Multi-view and Multi-level network for fault diagnosis accommodating feature transferability

Abstract

Various deep transfer learning solutions have been developed for machine fault diagnosis. The existing solutions mainly focus on domain adaptation by minimizing the data distribution discrepancy with certain metric, which emphasize the common features embedded in the data cross domains and neglect the unique features toward health condition classification in one specific domain. In these solutions, all the data for training have been forced to align in a common feature space and all the features for domain adaptation have been treated equally. However, there might exist domain specific features which are not appropriate for transfer but may contain essential information for classification in specific domain. In addition, due to the difficulty of collecting machine fault data, the number of machine fault samples is usually quite small or even zero. The traditional deep network structures and the training strategy are not the optimal choice in this occasion. To address these problems, a novel multi-view and multi-level network (MMNet) for fault diagnosis is developed. In MMNet, two network channels have been respectively constructed for cross domain common feature and domain specific feature learning to provide multi-view features. This architecture could implicitly differentiate the common features cross domains and the specific features only in one domain. In the channel of domain specific feature, a domain classifier and fault classifier are combined to learn the domain specific features. Multiple kernel maximum mean discrepancy (MK-MMD) is imposed on multiple layers of the common feature channel to implement domain adaptation and extract cross domain common features. The domain classification and fault classification together form a multi-level classification scheme. A classic few shot learning architecture with two modules respectively for feature extraction and relation computation is adopted as the backbone network. The relation score based classification mechanism enables zero shot fault classification in the target domain. Episode based few shot training strategy is employed to enhance the performance of MMNet with few labeled training data. Extensive experiments have demonstrated the state-of-the-art performance of MMNet on the involved transfer tasks.