One-Dimensional Residual GANomaly Network-Based Deep Feature Extraction Model for Complex Industrial System Fault Detection

Abstract

In the era of industrial big data, traditional shallow machine learning-based data analytical technologies can not handle complex industrial system fault detection issue effectively. In regard of this problem, this paper presents a deep learning-based fault detection method, called one-dimensional residual GANomaly (1DRGANomaly). This method builds a semi-supervised deep feature extraction mechanism where the unsupervised generator network is utilized to capture the latent data representation and the semi-supervised discriminator network is adopted to improve the reconstruction quality. Furthermore, one-dimensional convolution operation is applied to mine the local variable relationship and the residual learning blocks are designed for the efficient training. Based on the deep features and the reconstruction errors, the Bayesian inference is applied to construct two overall statistics for monitoring the process status. In order to locate the possible fault source variables, an input importance attribution method is developed by referring to the idea of the Shapley value, which gives a reasonable explanation for the detected fault. Finally, a benchmark industrial process, the Tennessee Eastman system, is applied to perform the method comparison and the application results demonstrate that the 1DRGANomaly method can detect the faults more effectively than the other existing methods.