This paper is concerned with fault prediction and maintenance for the hydraulic tube tester according to the process variables. For the two types of incipient faults of this machine, it is more difficult to predict the non-periodic incipient fault and make a predictive maintenance decision in comparison with the periodic incipient fault. For the non-periodic incipient fault, a new probabilistic fault prediction method is proposed using a multivariate principle component analysis and Bayesian auto-regression model. Then, a new predictive maintenance policy is developed, based on the results of the probabilistic fault prediction approach, which implies the evolvement of system deterioration caused by the non-periodic incipient fault. Simulation results demonstrate the effectiveness of the method proposed in this paper.