Abstract: ［Objective］ The operating environment of axle box bearings in high-speed train is complex and changeable, and the diagnostic accuracy of single-source signals for weak faults is insufficient. To improve the diagnostic accuracy of early weak fault of axle box bearing, it is necessary to study a diagnosis method for high-speed train axle box bearing slight fault driven by temperature-vibration feature fusion, in combination with the multi-source fault information of bearing temperature and vibration. ［Method］ First, an AE (auto encoder) driven bearing temperature feature extraction method is designed to obtain the abnormal bearing temperature features, and EMD (empirical modal decomposition) method is used to process the vibration signal，so as to obtain the statistical features of the effective vibration IMF (intrinsic modal function). Then, by optimizing the dimensionality reduction algorithm based on SAE (stacked auto encoder), an effective fusion method of temperature-vibration features is proposed to achieve nonlinear fusion and dimensionality reduction of temperature and vibration features. Finally, combined with BP (back propagation) neural network, a slight fault diagnosis model for axle box bearing based on temperature-vibration feature fusion is established. And the model is validated by the test data collected from the high-speed train rolling bearing test bench. ［Result & Conclusion］ Compared with single-source feature-driven fault diagnosis method, the fault diagnosis accuracy of the diagnosis method based on temperature-vibration feature fusion is higher, with an average diagnosis accuracy rate of over 99%. Compared with the PCA (principal component analysis) temperature-vibration model, the proposed temperature-vibration fusion bearing diagnosis model is more accurate and effective.