基于双分支时频融合的心冲击信号心率估计网络

Heart Rate Estimation Network for Ballistocardiogram Signals Based on Dual-Branch Time-Frequency Fusion

摘要: 文章提出了一种基于双分支时频融合架构的模型，用于从心冲击图(BCG)信号中实现高精度的心率估计。该模型以原始BCG信号为输入，通过并行的时间分支与频率分支分别提取时域波形特征与频域周期特征，其中频率分支采用快速傅里叶变换(FFT)获取频谱信息，并引入注意力机制自适应加权关键频段，使模型能够聚焦于心率基频及其谐波成分，有效提升了估计的准确性与鲁棒性。实验结果显示，该模型在平均绝对误差(3.10 BPM)、均方误差(11.23)及决定系数(0.72)等指标上均表现优异，显著优于传统单分支网络与时域回归方法。与近年来相关研究相比，本模型在BCG心率估计任务中展现了显著优势，突显了其在无接触心率监测场景中的实用价值。

Abstract: This paper proposes a model based on a dual-branch time-frequency fusion architecture for high-precision heart rate estimation from ballistocardiogram (BCG) signals. The model takes raw BCG signals as input and employs parallel temporal and frequency branches to extract time-domain waveform features and frequency-domain periodic features, respectively. In the frequency branch, the fast Fourier transform (FFT) is utilized to obtain spectral information, and an attention mechanism is introduced to adaptively weigh critical frequency bands, enabling the model to focus on the fundamental frequency and its harmonics of heart rate, thereby effectively improving estimation accuracy and robustness. Experimental results show that the model achieves excellent performance in terms of mean absolute error (3.10 BPM), mean squared error (11.23), and coefficient of determination (0.72), significantly outperforming traditional single-branch networks and time-domain regression methods. Compared with recent related studies, this model demonstrates significant advantages in BCG-based heart rate estimation tasks, highlighting its practical potential in non-contact heart rate monitoring scenarios.