Fault diagnosis for unknown nonlinear systems via neural networks and its comparisons combinations with RLS based techniques

An algorithm for the detection and diagnosis of faults in sensors and actuators in unknown nonlinear systems is presented. Firstly the normal operation of the nonlinear system is modelled via a feedforward neural network in order to produce a healthy model which is used later as a redundant relation. A residual signal is evaluated from the difference of the output of the healthy model and that of the system. The estimation of the fault is then computed by minimising the differences between the output of system and that of the healthy neuro model with the help of the gradient descent rule. The algorithm is able to estimate a fault with large magnitude accurately. However, if the fault is small, a linearised healthy model, together with the recursive least squares algorithm, can be used to estimate the fault. As a result, a combination of both methods will provide a powerful technique for accurate fault diagnosis.