Remaining useful life estimation of bearings under different working conditions via Wasserstein distance-based weighted domain adaptation

Various transfer learning methods have been applied in the remaining useful life estimation of bearings to reduce the data distribution discrepancy under different working conditions. However, the transferability of the sample (i.e., the sample quality) is always ignored. Low-quality samples caused by noise and outliers inevitably exist in the industrial data, which may negatively affect feature extraction and alignment. This article proposes a Wasserstein distance-based weighted domain adversarial neural network to utilize sample quality which is measured by the domain classifier. The feature extractor tends to learn the representations from the samples with cross-domain similarity. Feature alignment is fine-tuned according to the sample weights. The effectiveness of the proposed method is validated using IEEE PHM Challenge 2012 dataset. The comparison results prove the features extracted from the proposed approach are more domain-invariant.