Adaptive Nonlinear Recursive Model Updating Method for Quality-Related Process Monitoring

In data-driven nonlinear modeling, kernel partial least squares (KPLS) is a useful tool for quality-related process monitoring. However, in time-varying nonlinear processes, degraded equipment and parameter drift require nonlinear models to be updated regularly. If a model is not updated, it is difficult for the nonlinear model to track the changes in the system, and the process monitoring system has a high false alarm rate. To our knowledge, there is a lack of research on the adaptive updating of quality-related nonlinear models. To address this problem, an adaptive nonlinear recursive model updating method is proposed. In the proposed method, a nonlinear orthogonal regression algorithm, named NOKPLS, which is suitable for model updating, is first developed to construct a nonlinear model. Specifically, by constructing an orthogonal projection operator, the input space can be divided into mutually orthogonal quality-related subspaces and quality-unrelated subspaces, and false alarms for quality variables caused by quality-unrelated faults can be greatly reduced. Furthermore, on the basis of the NOKPLS model, an adaptive nonlinear model updating method is developed with low computational complexity. Finally, a quality-related process monitoring strategy for recursive nonlinear model updating is designed. A numerical simulation and the Tennessee Eastman process (TEP) are applied to demonstrate the effectiveness of the proposed method. Note to Practitioners—Monitoring and controlling the key performance indicators is crucial for ensuring the healthy operation of nonlinear processes. In practice, however, the aging of the components and the drift of the parameters often lead to slow time-varying system processes. Therefore, the nonlinear model needs to be updated regularly. Here, a recursive nonlinear PLS model adaptive updating method is proposed. The nonlinear orthogonal regression algorithm is first developed to reduce the number of false alarms caused by quality-unrelated faults and to prepare for model updating. Considering the high computational complexity of kernel-based nonlinear methods in modeling, a new model update matrix is constructed by using model parameters instead of historical data, which effectively reduces the computational complexity and improves the efficiency of model updates. Then, a complete quality-related nonlinear monitoring and adaptively updating strategy is designed. In the proposed approach, the number of false alarms induced by time-varying processes can be reduced, and unnecessary overhaul and maintenance can be avoided. In addition, the superiority and characteristics are discussed and studied experimentally.