The Fault Diagnosis Problem: Residual Generators Design Using Neural Networks in a Two-Tanks Interconnected System

In this work, a fault detection method based on a neural-network models bank to residual generation and a residual evaluation scheme using a fuzzy rules type is developed. The case of study is a nonlinear hydraulic system consisting of two interconnected tanks which is simulated, in normal conditions and fault conditions. In this case we use also its equivalent Takagi-Sugeno Model in discreet time. This way, the simulation provides the data to train each one of the neuronal models. The update the weights is based on the algorithm BP (Back Propagation) with a stage of scale applied to the training data in order to avoid over-training on the neural network, due to the asymptotic limits of the sigmoid function used. The results show a correct identification on the different fault scenes and it motivates us to the real implementation of faults diagnosis procedure.