Adaptive Compensation of Hardware Impairments in Digitally Modulated Radars Using ML-Based Behavioral Models

This work proposes novel compensation approaches for full radar transceiver (TRX) self-calibration of the IQ modulator, power amplifier (PA), and IQ demodulator without the need of dynamic channel conditions, a local oscillator (LO) frequency offset, or additional hardware. The methods employ behavioral models based on neural networks, which are trained with the radar sensor itself. Three different compensation methods at transmitter (TX), receiver (RX), as well as a distributed approach are presented, validated, and compared using a 78-GHz orthogonal frequency-division multiplex (OFDM) radar and computer simulations. The proposed distributed approach is shown to increase the radar's spurious-free dynamic range (SFDR) by 17.6 dB compared to an uncompensated case and by 8.5 dB in comparison to high input back-off (IBO) operation. Furthermore, among the studied compensation methods in this work, the distributed approach simultaneously offers additional significant benefits: the highest signal-to-noise ratio (SNR) for all targets in the scene and reduced out-of-band (OOB) emission.