In the context of big data, the effective extraction and monitoring of atypical fault features in bearing cages remain a significant challenge. Existing deep learning-based methods are often limited by poor interpretability and a strong dependence on training data, making them less effective in identifying rare or non-standard fault patterns. To address these issues, a fault feature detection method is proposed, integrating statistical indicators with Multivariate Variational Mode Decomposition (MVMD). In this approach, statistical features are first extracted from the signal using a sliding window. The extracted features are then modeled under a Gaussian distribution, allowing atypical anomaly segments to be identified. Subsequently, MVMD is applied to enhance the atypical components within the signal. Based on the enhanced results, the atypical fault features of the bearing cage are effectively extracted. The proposed method provides improved interpretability and robustness, making it suitable for intelligent fault diagnosis in bearing cage and complex mechanical systems.