A Multistage Deep Learning Network for Trace Explosive Residues Detection in SERS Chips

To address the challenges of relying on specialized personnel and incurring significant time costs in qualitative and quantitative analysis using surface-enhanced Raman scattering (SERS) technology for explosive residue detection, this article proposes a detection method for explosive residues based on a multistage deep learning network and SERS chip. To improve the qualitative analysis performance of the SERS spectrum, a novel fusion attention module-based residual neural (FAB-ResNet) is constructed through the integration of a modified attention mechanism into the ResNet network. In addition, for proper processing of long sequential data, the nested long short-term memory (NLSTM) network is selected for quantitative analysis with its powerful global information aggregating capability. Consequently, the NLSTM is incorporated into FAB-ResNet to construct a multistage hybrid network. Extensive experiments are carried out to prove the effectiveness of the proposed hybrid network. The qualitative results demonstrated the superiority of the proposed FAB-ResNet with its outstanding classification accuracy (100%). Meanwhile, by comparing quantitative results, the NLSTM network provides promising performance (<inline-formula> <tex-math notation="LaTeX">${R} ^{{2}}$ </tex-math></inline-formula> = 0.9835, root mean square error (RMSE) = 0.1653, mean absolute error (MAE) = 0.0916, and mean absolute relative error (MARE) = 2.7488%). Furthermore, the comparative results among other state-of-the-art networks confirmed the effectiveness of the proposed method as a means of explosive residue detection and analysis, which shows the potential for further application of SERS technology in explosive site detection.