Maintaining and deploying a fleet of aircraft with limited resources and various mission requirements is both immensely challenging and of primary importance. Traditional preventive maintenance methods are static and inflexible and are not equipped to consider the complex dynamics of aircraft (e.g., wear and age), which may lead to low fleet availability and high maintenance costs. In this paper, we propose an integrated learn-then-optimize framework for condition based predictive maintenance scheduling to support daily flight and maintenance planning by fusing data from multiple onboard sensors. The paradigm first predicts the remaining useful life for components of aircraft by using deep learning techniques, then models the fleet level optimization as a constrained mixed-integer programming problem that captures different failure modes of aircraft and the available maintenance facilities. We also propose valid inequalities to improve the computational efficiency of the optimization model. Finally, we conduct a series of simulated experiments to validate the performance of the proposed predictive maintenance model. The numerical results show that the predictive maintenance model outperforms the traditional preventive maintenance model with respect to the mission accomplishment rate, aircraft availability rate, and cost effectiveness.