Artificial intelligence Defect Detection Robustness inReal-time Non-Destructive Testing of Metal Surfaces

Artificial intelligence (AI) is revolutionizing defect detection by employing advanced computational techniques to enhance accuracy and efficiency. Through machine learning methods and deep neural networks, it is possible for AI systems to learn from diverse datasets and accurately identify defects across various applications. In addition to minimizing human intervention, AI-driven defect detection provides real-time analysis, offering an important assistance for industries. Metal surfaces play a critical role in a wide range of industries. Their structural integrity, durability, and functionality depend heavily on their quality, making surface inspection a vital process. Characteristics such as texture, finish, corrosion resistance, and the presence of defects like scratches, dents or cracks, can significantly impact performance and longevity. Advanced inspection techniques, such as non-destructive testing (NDT) and surface characterization methods, are often employed to evaluate metal surfaces for flaws. Real-time NDT of metal surfaces is essential for maintaining the integrity and reliability of metal components without causing any damage. This technique enables the immediate detection of defects that could compromise the performance and safety of metal structures. By incorporating advanced AI technologies, real-time NDT enhances accuracy, shortens inspection times, and delivers instant feedback, making it a vital tool for metal surface inspection. However, the robustness of AI-based defect detection presents uncertainties in the NDT process of metal surfaces, particularly in real-time scenarios under varied and challenging conditions. Factors such as surface variability, environmental influences, and defect types can significantly impact the performance of an AI real-time NDT system. In this study, we identify and summarize the key factors that potentially affect metal surface NDT. Given the diversity of defects, we propose a multi-model fused AI approach for surface detection. Additionally, we establish a real-time, cost-effective NDT platform to quantitatively assess the robustness of the AI model in practical, online settings. The primary robustness-affecting factors and AI model uncertainties are verified, providing insights for robust NDT system design.