Pattern modeling and fault detection based on dynamic controlled autoencoder

Industrial processes often exhibit significant nonlinear and dynamic characteristics. To effectively monitor these processes, this paper proposes a dynamic controlled autoencoder (DCAE) model for pattern extraction, which primarily consists of an autoencoder and dynamic mapping components. It is capable of simultaneously extracting the nonlinear structural relationships of process variables in static space and their nonlinear dynamics in the time domain, and in particular, establishing the dynamic causality between control input and pattern. The dynamic controlled pattern extracted using DCAE can sufficiently represent the operation information of the nonlinear process. Then, the relationships between DCAE modeling errors and model variables are explored, leading to the construction of error statistics for monitoring industrial processes and the development of a DCAE-based fault detection scheme. Finally, the case study of an industrial boiler combustion system illustrates the effectiveness and superiority of the DCAE model in extracting the pattern of industrial processes and performing fault detection.