Various sensors implemented in the railway transportation system brings opportunities in improving its safety and challenges in fault information mining. Extracting effective and synthetic fault-specific information from the over-rich data is one of the key challenges. The classical feature dimension reduction methods are mostly based on the statistical correlation among variables. Considering the cause-effect relationship may reflect the true influence of one variable on the other, this paper proposes three unsupervised feature extraction methods based on causal network. Precisely, after discovering the causal network among the monitoring variables in a rail transportation system, principle components related to the specific fault are extracted from the causal strength matrix or the full causal strength matrix constructed from the causal network. In comparison with the state-of-art correlation-based feature reduction methods, the effectiveness of the proposed methods is verified on two public datasets and a real dataset considering high-speed train braking system. In addition, the intrinsic working mechanism of the proposed methods is analyzed with respect to the constructed causal network, which improves the interpretability of the fault detection and diagnosis.