A multi-scale residual neural network (Res2APCNN) model combining data fusion and adaptive attention mechanism is proposed to monitor the health of aero-engine inter-shaft bearing under strong noise. Firstly, the bearing signals are converted into two-dimensional grayscale images by using the Gram angular difference field (GADF), Gram angular sum field (GASF) and Markov transfer field (MTF) methods. These three images are mapped to the RGB channels respectively to construct composite color images, thus enhancing the capture ability of time series information. Secondly, Res2Net module is introduced to extract multi-scale information through parallel convolution operation, filter noise interference and optimize information flow. Thirdly, the adaptive parallel feature fusion module is embedded to assign differentiated weights to feature dimensions, enabling the screening and amplification of key feature signals. Finally, the fault types of inter-shaft bearings are identified through a feature extraction and classification module. The proposed model is verified by using the bearing datasets of Polytechnic University of Turin in Italy and Harbin Institute of Technology, as well as the self-built test bench dataset. The experimental results show that the proposed Res2APCNN model demonstrates excellent fault diagnosis performance in a strong noise environment. Compared with advanced existing methods, the model achieves a 1.52% increase in accuracy over the IDRSN method on the Turin dataset, a 6.65% increase over the MC-CNN method on the HIT dataset, and a 2.35% increase over the Wen-CNN method on the self-built dataset. Furthermore, the diagnostic accuracy rate of this model exhibits the least fluctuation, indicating superior stability. Even under strong noise conditions, the Res2APCNN model can still maintain a high recognition accuracy and show good anti-interference ability.