Model-free fault diagnosis for nonlinear systems: a combined kernel-regression and neural networks approach

A novel way of using the kernel regression methodology in the context of model-free fault diagnosis for nonlinear systems is proposed. The basic qualitative idea is: when a fault occurs, some changes in the smoothness characteristics of the time-behaviors of the measurable variables may also occur. This changes are reflected in modifications to the typical features of the kernel smoother applied over some suitable temporal batch of the measurable variables, and this could be interpreted as a fault symptom to be fed into the decision scheme based on a neural classifier. The neural classifier may be trained off-line to associate the fault symptoms with some eventual critical behavior of the plant. We briefly describe the kernel smoothing technique in the context of dynamic systems. The statements of some basic definitions are also be provided