A Multi-Resolution Convolutional Neural Network Architecture for Ultrasonic Flaw Detection

In this work, a Convolutional Neural Network (CNN) architecture with a novel feature selector based on Wavelet Packets (WP) is introduced for ultrasonic flaw detection applications. CNN is among the key Deep Learning algorithms that have gained increased attention in recent years due to their success in machine vision, pattern recognition and classification tasks. CNN consists of multiple layers and it can extract features in a hierarchical manner to construct a high-level abstraction of the original data. We propose two CNN architectures (based on ID-CNN and LeNet models) for ultrasonic data (A-scan) using wavelet coefficients as feature inputs and investigate key topologies such as number of parallel convolution networks, number of filters and output classifiers. Optimal selection of wavelet subbands enables the CNN based model to detect the presence or the absence of flaws with an accuracy of up to 92% on experimental ultrasonic data.