KD-APFNet: Knowledge Distillation and Adaptive Prototype Fusion for Cross-Load Few-Shot Bearing Fault Diagnosis
Year: 2026
Type: article
Abstract: For bearing fault diagnosis under load-induced spectral drift and extremely scarce target-domain labels, this paper proposes the Knowledge Distillation and Adaptive Prototype Fusion Network (KD-APFNet). First, a lightweight time–frequency student network encodes vibration segments with short-time Fourier transform spectrograms, and learnable band masks initialized from bearing characteristic frequencies emphasize fault-related bands under load-induced spectral drift. Second, hierarchical dual-teacher distillation compresses a wide teacher into a compact student to transfer inter-class structure under scarce target labels. Finally, adaptive prototype fusion builds metric, student, and teacher prototype classifiers on one frozen checkpoint and arbitrates decisions by confidence; source-domain training and target-domain adaptation do not rely on a target unlabelled pool, and transductive prototype refinement is optionally available at inference. Primary experiments are conducted on Case Western Reserve University drive-end data. Under a strict no-leakage evaluation protocol, macro-averaged accuracy reaches 87.62% with one support sample per class, stably outperforming representative few-shot baselines across three variable-load tiers, with negligible difference between the two inference modes. Extension experiments on Paderborn University and Xi'an Jiaotong University data further show that the framework transfers across test rigs and that inference modes can be selected in a graded manner according to condition-shift severity.
Source: Engineering Research Express
Authors: Cunli Song, Wenjing Yu, Weiguo Shi
Institution: Dalian Jiaotong University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7223384344