loading page

Multi-classification generative adversarial network for streaming data with emerging new classes: method and its application to condition monitoring
  • Yu Wang ,
  • Alexey Vinogradov
Yu Wang
Norwegian University of Science and Technology

Corresponding Author:[email protected]

Author Profile
Alexey Vinogradov
Author Profile


The paper extends the application of Generative Adversarial Network (GAN) to streaming data with the emerging new classes (SENC) problem. Four challenges of the SENC problem are proposed in this work. To deal with these problems, a novel multi-classification GAN (MC-GAN) is developed and investigated. The proposed MC-GAN employs a Generator to learn the implicit information embedded in data and adopts a re-designed Discriminator for multi-classification and novelty detection. The training of the whole network is realised by a modified loss function. In contrast to previous studies, the MC-GAN integrates the procedure of novelty detection into multi-classification. Moreover, the realisation of novelty detection does not require any predefined threshold. Case studies on rolling element bearings show that the proposed MC-GAN can maintain high classification accuracy for known classes as well as emerging new classes.