Near-surface defect detection in ultrasonic testing using domain-knowledge-informed self-supervised learning

Abstract

Recently, significant research efforts have been made to enhance ultrasonic testing (UT) by employing artificial intelligence (AI). However, collecting an extensive amount of labeled data across various testing environments to train the AI model poses significant challenges. Moreover, conventional UT typically focuses on detecting deep-depth defects, which limits the effectiveness of such methods in detecting near-surface defects. To this end, this paper proposes a novel near-surface defect detection method for ultrasonic testing that can be employed without collecting labeled data. We propose a self-supervised anomaly detection model that incorporates domain knowledge. First, synthetic faulty samples are generated by fusing the measured UT signals with the back-wall UT reflection signals, to simulate real faulty features. Unlike the CutPaste method used for computer vision applications, this synthesis method adds the back-wall echo signal to random locations by incorporating the physical principles of the superposition of ultrasonic signals. Next, a de-anomaly network is devised to isolate subtle defect features within the measured UT signals. The presence of defects was determined using the three-sigma rule of the mean absolute value of the residual output. The defect depth is determined by a time-of-flight calculation from the residual output. The effectiveness of the proposed method was evaluated through the UT of aluminum blocks with near-surface defects of varying depths under different surface conditions. Both qualitative and quantitative comparison studies demonstrated that the proposed method outperformed existing methods in detecting the presence and depth of near-surface defects.

Keywords: Data synthesis; Denoising autoencoder; Diagnostics; Self-supervised learning; Ultrasonic testing.