Fault diagnosis (FD) is an important foundation for the maintenance of complex aerospace systems, such as satellite power systems, in which the attribute reduction has essential effect to eliminating data redundancy and improving diagnostic results. However, due to the difficulty and high cost of obtaining labels in some situations, especially early failures, FD based on unsupervised methods is of great significance but less commonly-studied. Moreover, with respect to FD preprocessing, unsupervised attribute reduction (UAR) usually applying clustering methods suffers from the need for cluster number, randomness, inability to handle non-spherical clusters, etc. Therefore, this paper proposes an unsupervised FD strategy including a knowledge acquisition method to mine the rules from the unlabeled data, a decision-making method to process the acquired knowledge, and a diagnosis decision for the fault identification. As for the preprocessing part, this paper proposes a wrapper UAR method (named DPC-UAR) based on the density peak clustering (DPC) and heuristic method, which can automatically identify the cluster centers and deal with the nonspherical data. Finally, experiments of attribute reduction performance on UCI data show that compared with other UAR methods, DPC-UAR has the greatest effect to improve performance of unsupervised learning algorithms, and plays a relatively good role in the supervised algorithm. Experiments on satellite power system fault diagnosis illustrated that the proposed FD strategy based on DPC-UAR has high accuracy, a high fault detection rate, and a low false alarm rate.