Kernel Regression and Neural Networks for Model-Free Fault Diagnosis

Correct and timely fault detection is of major importance in the field of system engineering, and constitutes a primary problem in a broad spectrum of cases, from industrial processes to high-performance systems and to mass-produced consumer equipment. A large number of methods can be found in the literature, and the recent use of non-parametric techniques and neural networks for solving fault-diagnosis problems in real industrial situations seems to be particularly promising. This paper describes a novel approach combining kernel-regression methods and neural networks to solving approximately some difficult fault-diagnosis (FD) problems. The method falls into the model-freeclass of FD techniques as no model of the plant is needed. The kernel-regressor makes it possible to detect changes in the plant dynamics, possibly due to some malfunction, whereas the neural network is used to build a decision scheme to actually diagnose the fault. A theoretical sufficient condition for fault detectability is also illustrated.