Enhancing weak Raman spectral fingerprints for multi-source walnut oil adulteration using multi-level feature fusion

Walnut oil is a high-value edible oil with complex lipid composition, but its adulteration with lower-cost vegetable oils remains difficult to identify because different edible oils share highly similar Raman vibrational fingerprints. In particular, low-level adulteration induces only weak and localized spectral variations, while the major lipid-related Raman bands are strongly overlapped. To address these challenges, this study developed RamanFusionNet, a multi-level feature fusion-enhanced Net for weak Raman spectral fingerprint analysis of multi-source walnut oil adulteration. Unlike conventional approaches that rely primarily on a single spectral representation or a single model, RamanFusionNet combines raw spectral-band features, local differential descriptors, multiscale statistical features, and fast Fourier transform-based frequency-domain features. These complementary representations characterize changes in peak intensity, peak shape, and local spectral profiles caused by low-level adulteration. Ensemble modeling was further used to reduce the sensitivity of individual models to spectral noise and data partitioning, thereby enabling the identification of adulterant oil types and the prediction of adulteration levels. Experimental results show that RamanFusionNet can effectively distinguish between pure walnut oil and various adulterated oils, with an AUC of up to 0.996, a Spec of up to 0.993, and optimal regression performance of RMSE = 0.021 and R²= 0.991. For samples adulterated at a low concentration of 1%, the model achieved an RMSE of 0.0328 and an MAE of 0.0278.