Data-driven soft sensing for raw milk ethanol stability prediction

Abstract To address the destructive limitations of conventional ethanol stability testing methods for raw milk and their declining applicability in industrial contexts, this study proposes a TabNet-based soft sensing model. By leveraging autoencoder-based feature reconstruction and a multimodal feature selection strategy, eight highly relevant attributes were systematically identified as model inputs: protein content, total solids, solids-not-fat (SNF), fat content, titratable acidity, lactose content, relative density, and raw milk temperature. A diffusion model was innovatively employed to overcome the constraint of class imbalance, enabling the development of a non-destructive model that predicts ethanol stability based on routinely monitored indicators. Validated on a three-year industrial-scale raw milk intake dataset, the proposed model achieved an accuracy of 92.57% and a recall of 90.26% in identifying ethanol-unstable samples, demonstrating substantial potential for real-world engineering applications.