Chinese researchers have created a lightweight artificial intelligence (AI) system that can help air-to-air missiles with thermal guidance identify advanced fighters. The system’s high accuracy was demonstrated during laboratory tests on simulated F-22 and F-35 targets.
Although the F-35 and F-22 are among the most advanced stealth fighters in the world, they cannot hide their thermal radiation. The thermal signatures generated by engine exhaust gases and air friction on the fuselage differ from the signals produced by thermal traps. Such signatures can provide valuable infrared information for target detection and recognition in future air battles, potentially reducing the effectiveness of thermal traps.
“Lightweight recognition models can be widely used in future air-to-air missiles as they can provide rapid recognition while maintaining high identification capabilities,” said An Jiangshan, the first author of the study, on August 10. The study was published in the Chinese peer-reviewed journal Journal of Electronic Measurement and Instrumentation. It involved researchers from the Beijing Institute of Technology, known for its military research and one of the few Chinese universities under U.S. sanctions.
Among the co-authors were researchers from the China Academy of Air-to-Air Missiles and the National Key Laboratory of Surface Detection. “The model provides effective classification and target recognition in onboard scanning infrared missile systems, achieving a recognition accuracy of 97.1% during testing,” the researchers wrote. The system is relatively inexpensive: the Zynq7020 platform used in the work costs only a few hundred yuan, indicating low equipment costs.
Optimization and accuracy
According to An, the corresponding set of test data for identifying fifth-generation fighters such as the American F-22 and F-35 contains confidential data and cannot be disclosed. Meanwhile, for simulated targets in the current dataset, the recognition accuracy can reach about 90%. The study focuses on onboard scanning infrared missile systems, where computational resources are limited by strict size, weight, and power consumption requirements.
The team developed a lightweight neural network because traditional deep learning models, despite their high accuracy, have a large number of parameters and high computational requirements, making their deployment on small embedded devices difficult. The new model reduced the number of parameters to 16.1% of previous methods and computational requirements to 19.2%. To achieve this, the researchers used structural optimization, batch normalization merging, and 8-bit quantization.
To adapt the AI system to missile platforms, the researchers developed a special AI accelerator. It included optimized convolution modules, parallel computing structures, and data buffering methods to improve efficiency. “The final system achieved a balance between recognition accuracy, processing speed, hardware power consumption, and hardware resource utilization,” wrote lead author Liu Ming and his team.
Testing was conducted using 3,245 infrared images of targets collected by an onboard scanning infrared system. The dataset included three categories of air targets: two types of aircraft simulating the F-22 and F-35, and a loitering munition. According to the test results, the hardware accelerator achieved a recognition accuracy of 96.4%, with an average processing time of about 1.5 milliseconds and total power consumption of 2.2 watts.
“This work mainly focuses on air-to-air missile fuzes for short-range applications, and its effectiveness for surface-to-air missiles has yet to be verified,” An noted. The researchers stated that in future work, they will focus on improving recognition accuracy and processing speed for more effective recognition of airborne infrared targets.
Source: South China Morning Post
