An Online Reduced KPLS Data-Driven Method for Fault Diagnosis of Nonlinear Processes

System security is a very important organizational task for the system to maintain proper functioning and to prevent modifications or hijacking of the system. Indeed, it is necessary to address any detected problem or defect to protect human beings, industry, and machines. So the identification, after the fault detection phase, of the variables correlated to the detected or occurred fault is a very important step. For this purpose, this paper proposes a nonlinear machine learning method for fault diagnosis. Indeed, the Reduced Kernel Partial Least Squares (RKPLS) is proposed as a processing method for the suitable localization of detected faults. The idea of this approach is to generate partial RKPLS models, using the principle of structured symmetry residues, with reduced sets of variables. On the other hand, the Fault Isolation (FI) using the online RKPLS method (ORKPLS) is developed in this article to generate indices of fault detection sensitive to certain faults and insensitive to others. Thus, a partial ORKPLS method, for fault isolation, is proposed to secure the systems and ensure a proper operation. The suggested approaches are applied for monitoring the continuous stirred tank reactor (CSTR) and the Air quality monitoring network (AIRLOR). The obtained results underscore the role of leveraging symmetry in designing fault.