Recent Projects
I've been invited by the Irvine Ranch Conservancy (IRC) to conduct a pilot project evaluating a multitude of environmental management and protection applications for my proprietary AI-driven sensor network technology. The program is focused on championing the detection, measurement, notification, and prediction of a variety of dangerous environmental factors in order to facilitate their mitigation and prevention. I am really excited about the opportunity to demonstrate my solution in a larger context, in collaboration with Dr. Gregory and his team, and expect the outcomes to be instrumental in future conservation efforts. Regular updates about this project will be provided.
Initial Inspiration
In November 2018 while traveling, I saw on TV the devastating fires burning through California hillsides very similar to where I live. I panicked and called my mom back home to see if she was ok. I then followed the news with dismay as 85 people were killed, and over 18000 structures destroyed by a single wildfire. I was shocked to find out that most of the damage (and almost all fatalities) happened within the first six hours. In high wind conditions, these fires can travel up to 10-15 mph, so early detection is crucial. Existing solutions, such as satellites only, or high-end cameras on geological stations, are too expensive and difficult to be widely deployed. Hence, they cannot be readily available at high-risk locations (near high voltage power company equipment/lines for example). As such, existing solutions suffer from the timeliness of response. By doing further research I realized what a global problem this is and is only getting worse with global warming. I became very motivated to come up with a solution for early wildfire detection and growth prediction that is low cost and easily deployable, and that can hopefully be used all over the world, including developing countries.
Proposed Solution
Early detection is achieved if fire detectors are deployed locally at high-risk locations. Real-time growth prediction, enabled by running Machine Learning at the edge of the network, could warn people in the path of the fire and allow for efficient deployment of fire-fighting resources.
My solution is a wireless-adhoc-mesh network of solar-powered sensors, comprised of two types of nodes: $20 Fire Detectors FD, and $60 Mini Meteorological Stations MMS. In normal operation, all sensor data is constantly sent to my Mobile App once every 60 seconds and can be viewed. Once any of the fire detectors detect a fire, a push notification is sent to the App specifying which sensor has detected a fire. At that point, the Micro Meteorological Station will proceed to predict fire growth map in real-time on the edge of the network, leveraging trained machine learning models and feeding it all sensor data. captured.
Real-world wildfire data was captured from Google Earth to train Machine Learning models and test various algorithms. Random Forest (RF) algorithm was found to be the best solution and can predict wildfire growth with >70% accuracy and false alarm <15%. RF also showed the best resilience to data loss which is possible in a remote/fragile setting. RF algorithm was run on MMS and shown to predict wildfire growth in <10 seconds. The fire location/growth-path are communicated to a custom-built Fire mobile app. A realistic network with 2 FDs and 1 MMS was built and tested in the outdoor environment. A variety of tests were performed to check its wireless network robustness and ability to detect fires minutes within the start.
Micro Meteorological Station & Fire Detector Architecture
I used the Raspberry Pi (R-Pi) as the foundation of my solution. R-Pi model 4 was used for MMS (since it runs the machine learning), and R-Pi model 0 was used for FD (since all they do is sensor fusion). The PiJuice power supply was utilized to enable solar operation. LTE cellular connectivity was added to MMS to allow for long-range connectivity. I developed/integrated all the Python code on the R-Pi.
Mobile App
The Mobile App can be used by firefighters or people who live in high-risk areas. In normal operation, all sensor data is constantly sent to the App once every 60 seconds and can be viewed. Once any of the fire detectors detect a fire, a push notification is sent to the App specifying which sensor has detected a fire. At that point, the Micro Meteorological Station will proceed to predict fire growth map in real-time on the edge of the network, leveraging trained machine learning models and feeding it all sensor data. captured.
Leveraging Machine Learning to Predict Growth of Wildfires
To train/test fire prediction models, actual wildfire data was captured for “Camp-Fire” of November 2018 from Google Earth Javascript API. 10 days were captured, 8 days for training 2 days for testing. Various independent variable data were captured from different satellite/terrestrial sources in Google Earth. For training, Neural Networks, Random Forest, SVC and Logistic Regression were tested. Models trained non-real-time and downloaded to RaspberryPi4 on MMS to run real-time.
Evolution towards a general hazard detection platform
When COVID hit, I started thinking about how can I utilize my platform to help. The network/platform architected is robust/modular and can be rapidly repurposed to deal with other emergencies. The fire-detection sensors were replaced with thermal cameras and the FDs were rebuilt and tested to function as a highly cost-effective network of Fever Detection cameras for COVID19 screening. Machine Learning predicts spikes throughout the network and can help with tracing/tracking. Furthermore, I realized by replacing sensors and software, my solution can be effectively utilized for early detection and growth prediction, Methane Gas Leaks, Water Pollution.....