As a common mechatronic special equipment, elevators provide significant convenience for people"s travel. However, the increased risk of elevator failures poses challenges and threats to passenger safety. To address this, this paper proposes a condition-based predictive maintenance strategy for traction elevators to enhance their operational reliability. First, the degradation patterns of critical elevator components and subsystems are analysed, the Wiener process is selected as their theoretical degradation model. Among them, the deep learning technology is integrated to assess the condition of components that can be monitored in real time. Considering factors in actual maintenance processes, a virtual service life model for key components is established based on imperfect maintenance. Subsequently, an intelligent Markov decision model is established to seek optimal maintenance strategies. In particular, a state probability space for crucial components and subsystems is constructed and the impact of maintenance strategies is described using state transition probability models. A reward feedback model is established to penalize maintenance decisions while comprehensively considering various cost factors in the maintenance process. Finally, a deep Q-network reinforcement learning approach is employed to solve the Markov decision model, determining the optimal maintenance strategy for critical components within the current state space. By utilizing degradation data and online operational data from several critical elevator components, and comparing the proposed method with the current periodic maintenance strategy for elevators, heuristic threshold strategies, and traditional Q-learning, the effectiveness and validity of the proposed method were verified.