Machine Learning enabled 2D-CFAR improvement in Integrated Sensing and Communications

Sammanfattning

A total of 24 UAV flights in the vicinity of integrated sensing and communications-enabled (ISAC) antenna systems, called base stations, were selected to investigate whether tracking can be improved using machine learning methods. This was done by inputting time series of range-Doppler (r-D) maps created from continuously calculating channel estimates of the bistatic environment around base stations to various neural network designs. The r-D maps were processed by a constant false alarm rate (CFAR) algorithm to detect potential candidate points with a higher signal-to-noise ratio. Patches surrounding the CFAR detection points were then extracted from the current time step and the three previous time steps and fed to a neural network to compute a probability score for each point. The probability scores were then used to classify the CFAR points as UAVs or clutter. The final neural network, reached by iterating through different designs, relied on three convolutional blocks to learn spatial relations in the patches along with a transformer block to employ attention between different sensing directions. This design managed to achieve 94.2\% accuracy and 98.9\% recall at a 50\% decision threshold, and 87.8\% accuracy and 99.4\% recall at a 30\% decision threshold. Furthermore, it required only 400 ms to process 6,400 CFAR detection points. Implementing similar task-specific neural networks in any radar system can enable the development of faster and more resource-efficient radars, not only in the ISAC framework.