Discriminative and Graph Knowledge Embedding Based on Incremental Confidence Labeling for Cross-Domain Fault Diagnosis

Domain adaptation (DA) has the promising potential to mitigate the distribution gap between source and target data in fault diagnosis. However, many DA methods only attempt to align the mean of the feature distribution, neglecting the covariance information about the distribution. Furthermore, most existing DA methods ignore the topological structures among different domains, potentially leading to negative transfer. To tackle these issues, a novel DA mechanism, named discriminative and graph knowledge embedding based on incremental confidence labeling (DGKE-ICL) is proposed. First, we develop a multiscale attention residual network that extracts important channel and cross-spatial semantic features from time-frequency images. Second, a new maximum mean and covariance discrepancy (MMCD) metric is designed to align marginal and class-conditional distributions of two domains. To accurately match class-conditional distributions, an incremental confidence labeling (ICL) strategy is employed to incrementally introduce confidence pseudo-labels for target samples in each iteration. Simultaneously, we incorporate a repulsive power term in the class conditional distribution to increase the interclass distance of features. Finally, the similarity graph-based manifold regularization (MR) is incorporated into the structural risk minimization (SRM) framework to encourage consistency between the manifold structure of data and predicted labels, thereby enhancing the generalization performance of the domain classifier. DGKE-ICL is validated on 22 transfer tasks using the vibration data from five bearing datasets. Experimental results demonstrate that DGKE-ICL outperforms other typical DA methods, achieving an average accuracy of 98.56% on cross-operating condition tasks and 98.23% on cross-machine tasks.