Mlp, pnn and fuzzy logic improved by genetic algorithms in fault detection and isolation for systems with parameter uncertanties

Abstract EN

In this paper a multilayer preceptor (MLP), a probabilistic neural network (PNN) and a fuzzy approach are proposed in on-line sensors and actuators fault detection and isolation for systems with parameter uncertainties.

The residuals obtained by analytical redundancy relations are used as input of the three systems.

The MLP is trained to present for each output signal I in the occurrence of a fault at the corresponding variable and o otherwise, PNN has one output which is trained to present the fault index (I in case of a free-fault context, 2 in case of a fault affecting variable I…).

these previous approaches are improved by the used of genetic algorithms (GA) to optimization the initial weights and bias in case of MLP, the spread in case of PNN and the membership functions parameter in case of the fuzzy approach.

MLP, PNN and fuzzy approach are compared through a hydraulic example.