A dual-channel time-frequency meta-learning approach for bearing fault diagnostics under variable operating conditions

Abstract In real industrial scenarios, owing to data scarcity and variable operating conditions, the fault diagnosis of rotating machinery is still very challenging. To address the above limitations, a few-shot cross-domain fault diagnosis method is proposed based on dual-channel time-frequency feature and model-agnostic meta-learning (DCTFF-MAML). First, a dual-channel feature fusion method is proposed, and a DCTFF diagram integrating the fault mechanism and time-frequency characteristics is obtained by independently processing high-frequency transients and low-frequency harmonics through continuous wavelet transform (CWT). Then, a multi-source two-stage ML method is proposed to enhance the cross-domain diagnosis accuracy of the model by aligning the cross-domain feature space with multi-source data during the meta-training process. Finally, the effectiveness of the proposed method is verified using two case studies. Experimental results demonstrate that DCTFF-MAML can achieve higher diagnostic accuracy and better generalization capability than several mainstream domain adaptive and ML methods.