VampNet-based BTT vibration reconstruction

Blade tip-timing (BTT) vibration measurement is seriously under-sampled due to few BTT probes, so it often needs to reconstruct true blade vibration characteristics. In practice, the characteristic of sparsity makes compressed sensing be useful for under-sampled BTT vibration reconstruction. Most existing algorithms still depend on prior sparse information. In recent years, deep learning methods have been investigated to deal with compressed sensing (CS) without prior information. However, traditional deep neural network architectures need many layers for accuracy, leading to more complexity. In order to cope with it, iterative reconstruction algorithms have been recently studied to modify traditional neural-network architectures. Thus, this paper proposes a VampNet-based BTT vibration reconstruction method by using the approximate message passing algorithm. Firstly, the angular-domain CS model of BTT vibration measurement is built. Based on it, the VampNet model of BTT vibration reconstruction is derived. Next, Matlab/Simulink simulations are done to generate BTT samples under variable speeds and then validate the reconstruction accuracy. The simulation results demonstrate that the proposed method can carry out BTT vibration reconstruction faster than existing deep learning methods.