Abstract:Current unsupervised domain adaptation methods for intelligent fault diagnosis focus on learning domain-invariant representations under covariate shift assumption. However, the covariate shift assumption is usually unsatisfied when each fault class in different domains consists of multiple modes with skewed proportions, which is common in industrial scenarios. Imbalanced data from multiple modes cause the presence of within-domain class imbalance and between-domain label distributional shift. This paper introduces a novel subpopulation shift that further considers the domain shift from a subpopulation perspective, i.e., that covariate shift and label distributional shift across domains are caused by shifts in the multiple modes. To address this issue, a label propagation-based unsupervised domain adaptation is proposed based on a realistic expansion assumption. We apply the theoretical analysis of the proposed method with a bi-level optimization strategy adapted from meta-learning. Using joint optimization of a teacher model and a student model, the label propagation-based model-agnostic meta-learning (LPMAML) not only propagates supervision information from the source to the target but also adjusts the teacher’s strategy throughout the student’s learning process. To alleviate the noise caused by label distributional shift, we integrate a sampling-based alignment method that aligns the empirical label distributions across the two domains into LPMAML. Experimental results on three bearing datasets show that the proposed method has impressive generalization ability under covariate, label distributional, and subpopulation shifts. The proposed method offers consistent improvements to unsupervised domain adaptation (UDA) methods. Compared with the vanilla UDA methods, the average diagnosis accuracies of the proposed method on the label distributional shift benchmark and subpopulation shift benchmark are improved by 8.21% and 7.63%, respectively.

Label Propagation with Multi-Stage Inference for Visual Domain Adaptation.

ADeLA: Automatic Dense Labeling with Attention for Viewpoint Shift in Semantic Segmentation

Multiscale Attention-Based Subdomain Dynamic Adaptation for Cross-Domain Scene Classification

Label transfer and propagation for domain adaptation

ADeLA: Automatic Dense Labeling with Attention for Viewpoint Adaptation in Semantic Segmentation

Multi-Source Domain Adaptation and Semi-Supervised Domain Adaptation with Focus on Visual Domain Adaptation Challenge 2019

Label propagation-based unsupervised domain adaptation for intelligent fault diagnosis

Unsupervised Domain Adaptation for Referring Semantic Segmentation

MultiDAN: Unsupervised, Multistage, Multisource and Multitarget Domain Adaptation for Semantic Segmentation of Remote Sensing Images

Gotta Adapt 'Em All: Joint Pixel and Feature-Level Domain Adaptation for Recognition in the Wild

Joint Feature and Labeling Function Adaptation for Unsupervised Domain Adaptation

Domain Adaptation for Remote Sensing Image Semantic Segmentation: An Integrated Approach of Contrastive Learning and Adversarial Learning

Multi-source unsupervised domain adaptation for object detection

Align and Adapt: A Two-stage Adaptation Framework for Unsupervised Domain Adaptation

Soft Labels Transfer with Discriminative Representations Learning for Unsupervised Domain Adaptation.

An Unsupervised Domain Adaptation Method for Multi-Modal Remote Sensing Image Classification

Heterogeneous Graph Attention Network for Unsupervised Multiple-Target Domain Adaptation

An Adversarial Domain Adaptation Network for Cross-Domain Fine-Grained Recognition

Multi-source domain adaptation handling inaccurate label spaces

Multi-view Based Multi-Label Propagation for Image Annotation

Domain-Agnostic Priors for Semantic Segmentation Under Unsupervised Domain Adaptation and Domain Generalization