Representation learning based adaptive multimode process monitoring

In this paper, a representation learning based adaptive monitoring method for multimode processes is proposed, in which mode identification and fault detection are integrated with an adaptive threshold strategy. Compared to conventional methods, the learned representations can integrate mode features with fault details here, which are formed by the composition of multiple non-linear transformations under a global modeling. To explore the expressive powers of AE net, a geometry interpretation based on the reformulated sigmoid function is presented. Moreover, take the significance of the dynamic characteristics into account, an adaptive thresholding scheme is proposed for the learned representations based on a modified exponentially weighted moving average (EWMA) control chart. Experiment results show that the proposed method not only improves the divisibility between multimode, but also exhibits superior performance of fault detection on an industrial benchmark of chemical process, Tennessee Eastman process (TEP).