We are developing lightweight machine learning methods to ensure real-time distributed autonomous sensing with environmental and health objectives. In this regard, and in order to first validate our drone network, we focused over the past year on an application that has an analogy with toxic gas leaks dispersion, which is the detection and localization of mobile wireless targets. We implemented our first algorithms and showed that our drone platform is fully operational and ready to accommodate the air pollution dedicated ML algorithms that we are currently developing.
Because gas sensors are expensive and gas leaks are rare events, we had to start this thrust with a sensing application that is viable for testing the robustness of our drones’ network. We have chosen to focus over the past year on the application of the detection and localization of a mobile wireless transmitter for three main reasons. First, the spectral signature of radio transmissions is analogous to gas dispersion in the air. Secondly, when the wireless target is moving, it mimics the impact of wind and in-air chemical reactions. Thirdly, the mobile target can hide, which is analogous to non-uniform pollution gradients.
Our ML algorithm operates in two phases: the search and learn phase and the swarm and track phase. During the search and learn phase, drones fly in accordance with pre-computed partitioned zones and planned paths obtained by solving a multiple travelling salesman problem. In particular, the zone partition will determine how to divide the search area into zones so that each drone takes charge of one zone, and the planned paths provide guidance for drones to fly within their corresponding zones. Each drone works independently during this phase and learns the signal propagation model parameters while trying to identify a potential target. Once a drone is in the range of a target, it informs all other drones. Subsequently, all drones will swarm to the target and switch to the tracking phase in which they work collaboratively to locate and track the mobile target. We performed over 500 hours of test flights to validate the robustness of our drones’ platform and evaluate the performance of our ML algorithms. We configured the drones to have a maximum velocity of 2 m/s, well below their peak capability of approximately 20 m/s, in order to focus on algorithmic dynamics in a relatively small test area, which was the football field of our campus. We launched the drones from the edge of the field. On average, our tests showed that a single drone can find the wireless target within 25s, whereas two drones can cut the average to about 15s. However, additional drones have only minimal marginal benefit in our test configuration. Nonetheless, the results show that when the size of our drones’ network increases, the network remains fully operational and the drones are able to coordinate autonomously to achieve their mission more efficiently.