Detection and classification of ultrasonic echoes using neural networks

Ultrasonic imaging techniques have been widely used for industrial applications. In particular, the detection of multiple interfering target echoes (e.g., related to cracks, defects, multiple layers) in the presence of scattering noise, and the classification of microstructure scattering echoes for material evaluation are recognized to be challenging problems. Neural networks are powerful tools for overcoming these challenging tasks due to their trainability and adaptability. This thesis presents neural network models which detect and characterize multiple target echoes that are in close proximity. Several circumstances have been explored utilizing neural networks: (i) deconvolution neural networks for detecting target echoes in the presence of noise; (ii) neural network filters for order statistic processing of multi-channel flaw signals; (iii) neural networks for classifying different types of materials using the microstructure scattering. Three design procedures have been developed in implementing deconvolution (i.e., deconvolution neural network, autoassociative deconvolution neural network, and probabilistic deconvolution neural network). Results obtained in the performance analysis of these algorithms indicate that multiple target echoes can be deconvolved and resolved accurately in the presence of noise. A method for detecting flaw echoes in large grain materials uses split-spectrum processing coupled with order statistic filters. A procedure has been developed utilizing neural networks to implement the sorting process (i.e., minimum order statistic neural network (MinNNet), median order statistic neural network (MedNNet) and maximum order statistic neural network (MaxNNet)). Both simulated and experimental results indicate that neural network order statistic filters offer desirable performance for sorting data and detecting flaw echoes. A design procedure for a novel application of neural networks has been developed to discriminate the frequency signatures inherent in ultrasonic microstructure scattering signals consisting of multiple unresolvable echoes. This method is called the grain power spectrum neural network (GPSNN). Overall, GPSNN achieves an average recognition performance of over 98%. Based on analytical and experimental observations, one can conclude that neural network models are encouraging and potentially useful techniques for ultrasonic nondestructive testing.