ABSTRACT The performance of rolling bearing fault diagnosis based on vibration signals has been steadily improving with the increasing sophistication of intelligent technologies. However, variations in the equipment’s operating condition cause fluctuations in the vibration signals, which significantly impact the robustness of the diagnosis. Besides that, differences in data characteristics between experimental setups and engineering applications restrict the application of fault diagnosis models in cross-domain transfer tasks. To address the challenges in diagnosis, a proposed transfer learning method named Model-Based Sampling and Wide-kernel Domain Adversarial Neural Network (MBS-WDDANN) has been designed in this study, which introduced the unlabelled sampling method and wide convolutional neural network into the domain adversarial neural network. This data pre-processing method is used to train the variable operating condition signal into sample aligned data in order to reduce the variability of samples. The wide convolutional kernel in the first layer of the WDDANN was used to optimise the extraction of short-term features. Through testing on CWRU data, the algorithm’s superiority over other intelligent fault diagnosis methods was validated. Moreover, further experiments on six datasets collected from MFS-MG bearing test bench demonstrated the ability and superiority in variable operating conditions.