CHEN Z G,ZHANG Z H,WANG Y X, et al．Few-shot bearing fault diagnosis based on HyperMAML［J］.Journal of Henan Polytechnic University( Natural Science) ，DOI: 10.16186/j.cnki.1673-9787.2024120023

DOI: 10.16186/j.cnki.1673-9787.2024120023

Received：2024-12-23

Revised：2025-06-07

Online：2025-07-30

Few-shot bearing fault diagnosis based on HyperMAML

Chen Zhigang, Zhang Zhihao, Wang Yanxue, Liu Jiale

(School of Mechanical and Vehicle Engineering, Beijing University of Civil Engineering and Architecture 100044, Beijing, China)

Abstract: Objectives To address the issues of prolonged diagnosis time caused by the need to ensure diagnostic accuracy under few-shot conditions and the low accuracy resulting from the inability to effectively extract fault features in rolling bearings, a fault diagnosis method, named HyperMAML, based on model-agnostic meta-learning(MAML) integrating hyper-networks was proposed. Methods Firstly, to more comprehensively represent the temporal information and fault features in the time series data, the time series were transformed into two-dimensional grayscale images using the Gramian Angular Summation Field (GASF). Then, the weights generated by the hypernetwork were used to update the weights in MAML, effectively avoiding the redundant gradient updates during a single iteration in MAML and thereby reducing training time. Finally, the updated MAML weights were applied to perform fault diagnosis on the dataset. Results Analysis of both public datasets and self-built test stand datasets indicated that on public datasets, HyperMAML achieved an accuracy of 96.62%, with training and testing times of 1.221s and 1.346s, respectively. MAML's accuracy was 91.06%, with training and testing times of 11.313s and 6.007s, respectively. RelationNet's accuracy was 92.18%, with training and testing times of 1.644s and 1.559s, respectively. ProtoNet's accuracy was 95.08%, with training and testing times of 1.646s and 1.422s, respectively. DCA-BiGRU's accuracy was 95.34%, with training and testing times of 1.021s and 0.940s. On the self-built test stand datasets, HyperMAML's accuracy was 97.89%, with training and testing times of 0.851s and 0.824s, respectively. MAML's accuracy was 97.89%, with training and testing times of 6.04s and 2.603s, respectively. RelationNet's accuracy was 98.00%, with training and testing times of 0.847s and 0.896s, respectively. ProtoNet's accuracy was 97.75%, with training and testing times of 1.032s and 0.875s, respectively. DCA-BiGRU's accuracy was 97.80%, with training and testing times of 0.744s and 0.692s, respectively. Conclusions The few-shot fault diagnosis model based on HyperMAML leveraged the capability of hypernetwork to generate dynamic weights and the meta-learning framework advantages of MAML. It effectively met the requirements for fault feature extraction under few-shot learning conditions, improved model training efficiency and diagnostic speed while ensuring diagnostic accuracy and stability. This model provideed an innovative solution for intelligent fault diagnosis under small-sample learning scenarios.

Key words: fault diagnosis; rolling bearings; few-shot learning; meta-learning; Gramian Angular Field