Deep Variational Autoencoder Classifier for Intelligent Fault Diagnosis Adaptive to Unseen Fault Categories

With the rapid development of artificial intelligence (AI) in recent years, fault diagnostics for industrial applications have leaped toward partially or fully automatic provided by the capability of analyzing massive condition monitoring data from sensors and actuators. Generally, AI-based fault diagnostics can achieve high accuracy when failure types appear in training dataset and testing dataset are the same. These diagnostic methods could be invalidated for applications dealing with unprecedented faults because the pretrained classifier for diagnostics tends to misclassify the novel instances into existing known classes. In order to address these limitations of conventional diagnostic approaches, we propose a unified diagnostics framework that can achieve novel fault detection and known fault classification tasks together. Through jointly training a variational autoencoder and a deep neural networks classifier, we convert the original entangled raw data into latent variables with Gaussian probabilistic distributions in the latent space and utilize the probabilistic latent variables to detect novel samples against known fault classes or classify them into one of the existing fault classes if they are not novel. The effectiveness of our proposed joint-training framework is validated through experimental studies on two different bearing datasets. Compared with the state-of-the-art methods in the literature, our unified framework is able to not only accurately detect the novel fault classes but also achieve high classification accuracy of known fault classes.