Alternatywne podejscie do wykrywania uszkodzeń w układach dynamicznych na podstawie porównania wag wyjsciowych neuronowych sieci typu RMLP

The complexity of technological processes needs the study and development of computer based fault detection and diagnosis method enabling process faults be detected and localized during normal plant operation. In this paper we propose fault detection method based on a simple arithmetic relations of output layer weight values of the model RMLP (Recurrent Multilayer Perceptron) networks, assuming each of the model neural networks possesses only one output layer neuron. We build a neural model bank of model neural networks designed and trained on the different operating points of an arbitrary assumed dynamic system. We consider 5 different operating points, where the first state is taken to be the normal operation point (no fault) of the system and the rest four states are different faulty states of the same system. For each of these operation points a neural network is designed and trained. After the training, the output layer weight values of each of the trained neural networks are registered to be used as inputs to calculate a certain value. Based on the comparison of the values, we make conclusion to which of the 5 pre-defined states does a new assumed unknown system may belong.