Condition Monitoring Of Machine Components From Drive Data Using Semi-Supervised Anomaly Detection Methods

Abstract: The mission of the machine manufacturer is to gain insights from machine data to increase their machines' efficiency and sustainability. Continuously monitoring the machine data with machine learning helps to detect emerging mechanical problems and prevents unexpected failures. The current manual fault detection system, which relies on expert knowledge and static rules, has proven inadequate in identifying faults, necessitating the development of an automated and data-driven solution.  This thesis explores the potential of utilizing drive data and anomaly detection methods to monitor the condition of machine assets continuously. Four model candidates, namely HBOS (statistical), OCSVM (clustering), autoencoder (reconstruction-based), and ARIMA (time-series forecasting), are selected, trained in a semi-supervised manner, and evaluated based on their ability to detect faulty behavior in a servo motor. To simulate the behavior of an imbalanced motor and its load, a dedicated testbed is designed to generate labeled drive data (Healthy vs. Unhealthy), replicating a common cause of machine failure.  The candidates have all demonstrated outstanding detection accuracy (100% F1-score) when identifying motor load imbalances on a testbed, using current consumption as input. For real-time applications, the ARIMA and Autoencoder models stand out for their ability to make rapid predictions without requiring feature extraction. The thesis suggests implementing ARIMA as a forecasting model given its ease of implementation, consistent performance with minimal training data, and speed in detecting faults - even at a reduced sample rate of 10 Hz.