Self-supervised Learning for Time Series Analysis in Predictive Maintenance Application
With the advancement of the Industry 4.0, the endorsement of artificial intelligence methods, the dominant development of the Internet of Things (IoT), operational industries have adapted preventive maintenance approach by monitoring machines health closely via multiple sensors. Machine condition monitoring system for rotating machines, is a very effective tool to enhance machine reliability, reduce maintenance cost and prevent rotating equipment catastrophic failure during operation. For rotating machines, rolling bearings are the most common cause of machine failures. Therefore, early detection of bearing failures and fault diagnosis can be done through monitoring and analyzing the vibration signals during machine operation by various deep learning methods. Unlike model-based vibration analysis tools that require expert knowledge, deep learning is a powerful data-based tool for bearings fault diagnosis. The primary goal of the thesis is to present a benchmarking of deep learning unsupervised anomaly detection models for early bearing failure detection using raw vibration data, employing Autoendcoders and TadGAN framework. The second goal is to classify bearing defects using an innovative self-supervised framework trained by Siamese networks. Experimenting on CWRU and Paderborn bearing datasets, SSL models have demonstrated an improved accuracy in classification with less labeled data.
A.H.H. Alhashmi, "Self-supervised Learning for Time Series Analysis in Predictive Maintenance Application", M.S. Thesis, MBZUAI, Abu Dhabi, UAE, 2022.