A prominent challenge physicist and computer scientist faces is developing ways to accurately rebuild spatial fields from data gathered by sparse sensors. This process becomes even more challenging when sensors are sparsely positioned in a seemingly random or unstructured fashion.
Traditional linear theory-based methods have shown to be inefficient in reconstructing global fields for complicated physical systems or processes, especially when only a small amount of sensor data is available or sensors are randomly positioned. As a result, computer scientists have been investigating the potential of alternate methods for global field reconstruction, such as deep learning models.
A team of researchers from Japan’s Keio University, the University of California-Los Angeles, and other institutions in the United States has introduced a novel deep learning technique that can accurately rebuild global fields without requiring large amounts of well-organized sensor data.
Researchers studying complicated physical systems frequently have limited access to data collected by a small number of sensors arranged in an unorganized manner. These sensors sometimes may also move and go offline for short periods.
The lack of suitable sensor data has made reconstructing global fields for these complicated systems problematic. Although deep learning approaches have yielded some promising results, putting them into practice may be costly and time-consuming.
The team created a global field reconstruction technique that combines deep learning with Voronoi tessellation, a method of representing and characterizing biological structures and physical systems. Voronoi tessellations or diagrams have been employed in various fields of science and engineering in the past.
The proposed approach combines sparse sensor data into a CNN and approximates local information onto a structured representation while keeping data relating to the sensor location. To accomplish this, it first creates a Voronoi tessellation of the unstructured information. Then it adds the input data field corresponding to the sensor locations as a mask.
This method for global field reconstruction has two advantages:
- It may be used with an arbitrary number of sensors
- It’s compatible with DL-based image processing algorithms that have shown promise in the past.
The researchers have already proved the efficacy of their method by reconstructing global fields using three different kinds of sensor data, namely unsteady wake flow, geophysical data, and 3D turbulence data.
This new approach operates with data collected from a random number of moving sensors. Therefore, it can be applied in numerous applications, such as estimating real-time global fields for various physical systems, even when sensors are positioned in a disorganized manner.