Bearings are critical bogie components, making their early fault detection particularly important. This paper proposes an early fault detection method for bearings based on a relation network (RN). A health status detection relation network model is designed to effectively extract bearing condition features and measure the nonlinear distance between these features. In the offline modeling phase, normal samples from the bearing are collected for training, allowing the model to learn the nonlinear distances among the healthy state sample features. During the online monitoring phase, samples from the current operating state are acquired, and a relation score is obtained as a health indicator for the bearing condition. The 3σ criterion is then applied to determine the health indicator threshold for detecting the bearing health status and identifying faults promptly. Experiments were conducted on the XJTU-SY rolling bearing full-lifecycle dataset. Results show that, compared to methods like root mean square, kurtosis, and stacked autoencoders, the health indicator of the proposed method is more sensitive to early faults and exhibits better monotonicity and trend. Furthermore, in comparison with methods such as Isolation Forest, Support Vector Machine, and stacked autoencoders, the proposed method detects the first fault occurrence earlier, demonstrating considerable practical value.