A digital twin-enhanced semi-supervised framework for motor fault diagnosis based on phase-contrastive current dot pattern

Motor plays a core role in most industrial equipment. Accurate fault diagnosis of motor is a critical task and intelligent data-driven methods have gained significant advances. However, to obtain sufficient labeled data to train the models is expensive and laborious in industrial applications, and how to utilize three-phase current signals efficiently is a challenging task. To deal with these problems, a digital twin-enhanced semi-supervised framework is proposed for label-scarce motor fault diagnosis. First, a precise motor digital twin model is established based on multi-physics simulation and knowledge transfer is performed from the virtual space to the physical space. Second, a novel phase-contrastive current dot pattern (PCCDP) representation is proposed to transform three-phase motor stator current to a gray-scale image with an ordered arrangement and then characteristics of three phases can be contrasted in tight regions for efficient processing. Third, inter-space sample generation is proposed for continuous feature manifold learning to tackle discrepancy between spaces. Finally, intra-space sample generation and a clustering-based metric learning are also introduced to improve semi-supervised fault diagnosis performance. An induction motor fault experiment is conducted and a digital twin model is built correspondingly. Experiments verify the effectiveness and superiority of the proposed framework.