A deep feature fusion network based on multi-scale kernel construction for filling wing stress field data

Stress monitoring of aircraft wings is crucial for ensuring flight safety and maintaining the structural reliability of aircraft. In practical engineering applications,sensors must be installed at sparse and fixed locations,which limits their ability to capture stress data across the entire wing structure. As a result,comprehensive monitoring of the wing’s structural health cannot be ensured. To address this issue,this paper proposes a deep feature fusion network based on multi-scale kernel construction to reconstruct the wing stress field in cases where sensor placement is spatially sparse or incomplete. First,based on the similarity among stress fields,the data is clustered into several stress field groups,and the field closest to the center of each cluster is selected as the reference set. Then,by extracting compo⁃ nents of varying scales from each reference stress field,multi-scale convolution kernels are constructed to perceive the missing data points from different directions,enabling the capture of multi-scale features conducive to information completion. Finally,a parallel channel attention module is employed to adaptively select salient features captured by the convolutional kernels and map them into a unified feature space for fusion,thereby generating the imputed values for the missing points. Additionally,comparative experiments were conducted on wing stress field datasets with differ⁃ ent missing ratios. The proposed method outperforms mainstream approaches in terms of MAE,RMSE,and MAPE, achieving the best overall filling performance.