Incipient faults have low amplitudes and can be easily covered by system disturbances and noises. As timely incipient fault detection is the key to guarantee operation safety and suppress fault deterioration. In this paper, a novel incipient fault detection method based on robust support vector data description (RSVDD) is proposed. On the basis of traditional SVDD, both normal samples and faulty samples are introduced for RSVDD modeling and the computation of sphere radius is improved. Furthermore, a novel self-learning method based on restricted Boltzmann machine (RBM) and probabilistic neural network (PNN) is proposed for fault identification. The benefits of the proposed RSVDD and RBM-PNN scheme are illustrated by Tennessee Eastman benchmark. (C) 2019 Elsevier Ltd. All rights reserved.