Detection and Classification of Defects in Ultrasonic Testing Using Deep Learning

Abstract

Automated inspection systems are gaining importance for quick inspection and maintenance purpose for large volume of data in welding fields. Welding is very complex process and during the process defects may arise in weld zone. Identifying those defects requires skill and is impractical for large volume of data. This paper focuses to solve the problem by implementing deep learning model which is part of industrial 4.0 revolution, to classify seven common defects and locate them automatically for ‘v’ butt joint weld of 12 mm plate thickness for 19 angles from 2 groups by using ‘A’ scan signal of non-destructive testing method, phased array ultrasonic testing (PAUT). For feature extraction, discrete wavelet transform (DWT) is used with selection of proper mother wavelet. For supervised training, two architectures namely feed forward neural network (FFNN), convolutional neural network (CNN) are used. FFNN used features from DWT and for CNN we used both features and whole signal for training. Finally we discussed about implementation details for different models. It is observed that due to distribution of training data all the models achieves about 80% accuracy but for simple applications FFNN can give satisfactory results but if output is complex then CNN should be implemented.