Industrial Data Analytics from IoT sensors: an explorative study on coffee machines

The ability of collecting data from any kind of device is becoming more and more important every day. Many companies have realised that data acquisition and analysis provide a way to find new and more efficient solutions to their problems and to open up new perspectives. In order to optimize costs and maximize results, an initial exploratory analysis is necessary: in this phase, the interaction and exchange between domain experts and data analysts is fundamental to guide the analysis towards the company's objectives and to correctly interpret the results obtained from the data. For this reason, the thesis, conducted in collaboration with Lavazza, explores the data obtained from the telemetry sensors installed by the company inside their bar coffee machines. The study is focused on analysing the feasibility of predicting coffee quality, predictive maintenance and identifying the variables mostly related to the two previous objectives. The analysis was developed in two phases: a first phase for studying the data provided by tests performed in the laboratory, in which both the data collected by the telemetry sensor and those collected from the cup were available. Afterwards, a second phase of analysis of data from coffee machines on the market has been performed, thus including only data from telemetry sensors. Laboratory tests were structured in such a way that the overall external factors were reduced as much as possible: only double brewings (more stable than single brews) and always with the same type of blend were performed. In both phases, a preliminary analysis was necessary which showed that the three main parameters that characterise the coffee brewing process are: brewing time, flow-rate and quantity of coffee delivered in the cup. About these, Lavazza's domain experts provided initial quality thresholds, which were then evaluated and updated as a consequence of the analyses of the thesis. The results of the analyses are promising. Data-driven results confirmed known and hypothesized behaviours, and also revealed new possibilities. In fact, in some cases, the variation of the three external variables (dose, grinding and pressure) influenced flow, time and quantity of coffee in an evident way. In others, instead, the compensation effect covered up the consequences of the variations. Thanks to the identification of the correlation between the variables measured from the cup and those extracted from the telemetry, the study was extended from the laboratory to real-world coffee machines in bars. The latter revealed the limits of the data currently available in relation to the planned objectives. However, the analyses performed so far are a solid starting point. A promising approach is represented by the use of the time series: through a process of feature engineering, four new variables have been designed, related to the single coffee brewing. From the first analyses conducted, the addition of this information, on one hand, is potentially useful to deepen the intuitions already identified, on the other hand it marginally increases the amount of data sent by the sensors.