Multi-block domain adaptation with central moment discrepancy for fault diagnosis

Despite the recent advances of deep neural network-based domain adaptation for fault diagnosis under different working conditions, the degradation of feature transferability in top layers and incomplete elimination of mismatch between different working conditions hinder its further development. Accordingly, this study develops a new deep transfer method, i.e., multi-block domain adaptation, which embeds the central moment discrepancy adaptation at multiple dense blocks to work collaboratively with densely connected convolutional networks to improve the transferability of learned features. First, a densely connected convolutional network is constructed as the main architecture. Second, the central moment discrepancy is adopted for minimizing the domains discrepancy at each dense block inside the network. Finally, the training datasets are fed into the model for training, then the learned model is used to perform the fault diagnosis. The experiment on the planetary gearbox is conducted to validate the effectiveness of the developed method.