Nonlinear process monitoring based on decentralized generalized regression neural networks

•A novel nonlinear process monitoring method based on VGRNN is proposed.•The feasibility of applying GRNN in nonlinear process monitoring is investigated.•Comparisons demonstrated the superiority and effectiveness of the VGRNN method. Given that the main task of process monitoring (i.e., fault detection) is actually a classical one-class classification problem, the generalized regression neural network (GRNN) is directly inapplicable for handling process modeling and monitoring issues. Through the selection of only one variable to be the output while the others serve as the corresponding input, a GRNN model can then be constructed to approximate the nonlinear input to output relationship. The residuals, signifying the inconsistency between the actual measurement and the predicted output from the GRNN model, could be a good indicator for online fault detection. The proposed nonlinear process monitoring approach is termed decentralized GRNN (DGRNN), which applies the GRNN in an extremely decentralized manner and utilizes the squared Mahalanobis distance for the online monitoring of the abnormalities captured by the generated residuals. The effectiveness and superiority of the DGRNN-based nonlinear process monitoring approach over other state-of-the-art nonlinear process monitoring methods are investigated by comparisons in two nonlinear processes.