A Novel Fault Detection Method Based on Reconstruction Error and Clustering of Latent Variables

Abstract

Traditional fault detection methods based on AutoEncoder (AE) usually complete fault detection by comparing reconstruction errors, and ignore a lot of useful information about the distribution of latent variables. In this paper, we propose a novel unsupervised fault detection method named One Dimension Convolutional Adversarial AutoEncoder (1DAAE), which introduce two new ideas: 1D convolution layers for autoencoder to get better features and the adversarial thought to impose the latent variables z to cluster into a prior distribution. Then two anomaly scores are proposed to detect fault samples, one is based on reconstruction errors, the other is based on latent variables distribution. Finally, it is shown by experiments that the proposed method outperforms traditional AE-based, Adversarial AutoEncoder (AAE)-based, One Dimension Convolutional AutoEncoder (1DAE)-based, and 1DAAE-based algorithms on Tennessee Eastman Process.