Knowledge and data dual-driven transfer network for industrial robot fault diagnosis

Abstract

Various deep transfer learning solutions have been developed for machine fault diagnosis, which are purely data driven. Plenty of prior knowledge on the fault of different machinery parts has been summarized in previous research. The value of these prior knowledge has not been ever deeply explored in the existing deep transfer solutions, which might greatly facilitate the training progress and improve the diagnosis performance of network models. However, how to represent and incorporate prior knowledge into deep models remains a challenge. To address this problem, a knowledge and data dual-driven transfer network for fault diagnosis is developed in this paper. The prior frequency knowledge of fault signal is extracted by envelope analysis and pass band selection. Based on which, a band-pass filter based embedding method is proposed to incorporate the prior knowledge by imprinting the convolutional kernels with the designed filters. A dual channel weight shared deep adaptation network is constructed to perform the prior knowledge embedding and knowledge transfer across domains. Multi-kernel maximum mean discrepancy (MK-MMD) is adopted for domain adaption. A symmetry constraint is used to reserve the linear phase property of the band-pass kernels. The constructed network combines data driven and knowledge driven mechanism which is termed as Knowledge and Data Dual-driven Transfer Network (KDDT Network). Extensive experiments have been performed on an industrial robot rotate vector (RV) reducer dataset collected in our laboratory and some bearing benchmarks. Comparisons results with the state-of-the-art methods have shown the superiority of the proposed method.