We investigate the decision-making problem of condition-based maintenance for multi-unit series systems with non-identical units subject to internal deterioration and external shocks. Two types of interdependencies are considered: (1) stochastic and economic dependencies among units, where the deterioration of one unit affects others, and simultaneous maintenance reduces costs; and (2) interactions between internal deterioration and external shocks, where accumulated deterioration increases susceptibility to shocks, and shocks accelerate deterioration. These interdependencies and the high-dimensional nature of the decision-making problem make traditional algorithms, such as value iteration and policy iteration, computationally infeasible for large-scale systems. To address this challenge, a deep reinforcement learning (DRL) algorithm is adopted. The DRL-based approach effectively captures these interdependencies and derives maintenance policies, demonstrating its adaptability to real-world systems with complex dependencies and high-dimensional state spaces. Furthermore, the DRL-based policy achieved lower maintenance costs and downtime due to failures than the alternative policy, which optimizes individual unit policies. This study contributes to advancing the application of DRL in reliability engineering and maintenance management. The study concludes with a discussion of the limitations of the proposed policies and suggestions for future research.

View full abstract

Even if repeated measurements are performed under the same assumptions, the results will vary. This variation is called measurement precision. The dose-response relationship refers to the relationship between the dose of a substance, such as a drug, and the response of an organism. It is useful for pharmaceuticals and safety evaluations. However, while a methodology, ISO 5725, exists for quantifying measurement precision, it only applies to quantitative data, such as length and weight, and not to qualitative data, such as the presence or absence of a response, which is the subject of this study. Therefore, the authors propose an accuracy evaluation model based on logistic regression that considers the dose-response relationship. In the model, the between-laboratory variance can be quantified by two indices, the position and slope of the dose-response curve, to determine which variance is relatively larger. The consistency between true values and estimated values was confirmed by simulation. and the within-laboratory variance can be estimated to obtain the variance of residuals in single regression analysis. By applying the proposed model to actual data, the measurement precision could be quantified and demonstrated. This model can improve the reliability of measurement precision evaluations in qualitative data considering the dose-response relationship.

View full abstract