Abstract: To guarantee continuity and decrease maintenance cost of a complex mechanical assembly system under uncertain environment, a dynamic selection method of maintenance policy based on Markov decision is put forward. On the basis of the comprehensive consideration to the cost of operating, inventory, maintenance and the cost due to the lost production, a reliability-cost model is established. The model takes the secondary production system as the research object, equipment condition and the buffer inventory as independent variables, and reliability-cost as objective function. The best maintenance policy in different conditions of production system is derived by simulated annealing algorithm and fuzzy nonlinear programming. The method decreases downtime, reduces equipment maintenance frequency, and provides basis for designing the production line and scheduling the maintenance plan. The calculation analysis of samples proves the effectiveness of the model.