Deep Learning for Compressed Sensing-Based Blade Vibration Reconstruction From Sub-Sampled Tip-Timing Signals

Abstract

Blade tip-timing (BTT) signals are always seriously under sampled, so reconstruction is much needed for vibration analysis. Blade vibration responses are sparse in order domain and classical compressed sensing (CS) algorithms are difficult to reconstruct vibration orders due to lack of prior sparse information under variable speeds. In order to address this issue, this paper introduces deep learning (DL) into BTT vibration reconstruction and proposes an end-to-end deep compressed sensing (DCS) method. Firstly, a multi-coset BTT measurement model is built under variable speeds and the DCS model is derived in order domain, where a specific convolutional neural network (CNN) is designed. Next, a Simulink model is built to generate training and testing samples. The simulation results show that the convolution layer with the rectified linear unit (ReLU) layer placed after the batch normalization (BN) layer can improve the reconstruction performance and the proposed method has better reconstruction accuracy and efficiency than classical CS algorithms. Finally, experiments are done and the results demonstrate that blade vibration orders can be recovered accurately by the proposed method, which will provide a novel way of BTT signal analysis.