In recent years, intelligent transfer models have focused on narrowing the gap between the source domain and target domain data to improve diagnostic effectiveness. However, collecting unlabelled target domain data in advance is challenging, leading to suboptimal performance of domain adaptation models for unknown target domain data. To address this issue, this paper proposes a deep physical information consistency embedded network (DPICEN) for tackling unknown domain bearing fault diagnosis problems. First, a physical information encoder (PIE) is constructed to encode physical information into tensors with values of 0/1. Second, fault samples and their encoded tensors are embedded into a physically consistent space, and the mean squared error (MSE) is employed to reduce the distance between data feature embeddings and physical information embeddings. Subsequently, to further constrain the distribution differences of unknown domain data, a plug-and-play multiple sparse regularization (MSR) algorithm is proposed. Finally, the embedded features are input into a classifier with MSR to achieve bearing fault diagnosis. The results demonstrate the effectiveness and advancement of DPICEN in comparison with 16 related methods in 13 unknown domain fault diagnosis tasks in three bearing datasets. The code can be found at https://github.com/John-520/Models-for-DPICEN.