Frequency discrimination of ultrasonic signal using neural networks for grain size estimation

Abstract

A neural network model has been developed to discriminate the frequency signatures inherent to ultrasonic microstructure scattering signals consisting of multiple irresolvable echoes of random amplitude and arrival time. A practical method which is called the grain power spectrum neural network (GPSNN) has been studied. This model is also compared with two other neural network models, called grain autocorrelation neural network (GACNN) and the grain amplitude neural network (GANN). The materials tested for grain size discrimination were three steel blocks type 1018 (two blocks were heat-treated at 1600 and 2000 degrees Fahrenheit for 4 hours) with grain sizes of 14 microns (ASTM No. 9), 24 microns (ASTM No. 7) and 50 microns (ASTM No. 5). Experimental grain signals were obtained using a broadband transducer with a 5 MHz center frequency and the measurements were made in the Rayleigh scattering region. A set of 2565 training sequences was utilized to train the neural network. A new set of 855 testing sequences was acquired to test the GPSNN, GACNN and GANN performance. Overall, GPSNN and GACNN achieved an average recognition performance of 94% and 90% respectively. This high level of recognition suggests that the GPSNN is a promising method for ultrasonic nondestructive testing and grain size estimation. In contrast, GANN failed to sort the grain scattering signals and was able to correctly classify the signal only 50% of the time.