Sparsity Prediction-based Adaptive Deep Compressed Sensing for MEC-enabled IoT Systems

Deep compressed sensing has been proposed to reduce the amount of transmitted data for Internet of Things (IoT) systems. Although the data compression and reconstruction have been deeply studied, the reconstruction of IoT data with dynamic characteristics of sparsity is still a challenge. To ensure the accuracy and timeliness of data reconstruction in IoT systems, we propose a novel sparsity prediction-based compressed sensing framework assisted by multi-access edge computing for both adaptive measurement (compression) and accurate reconstruction. Specifically, adaptive measurement is to dynamically adjust the compression ratio in time according to the feedback of predicted sparsity from the edge server. For accurate reconstruction, we design a space-time multi-scale deep compressed sensing network (STDCS-NET) that is composed of a sampling network and a reconstruction network, which are jointly optimized. The reconstruction network integrates multi-scale convolutional and long short-term memory networks, which can fully mine the multi-scale characteristics of low-dimensional data in space and time domains. Extensive numerical studies have been carried out on two real sensor data sets. The results prove the superiority of the proposed scheme in sparsity adaptation and reconstruction accuracy.