A data-driven approach to detect air leakage in a pneumatic system

Abstract

Maintenance practices in production systems have for the past decades followed a reactive approach. With the production domain currently shifting towards a new era, Industry 4.0, maintenance practises have experienced an incremental drift towards predictive approach. With key enabling technologies such as BigData, Industrial Digitalization, and Machine Learning that enables data-driven decision making, the field of production is now more ready than ever to establish new practises in maintenance. The fast-emerging area of data-driven decision making with key enabling technologies is where this thesis finds its roots. This thesis work aims to develop a data driven approach based on machine learning to identify an early stage of a pneumatic leakage in a production process, before the leakage is so severe that it would impact the process adversely and cause machine breakdown. Through two experiments and extraction of historical data coming from sensors, a supervised machine learning model was built on the extracted significant statistical features. Adding on, an unsupervised model was developed to distinguish separate clusters representing the normal working state and leaking state of the machine for another process in the same production line. The supervised machine learning model is successful in detecting the early stage of leakage with an accuracy of 98.2%, thus giving plenty of time to perform appropriate maintenance on the equipment. The unsupervised model with an accuracy of 99.3%, is successful in detecting the physically evident stage of leakage. The presented thesis makes salient recommendations for deployment of this model. Moreover, the provided valuable insights pertaining to features which were previously believed to be insignificant for modelling purposes, provide a standardized work methodology and a concrete platform for future study in the area.