A Novel Diagnosis Framework for Maglev Levitation Control Faults Based on Contrastive Self-Supervised Feature Learning

The performance of levitation control is critical for maglev trains during operation. However, it is highly susceptible to external interferences, leading to faults. It is traditionally assessed solely based on levitation gap data which is often unavailable. In contrast, an acceleration-based, data-driven fault diagnosis presents a more comprehensive and reliable approach. However, such data-driven diagnosis faces challenges due to feature learning difficulties, including large datasets, limited labeled data, and experimental constraints. Consequently, a novel diagnosis framework for maglev levitation control faults is developed utilizing a contrastive self-supervised learning algorithm. The accuracy of the proposed framework is validated using accelerations collected from an in-situ condition monitoring system. Furthermore, the framework effectively extracts discriminative features associated with normal and fault conditions caused by load changes, rail irregularities, and resonance. The contrastive self-supervised learning algorithm additionally offers insights into the underlying features of the data, and the model classifies with accuracy comparable to that of fully supervised models, facilitating feature learning from large-scale unlabeled datasets.