In this paper, we consider a system which deteriorates stochastically from one time unit to another. The state of the system, which is a continuous random variable, can be observed only by inspection. The goal of this paper is to derive a predictive maintenance policy which indicates, at each inspection and according to the observed value, whether a preventive maintenance is necessary and when the next inspection should be done. The objective is to minimize the long-run average cost incurred by inspections, preventive maintenance and unexpected breakdowns. The problem is modelled by a semi-Markov decision process. Analytical properties are derived, and based on these properties, numerical methods are constructed to compute an optimal policy.