A balanced and weighted alignment network for partial transfer fault diagnosis

Domain adaptation techniques have attracted great attention in mechanical fault diagnosis. However, most existing methods work under the assumption that the source and target domains share the identical label space. Such methods are unable to handle a practical issue where the target label space is a subset of the source label space. To tackle this challenge, a balanced and weighted alignment network is proposed for partial transfer fault diagnosis. The proposed method views this issue from a new angle by augmenting the target domain to make the classes of two domains balanced and shortening class-center distances to reduce conditional distribution shifts. Meanwhile, a weighted adversarial alignment is developed to filter out the irrelative source samples and minimize marginal distribution discrepancy. As such, negative transfer can be avoided, and positive transfer can be enhanced. Comprehensive experiments on two test rigs demonstrate that the proposed method achieves promising performance and outperforms state-of-the-art partial transfer methods.