Multirate Dynamic Variational Autoencoder for Fault Detection in Nonlinear Industrial Processes

Abstract

The variational autoencoder (VAE) has proven highly effective in monitoring nonlinear stochastic processes, primarily under the assumption of complete and temporally independent data. However, real-world industrial data often exhibit additional complexities, such as multiple sampling rates and dynamic behavior, which complicate the development of efficient latent variable models for nonlinear systems. To address these challenges, this study proposes a novel multirate dynamic VAE (MDVAE) model, specifically designed for dynamic industrial fault detection with incomplete dataset. In MDVAE, both the encoder and decoder are adapted to a multirate structure, enabling the latent variables to capture nonlinear correlations across varying sampling rates. Additionally, a first-order Markov chain is applied to the latent variables to represent the dynamic behavior of multirate nonlinear systems. The effectiveness of MDVAE is demonstrated using the Tennessee Eastman and coal-fired power generation processes (CFPs). The experimental results show that the proposed model outperforms comparable approaches in addressing the coexisting challenges of multirate sampling and nonlinear dynamics.