Elevator Brake Fault Diagnosis Method Based on Active Learning of Temporal Ensembling Gradient Vector

To address the high annotation costs in elevator brake fault diagnosis, this study proposes an active learning query strategy based on temporal ensembling gradient vectors. By integrating historical prediction information with current gradient dynamics, the method accurately captures model uncertainty for unlabeled samples. Additionally, it employs the k-means++ algorithm to optimize the sample selection process, significantly enhancing sample diversity and avoiding information redundancy. Experimental results demonstrate that the proposed method achieves a diagnostic accuracy of 99% on an elevator brake fault diagnosis dataset, requiring only 320 annotated samples—a 54.54% reduction in annotation costs compared to random sampling. Furthermore, an analysis of learning difficulty across different fault categories shows that the method effectively allocates annotation resources by prioritizing samples from harder-to-learn categories, thereby improving both diagnostic performance and model generalization. The findings provide a novel approach for applying active learning to fault diagnosis of elevator brakes and other mechanical systems, offering significant implications for resource optimization in industrial model development.