Deep Learning Techniques for Flaw Characterization in Weld Pieces from Ultrasonic Signals

Computer aided Interpretation of Ultrasonic signals depicting flaws in weld pieces is depicted in this work. In this work, feasibility of Long Short Term Memory (LSTM) for flaw characterization is studied. Owing to the advantage of LSTM, the first technique involves training LSTM directly with the signals as inputs and testing its ability to characterize the flaws from the input signals. Due to wide variation in the length of input sequences, which introduced sparseness in other sequences, overall accuracy is affected. Hence in the second technique, LSTM are trained with features of the signals and it is found that the overall accuracy for test data is 67.64%. These features are statistical parameters obtained from the approximation co-efficient of the input signals. The input signals are decomposed with a novel wavelet template.