NOVELTY DETECTION FOR PREDICTIVE MAINTENANCE

Abstract

Since the advent of Industry 4. 0 significant research has been conducted to apply machine learning to the vast array of Internet of Things (IoT) data produced by Industrial Machines. One such topic is to Predictive Maintenance. Unlike some other machine learning domains such as NLP and computer vision, Predictive Maintenance is a relatively new area of focus. Most of the published work demonstrates the effectiveness of supervised classification for predictive maintenance. Some of the challenges highlighted in the literature are the cost and difficulty of obtaining labelled samples for training. Novelty detection is a branch of machine learning that after being trained on normal operations detects if new data comes from the same process or is different, eliminating the requirement to label data points. This thesis applies novelty detection to both a public data set and one that was specifically collected to demonstrate a its application to predictive maintenance. The Local Optimization Factor showed better performance than a One-Class SVM on the public data. It was then applied to data from a 3-D printer and was able to detect faults it had not been trained on showing a slight lift from a random classifier.