Milk powder adulteration identification and quantification based on shared encoder features using Raman spectroscopy.

Detecting adulterants in milk powder is challenging at low concentrations and in multi-adulterant mixtures. Therefore, precise detection is essential for milk-powder safety and quality control. This study proposed a Raman-based approach that combined data-driven soft independent modeling of class analogy (DD-SIMCA) with a multilayer perceptron (MLP). First, DD-SIMCA performed nontargeted screening to distinguish authentic from adulterated samples at low concentrations and in multi-adulterant mixtures, without prior knowledge of adulterant types. Next, a MLP with a shared encoder was applied to samples flagged as adulterated by DD-SIMCA. The encoder extracted features required for adulterant-class classification and concentration prediction, enabling accurate classification and precise quantification. Experiments showed that DD-SIMCA achieved 100 % accuracy for adulteration detection. The MLP achieved 99 % accuracy in adulterant-class classification. For concentration prediction, it achieved an R2P of 0.99 with RMSEP below 0.6, outperforming conventional methods such as SVM, RF, and PLSR. These results demonstrated that the proposed approach addressed low-concentration conditions and multi-adulterant mixtures. It provided a practical pathway for rapid detection and on-site deployment in milk-powder adulteration screening.