A recurrent attention semi-supervised domain adaptive network for cross-domain diagnosis of resistance spot weld nugget quality

The recent advancements in intelligent spot-welding quality diagnosis during automobile production necessitate labeled data with enhanced quality. However, the scarce samples in the target domain under new working conditions often lead to inaccurate diagnoses of body-in-white (BIW) weld nugget quality by the existing models. Thus, this paper presents a cyclic attention semi-supervised domain adaptive network for cross-domain quality diagnosis of BIW resistance spot weld nuggets. The proposed network synthesizes a standard curve metric via wavelet analysis reconstruction to measure welding process stability and integrates it into the domain adaptation process as prior knowledge. In addition, the feature extractor combines long short-term memory and channel attention mechanisms to dynamically adjust feature weights for enhanced representativeness. During semi-supervised domain adaptive learning, the maximum mean discrepancy based on the Gaussian kernel is utilized to measure the feature distribution difference between the source and target domains. Meanwhile, target domain labeling is leveraged to guide the parameter adjustment of the output layer. These two measures in combination facilitate feature alignment and knowledge migration, ensuring the ultimate generalization capability of the model in feedback adjustment. The results demonstrate that the proposed network achieves a correct diagnosis rate exceeding 90% under new working conditions, affirming its robust generalization ability. This cross-domain diagnosis method offers valuable insights for advancing BIW welding quality diagnosis in automotive production lines.