Remaining Useful Life Estimation Under Variable Failure Behaviors via Transferable Metric Learning

A reasonable estimation of remaining useful life (RUL) is able to effectively decrease the costs and prevent mechanical accidents. However, the failure modes of different bearings may bring about domain shift, which results from different failure behaviors. Traditional approaches for solving the domain shift problem attempt to infer domain-invariant representations but fail to consider different domain relations of unknown samples, resulting in limited performance. To solve the above challenging issues, a cross-domain adaptation method based on deep transferable metric learning is proposed for bearing RULs prediction. Temporal convolutional network adaptively extracts hidden features. To minimize the domain discrepancy, a new cross-domain adaptation architecture is developed to extract the domain-invariant representations in which the contrastive loss of metric learning is used to enhance the transformation invariance of the maximum mean discrepancy to make a contribution to the common scenarios in RUL estimation. Extensive experiments have shown that our approach is improved by more than 17% and 21% in two different evaluation metrics, which is obviously superior to the other methods compared.