Fault diagnosis of wheelsets under non-periodic fragmented signals: a single-sensor approach based on phase space reconstruction

Abstract In the metallurgical and port industries, the operational safety of multi-wheelset equipment, such as overhead cranes, is critical. Traditional monitoring approaches requiring sensors on every wheelset incur prohibitive costs and complex wiring. Alternatively, deploying a single track-side sensor offers a cost-effective solution but generates vibration signals that are inherently non-periodic and fragmented, rendering conventional time–frequency analysis methods ineffective. To address this engineering challenge, this paper proposes a novel fault diagnosis framework based on phase space reconstruction (PSR). The proposed method first employs empirical mode decomposition to denoise raw signals and extract intrinsic mode functions, selecting the top four components based on energy and correlation criteria. Subsequently, a comprehensive 328-dimensional feature vector is constructed by fusing two-dimensional PSR geometric features and high-dimensional PSR statistical features derived from the reconstructed trajectories. Finally, uniform manifold approximation and projection is utilized for dimensionality reduction and visual classification of fault states. Validation experiments performed on 1000 sets of on-site vibration signals from a 150 ton overhead crane demonstrate the framework’s superiority. The method achieves perfect clustering performance (adjusted rand index of 0.9966 and adjusted mutual information of 0.9941) and distinct fault separation, significantly outperforming traditional time–frequency analyses and convolutional neural networks. These results confirm the framework’s capability to capture dynamic features from fragmented signals, providing a robust and economical solution for the predictive maintenance of heavy-duty industrial equipment.