Machine learning and electrochemistry techniques for detecting adulteration of goat milk with cow milk

A semi-supervised deep learning approach combining autoencoder feature extraction and multilayer perceptron regression networks was developed for rapid on-site detection of adulteration in goat milk using voltammetric fingerprints. The compact ten-electrode microsensor array differentiated compositional profiles between pure goat and cow milk with 2–5% precision. Model training on 88 labeled spiked samples augmented by unlabeled surrogates improved test prediction accuracy by 4.3% over standalone supervised learning, achieving 2.21 RMSE down to 1% adulterations. Independent validation across 0-100% cow milk ratios showed under 2% relative error up to 90% blending before declining fidelity, successfully identifying adulterants up to legal limits. Extreme case studies with as few as 4 known sensors demonstrated the semi-supervised framework maintaining 26% superior test performance owing to transferable invariant feature extraction, confirming real-world field reliability amidst unseen volatility. Sweeping labeled/unlabeled ratios indexed tradeoffs between uncertainty range, discrimination power and complexity to guide customizable deployments optimizing domain-specific accuracy, affordability and robustness priorities. The integrated data science-instrumentation solution provides a deployable blueprint for milk provenance assurance through democratized quality inference. It introduces scalable deep learning to hitherto error-prone chemical methods for functional adulterant quantification at the point-of-origin.