<p>Bearings represent critical rotating components in aero-engines, where incipient fault detection is essential for ensuring safety. To address the challenges of incomplete feature extraction and limited adaptability to cross operating conditions in bearing incipient fault diagnosis, this paper proposes a novel time-frequency feature fusion cross-domain diagnosis method based on acoustic emission signals. The proposed method employs a time-frequency feature fusion network (TFFN) for feature extraction, which utilizes a dual channel parallel convolutional neural network (CNN) integrating long short term memory networks (LSTM) and a temporal attention mechanism. This design enables the effective capture and adaptive fusion of time-frequency domain features in acoustic emission signals. Additionally, a mean-covariance domain adaptation (MCDA) criterion is proposed for joint distribution alignment, effectively enhancing the model’s generalization capability under cross-domain bearing fault diagnosis. Experimental validation was conducted using both an aero-engine rotor test bench and the publicly available JUST slewing bearing dataset. The experimental results demonstrate that the proposed method achieves better fault diagnosis accuracy and superior domain adaptability in both cross-domain and limited samples.</p>