Temporal Decomposition and Attention-Based Deep Learning for Magnetocardiography Signal Denoising

This article proposes a temporal decomposition and attention-based denoising network (TADNet) for magnetocardiography (MCG) signal denoising. The raw time-series data are decomposed into trend and seasonal components. The trend component is modeled using a multilayer perceptron (MLP) to capture the trend of the data. For seasonal components, a multilayer convolutional neural network (CNN) is employed to perform segmentation and extract local temporal features. These fine-grained features are processed by a Transformer-based self-attention mechanism to capture long-range temporal dependencies. Finally, deconvolution operations are applied to reconstruct the signal, achieving effective noise reduction for MCG data. This article conducts in-depth studies on the effects of time, wavelet, and Fourier-style attention on model performance, enhancing the interpretability of the model. Extensive experiments on clinical MCG data under various noise conditions demonstrate that the TADNet outperforms traditional algorithms, proving its effectiveness and robustness. The research results indicate that the TADNet provides a highly efficient and accurate solution for MCG signal denoising.