Many cross-domain bearing fault diagnosis methods based on deep transfer learning have been emerged over the past few years. Nevertheless, most of these existing diagnostic methods make a demand on the sufficient labelled data and the involvement of the target domain for training, which is difficult to meet in engineering scenarios. Aimed at these abovementioned challenges, this paper proposed a semi-supervised proxy contrastive generalization network (SSPCGN) for bearing fault diagnosis towards previously unobserved operating conditions under the absence of partial label information. In SSPCGN, a proxy-based contrastive loss is intended to excavate cross-domain transferable features though the contrastive relations among proxies and samples. Meanwhile, a semi-supervised cross-entropy loss is utilized to make full use of the unlabeled samples by a pseudo-labeling coefficient. The experiment of Paderborn bearing dataset has proved the outstanding diagnostic performance of SSPCGN in semi-supervised domain generalization.