Unsupervised deep generative adversarial based methodology for automatic fault detection

System health management is of upmost importance with today's sensor integrated systems where a constant stream of data is available to feed information about a system's health. Traditional methods to assess this health focus on supervised learning of these fault classes. This requires labeling sometimes millions of points of data and is often laborious to complete. Additionally, once the data is labeled, hand-crafted feature extraction and selection methods are used to identify Which are indicators of the fault signals. This process requires expert knowledge to complete. An unsupervised generative adversarial network based methodology is proposed to address this problem. The proposed methodology comprises of a deep convolutional Generative Adversarial Network (GAN) for automatic high-level feature learning as an input to clustering algorithms to predict a system's faulty and baseline states. This methodology was applied to a public data set of rolling element vibration data from a rotary equipment test rig. Wavelet transform representations of the raw vibration signal were used as an input to the deep unsupervised generative adversarial network based methodology for fault classification. The results show that the proposed methodology-is robust enough to predict the presence of faults without any prior knowledge of their signals.