부산대학교 도서관

As advanced measurement technologies and new generation industry intelligence develops rapidly, the methods based on machine learning, especially on deep learning, have made phenomenal achievements and applications. However, few researches focus on the task of decoupling compound faults under different speeds and loads conditions. To address such a problem, a novel deep transfer learning method called transferable capsule network (TCN) is proposed for decoupling compound fault of machinery under various working conditions. First, a capsule network (CN), consisted of 3 parts: feature learning model, capsule layers and decoupling classifier, are constructed for learning the transferable representations from raw signals and make further diagnosis for the single and compound faults. Second, the maximum mean discrepancy (MMD) are introduced into the final layer of feature learning model and capsule layers during the training process, which can adapt the distribution of the learned representations from the source domain samples with labels to the target domain samples without labels. Finally, after the above 2 stages finished, the TCN model can be used to diagnosis and decoupling the compound fault for machinery operating at different working conditions. The experimental results based on an automobile transmission (AT) dataset demonstrate that the TCN achieves a superior diagnosis performance with the average accuracy of 99.37%, which outperforms that of the deep decoupling convolutional neural network (DDCNN).