An unsupervised approach to early fault detection and performance degradation assessment in bearings

Early fault detection and performance assessment are critical components of machine health management, with the primary goals being the detection of incipient faults and the development of a health index (HI) to monitor degradation. However, current methods typically require significant domain exper...

Descripción completa

Detalles Bibliográficos
Autores: Wang, shun, Vidal Seguí, Yolanda|||0000-0003-4964-6948, Pozo Montero, Francesc|||0000-0001-8958-6789
Tipo de recurso: artículo
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/439334
Acceso en línea:https://hdl.handle.net/2117/439334
https://dx.doi.org/10.1016/j.aei.2025.103620
Access Level:acceso abierto
Palabra clave:Unsupervised fault detection
Autoencoder
Health index
Degradation assessment
Descripción
Sumario:Early fault detection and performance assessment are critical components of machine health management, with the primary goals being the detection of incipient faults and the development of a health index (HI) to monitor degradation. However, current methods typically require significant domain expertise or labeled faulty data, and often face challenges in simultaneously achieving accurate early anomaly detection and maintaining a clear performance degradation trend. To address these limitations, in this work, a simple unsupervised fault detection framework is stated for early fault detection and performance assessment. The proposed method is grounded in spectrum analysis, where the logarithmic envelope spectrum (logES) is first introduced to enhance fault signatures and ensure a consistent representation of degradation. Subsequently, a variational autoencoder (VAE) is employed for anomaly detection, using the logES as input. The VAE learns the spectral distribution of healthy operational data, and its reconstruction error is used as the HI. This HI not only achieves precise early fault detection but also exhibits strong monotonicity and robustness in tracking performance degradation over time. Extensive experiments on four bearing run-to-failure datasets validate that the proposed framework delivers highly accurate incipient fault detection and generates a reliable HI. Compared to other established HI methods, the proposed framework accurately characterizes the health status, establishing itself as a powerful tool for machine health management without the necessity for labeled fault data.