Demonstration of using signal feature extraction and deep learning neural networks with ultrasonic data for detecting challenging discontinuities in composite panels

For the ultrasonic inspection of large composite structures, certain inspection scenarios can be difficult to interpret due to the superposition of reflected signals from the panel surface and the discontinuities in the material. In recent years, impressive advances have been made in the field of machine learning; however, challenges exist with transitioning emerging deep learning neural network (DLNN) algorithms for NDE applications. This work demonstrates the potential of an approach using chirplet feature extraction with the training of DLNN models to classify certain challenging ultrasonic indication calls. For this case study, the 2D raster scan of a test panel with known discontinuities was segmented into 6 x 6 A-scan groups for indication analysis and training purposes. Several data enhancement schemes were applied to increase the size of the training set. Using this approach, a set of challenge indications were successfully called by the DLNN based classifier and found to be clearly separable from the un-flawed regions of the scans.