Incremental Learning with Maximum Dissimilarity Sampling Based Fault Diagnosis for Rolling Bearings

Traditional fault diagnosis methods for rolling bearings require retraining the model from scratch when faced with new fault signals, consuming significant computational resources and exhibiting lower efficiency. Class incremental learning can be applied to online fault diagnosis, effectively reducing resource consumption. However, the model often experiences catastrophic forgetting during the training process, a problem that can be addressed by replaying data. Nonetheless, the subsets obtained by the current sampling strategy often fail to cover the distribution of classes, leading to localized information gaps and subsequent degradation in the performance of the fault diagnosis model. To address this, this paper introduces Maximum Dissimilarity Sampling (MDS). Dissimilarity refers to the Euclidean distance between feature vectors of different samples. MDS iteratively selects samples with the maximum dissimilarity to the already chosen subset, aiming to maximize the dissimilarity between samples within the subset and comprehensively cover the class distribution. Additionally, the model tends to favor new classes in decision-making due to the imbalance in class quantities. This paper recommends using Logit Balanced Cross-Entropy Loss (Laloss) to mitigate this issue. Experimental results on the CWRU dataset demonstrate that MDS can comprehensively cover the class distribution and better retain the model’s ability to recognize previous tasks. Compared to classical algorithms, our method exhibits superior performance on both the CWRU and MFPT datasets.