Predictive diagnosis of research reactor primary pump failures using artificial neural network and vibration data

Abstract

Neural networks are seldom employed in research nuclear reactors, particularly regarding applications involving vibration measurements. The existence of a vibration database, combined with expert analysis of these data, inspired the development of an artificial intelligence technique designed to automate the fault diagnosis process in the pumps of the IEA-R1 reactor. A new predictive diagnostic method for identifying failures in the primary cooling circuit pumps of the IEA-R1 research nuclear reactor is proposed. An Artificial Neural Network (ANN) is applied to the raw pump vibration signals both in the time and frequency domain. The Neural Network architecture is a feedforward network with 12 inputs employing a sigmoid transfer function and softmax in the output layer. The effectiveness of the proposed method is validated and compared with failure history during scheduled pump maintenance. The results fully agree with the neural network predictions.