电弧增材制造中融合物理约束的小样本多层成型缺陷识别方法

Wire arc additive manufacturing (WAAM) has become a key technology in the fabrication of electrical equipment due to its high deposition rate and excellent material utilization efficiency. However, this process is inherently influenced by coupled multi-physical fields, making it prone to the formation of structural defects during multilayer deposition, which can significantly compromise the forming quality and structural reliability. To address the challenges of defect recognition in multilayer deposition, including the scarcity of labeled samples, the high dimensionality of spectral inputs, and the limited physical interpretability of existing models, this paper proposes a spectral information-driven optimization method that integrates physical spectral mechanisms with machine learning strategies to tackle high-dimensional defect recognition under small-sample constraints. First, key spectral regions of interest (ROI) associated with metal evaporation and gas excitation processes are identified based on elemental radiation mechanisms and empirical knowledge, enabling initial feature compression. Subsequently, a physically-constrained genetic optimization framework is constructed, where spectral feature selection and model parameter tuning are jointly encoded. A fitness function incorporating spectral retention ratio is further designed to guide each individual in the population toward a solution that balances classification performance with physical consistency. Finally, a lightweight extreme gradient boosting (XGBoost) classifier is employed for model training and prediction, ensuring both robustness and deployment efficiency. Experimental results demonstrate that the proposed method achieves an accuracy of 93.21% in recognizing multi-layer defects under limited sample conditions, exhibiting notable advantages over comparative methods in terms of accuracy, stability, and interpretability.