As an emerging solid-state joining process, simultaneous double-sided friction stir welding (SDS-FSW) offers significant potential for enhancing the quality and efficiency of traditional FSW aluminum alloys. However, its complexity and cost restrict extensive experimentation, thereby limiting mechanical performance prediction. To address this limitation, data generation and model prediction strategies are explored. First, a full-factorial experimental design approach is employed to conduct SDS-FSW experiments under three key process parameters: pin offset distance, rotation speed, and welding speed. The welding process temperatures are collected, and the ultimate tensile strength (UTS) of the welded joints is measured. Second, a small-sample data generation model is proposed, leveraging feature engineering techniques with Variational AutoEncoders and Generative Adversarial Networks to synthesize data from original values, effectively addressing the issues of data scarcity and imbalance. Finally, a prediction model for the UTS based on the process parameters and temperature is developed using ensemble learning techniques. Experimental results demonstrate that the proposed small-sample data generation model reduces the Wasserstein distance to 1.0027 and achieves a Kolmogorov-Smirnov statistic of 0.0482. Furthermore, the proposed prediction model attains a root mean square error and mean absolute error of 0.6 MPa and 0.4 MPa, respectively, on the original data, achieving 50% and 55.6% reductions over TabNet. A robustness advantage of 10.42% is also observed under a noise perturbation of 1.9 μm. This work establishes the first data-driven framework for SDS-FSW joint strength prediction under small-sample constraints, providing a foundation for industrial deployment.
Xu et al. (Wed,) studied this question.