A Novel Fault Detection Method Based on One-Dimension Convolutional Adversarial Autoencoder (1DAAE)

Abstract

Fault detection is an important and demanding problem in industry. Recently, many researchers have addressed the use of deep learning architectures for fault detection applications such as an autoencoder. Traditional methods based on an autoencoder usually complete fault detection by comparing reconstruction errors, and ignore a lot of useful information about the distribution of latent variables. To deal with this problem, this paper proposes a novel unsupervised fault detection method named one-dimension convolutional adversarial autoencoder (1DAAE), which introduces two new ideas: one-dimension convolution layers for the encoder to obtain better features and the adversarial thought to impose the latent variable z to cluster into a prior distribution. The proposed method not only has powerful feature representation ability than the traditional autoencoder, but has also enhanced the discrimination ability by imposing a prior distribution of the latent variables to cluster. Then, two anomaly scores for 1DAAE were proposed to detect fault samples, one based on reconstruction errors, and the other based on latent variable distribution. Finally, it was shown by the experiments that the proposed method outperformed the autoencoder-based, adversarial autoencoder-based, one-dimension convolutional autoencoder-based and generative adversarial network-based algorithms on the Tennessee Eastman process. Through the experiments, we found that the both one-dimension convolution layers and the latent vector distribution are helpful for fault detection.