Nonlinear signal processing for ultrasonic target detection

Ultrasonic flaw detection is an important problem in the nondestructive evaluation (NDE) of materials. In order to successfully detect and classify flaw echoes from high scattering grain echoes, an efficient and robust method is required. This paper utilizes split-spectrum processing (SSP) combined with a neural network (NN) to develop a dedicated ultrasonic detection system. SSP displays signal diversity (i.e., correlated or uncorrelated), and the neural network (NN) performs highly complex nonlinear mapping by which signals can be classified according to their feature vectors. Therefore, the combination of SSP and NN (SSP-NN) presents a powerful technique for ultrasonic NDE. In this paper, SSP is achieved by using Gaussian bandpass filters. Then, an adaptive three layer neural network using a backpropagation learning process is applied to perform the classification processing of frequency diverse data. Both simulated and experimental data are used to test this method, and results show that SSP-NN shows an excellent sensitivity in detecting and separating adjacent flaw echoes and is able to detect flaw echoes when the flaw-to-clutter ratio is about 0 dB.