Open-set deep learning enabled LIBS sensor for label-free and on-site identification of unknown pathogen powders

On-site and label-free identification of pathogen powders is critical for effective biosecurity and bioterrorism response. Laser-Induced Breakdown Spectroscopy (LIBS) and deep learning have demonstrated significant potential for the on-site analysis of such powders. However, existing deep learning models are incapable of rejecting input from unknown classes not included in the training set, which limits their reliability in real-world applications. In this study, we developed a novel LIBS sensing technique driven by open-set deep learning for pathogen powder identification. By integrating the Classification-Reconstruction Open Set Recognition (CROSR) strategy into a Residual Network (ResNet), we constructed an improved open-set deep learning model. The results show that the ResNet model using CROSR not only maintains high classification accuracy for known classes but also significantly enhances its ability to reject unknown samples. It achieved an identification accuracy of 89.7% on the test set containing five known target pathogens and 86.6% for nine types of unknown samples, including biological samples, culture media, and non-biological interferents. Compared to the traditional open-set strategy OpenMax, the proposed model improved the identification accuracy of target pathogens by 11.7% and of unknown classes by 11.9%. This study provides a promising solution for on-site, label-free identification of pathogen powders, with broad application prospects.