Generalized MAML for few-shot cross-domain fault diagnosis of bearing driven by heterogeneous signals

Despite a few recent meta-learning studies have facilitated few-shot cross-domain fault diagnosis of bearing, they are limited to homogenous signal analysis and have challenges to flexibly extract generic diagnostic knowledge for multiple meta-tasks. In order to solve these problems, this paper presents generalized model-agnostic meta-learning (GMAML) for few-shot fault diagnosis of bearings cross various operating conditions driven by heterogeneous signals. The proposed method involves constructing a channel interaction feature encoder using multi-kernel efficient channel attention, which allows for focusing on mutual fault information and enabling effective extraction of general diagnostic knowledge for multiple diagnostic meta-tasks. Additionally, a flexible weight guidance factor is designed to adjust the training strategy and optimize the inner loop weights for different diagnostic meta-tasks, improving the overall generalization performance. This method is applied to analyse the acceleration and acoustic signals of bearings, and its extensiveness and effectiveness are verified through various few-shot cross-domain scenarios.