Fault Detection Based on Variational Autoencoders for Complex Nonlinear Processes

Deep learning models have been proved to outperform shallow methods for industrial process fault detection because of their high capacity for complex nonlinearity. However, typical deep models applied to monitoring processes are conducted in a deterministic manner. They are unable to provide a confidence level for each decision. Also, most deep learning methods often need to integrate prior conditions, such as orthogonal latent variables, constraints, and some given distributions. The consequences of these issues cause lots of trials and errors as conventional deep models are built based on experiences. In this paper, a variational auto-encoder is used to set up a framework to tackle these problems. The learned latent variables, which would be orthogonal to each other, are constrained under the specified and optimized objective. Simultaneously, considering uncertainty in data, probability density estimates of latent variables and residuals instead of point estimates are given to design distribution based monitoring indices. A numerical example validates the effectiveness of the proposed method.