Deconvolution neural networks for ultrasonic testing

In this study, three novel design procedures have been developed in implementing deconvolution using neural network algorithms. The first method is called the Deconvolution Neural Network (DNN), the second method is named the Autoassociative Deconvolution Neural Network (ADNN), and the third method is referred to as the Probabilistic Deconvolution Neural Network (PDNN). The DNN trains the network by employing brute force and by exposing the network to a set of target echoes with and without noise. The ADNN processes the data for signal-to-noise ratio enhancement using an autoassociative neural network, and then applies the deconvolution neural network. The PDNN consists of two processing stages. The first stage estimates parameters using Gram-Charlier approximation to describe the probability density functions corresponding to target echoes and scattering noise. Then, in the second processing block, these parameters are used to classify and detect multiple target echoes. Results obtained in the performance analysis of these algorithms indicate that multiple interfering target echoes can be deconvolved and resolved accurately in the presence of noise.