TCN–GRU With Multihead Attention (TCN–GRU–MATT) for Quantitative Regression of NO2 From SERS Spectra

Abstract

Accurate quantification of NO₂ in construction machinery exhaust remains challenging because surface-enhanced Raman scattering (SERS) spectra are susceptible to fluorescence background, baseline drift, and random noise. In this study, an end-to-end hybrid deep learning model integrating a temporal convolutional network (TCN), a gated recurrent unit (GRU), and multihead attention mechanism (MATT) was developed for NO₂ concentration prediction from SERS spectra. A dataset containing 7500 spectra at six concentration levels was established using the FL_ArraySERS sensing platform. The TCN module was used to extract local multiscale spectral features, the GRU module to capture sequential dependencies across spectral bands, and the attention module to enhance informative regions while suppressing irrelevant interference. Compared with TCN, GRU, and TCN–GRU models, the proposed TCN–GRU–MATT model achieved the best predictive performance, with MAE, RMSE, and R² values of 0.5199, 0.8073, and 0.9996, respectively. Residual and spectral-band importance analyses further confirmed the robustness and interpretability of the proposed model. The results demonstrate that the proposed framework provides an effective strategy for quantitative analysis of NO₂ in complex exhaust environments.