Attribute fusion transfer for zero-shot fault diagnosis

Abstract

In most industrial processes, it is impossible to collect samples of all fault modes, which greatly limits fault diagnosis applications in industrial fields. The fault diagnosis for the inaccessible fault modes having no labeled samples, termed zero-shot fault diagnosis, is quite challenging. This paper proposed an Attribute Fusion Transfer (AFT) for zero-shot fault diagnosis. First, we proposed a multi-Gaussian assumption for the posterior probability calculation in fault attribute transfer process and theoretically and experimentally proved its feasibility and effectiveness. Then, considering the ability difference of various attribute transfer methods to predict the same attribute, we proposed an attribute fusion network, which fuses the Bayesian fault attribute transfer and the attribute transfer under the multi-Gaussian assumption to attain better attribute transfer ability. Finally, AFT transfers fault attribute knowledge from accessible fault modes to inaccessible fault modes, thereby getting the ability to identify inaccessible fault modes. A series of ablation experiments and comparative experiments on the Tennessee–Eastman process were conducted to demonstrate the rationality and effectiveness of AFT fusion structure and the superiority of AFT. Furthermore, we conducted attribute-level and data-level interpretability analyses to reveal why our proposed AFT can obtain excellent zero-shot fault diagnosis ability. Codes are available at https://github.com/foryichuanqi/ADVEI-Paper-2023.11-Zero-shot-fault-diagnosis-by-attribute-fusion-transferGithub.