Abstract
On-load tap changers (OLTCs) are critical to reliable transformer operation, yet mechanical fault samples are difficult to obtain in practice, limiting the feature representation and diagnostic performance of conventional data-driven methods under limited-sample conditions. To address this issue, this paper proposes a Semantic-Enhanced Fuzzy Prototypical Network (SFPN) for few-shot OLTC fault diagnosis. Three parallel encoders are constructed in the time, frequency, and time–frequency domains to extract complementary vibration features. Fault descriptions are encoded using a pretrained text embedding model and fused with vibration features through cross-modal attention. Fuzzy prototypes are then constructed according to the class memberships of support samples to reduce the influence of uncertain samples on prototype estimation. Experiments are conducted using vibration signals collected from a self-built OLTC mechanical fault simulation platform. Under a closed-set 4-way 5-shot episodic setting, where the fault classes are observed during meta-training and 5-shot denotes five support samples per class in each episode, SFPN achieves an average diagnostic accuracy of 97.24 ± 0.48%. SFPN also maintains superior performance across different support-sample sizes and noise levels, demonstrating its effectiveness for closed-set OLTC fault diagnosis under limited-support conditions.