Data-driven sensor placement with shallow decoder networks

Sensor placement is an important and ubiquitous problem across the engineering and physical sciences for tasks such as reconstruction, forecasting and control. Surprisingly, there are few principled mathematical techniques developed to date for optimizing sensor locations, with the leading sensor placement algorithms often based upon the discovery of linear, low-rank sub-spaces and the QR algorithm. QR is a computationally efficient greedy search algorithm which selects sensor locations from candidate positions with maximal variance exhibited in a training data set. More recently, neural networks, specifically shallow decoder networks (SDNs), have been shown to be very successful in mapping sensor measurements to the original high-dimensional state space. SDNs outperform linear subspace representations in reconstruction accuracy, noise tolerance, and robustness to sensor locations. However, SDNs lack principled mathematical techniques for determining sensor placement. In this work, we develop two algorithms for optimizing sensor locations for use with SDNs: one which is a linear selection algorithm based upon QR (Q-SDN), and one which is a nonlinear selection algorithm based upon neural network pruning (P-SDN). Such sensor placement algorithms promise to enhance the already impressive reconstruction capabilities of SDNs. We demonstrate our sensor selection algorithms on two example data sets from fluid dynamics. Moreover, we provide a detailed comparison between our linear (Q-SDN) and nonlinear (P-SDN) algorithms with traditional linear embedding techniques (proper orthogonal decomposition) and QR greedy selection. We show that QR selection with SDNs enhances performance. QR even out-performs our nonlinear selection method that uses magnitude-based pruning. Thus, the combination of a greedy linear selection (QR) with nonlinear encoding (SDN) provides a synergistic combination.