Low-Complexity On-Demand Reconstruction for Compressively Sensed Problematic Signals

Abstract

Compressed Sensing (CS) is a revolutionary technology for realizing low-power sensor nodes through sub-Nyquist sampling, and the CS reconstruction engines have been widely studied to fulfill the energy efficiency for real-time processing. However, in most cases, we only want to analyze the problematic signals which account for a very low percentage. Therefore, large efforts will be wasted if we recover uninterested signals. On the other hand, in order to identify the high-risk signals, additional hardware and computation overhead are required for classification other than CS reconstruction. In this paper, to achieve low-complexity on-demand CS reconstruction, we propose a two-stage classification-aided reconstruction (TS-CAR) framework. The compressed signals can be classified with a sparse coding based classifier, which provides the hardware sharing potential with reconstruction. Furthermore, to accelerate the reconstruction speed, a cross-domain sparse transform is applied from classification to reconstruction. TS-CAR is implemented in electrocardiography based atrial fibrillation (AF) detection. The average computational cost of TS-CAR is 2.25× fewer compared to traditional frameworks when AF percentage is among 10% to 50%. Finally, we implement TS-CAR in TSMC 40 nm technology. The post-layout results show that the proposed intelligent CS reconstruction engine can provide a competitive area- and energy-efficiency compared to state-of-the-art CS and machine learning engines.