Data-driven individual-joint learning framework for nonlinear process monitoring

Industrial processes are generally characterized by significant nonlinearity, and the monitoring of nonlinear processes remains a challenge. This study proposes a novel data-driven individual-joint learning (IJL) framework for achieving efficient nonlinear process monitoring. First, individual learning is performed by separately establishing stacked autoencoders in process input or output variables to characterize the variable relationship within process input or output. Second, joint learning is performed between the input and output variables to characterize their relationship and extract deep correlated representations. Subsequently, fault detection residuals and statistics are constructed to examine the process status. Given the superiority of deep neural network in representation learning, the complex relationship among process variables can be efficiently characterized, and satisfactory monitoring performance is then obtained. IJL monitoring is tested on the Tennessee Eastman benchmark process and applied on a glycerol distillation process, through which its effectiveness and superiority are demonstrated.