An Unmasking-based method of anomaly detection for rotating devices

In anomaly detection, detecting anomalies in rotating devices with abnormal data deficiency presents a significant challenge. This paper proposes a novel Unmasking-based approach for anomaly detection in rotating devices. Assuming that the separability of similar data concentrates on a small number of features, while the separability of dissimilar data is scattered across most features. By continuously removing important features and observing the changes in indicators, we can assess the differences between data. A scoring method of anomaly degree is proposed. The evaluation indicator has been improved by using the Area Under the Curve (AUC) instead of Accuracy (ACC) to achieve higher accuracy and fault tolerance. The method is validated with 2 datasets, showcasing its ability to identify abnormal data accurately without specific training on anomalies. A testing framework based on Unmasking has been proposed and demonstrated to be effective and accurate using data from multiple operating conditions.