Indirect Estimation of Limescale Accumulation in Coffee Machines Using Thermal Dynamics and Machine Learning

Predictive maintenance is an use of Artificial Intelligence (AI) and Internet of Things (IoT). Predictive maintenance lets us decide when to maintain a system based on its condition instead of on a fixed schedule.In coffee machines, limescale accumulation inside the hydraulic circuit is one of the main reasons that performance gets worse. Limescale affects the machine in ways than just by how much is deposited; it also matters where the deposits form because they can change heat transfer, hydraulic resistance and the final brewing result. This thesis, made with Zoppas Industries looks at a way to watch the effects of limescale using signals that are already part of the coffee machine. The warm‑up phase gets focus. In the warm‑up phase the heating-element heats up before any water flows. The warm‑up phase keeps external factors, such as the hydraulic resistance from the capsule or the chosen beverage out of the picture. A special experimental set‑up was built to create ageing while keeping normal machine conditions. Thermal, hydraulic, electrical and operating signals were taken over run‑to‑failure calcification cycles. Those signals were fed into a synchronization and preprocessing pipeline. The experiment showed that changes in thermal dynamics are linked to the slow increase in hydraulic degradation, which shows up as a lower water flow. Many modelling ideas were tried, from custom indicators and change‑point methods to direct time‑series learning. The final method uses a two‑stage design. First a 1D Convolutional Neural Network takes the dynamics and operating data and guesses the hydraulic flow. Second another neural network takes that guess and life‑cycle data to estimate the Remaining Useful Volume (RUV). RUV is the amount of water that can still be poured before the calcification cycle ends. Finally the predicted RUV is turned into a warning and a critical alert and this plan is compared with static maintenance rules. The data‑driven plan gives alert coverage on every tested cycle. Compared with a static rule that gives the same coverage the data‑driven plan lets a lot more coffee be poured before an intervention is needed. On cycles the critical alert appeared on average about 15.6 L later than the conservative rule. These results show that combining sensing, machine learning and RUV estimation can move us from fixed maintenance limits to condition‑based descaling, for connected coffee machines.