A Physical Prototype-Guided Self-Supervised Learning Method for Rolling Bearing Fault Diagnosis
Year: 2026
Type: conference-paper
Abstract: Rolling bearing fault diagnosis typically relies on fully annotated samples to train deep models. However, in industrial scenarios, acquiring fault samples is costly and operating conditions vary significantly, making supervised models prone to overfitting under limited label conditions. To address this issue, this paper proposes a raw vibration signal diagnosis method based on physical prototype-guided self-supervised learning. The method directly takes raw vibration windows as input. First, an augmentation consistency self-supervised task is employed to learn impact, modulation, and local transient structures from unlabeled vibration segments. Subsequently, fault physical descriptors are constructed according to bearing fault types and damage sizes, and a prototype alignment loss is introduced to constrain the deep feature space, ensuring that sample representations remain consistent with fault mechanisms such as inner race, rolling element, and outer race faults. Experimental results on the CWRU bearing dataset show that when only 10% of labeled training samples are used, the test accuracy of the plain supervised Raw-CNN is 35.00%, while the proposed method achieves 66.25%. The Macro-F1 score improves from 24.97% to 59.89%. When the labeling ratio increases to 30%, the Macro-F1 of the proposed method reaches 68.26%, still outperforming the purely supervised baseline. These experiments demonstrate that self-supervised pre-training combined with physical prototype constraints can alleviate the problem of insufficient representation in low-label diagnostic tasks.
Authors: Xiao Xiang, Minglu Xu, Peng Shu, Shuying Ma
Institution: Chengdu Technological University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202202425