Abstract:Bearing is recognized as one of the most critical components of space flywheels, making its operational monitoring essential due to the high frequency of failures. Given the limited availability of fault samples and the variety of fault modes, employing unsupervised methods for state monitoring is expected to address challenges related to sample scarcity and incomplete fault information. A method based on an improved ensemble auto-encoder is proposed to enhance the accuracy and robustness of unsupervised monitoring models for space flywheel bearings. First, monitoring models for flywheel bearings are developed using convolutional, denoising, and sparse Auto-encoders, which utilize local feature extraction, sparse representation, and strong generalization capabilities based on the concept of reconstruction error. Next, a scoring method is designed that reflects the degree of differences in sample states, combining reconstruction error with specific criteria. The model's robustness is then assessed through the permutation entropy of the anomaly scores from the sample set, which is used to allocate scoring weights and construct a comprehensive state index. Finally, this method is validated with ground experimental data from various models of space flywheel bearings, even under low-quality data conditions. The results indicate that the proposed method achieves high accuracy in detecting fault states, demonstrating improved robustness and a lower false alarm rate compared to traditional auto-encoders.