Wisdom via Multiple Perspectives: A Multigranularity Clusters Fusion Approach for Fault Diagnosis With Noisy Labels

Fir Dunkin, Xinde Li, Guoliang Wu, Chuanfei Hu, Le Yu, Xiaoyan Lu, Shuzhi Sam Ge

Abstract: Deep neural networks (DNNs) have shown excellent performance in fault diagnosis, but this heavily relies on training datasets with high-quality labels. However, both manual annotation and automatic annotators inevitably introduce noisy labels into the dataset, which mislead the model during training and negatively affect generalization, especially in strong noisy environments. Therefore, this article proposes a multigranularity label construction approach via multigranularity cluster fusion (MgCF), aiming to more efficiently suppress the misleading effect of noisy labels. MgCF first constructs a granular cluster for each sample in the latent feature space based on the estimation of the noise intensity of the training dataset, and estimates the membership relationship of the sample to each category based on the distribution of labels in the cluster. Then, by fusing the membership relationship with the original observation label, a corrected multigranularity label is obtained. Finally, the fusion label replaces the original label to complete the training process. A series of experimental results and theoretical analyses confirm MgCF's effectiveness, offering a robust training strategy for intelligent fault diagnosis even with low-quality, cost-effective datasets. MgCF has the potential to expand the practical application of such datasets, advancing the development of intelligent mechatronic systems.

External IDs:doi:10.1109/tmech.2025.3558839