Deep Learning-Based Defect Detection from Ultrasonic Testing Imaging of Stainless Steel Plate

Abstract

Ultrasonic Testing is widely applied to inspect product and building structure with unseen defect in manufacturing industries. Furthermore, the advancement of the test changes from waveform signal into image which clarifies the unseen defects. The clarity of the unseen defects causes it is widely accepted in the industries. Though, ultrasonic images screening by operators the to detect the defect are prone to misjudgments. Therefore, this paper aims to automate the test using Deep Learning based approach using SqueezeNet model. Besides, the automated system is tested stainless steel plate with artificial defects from ultrasonic test image. The image is designed in two classes – 100 images with defect and 100 images without defect. Then the total number of 200 images is labelled and classes into 70% of learning data, 20% for testing data, and 10% for validating data for the following system modelling stage. The detection rate stands at 81.67% based on testing and validating data. Moreover, for the all evaluation measures resulted above 80% – Precision, Recall, and Accuracy. Statistically, the best performance of the model rated by F-score at 81.36%. In addition, by margin of error value, the used data and by chance to get the similar result at above 73% with 99% confidence level. These findings suggest the used model is fit to detect defects based on the ultrasonic test imaging image at best performance.