Predictive maintenance policy optimization by discrimination of marginally distinct signals

The necessity of discrimination of marginally distinct measured signals is one of the main problems in the creation of maintenance policies. Applying classical methods of statistical classification of observations to the solution of this problem entails considerable difficulties caused by the need to discriminate signals in a gray area between absolutely healthy and fully degraded system. In the gray area, the difference between population parameters inferred from samples is hardly noticeable. Instead of the classical discrimination criteria, a discriminant function that minimizes the expected sum of losses (for example, losses of time) relevant to system preventive maintenance and recovering of the system after its failure is used. This discriminant function is developed on the basis of the representation of the observed system degradation process as a Discrete Parameter Markov chain. The extremum of this function determines the discrimination boundary and the optimal time for maintenance. The requirements for the possible deviation of the experimentally obtained Markov process parameters that do not invalidate the obtained optimal rule of maintenance are specified. The developed methods are illustrated on synthetic data reminiscent of the operation of a database management system.