The reliability of rolling bearings is one of the crucial guarantees for the continuity and safety of industrial production. However, due to the high temperature, high pressure, and long duration characteristics of their operating environment, the vibration signals often exhibit certain non-stationarity and non-linearity. Additionally, the lack of fault samples makes it challenging to apply data-driven diagnostic methods. This paper proposes a fault diagnosis method for rotating machinery under variable operating conditions based on multi-feature and transfer learning. Firstly, dimensionless features of bearing vibrations are extracted to address the non-linearity of bearing information. To tackle the issue of missing fault samples, an improved convolutional neural network transfer model is proposed to transfer large-scale data models to small-sample models. Validation on the bearing experiment platform of Case Western Reserve University shows that the proposed method achieves an average diagnostic accuracy of 95.9%, providing a theoretical basis for the fault diagnosis of rolling bearings. © 2024 IEEE.