面向可穿戴设备的压缩感知关键技术研究

Due to the low sampling rate and a number of strong noise and interference in Wearable Devices, the collected raw physiological signals are non-sparse in time domain or transform domain. It is difficult to recover the physiological signals with high reconstruction quality. Most Compressed Sensing (CS) algorithms fail in directly reconstructing such non-sparse signals. However only several CS algorithms, which can recover non-sparse signals, can’t be efficiently satisfied for high reconstruction quality and speed need of the physiological signals in Wearable Devices. Therefore, it is a challenge for research on research key technologies of CS based on the non-sparse signals in Wearable Device based Health Telemonitoring Systems. The main research works are as follows:(1) the dictionary learning algorithm, which is suitable for the non-sparse physiological signals, is designed so as to obtain the best sparse representation of the signals;(2) based on the best sparse dictionary, an effective sensing matrix is constructed to ensure that the important information is not missed in the compressed signals; (3) a novel Variational Inference (VI) Bayesian learning based CS reconstruction algorithm with better recovery quality and high iterative convergence speed, is proposed for an application need of the Wearable Device based Health Telemonitoring Systems. Because a VI method of Machine Learning field is used into the framework of block Sparse Bayesian Learning.

可穿戴设备由于较低采样率和噪声干扰，导致采样的生理信号的稀疏性较差。大多数压缩感知算法对不足够稀疏的信号重构效果不理想，而目前能对不足够稀疏信号重构的仅有的少数压缩感知算法，不能有效满足可穿戴远程健康监护系统同时对重构精度和速度的需求。因此，面向可穿戴设备的压缩感知关键技术研究具有一定的挑战性。本项目的主要研究内容：1.针对不足够稀疏的生理信号，设计高效率的字典学习算法，以获得更为有效的稀疏字典，解决生理信号的最佳稀疏表示；2.设计与最佳稀疏字典满足非相干条件的观测矩阵，确保可穿戴设备以高压缩率采样后的数据不丢失微弱的重要生理信息，为精确重构信号提供保障；3.将机器学习领域的变分推断方法与稀疏贝叶斯学习理论相结合，提出基于变分贝叶斯推断的重构恢复算法，以满足可穿戴健康监护系统同时对快速和高精度重构信号的需求。

本项目按预定计划，完成了文献调研，理论研究，算法设计，数据处理和实验验证等工作。首先，我们针对不足够稀疏信号的最佳稀疏表示问题，充分挖掘不足够稀疏生理信号隐含的相关性的结构特征，提出了一种最小二乘字典学习算法，解决了传统字典学习算法所设计得到的字典不能获得非稀疏信号的最佳稀疏表示导致重构算法精度低的问题。其次，针对不足够稀疏信号在不同压缩率下信号的稀疏性不同，导致压缩感知重构算法精度和速度差的问题，课题组提出了一种基于交替方向乘子法和块稀疏贝叶斯学习的压缩感知重构方法。最后，我们利用块稀疏模型、稀疏表示、贝叶斯重构算法实现了可穿戴远程健康监护原型系统，并在公开的标准数据集和实际应用场景中验证了算法的有效性和先进性。本项目圆满完成了预定的研究目标和成果目标，代表成果包括了5篇SCI检索论文，2篇EI检索论文，同时在成果应用转化和交叉学科的研究中有所突破，包括2个国家发明专利申请，1个国家发明专利授权。