Integrated maintenance and spare parts inventory optimization with transshipments for multi-fleet systems

Maintenance of distributed engineering asset fleets frequently relies on multi-location spare parts inventory systems that incorporate reactive transshipment mechanisms. The intrinsic interdependence between maintenance activities and cross-warehouse spare parts inventories necessitates the joint optimization of maintenance and inventory policies across the entire multi-fleet system. This joint optimization problem poses substantial computational challenges due to its high-dimensional state and action spaces. To tackle this complexity, this study adopts a multi-agent deep reinforcement learning framework that decomposes the monolithic problem into manageable subproblems. The system is modeled as a decentralized partially observable Markov decision process (Dec-POMDP) and solved using the value decomposition twin delayed deep deterministic policy gradient (VDTD3) algorithm. Furthermore, the learning process is guided by a teacher policy generated via an approximate method. Additionally, the local interpretable model-agnostic explanations (LIME) are utilized to elucidate the policies derived from VDTD3. Numerical experiments indicate that the VDTD3 approach consistently surpasses benchmark methods in terms of both solution quality and computational efficiency. Moreover, LIME-based analysis validates that the resulting maintenance and inventory policies conform to operational intuition.