Cleaning-In-Place of heat exchangers in Ultra-High Temperature dairy production remains a resource-intensive process, typically operated without explicit information on the underlying cleaning mechanisms. In laboratory experiments from the literature, grayscale image sequences of soil removal were used to extract statistical features describing local and temporal variations in pixel intensity. Building on this framework, the present work investigates the feasibility of identifying cleaning mechanisms directly from industrial dairy process data using machine learning models retrained on image-based features. A fully connected neural network was trained to classify three dominant cleaning mechanisms – cohesive separation, adhesive detachment, and viscous shifting – achieving a classification accuracy of 97.9 % on independent experimental test data. The same feature definitions were subsequently transferred to industrial cleaning-run data of a shell-and-tube heat exchanger. When applied to more than 800 real cleaning runs of mainly pudding, the models consistently identified adhesive detachment as the dominant mechanism across the two main cleaning stages. Additionally, laboratory cleaning experiments conducted with the corresponding pudding products under near-industrial conditions supported these findings by reproducing similar phase-dependent removal dynamics. The study demonstrates the potential of mechanism-resolving machine learning for interpretable, data-driven monitoring of cleaning efficiency in industrial heat exchangers.