Deep neural networks present very competitive results in mechanical fault diagnosis. However, training deep models require high computing power while the performances of deep models often suffer from the lack of sufficient training data. In this paper, a transferable convolutional neural network (TCNN) is proposed to improve the learning of target tasks. Firstly, one-dimension CNN is constructed and pre-trained based on large source task datasets. Then a transfer learning strategy is adopted to build a deep model on target tasks by reusing the pre-trained network. Thus, the proposed method not only utilizes the learning power of deep network but also leverages the prior knowledge from the source task. Four case studies are performed, and the effects of transfer layers and training sample size on classification effectiveness are investigated. Results show that the proposed method exhibits better performance compared with other algorithms.

Recently, deep learning-based intelligent fault diagnosis techniques have obtained good classification performance with amount of supervised training data. However, domain shift problem between training and testing data usually occurs due to variation of operating conditions and interferences of environment noise. Transfer learning provides a promising tool for handling the cross-domain diagnosis problems by leveraging knowledge from source domain to help learning in target domain. Most existing studies attempt to learn both domain features in a common feature space to reduce domain shift, which are not optimal on specific discriminative tasks and can be limited to small shifts. This paper proposes a novel Domain Adversarial Transfer Network (DATN), exploiting task-specific feature learning networks and domain adversarial training techniques for handling large distribution discrepancy across domains, Firstly, two asymmetric encoder networks integrating deep convolutional neural networks are designed for learning hierarchical representations from source domain and target domain, respectively. Then, the network weighs learned in source tasks are transferred to improve training on target tasks. Finally, domain adversarial training with inverted label loss is introduced to minimize the difference between source and target distributions. To validate the effectiveness and superiority of the proposed method in the presence of large domain shifts, two fault datasets from different test rigs are investigated, and different fault severities, compound faults and data contaminated by noise are considered. Experimental results demonstrate that the proposed method achieves the average accuracy of 96.45% on the bearing dataset and 98.92% on the gearbox dataset, which outperforms other algorithms.