Optimal Compressed Sensing Reconstruction for Vibration Monitor Data Using Deep Learning

Performance degradation and failure of rotating components such as flywheel, momentum wheel, and solar panel drive mechanism have become the key bottlenecks restricting satellite life and mission efficiency. In the process of condition monitoring of rotating components, due to the limitation of communication bandwidth, the complete data cannot be transmitted to the ground-receiving station, but compressed sensing technology is needed to recover the complete signal from the highly undersampled data. As the incoherence loss continuously decreases, the proposed method expands the range of sparsity levels for which the original signal can be accurately reconstructed. First, a new objective function is introduced through theoretical analysis. Second, aiming at the newly proposed objective function, a joint optimization solution of the 1-D alternating direction method of multipliers (ADMM) expansion algorithm and the neural network backpropagation algorithm is designed under the framework of deep learning. Finally, as the incoherence loss continuously decreases, the proposed method expands the range of sparsity levels for which the original signal can be accurately reconstructed. While the reconstruction effect is continuously optimized, the sparsity requirement is gradually reduced. Numerical experiments executed on actual control moment gyroscope (CMG) vibration data and bearing data with weak fault validate the effectiveness of the new method. The experimental results show that compared with the existing reconstruction methods based on a fixed basis, dictionary learning basis matrix, and network learning orthogonal basis matrix, our proposed method has better reconstruction performance at low sampling rates. The proposed method is real time due to its reconstruction time of only 0.1 s, thus meeting the needs of practical production applications.