Data-Driven Method for Real-Time Reconstruction of the Structural Displacement Field

Abstract

Accurate real-time displacement field reconstruction based on limited measurement points is crucial for spacecraft on-orbit monitoring. This study proposes a data-driven displacement field reconstruction method called stacked convolutional autoencoder with denoising autoencoder and filter. Precise reconstruction of the structural displacement from a small number of local strains was made possible by the two primary components of the method: low-resolution displacement field reconstruction and result optimization. Given the significant imbalance between the limited strain information input and the structural displacement field output, a deep learning model with multiple deconvolution layers was built in the low-resolution displacement field reconstruction part using the layer-wise training property of a stacked autoencoder and the sparse mapping property of a convolutional neural network. The result optimization part utilized a denoising autoencoder and a linear density filter to effectively alleviate the checkerboard phenomenon and displacement field discontinuity caused by the deconvolution operation. The results of the case study indicate that the proposed method can accurately reconstruct the structural displacement field of both simple regular geometric structures and irregular geometric structures with complex boundaries without prior information. Additionally, the method exhibits excellent robustness to unavoidable measurement noise, providing a new implementation approach for real-time monitoring of spacecraft.