| Abstract Scope |
Pinch welding, a specialized resistance spot welding technique, is widely used for sealing stainless steel tubing in critical applications such as hydrogen containment, where weld integrity and microstructural homogeneity are paramount. Traditionally, weld quality assessment relies heavily on destructive testing, which is both resource-intensive and time-consuming. This study explores the use of advanced data analysis and machine learning methods to predict weld quality from process controller data, aiming to minimize destructive testing requirements. Raw process signals—including force, current, displacement, and voltage—were collected from pinch welding operations and subjected to fast Fourier transform (FFT) analysis for noise reduction and feature extraction. Supervised learning algorithms, such as regression and classification models, were trained on labeled datasets to predict key weld qualification metrics based on real-time process data. In parallel, unsupervised learning techniques, including clustering and principal component analysis, were utilized to segment welding cycles and reveal underlying patterns in process variability. The integration of FFT-based signal processing with machine learning enabled accurate identification of weld cycles and robust prediction of weld quality metrics, such as bond strength and weld dimensions. The results demonstrate a strong correlation between process parameters and weld outcomes, supporting the feasibility of in-line, data-driven weld qualification. This approach not only improves process monitoring and control but also offers significant potential for reducing reliance on destructive testing, thereby enhancing efficiency and reliability in critical pinch welding applications. |