Machine learning-based quality prediction for aluminum alloy laser welding using hybrid sensing

Abstract Aluminum alloy is a key material for automotive lightweighting, and its welding quality directly affects structural safety. Although laser welding offers high efficiency, the process for aluminum alloys is unstable with significant quality variations. Traditional post-weld inspection is severely delayed, while existing online methods often rely on single signals, making it difficult to predict internal weld morphology and mechanical properties in real time with accuracy. To address this, this study proposes a real-time prediction method based on multi-source information fusion. For dual-beam coaxial hybrid laser lap welding of 6061 aluminum alloy, optimized infrared temperature signals and 3D profile data acquired by line-laser scanning were synchronously collected. A weld quality prediction model was established by integrating backpropagation (BP) neural network and random forest (RF) algorithm. By defining key characteristic parameters such as weld width (A, L), penetration depth (H, D), and cross-sectional area, strong linear correlations were revealed among real-time temperature, surface morphology, weld geometry, and tensile-shear strength. The results show that the neural network optimized by RF significantly improves prediction accuracy and stability, effectively eliminates high-error predictions, and enhances generalization capability. This research provides an effective approach for online assessment and closed-loop control of weld quality in aluminum alloy laser welding.