Hybrid model-based predictive maintenance: a granulator case study

This dissertation shows a practical case in which there are no historical data available to perform predictive maintenance on a machine, but it is still possible to estimate the behavior of a machine when a component is affected by an increasing degradation in time. Typical predictive maintenance model methods are data-driven and use machine learning algorithms, trained with large numbers of data: the problem arises when the system is new, so there are no data enough to perform an only data driven approach, so it is necessary to start from a physical model. As usual, a model is a simplification of reality, then some parameters are not certain: to solve the possible lack of consistency, the data driven approach allows to identify them. The physical model development has been done with the 20-Sim® software using the bond-graph technique, which allows to represent multi-domain engineering systems in a simple way, starting from power transmission and kinematic relations instead of dynamics. In addition to it, in the 20-Sim® model have been developed also the physical elements that bring the system towards a fault state and, in particular, have been developed both a draft model with one degree of freedom (angular velocity) and the final one with two degrees of freedom (horizontal and vertical velocity). It is then presented the FFT analysis of the faulty state compared to the healthy one, focusing on the fault occurring on a specific component, interesting for the heavy impact that has on the machine and both for its economical cost. Given the impossibility to implement a reliable model for predictive maintenance, have been provided the guidelines to achieve it. In conclusion, possible improvements and future developments of this model are presented, along with extensibility to similar machines.