Open Access Open Access  Restricted Access Subscription or Fee Access

Fault Diagnosis Method Based on Transfer Learning and Deep ConvNet



With the development of Internet of Things (IoT) and Cyber-Physical System (CPS), a massive amount of real-time condition monitoring data of equipment are collected. Thus data-driven fault diagnosis methods have become more and more important in modern industry. However, traditional methods are based on handcrafted features which rely on the expert knowledge and experience. In addition, the distribution of training dataset is different from the distribution of testing dataset in many real-world fault diagnosis problems, which has not been taken into consideration in most existing data-driven methods. In this paper, we propose a novel fault diagnosis approach based on transfer learning (TL) and deep convolutional network (ConvNet). A convolutional neural network (CNN) structure is established to learn the representative features automatically from the converted images with abundant fault information and classify fault types. Furthermore, the performance for fault diagnosis on the target domain is improved by reusing parameters trained on the different but related distribution dataset on the source domain. The proposed approach is validated on the famous bearing dataset, constructed by the Case Western Reserve University (CWRU). The experimental results indicate that the proposed approach confirms faster convergence rate and higher accuracy compared to fault diagnosis based on deep traditional learning methods.

Full Text:



  • There are currently no refbacks.