Structured Low-Rank Tensor Completion for IoT Spatiotemporal High-Resolution Sensing Data Reconstruction

Due to various restrictions, some Internet of Things (IoT) sensing layers can only deploy a small number of sensor nodes for spatiotemporal low-resolution environmental information sensing, making the urgent issue of how to recover the spatiotemporal high-resolution sensing data (SHD). Existing methods mainly focus on the reconstruction problem of random data loss in densely deployed nodes, while continuous data loss in spatiotemporal low-resolution sensing data (SLD) can severely degrade their reconstruction performance. In this work, an structured low-rank tensor completion method is proposed to avoid the impact of continuous data loss and further enhance the spatio-temporal correlation. The SLD is arranged in a third-order tensor, where horizontal and vertical directions are node location indexes, and tubal direction is the time index. To avoid continuous data loss and enhance the spatial correlation of data, each frontal slice of the tensor is divided into a group of overlapping patches and then concatenated into a third-order spatial structure tensor. To further ensure the stricter low-rank prior, the spatial structure tensors are divided into two groups and linearly mapped to a third-order tensor with Hankel structure to exploit the spatiotemporal correlation among the data, and then the two Hankel tensors are concatenated into a three-order tensor for exploiting inter-Hankel tensor temporal correlation. Experimental results on real and simulated IoT data show that the proposed method can reconstruct SHD with high accuracy and the normalized mean absolute error is lower than 0.0172 and 0.0124, respectively, when only 12% of the data is observed.