Tri-Modal LIBS Signal Fusion with DOWA-Extracted LIPA for Enhanced Water Pollution Identification

The accuracy and robustness of laser-induced breakdown spectroscopy (LIBS) in water pollutant identification are limited, mainly due to its reliance on single-source signals. To address this issue, this study innovatively proposes an optimized tri-modal fusion approach that integrates laser-induced plasma acoustic (LIPA) signals, images, and spectral data. A dynamic overlapping window algorithm (DOWA) is designed to extract features from LIPA signals, and the resulting model is named DLIPA (DOWA-Extracted LIPA). Meanwhile, a VGG16 convolutional neural network combined with principal component analysis (PCA) is developed to extract spatial structural features from images. These are further integrated with spectral features into a tri-modal signal fusion model, named LIBS-Tri-Fusion, which is trained using the random forest (RF) algorithm. Experimental results show that the proposed model achieves an identification accuracy of 0.954 ± 0.017, significantly outperforming the single-spectrum model (0.854 ± 0.044), with superior performance in both recall and F1-score. These findings validate the potential of multimodal fusion to enhance detection performance and provide an accurate and robust solution for water quality monitoring, demonstrating promising application prospects.