Supervised Regression Learning for Maintenance-related Data

Maintenance is among highest operational expenses in manufacturing companies, where production assets can be extremely complex and expensive. It is very difficult to collect fault related data in usual manufacturing environment, where production stops are to be avoided. For this reason, several researchers focused on the development of data sets made of signal acquired from machines led to faults. Among these sets of data there is the renown Milling Data Set (NASA Prognostic Center of Excellence), that is used in this work for the evaluation of possible data-driven Predictive Maintenance modeling attempts. Starting from a literature review dealing with the Milling data set, in this paper a revised work is iterated and then enhanced by the proposed approach based on features extraction, and on supervised regression learning models exploiting these features. Results achieved show how much time-domain features are important in the predictive maintenance domain, since best performances achieved are always connected to features extraction procedure where time-domain also have been extracted. Moreover, critical issues of Deep Learning with respect to data availability and model's input will be discussed as a general takeaway of this work.