Search

• The proposed unsupervised method merges anomaly detection and fault tracing. • Signal decomposition-based energy features are extracted to make the diagnosis more robust. • The model training process only requires normal data. • The diagnostic process can be performed without labeled fault samples. The absence of detailed fault data presents a significant challenge in diagnosing faults of rotating machinery. This paper focuses on addressing abnormal detection and fault localization challenges in drivetrain components under the assumption that only normal data is available. Specifically, this paper proposes an unsupervised hypersphere description (UHD) framework by seamlessly integrating abnormal detection and localization. By constructing a one-class description hypersphere, this UHD method achieves fault detection without special fault samples. Abnormal samples can be detected based on their distance from the center of hypersphere, enabling accurate localization of the underlying abnormality. Validation of experimental data demonstrates that the UHD approach effectively detects anomalies and accurately localizes faults within specific locations. Remarkably, the UHD approach surpasses several mainstream unsupervised methods in terms of fault tracing accuracy. This shows the potential of the UHD approach in abnormal detection and localization in a drivetrain, particularly when only normal data is available. [ABSTRACT FROM AUTHOR]