Industrial Process Modeling and Fault Detection with Recurrent Kalman Variational Autoencoder

This article proposes the recurrent Kalman variational autoencoder, an end-to-end trainable framework for process modeling and fault detection. Based on the backbone of variational autoencoder, our research focuses on the dynamic and nonlinear properties that implied in the industrial process. For the dynamics describing, the Kalman filter is integrated to estimate the hidden state uncertainty due to process noise, and hence to identify the time-domain correlation in the hidden space. While for the nonlinear evolution in the hidden mapping, the dynamic parameter network is constructed for recurrent updating, so as to further improve the inference accuracy. Finally, for process monitoring, a fault detection panel is designed in the latent and residual domains generated by the deep model. Through the simulation of the industrial benchmark Tennessee Eastman Process, the superiority of the proposed method is evaluated and discussed.