A temporal-frequency contrastive learing method for acoustic-based mechanical fault detection in gas-insulated switch-gear

Although data-driven acoustic-based fault diagnosis has achieved satisfactory results for gas-insulated switchgear (GIS) under the massive labelled datasets built in laboratory conditions, it remains challenging to ensure accurate and robust diagnosis when facing limited fault samples and operational variability in real-world scenarios. To address this, this paper proposes a novel time-frequency self-supervised contrastive learning method for GIS fault diagnosis under small-sample and diverse operational conditions. The method adopts a pre-train and fine-tune framework to extract shared discriminative features across different domains. In the pre-train stage, an unsupervised contrastive learning strategy is used to learn the physical consistency of information in the time-frequency domain of mechanical faults. In the fine-tune stage, only 5% of the labelled target data is required to adapt the model using the shared representations from pre-training. Experiments on laboratory and field data validate the effectiveness of the proposed method, which achieves 97.64% accuracy and shows competitive generalisation and robustness compared to existing methods.