A Data-Driven Approach to Air Leakage Detection in Pneumatic Systems

During the transition phase of traditional manufacturing companies towards smart factories, they are likely to experience challenges like lack of prehistoric data recordings or events on which the machine learning models need to be trained. This paper introduces a novel approach of artificially induced anomalies for data labelling. Moreover, for newly installed systems or a setup, which has not seen any kind of malfunction yet, the combination of artificially induced anomalies by experiments and machine learning model help to proactively prepare for any kind of future hindrance of the production systems. Two experiments were performed for detection of air leakage. The first one was designed to identify ‘sensitive feature’ and understand the behaviour of the sensor readings with respect to different state of the machine. The second one was performed to capture more data points pertaining to leaking state of machine on a normal production day since the first one was conducted on a maintenance break). RUSBoosted bagged trees model was built as a supervised machine-learning model, which was yielded 98.73% accuracy, 99.40% precision, recall of 99.21%, and F1 score of 99.30% on test data for detecting pneumatic leakage. As a conclusion, previously unknown hidden patterns and insights regarding temperature feature along with a standardized and systematic methodology are the important deliverables of this study.