A High-Generalization Variational Denoising Autoencoder for Micronewton Thrust Signal Noise Removal and Step Reconstruction

Removing noise and recovering the micronewton thrust signal are of great significance in high-precision static thrust measurements. Typically, the micronewton thrust signal is in the shape of a staircase signal. Existing methods have limitations in decoupling sharp step edges and flat regions from noisy signals while ensuring the accuracy of step amplitude reconstruction. In this study, we have developed a novel generative denoising method, named variational denoising autoencoder (VDAE), based on a unique deep-learning-based Bayesian framework. Specifically, the encoder-parameterized approximate posterior maps the distribution of essential features (i.e., thrust step amplitudes) of limited training samples to a latent space with a Gaussian distribution. This distribution transformation gives the latent space the ability to describe complete continuous step amplitudes. VDAE inherits the excellent generalization ability of the generative model and greatly improves the amplitude accuracy of the denoised signals. In addition, considering the different scale features in the clean staircase signal, a trend feature disentangler (TFD) is introduced in the encoder. The TFD adaptively extracts ultrahigh-frequency sharp step edge features and ultralow-frequency flat region features. Furthermore, to address the issue of recovering sharp step edges, total variation (TV) sparse representation is introduced into the loss function, guiding the decoder to reconstruct the thrust step. Extensive simulations and experiments were carried out to demonstrate the effectiveness and superiority of the proposed method in micronewton thrust step reconstruction and measurement noise removal.