WSN optimization for sampling-based signal estimation using semi-binarized variational autoencoder

This study focuses on optimizing the sampling strategies for WSNs to estimate spatiotemporal signals. Existing deep-learning-based approaches for signal estimation tend to collect samples from pre-determined sensing locations, due to which the performance of signal estimation relies heavily on selected handcrafted features. Instead of fixing sensing locations, we propose a semi-binarized variational autoencoder to simultaneously optimize the sampling strategy and evaluate the signal estimated from the sampled sensing locations. The proposed framework is composed of a backpropagatable binarized encoding layer to optimize sensing locations and a generative model to estimate the complete signal from these sparse samples. Moreover, a feature-level discrepancy was proposed to further optimize the sampling locations with respect to the estimation error. The experiments were conducted using four publicly available datasets with three evaluation metrics (mean square error, standard deviation, and peak signal-to-noise ratio). The 10-fold cross-validation and two-sample t-test were utilized to analyze the experimental results, which demonstrate the significant improvement achieved by the proposed method.