A Comparative Study of Semi-Supervised Anomaly Detection Methods for Machine Fault Detection

Industrial automation has extended machines’ runtime, thereby raising breakdown risks. Machine breakdowns not only have economic and productivity consequences, but they can also be fatal. Thus, the early detection of fault signs is essential for the safe and uninterrupted operation of machinery and its maintenance. In the last few years, machine learning has been widely used in machine condition monitoring. Most existing approaches rely on supervised learning techniques, which face challenges in real-world scenarios due to the lack of enough labelled fault data. Additionally, models trained on historical fault data might struggle to detect new and unseen faults accurately in the future. Therefore, this research uses semi-supervised Anomaly Detection (AD) techniques to detect abnormal patterns in machines’ vibration signals. As semi-supervised techniques are trained on normal data only, they do not require faulty samples and abnormal patterns are detected based on their deviations from the learned normal pattern. We compared the effectiveness of seven state-of-the-art AD methods, ranging from traditional approaches such as isolation forest and local outlier factor to more recent Deep Learning (DL) approaches based on autoencoders. We evaluated the effectiveness of different feature types extracted from the raw vibration signals, including simple statistical features like kurtosis, mean, peak-to-peak, and more complex representations like the scalogram images. Our study on three public datasets, with unique challenges, shows that the traditional methods based on simple statistical analysis have shown comparable and sometimes superior performance to more complex DL approaches. The use of traditional approaches offers simplicity and lower computational needs. Thus, our study recommends that future researchers start with the traditional approaches first and then jump to DL methods if necessary.