Machine learning (ML) is a branch of artificial intelligence (AI) that specializes in studying how computers simulate human learning behavior, summarize patterns from experience, and develop decision-making and analytical capabilities. It has numerous applications in the industrial field, particularly in areas such as quality control, defect detection, and process optimization. Laser Beam Welding (LBW) is an advanced welding technique that utilizes a high-energy density laser beam to fuse materials together. It has the advantages of high precision, high efficiency, minimal heat-affected zone, and low welding deformation, making it widely adopted in aerospace, automotive manufacturing, electronics, medical devices, and other modern industrial fields.
ML can be applied to LBM for automated defect detection and process optimization. This thesis explores two applications in LBM: solidification crack detection and strain estimation, both helping prevent defects and ensure structural integrity of welded components. Solidification cracks result from internal stresses caused by shrinkage and temperature changes during solidification. Convolutional Neural Networks (CNNs) can analyze high-speed welding videos to detect crack formation and predict potential defects in real time. Strain estimation reflects deformation and residual stresses in a welded material, as excessive strain can lead to distortions and cracks. Compared to traditional measurements, ML-based methods offer faster and more efficient strain predictions.
However, integrating ML into industrial systems introduces security challenges such as data poisoning attacks, adversarial attacks, model stealing attacks, and membership inference attacks. To address these concerns, we investigate the reliability and robustness of the crack detection system. Specifically, we design a backdoor attack, a type of data poisoning attack that injects malicious data into the training set, causing the ML model to misclassify welding defects and fail to detect their occurrence. To ensure safe and reliable deployment of AI in welding applications, security problems must be effectively addressed. To counter backdoor attacks, we propose a novel defense strategy, NT-ML, which defends against more robust backdoor attacks compared to existing methods.
