Hydroxylated polycyclic aromatic hydrocarbons (OH-PAHs), as widespread environmental carcinogens, required rapid and accurate detection for exposure monitoring. Traditional chromatographic methods faced limitations in sensitivity and pre-processing complexity, particularly for trace-level OH-PAHs. This study proposed a Dynamic Weighted Voting Fusion Network (DWVF-Net) that integrated surface-enhanced Raman spectroscopy (SERS) with density functional theory (DFT) calculated spectra to enhance classification accuracy and noise robustness. The framework combined cross-modal attention fusion (CMAF), which aligned SERS and DFT data using cosine similarity to weight key spectral regions, and dynamic weight voting (DWV), which adaptively prioritized high-performing models. Validated on four OH-PAH derivatives, DWVF-Net achieved 97.14% classification accuracy, surpassing base models (KNN, GBM, Ensemble) by 6–11%. Under 20 dB Gaussian noise, accuracy declined by only 1.8%, outperforming equal-weight strategies (3.2% degradation). ROC analysis confirmed stability with AUC values above 0.98. The fusion of SERS specificity, DFT validation, and adaptive machine learning established a robust platform for pollutant detection without chromatographic separation, demonstrating potential for portable environmental monitoring systems.