Accurate noise suppression is vital for precise micro-thrust calibration and measurement. Conventional methods often fail to recover both nonlinear transitions and smooth trends within the signal effectively. In this research, we present an innovative U-Net-based contrastive blind denoising method that operates without needing a reference clean signal. Our method introduces contrastive representation learning combined with self-supervised blind denoising, forming a multi-task joint learning framework. This joint learning framework compels the network to extract robust content-invariant and disentangled features, i.e., clean signal features). Experiments validate the proposed method's superior performance in recovering both nonlinear transitions and smooth trends in the signal, outperforming traditional methods.

micro-thrust measurement,blind denoising,contrastive learning,U-Net