In the field of fault diagnosis for rolling bearings under variable working conditions, significant progress has been made using methods based on unsupervised domain adaptation (UDA). However, most existing UDA methods primarily achieve identification by directly aligning the distributions of the source and target domains, often overlooking the relevance of samples between different domains, which may result in incomplete extraction of deep features and alignment of feature distributions. Therefore, this study proposes a novel domain adaptation network based on Gaussian prior distributions, aiming at solving the challenges of cross working conditions bearing fault diagnosis. The method consists of a feature mining module and an adversarial domain adaptation module. The former effectively extracts deep features by stacking multiple residual networks (Resnet), while the latter employs an indirect latent alignment strategy, using Gaussian prior distributions in the latent feature space to indirectly align the feature distributions of the source and target domains, achieving more precise feature alignment. In addition, an adaptive factor is introduced to dynamically assess the method’s transfer and discriminative capabilities. Experimental data from two bearing systems validate that the method can effectively transfer source domain knowledge to the target domain, confirming its effectiveness.