Rotor system fault detection utilizing semi-supervised and unsupervised machine learning

Detecting multiple simultaneous faults in rotor systems is challenging, especially when labelled data is limited. This paper presents a novel framework combining unsupervised and semi-supervised machine learning to enhance fault diagnosis in rotor systems with various fault types. Using finite element method simulations, 100 vibration signal observations were generated for rotor systems under three fault conditions: imbalance, imbalance with shaft bending, and imbalance with cracking. Features were extracted via a multi-layer autoencoder in an unsupervised manner, followed by sequential feature selection to identify the most informative attributes. Two classification approaches were then applied: k-means clustering for unsupervised fault detection and a semi-supervised model with a Softmax layer for classification. The semi-supervised method achieved over 95% accuracy using only three selected features, effectively distinguishing different fault types. In contrast, the unsupervised approach proved better suited for anomaly detection rather than precise fault identification. These results demonstrate the potential of integrating unsupervised feature extraction with semi-supervised classification for reliable fault diagnosis in rotor systems with scarce labelled data.