EDGE real-time predictive maintenance framework
The term “predictive maintenance” is increasingly common in today’s industrial language. In the past, the preventive maintenance approach was used to substitute components after specific time periods. Sometimes this substitution was useless, throwing away the machine’s component even though it might work without problem. Modern maintenance approaches work in this direction, solving this wastefulness by warning the user only when a real problem appears or will appear on a specific component.
The proper term to identify a problem that is occurring at that moment is “anomaly detection.” This kind of detection does not compute any prediction but merely alerts the maintainer of the appearance of a problem.
On the other hand, predictive maintenance computes a prediction of the failure before it occurs, giving time to the maintainer to substitute that component without interruption. Following the Industry 5.0 paradigm, the predictive maintenance direction covers the wastefulness reduction issue well.
Furthermore, a Predictive Maintenance framework could be developed by monitoring the machine in real-time and by using existing data collected in the past about problems already encountered. Often this collection does not exist, creating problems in the framework application. An innovative and modern approach has been studied in these years to solve this inconvenience, calling it “novelty detection”. This methodology focuses on detecting when a novelty appears in the data. This algorithm can help companies use a sort of predictive maintenance framework even without past data collection.
My doctorate is based on all the elements within the predictive maintenance context. In particular, EDGE real-time novelty detection is the main topic I am working on. “EDGE” due to the exigency to load algorithms in a local and stand-alone application. Real-Time due to the need to detect a novelty (e.g., a possible failure) as soon as possible. The main advantage of this approach is its scalability in different domains such as the industrial (e.g., machines, complex systems), automotive (e.g., cars), and energy saving (e.g., accumulators, batteries).
Searching and studying other new methodologies (AI-based): current state-of-art gives many solutions that sometimes cannot be applicable in a company due to their complexity and low accuracy in the detection/prediction;
Deployment of algorithms on an embedded device such as low-power MCUs: the solutions proposed in recent works often provide algorithms that are too complex to be implemented on an embedded device such as a low-power MCU. All algorithms that will be studied have the lightness property as their basis, an uncommon feature of modern solutions;
Developing custom intelligent parts (also using Additive Manufacturing/3D-MID technology) to substitute “passive” components and better monitor the machine’s behavior: from the application point of view, implementing the smart sensor directly inside the mechanical component is new to the company’s visio
Department: PhD in Eletrical, Eletronics and Communication Engineering
Supervisor: Prof. Marcello Chiaberge
