Abstract
Globalized manufacturing and service systems depend on complex international supply networks that must provide the right spare parts at the right time to geographically dispersed maintenance sites. Traditional maintenance strategies that rely on fixed-interval schedules or corrective interventions after failure can generate lumpy, poorly forecasted demand for spare parts, increasing the risk of local stockouts and reactive emergency logistics. At the same time, advances in sensing, connectivity, and data storage have created large multivariate time series describing condition, usage, and environment for critical equipment. Machine learning techniques can exploit these data to estimate the probability of failure or remaining useful life, offering the possibility of transforming maintenance demand for spare parts from a largely stochastic phenomenon into a partially controllable process. This paper examines how machine learning enabled predictive maintenance can be mathematically integrated with multi echelon international spare parts networks to reduce parts shortages. The study develops a linear algebraic representation linking failure risk predictions, maintenance scheduling decisions, and inventory state across locations, and combines this representation with probabilistic machine learning models for component degradation. A simulation based evaluation with stylized yet realistic network structures illustrates how the coupling between predictive models and inventory control policies influences stockout frequency, fill rate, and inventory investment under varying lead times and forecast error profiles. The discussion highlights conditions under which predictive maintenance most strongly affects spare parts availability, emphasizes the role of uncertainty quantification in the predictions, and outlines practical considerations for embedding such models into existing planning processes.