Resistance Spot Welding: Surface Appearance Image-Based Weld Quality Prediction using HCLAN

Objectives: This study enhances resistance spot welding (RSW) quality inspection by proposing an automated method to predict weld quality, minimizing reliance on manual evaluation. Methods: A Hybrid CNN-LSTM with Attention Network (HCLAN) was developed, leveraging heat trace (HT) images of the weld surface for quality forecasting. Experiments used advanced high-strength steel (AHSS, 980 MPa) and uncoated cold-rolled (CR) steel, measuring tensile shear strength (TSS), weld fracture mode, nugget diameter, and ejection occurrence. Findings: The HCLAN model achieved a nugget diameter prediction error of 2.8% and TSS prediction error of 2.3%. For weld fracture mode and expulsion detection, it achieved 100% classification accuracy. Novelty: The integration of CNN, LSTM, and attention mechanisms in HCLAN demonstrates a novel approach to automating weld quality assessment. This method significantly reduces inspection errors and enhances efficiency in industrial welding processes. Keywords: Resistance spot welding, Tensile shear strength, Nugget size, Surface appearance image, Weld quality, CNN-LSTM