Acoustic Based Cavitation Detection of Centrifugal Pump by Neural Network

Abstract

A neural network simulator built for prediction of cavitation in centrifugal pump is discussed. A back propagation learning algorithm and a multi-layer network have been employed. The layers are constituted of nonlinear neurons and an input vector normalization scheme has been built into the simulator. This paper concentrates on a procedure for prediction of cavitation using acoustic signals and multi-layer perceptron neural network. Acoustic signals carry rich information about the condition of the system. Fast Fourier transform, proved itself a fast and strong technique for frequency analysis in the steady conditions and constant speed such as this work. Two classes were defined, namely i) healthy ii) cavitation. The extracted features were employed to feed the ANN classifier for cavitation detection. In developing the ANN models, different ANN architectures, each having different numbers of neurons in hidden layer, were evaluated. The optimal model was selected after several evaluations based on minimizing of mean square error (MSE) and correct classification rate (CCR). The network structure with 11 neurons in hidden layer and 11-11-2 structure is the best network. The mean square error achieved during training is 8.7204*10 -8 . Correct classification rate in system was 98.57%.