Compressive sensing (CS) is a promising technology for realizingenergy-efficient wireless sensors for long-term health monitoring. However,conventional model-driven CS frameworks suffer from limited compression ratioand reconstruction quality when dealing with physiological signals due toinaccurate models and the overlook of individual variability. In this paper, wepropose a data-driven CS framework that can learn signal characteristics andpersonalized features from any individual recording of physiologic signals toenhance CS performance with a minimized number of measurements. Suchimprovements are accomplished by a co-training approach that optimizes thesensing matrix and the dictionary towards improved restricted isometry propertyand signal sparsity, respectively. Experimental results upon ECG signals showthat the proposed method, at a compression ratio of 10x, successfully reducesthe isometry constant of the trained sensing matrices by 86% against randommatrices and improves the overall reconstructed signal-to-noise ratio by 15dBover conventional model-driven approaches.
展开▼