| Popis: |
Traditional intelligent fault diagnosis works well when the labeled training data (source domain) and unlabeled testing data (target domain) are drawn from the same distribution. However, in many real-world applications, the working conditions can vary between training and testing time. In this paper, we address the issues of intelligent fault diagnosis when the data at training and testing time do not come from the same distribution as a domain adaptation problem using domain adaptive convolutional neural networks (DACNN). Our proposed DACNN consists of three parts: a source feature extractor, a target feature extractor, and a label classifier. We adopt a two-stage training process to obtain strong fault-discriminative and domain-invariant capacity. First, we obtain fault-discriminative features by pre-training the source feature extractor with labeled source training examples to minimize the label classifier error. Then, in the domain adaptive fine-tuning stage, we train the target feature extractor to minimize the squared maximum mean discrepancy between the output of the source and target feature extractor, such that the instances sampled from the source and target domains have similar distributions after the mapping. Furthermore, to enable training efficiency in domain adaptation, the layers between the source and target feature extractors in our DACNN are partially untied during the training stage. Experiments on the bearing and gearbox fault data showed that DACNN can achieve high fault diagnosis precision and recall under different working conditions, outperforming other intelligent fault diagnosis methods. We also demonstrate the ability to visualize the learned features and the networks to better understand the reasons behind the remarkable performance of our proposed model. |
