Enhanced data-driven fault diagnosis for machines with small and unbalanced data based on variational auto-encoder

Abstract

Abstract Deep learning (DL) has become a popular option for data-driven fault diagnosis, because it can avert the influence of subjective factors in an artificial feature extraction process. However, it also suffers from the adverse effects accompanied with small fault sample and unbalanced data, resulting in limited accuracy improvement. For the aforementioned problem, this paper introduces a variational auto-encoder (VAE) into a fault diagnosis framework to realize data amplification by vibration signal generation, then an enhanced fault diagnosis approach is proposed combining with a convolution neural network. The well-trained VAE can realize the infinite generation of vibration signal by using the hidden variables sampled from Gaussian distribution, and then the generated artificial signals are mixed with the real original signals to form an enhanced training set, which can be utilized for classifier training to realize fault identification. Experimental results show that the generated artificial signals have similar time-frequency characteristics compared with the original real ones, and the enhanced fault diagnosis method holds a higher and more stable recognition accuracy than the unenhanced version and other typical methods.