All flying objects in the Indian airspace — commercial aircraft, missiles, drones, fighter planes and others — are detected and tracked by ground-based radars produced by public sector unit B harat Electronics Limited (BEL).
The current system used by the Indian Air Force relies on a Multi-Sensor Tracking (MST) mechanism via radars located in different parts of the country with information such as location and velocity coordinates captured by each radar collated to provide the Air Situation Picture (ASP).
ASP is a detailed listing of all aircraft in the airspace along with their corresponding flight numbers and flight plans. But, overlapping radars and delay in communication between sensors lead to a ‘merging’ error – a situation where multiple aircraft in close proximity to each other in the airspace are incorrectly identified as one.
Another is ‘splitting’ — where a single aircraft is sensed as multiple and erroneously flagged as a threat. From a security point of view, threat detection and evaluation need to be spot-on but due to these errors, the ASP generated is not 100% accurate.
A team from BEL-Ghaziabad comprising Roshan Kumar, Rohit Singh, Aravind Kumar and Manoj Tyagi approached the International Institute of Information Technology (IIIT)-Hyderabad, for the development of an automated solution to address these issues.
The institute’s research team led by Praveen Paruchuri of the Machine Learning Lab, with Masters student Anoop Dasika after due research, came out with a AI-assisted tracker model demonstrating a 5% improvement in the original tracking mechanism, providing an accuracy of upto 91%.
The software has been transitioned to BEL and is currently being tested out in their simulation environment. The algorithm accurately identifies the flights – that it is a single flight in case of split flights and that there are more than one in case of merged flights, said an IIIT-H online post on Friday.
In both cases, the corresponding global tracking number of the flights is obtained. “Instead of the MST system directly transmitting information to the operators’ screens, now data from MST goes into our AI system where it is preprocessed and the original global track number(s) is identified before relaying it back to the original server. MST now transmits it to the operators’ screens to decide the most suitable course of action,” explained Mr. Anoop.
Unlike other AI models developed based on a one-time training of data, the reseachers advocate periodical training every fortnight. “Periodic training is useful to catch any new or emerging patterns due to old radars getting decommissioned, newer ones installed and so on,” said Mr Paruchuri.
“Our model also helps in a detailed radar analysis to prioritise replacement or repair of the radars,” he added. Research findings were published in a paper titled, “CB+NN Ensemble To Improve Tracking Accuracy In Air Surveillance” and it is slated be presented at the 34th annual conference on Innovative Applications of Artificial Intelligence (IAAI-22) to be held virtually between February 24-26.
