TAMTL: A Novel Meta-Transfer Learning Approach for Fault Diagnosis of Rotating Machinery

Yuheng Wu, Qingyu Yang, Donghe Li, Pengtao Song

Abstract:

This paper proposes a novel fault diagnosis scheme
for rotating machinery based on meta-transfer learning and
test augmentation. The model-agnostic meta-learning (MAML)
framework is applied to the fault diagnosis problem by dividing
the training and testing tasks according to different operating
conditions, which allows the user to arbitrarily select the appro
priate basic model according to the task requirements. Then, an
additional pre-training phase based on meta-transfer learning is
designed to improve the comprehensive performance, and a test
ing stage is introduced to evaluate the generalization performance
of the hyperparameters of fine-tuned model. Experimental results
on the CWRU dataset demonstrate that the proposed scheme can
achieve high accuracy, stability, and efficiency in fault recognition
under cross-condition scenarios.