Research on State Recognition in Aircraft Skin Laser Paint Stripping Based on the Fusion of LIBS Spectra and Surface...

To address the recognition challenges caused by blurred state boundaries and the limitations of single monitoring modalities during aircraft skin laser paint stripping, this study proposes a multimodal data fusion method for state recognition based on laser-induced breakdown spectroscopy (LIBS) and surface imaging. By constructing a synchronous monitoring platform, a dataset covering five key physical states, namely topcoat (Tc), topcoat-primer transition (Tc-Pr), primer (Pr), primer-substrate transition (Pr-As), and substrate damage (As), was established. The proposed gated weighted multimodal fusion network (PGMF-Net) employs SE-ResNet1D to capture variations in elemental composition features from the spectra and integrates ResNet18 to extract changes in surface morphology from the images. The experimental results show that the proposed model outperforms the single-modal methods as well as the compared early-fusion and late-fusion methods, achieving a recognition accuracy of 94.12% on the test set and an average accuracy of 94.87% in stratified cross-validation. The bootstrap-based confidence interval analysis further verifies the stability of this method under the current dataset conditions. Further analysis indicates that the single-spectrum model has difficulty effectively distinguishing coating transition states because different transition states contain identical or highly similar characteristic peak information. The single-vision model, however, shows insufficient sensitivity to subtle substrate damage, whereas multimodal fusion enables complementary representation of material composition information and surface morphological information. Experimental validation under different power conditions further confirms that the model outputs are generally consistent with the macroscopic morphological evolution observed on the sample surface. This method compensates for the limitations of traditional single-source monitoring and provides a methodological foundation for online monitoring and state feedback during the laser paint stripping process.