低场核磁共振（LF-NMR）弛豫时间谱检测技术在石油、地质勘探等领域应用广泛．然而，常规的NMR弛豫时间谱反演算法要求采集到的NMR时域信号具有较高的信噪比，这在很多实际应用场景（如现场岩心分析）下难以实现．因此，亟需发展能够对低信噪比信号实现准确稳定反演的新算法．针对这一问题，本文提出了基于深度学习的结合U-Net降噪和多尺度交叉注意力的反演网络（UDMCA），在该网络中创新性地融入了多尺度交叉注意力机制，针对NMR时域信号的衰减特性进行对应尺度的弛豫特征提取并交叉融合全局特征信息，以提升网络对低信噪比多组分信号的反演准确性．同时，在反演模块前引入U-Net去噪模块，进一步提升网络的抗噪鲁棒性.实验结果表明，该算法对信噪比为10~50的低信噪比信号能够实现准确且稳定的反演，并在岩心NMR信号反演中表现出良好的适用性，为低场低信噪比NMR数据的处理与反演提供了一种有效新方案．

Low-field nuclear magnetic resonance (LF-NMR) relaxation time spectrum analysis has been widely applied in petroleum engineering and geological exploration. However, conventional NMR relaxation time spectrum inversion algorithms require time-domain signals with high signal-to-noise ratios (SNR), which are often difficult to obtain in practical scenarios such as on-site core analysis. To achieve accurate and stable inversion of low-SNR NMR signals, this paper proposes a deep learning-based inversion network named UDMCA (U-Net Denoising and Multi-scale Cross Attention Inversion Network), which incorporates a multi-scale cross-attention mechanism to extract and fuse relaxation features at multiple scales according to the decay characteristics of NMR signals, thereby improving inversion accuracy for low-SNR multi-component signals. In addition, a U-Net denoising module is introduced to further enhance noise robustness. Experimental results show that the proposed method achieves accurate and stable inversion for low-SNR signals and demonstrates good applicability in core NMR signal inversion, providing an effective solution for low-SNR LF-NMR data processing and inversion.