A Novel Fault Detection Method Based on Adversarial Auto-Encoder

In this paper, we propose a novel unsupervised fault detection method based on adversarial auto-encoder (AAE). The proposed method not only has powerful feature representation ability as auto-encoder (AE), but also has enhanced discrimination ability by imposing prior distribution for the latent variables. The network architectures of AAE include three parts: the encoder, the decoder and the discriminator for latent variables. The original operational data can be considered as sampling from a low-dimensional latent variables subspace. The encoder tries to learn the latent variables subspace from high-dimensional operational data, the decoder tries to reconstruct the operational data from the subspace. The discriminator helps AAE impose the latent variables subspace to follow a prior distribution. Then, a novel fault detection method based on AAE is proposed. Finally, it is shown by experiments that the proposed method outperforms traditional AE on Tennessee Eastman Process.