Manutenzione predittiva nella produzione: un caso di studio della società SKF

As time goes on, more and more manufacturing firms are developing new technologies that enable them to gather enormous volumes of data in real time through autonomous sensors that record measurements about the status of machinery. Operators can use this feature to observe how the production channel behaves and detect any anomalies that can call for their intervention. Also, business management can utilize this feature to take a data-driven approach to decision-making by using informative statistics, graphs, and reports. The most significant benefit is that gathering data enables the use of machine learning, which can assist operators in challenging tasks like anomaly detection, predictive maintenance, and product quality forecasts. The most recent advancements in cloud computing offer effective ways to implement completely automated machine learning services, making them widely available and enabling collaboration and maintenance. This will further improve the situation. The purpose of this thesis is to report my contribution to project Beat 4.0 , carried on by SKF and ALTEN Italia S.p.A. SKF is a multinational Swedish company that specializes in bearing manufacturing. ALTEN is a leading European IT consulting group that works with SKF on the project development. The project’s goal is to digitally innovate the SKF plant in Cassino (FR, Italy) by using machine learning techniques to improve the efficiency of bearing manufacturing and support decision making with data. The consulting’s goal has been to provide a complete system for making real time prediction maintenance of machinery in the production line. The goal has been to develop a support system for workers that provides real-time information about the state of health of the machinery and, if necessary, alerts them to intervene with maintenance in order to avoid any malfunctioning or break-down. To reach this aim, algorithms of anomaly detection and change point detection have been used. In particular, IsolationForest algorithm and algorithms which detect changes in mean in time series have been explored. These latter have been useful in distinguishing time intervals when a machine is operating normally from others when it is operating abnormally. After that, the idea has been to create data giving information about 24 hours and classification models have been trained on them, in order to predict in the future the state of health of the machinery and intervene when the prediction will result anomalous.