Improved Fault Detection in Ultrasonic Testing Using Self-supervised Learning

Abstract

This paper presents a novel self-supervised learning-based ultrasound testing method for defect detection in industrial applications. The integration of artificial intelligence (AI) into ultrasound examinations aims to address the variability in results dependent on human examiners. However, challenges arise from the sensitivity of AI to specimen surface conditions and the impracticality of obtaining labeled data in the same state and acquisition environment as the test object. To overcome this, we propose a method involving the introduction of defects by enhancing the floor reflection signal, enabling defect characteristic extraction. A residual model based on a denoising autoencoder is developed to isolate defect features within the inspection object. The proposed method is validated through ultrasonic-based B-scan inspections on aluminum blocks with uneven surface conditions, including defects near the surface. Results demonstrate superior defect detection performance compared to existing inspection methods. This study is significant as it eliminates the need for separate label data, accurately detects defects within specimens, and robustly accounts for surface states, model complexity and model structure. The self-supervised learning approach introduced holds promise for enhancing the reliability and efficiency of ultrasonic inspections in industrial settings.