New contributions to non-linear process monitoring through kernel partial least squares

• An efficient methodology for synthesizing a reliable KPLS model is proposed.
• An extended KPLS modeling is used for decomposing the process measurements.
• Four independent statistics are proposed for monitoring a non-linear process.
• The proposed diagnosis tool is able to discriminate among several types of faults.
• A prediction-risk assessment of the online KPLS soft-sensor is presented.

The kernel partial least squares (KPLS) method was originally focused on soft-sensor calibration for predicting online quality attributes. In this work, an analysis of the KPLS-based modeling technique and its application to non-linear process monitoring are presented. To this effect, the measurement decomposition, the development of new specific statistics acting on non-overlapped domains, and the contribution analysis are addressed for purposes of fault detection, diagnosis, and prediction risk assessment. Some practical insights for synthesizing the models are also given, which are related to an appropriate order selection and the adoption of the kernel function parameter. A proper combination of scaled statistics allows the definition of an efficient detection index for monitoring a non-linear process. The effectiveness of the proposed methods is confirmed by using simulation examples.