Nonlinear dynamic process monitoring based on latent mapping embedding deep neural networks

In industrial processes, complex nonlinearity and dynamics generally exist, making it challenging to achieve good results using conventional process monitoring methods. In this paper, a latent mapping embedding neural network method (LMNN) is proposed for efficient monitoring of nonlinear dynamic processes. First, a deep neural network (DNN) is employed to acquire features of state variables from nonlinear process data and expand them along with the input to a new feature subspace. Second, a latent mapping (LM) method is used to map the high-dimensional feature subspace to a low-dimensional subspace that includes the most beneficial time series information. Then the entire neural network and regression parameters are obtained through an end-to-end learning manner, through which the nonlinearity and process dynamics are well characterized. Subsequently, prediction error-based residual is generated and the monitoring model is established. The performance of the proposed method is verified through a simulation of penicillin production process and an actual fermentation process of penicillin. Comparisons with state-of-the-art methods are carried out, and results validate the effectiveness and superiority of the proposed method.