Numerical and experimental studies on unsupervised deep Lagrangian learning based rotor balancing method

Rotor balancing is essential to rotor dynamic analysis. To make the balancing process convenient and costless, a balancing method using unsupervised deep Lagrangian network without weight trail is proposed. In the proposed network, a Lagrangian layer is applied to the network to introduce the physical prior knowledge. Compared to traditional balancing method, trail weight process is not necessary. Meanwhile, parameter sharing mechanics in baseline design or Lagrangian layer are applied to identify the unbalanced force without labeled data. Both numerical case study and corresponding experiment are conducted to validate the method. Both experimental and numerical results find that the proposed rotor balancing approach gives reasonable and comparative results with the considerations of both cost and accuracy. Compared with the baseline, to which no physical prior is applied, the balancing method with Lagrangian mechanism involved could achieve better performance. This proposed rotor dynamic balancing method gives out an alternative approach of rotor balancing methods.