Intelligent Quality Monitoring of Edible Oil Production Using MambaOut-CA-Based Raman Spectral Recognition

Abstract

Raman spectroscopy can be used quickly and without damage for edible oil identification; however, precise differentiation of various types of adulteration at a low level and accounting for spectral changes among different production batches are still unresolved issues. MambaOut-CA is a compact neural classifier proposed in this paper that combines MambaOut-style gated spectral mixing with coordinate attention for one-dimensional Raman fingerprints of edible oils. A controlled dataset with five real oils and four adulteration paths was built, including 0, 2, 5, 10, 20, and 40% adulterant ratios across 1,920 spectra. The proposed workflow adds a baseline correction, standard normal variate normalization, first-derivative enhancement and band-aware patch embedding before supervised classification. MambaOut-CA achieved the highest overall accuracy of 98.72% and macro-F1 of 98.41% at the 2% adulteration level among PLS-DA, SVM, ResNet-1D, Transformer-1D and a MambaOut model without attention. Ablation results show that Coordinate attention increased the macro-F1 score by 1.37% and reduced low-level false alarms by 28.6%. The band assignments are concentrated at 1,265, 1,440, 1,650, and 1,745 cm⁻¹, which are consistent with Raman responses related to fatty chain deformation and carbonyl groups. The results show that gated local-global spectral mixing with lightweight attention is suitable for robust edible oil adulteration classification under practical spectral variations.