Bayesian approach in the predictive maintenance policy

In Industry, the maintenance policy is devoted to avoid sudden failures that can cause the stop of the system with a consequent loss of production, or - at least - to the minimization of the failure probability and/or the preservation of this probability under a fixed value. In such systems, the use of sensors for the monitoring of their degradation level is very useful. This gives the possibility to follow the time history of the component and to identify the most appropriate time for the maintenance activities, making possible the exploitation of the component for almost its whole useful life. The traditional preventive maintenance policy makes use of the a priori information on the population by assuming a probability distribution function and by estimating the involved statistical parameters [1]. By a monitoring system further information on the stochastic degradation process of the particular component belonging to the population can be available. Nevertheless, such sensors add new costs and exhibit inaccuracy in tracking the stochastic process. This inaccuracy implies an uncertainty in the supplied information. This occurs whether the degradation is defined as a geometric characteristic of the component or as the exhibition of a particular effect. For example, in a cutting tool, wear changes the geometrical characteristics causing an increase of superficial roughness on the machined parts. If a maximum value of roughness is accepted, the condition of failed cutting tool corresponds to the reaching of such value. In this case, the vibration signal is not a correct fault indicator because it is not suitable for tracking the degradation process. For these reasons, a predictive maintenance policy presupposes the identification of a signal well correlated to the degradation process and a high precision monitoring system. Components whose sudden failure can produce dramatic consequences on the system availability are considered. They must operate with a high required degree of reliability and the maintenance policy must assure a reliability level not lower than a pre-defined value. This paper is the second part of two [2], presenting an algorithm for the implementation of a sensor-driven predictive policy based on a Bayesian approach. Simulation results are supplied.