A Multi-sensor Dynamic Task-weighted Cross-Domain Meta-learning Network (MDTMN) is proposed to address the challenges of multi-source vibration signals collected by multiple sensors for rolling bearings in industrial scenarios, scarce fault samples caused by drastic working condition changes, negative transfer easily induced by uniform gradient optimization of traditional meta-learning, and insufficient mining of implicit spatial topological features of multi-channel signals. Firstly, a Chebyshev graph convolution composite representation module is constructed to map one-dimensional multi-channel vibration signals into graph topological structures and mine topological correlations among diagnostic samples. An adaptive task sampling strategy integrating three indicators of task diversity, signal kurtosis and task difficulty is designed to dynamically assign sampling probabilities to each meta-task, alleviating model oscillation and negative transfer brought by heterogeneous working condition tasks. Cross-working condition comparison, ablation and visualization experiments are carried out based on the self-built multi-speed bearing experimental dataset. The results show that the proposed MDTMN still achieves high diagnostic accuracy under 1-shot extreme few-sample and large-span domain shift scenarios, and has stronger generalization ability compared with MAML, graph convolution benchmark model and traditional multi-channel fusion network, which can meet the demand of few-shot cross-domain fault diagnosis of complex industrial equipment

Rolling bearing; Few-shot fault diagnosis; Meta-learning; Graph convolution; Adaptive task sampling; Negative transfer suppression