Noise-Reduced Anomaly Attention Transformer for Intelligent Microscale Defects Detection in Metal Materials

The ultrasonic microscope detector, capitalizing on the characteristics of ultrasonic nondestructive testing, is an indispensable instrument for industrial material quality inspection and failure analysis. Intelligent detection based on artificial intelligence represents a critically demanded analysis and evaluation technology, capable of efficiently and conveniently assessing material quality. However, ultrasonic signals are susceptible to random noise, leading to weak representation of microscale defect signals. Consequently, anomalous areas are often obscured by noise signals, posing challenges to extraction and detection. To address these issues, this study first constructed a high-sensitivity immersion ultrasonic microscope detector, scanning heavy steel samples and acquiring data. Subsequently, we introduced a dual-branch transformer structure based on a noise-reduced anomaly attention mechanism, termed as the noise-reduced anomaly attention transformer (NRAAT). By incorporating noise mask (NM), continuity segmentation mask, and a min–max training strategy, this approach mitigates the impact of noise while enhancing the contrast between weak defects and the background. Finally, the method was validated on a custom ultrasonic signal dataset and two public anomaly detection datasets, achieving high-precision extraction and detection of noisy microscale defect ultrasonic signals. It demonstrated universality across different types of input anomaly signals and superiority over traditional defect detection methods and state-of-the-art anomaly detection technologies. In summary, this method offers an intelligent and efficient solution for the microscopic inclusions detection inside metal materials. It has the potential to significantly improve quality control and inspection processes in the metalworking industry.