Mutual Information-Dynamic Stacked Sparse Autoencoders for Fault Detection

Data-based process monitoring is gaining increasing attention, especially deep learning modeling methods. Given that process data are inherently dynamic, the dynamic relationship of process data needs to be incorporated into the monitoring model. Measured variables have diverse characteristics; some variables have strong autocorrelation, but some have no or weak autocorrelation. Hence, a new nonlinear dynamic fault detection model based on mutual information (MI) and stacked sparse autoencoders (SSAE) was proposed. With input data taken as time series data, variables with a strong autocorrelation and their lag time of maximum autocorrelation were selected. Then, dynamic augmented data that contain the values of all measured variables at the current moment and the selected data were used as input of SSAE to extract features of cross-correlation and autocorrelation between variables. T-2 and Q statistical quantities were constructed to implement fault detection. The proposed MI-dynamic SSAE can not only automatically learn the key features of the present measurements but can also learn the latent information contained in the dynamic process data. Simulation studies on the Tennessee Eastman process were conducted to evaluate the performance of the proposed monitoring model.