Multilayer Perceptron Neural Networks for Grain Size Estimation and Classification

Abstract

Grain size estimation nondestructively as a method for characterizing the mechanical and structural integrity of materials has been long recognized. The scattering and attenuation of ultrasonic echoes depend on the frequency of the sound and grain size distribution. It is of high interest to estimate grain size and classify materials based on the scattering properties of the specimen microstructure. In this study, an ultrasonic NDE system is used for acquiring ultrasonic scattering signals. Backscattered signals collected from three steel blocks with different grain sizes are used to train the neural network for material classification and quality control. Scattering signals in time domain, frequency domain, and time-frequency distribution are applied to the neural network for grain size characterization and classification. The validation accuracy of the trained network is as high as 99% for grain size classification.