A self-learning neural net for ultrasonic signal analysis

Abstract This paper deals with ultrasonic signal analysis using artificial neural nets. In particular, it investigates the classification of ultrasonic inspection data using a backpropagation (BP) neural network. The traditional BP learning algorithm requires a set of signals that have been classified a priori for training the net. To eliminate the uncertainties involved in selecting the training set, a self-learning algorithm is developed. A BP net is trained using the self-learning algorithm and it is used to classify ultrasonic inspection data. The classification is performed in both the time and the frequency domains. The classification results are compared with the results obtained by conventional C-scan methods, and it is found that 95% of all the signals with flaws are classified correctly by the unsupervised BP net. The self-learning algorithm developed for classification of ultrasonic data is also detailed in this paper. It utilizes the correlation information computed by the first hidden layer of the BP net for the generation of the initial training set consisting of one signal per each class. Additional signals in each class are selected by the net using the nearest neighbour approach. An estimate of the number of different classes present in the data is made using the damage profile of the sample being investigated by Kohonen's learning vector quantization (LVQ) algorithm.