specific action rules were proposed for the joint decision-making of job selection and machine assignment. Then, considering the coordinated optimization of multiple objectives, a reward function was defined based on a combination of multiple state features. Finally, experiments on 18 simulated production instances demonstrated that the DDQN framework outperformed other methods, showing effectiveness and generalizability.