Optimal preventive maintenance schedules using specific genetic algorithms and probabilistic graphical model
Résumé
Equipments used in industrial environments such as production lines, engineering or mass transport system, are generally complex, multi-components and multi-states systems (MSS). These equipments are subject to degradation mechanisms caused by operating conditions/environment (temperature, vibrations). In addition to these degradation mechanisms, the deployed maintenance policy affects directly the dynamics of the occurrence of failure states. Given this situation, we should consider establishing preventive maintenance (PM) strategies to ensure an adequate trade-off between system availability and its maintenance costs. Solving this issue requires a prior modeling of the system degradation. This modeling must represent faithfully the evolution of the operating states of a multi-components system, during time. Given that, an evaluation model of PM policies can be considered, in order to look for the optimal schedules of the PM. A common way, in PM optimization, is to assume algorithms relying on classical degradations modeling approaches like deterministic models, stochastic processes or Markov chains. These approaches allow optimization algorithms to go faster, but they require, necessary, exact knowledges of degradation processes or strong assumptions about sojourn-time distributions. This limitation can be overcomed by the use of a particular structure of Probabilistic Graphical Temporal Model named Graphical Duration Models (GDM). GDM allows to represent duration models of MSS, regardless of the exact nature of their sojourn-time distributions. This work can be divided mainly into two steps. The first one proposes an utility model for the evaluation of maintenance policies. This model is mainly based on the GDM. It involves essential parameters like maintenance probabilities, utilities or system availability. In the second step, two Genetic Algorithms (GAs) were developed. They seek for the optimal PM schedules. The 1st one named GA1 seeks for periodic schedules contrarily to the 2nd one, GA2, which look for non-periodic schedules. GA operators are redesigned according to the specific characteristic of the problem. To assess the proposed methodology, a Distribution Fluid System (DFS) has been chosen as case study. This approach provides good results and allows a specific analysis of the weight of the different parameters of the utility function.