Multi-scale fusion SDAM-net empowered by physics-informed conditional spectral augmentation enabling spectral feature assignment for the concurrent detection of nitrogen and sulfur in ship fuel oil.

The quantitative detection of nitrogen (N) and sulfur (S), the primary pollutant elements in marine fuel oil, is essential for global atmospheric environmental protection. However, the efficient in situ detection of these elements is hindered by the reliance of current methods on shore-based technical support and the severe spectral entanglement within complex fuel matrices. Herein, we demonstrate a multi-scale feature-fusion multi-task network (SDAM-Net) that decouples the fluorescence spectra of tin dioxide quantum dots (SnO2 QDs) for the efficient in situ detection of N and S in ship fuel oil. A physics-informed conditional Wasserstein generative adversarial network (CWGAN) is proposed to reconstruct a continuous spectral manifold to overcome the limitations of insufficient training samples. This conditional generative strategy mitigates the combinatorial sparsity of N and S concentrations in the experimental dataset, thereby resolving predictive instability caused by sample-scarce regions. The incorporation of an interpretability module based on SHapley Additive exPlanations (SHAP) assigns the extracted salient features to specific microscopic spectral responses. This feature assignment confirms the successful separation of entangled spectra and elimination of chemical crosstalk, validating the model decoupling strategy. Performance evaluation on the test dataset confirms that the SDAM-Net-based framework achieves high detection accuracy for N and S through spectral decoupling, with root mean square errors (RMSEs) reaching 0.0454% and 0.0376%, respectively. This method establishes a robust paradigm for the efficient in situ detection of N and S in marine fuel oil, providing a crucial detection platform for atmospheric environmental protection.