We are developing lightweight machine learning methods in order to ensure real-time distributed autonomous sensing with environmental and health objectives, mainly in the case of extreme events. In this regard, we designed, implemented and experimentally evaluated 2 main drone mission planning algorithms. In the first one, the objective is to provide a fine-resolution characterization of air pollutants, which would allow us later to determine the boundary of high pollution concentrations in case of a gas leak. The second mission planning algorithm aims to assist first responders in localizing trapped and at-risk individuals during an extreme event.
Our design focus is on mission planning algorithms that are well adapted to the case of extreme events when a hazardous gas is released in the air. We aim to provide efficient and lightweight algorithms that allow a fine-resolution characterization of the pollution concentrations and therefore identify at-high-risk areas. We also design mission planning algorithms that guide the drones in rescue operations. Our design choices compared to the literature are based on the current status of sensing technologies that are highly noisy due to drone vibrations, which could lead to bad mission planning decisions if they are not considered.
Efficient drone networks’ mission planning for gas concentrations’ characterization: The proposed mission planning approach operates as follows: after a training phase prior to the flight mission in order to quantify the quality of aerial measurements of pollution concentrations, drones are first sent to uniformly distributed locations in order to characterize the spatial correlations of air pollution concentrations. Then, these spatial correlations are used together with the inferred aerial measurements’ quality in order to optimize the following measurement locations of the drones. The optimal measurement locations of each drone are obtained by minimizing the overall variance of the interpolated concentrations’ errors while taking into account the aerial sensing constraints (the dynamic sensing error and the response time of pollution sensors, the speed of the drone and the drone’s battery capacity).
We evaluate our mission planning approach using a set of 30 pollution maps of aerial and ground measurements of VOC pollutants collected in February 2020 using our custom-built drones and ground sensors. We compare our mission planning algorithm to traditional solutions and show that due to the dynamic and non-homogeneous nature of aerial sensing quality, traditional solutions fail to send the drones to the most informative locations and select instead measurement locations that are almost uniformly distributed in the monitoring region. This results in over x2.5 improvement factor in our experimental evaluation scenario and is expected to be larger depending on the size of the monitoring region.
Efficient drone networks’ mission planning for the localization of wireless RF targets: (a wireless target can be the phones that are carried by trapped and at-risk individuals during an extreme event).
Here, we leverage existing Wi-Fi technology and its recent Fine Time Measurement (FTM) protocol to realize the first FTM sensing drones that can dynamically range targets in a mission. Based on that, we propose a new mission planning strategy to simultaneously approach and localize a target to enable higher resolution sensory measurements and to realize approaching-critical tasks in a mission in addition to localization and tracking. For that, we propose to jointly exploit drones’ diversity of observation and dynamics of approaching the target, and dynamically adjust the intensities of approaching and observation based on a mission requirement.
In order to evaluate our approach, we consider as a baseline a Bio-inspired scheme that encourages drones to move directly towards the latest estimated target location. We find that, with just two drones in the network, our mission planning algorithm can realize 2.2× average angular spread gain at the expense of only 27% additional average travel distance compared to Bio-inspired scheme. We also show that, compared to the Bio-inspired scheme, our approach consistently and rapidly acquires information about the target position due to its diverse observation feature, and localizes a target 2× more accurately and in 30% less time.
Publications:
Yue Wang, Ziyu Jiang, Xiaohan Chen, Pengfei Xu, Yang Zhao, Atlas Wang, and Yingyan Lin, “E2-Train: Training State-of-the-art CNNs with Over 80% Energy Savings”, Thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019).