Instantaneous Signal Collision Detection Using In-Band Full-Duplex: Machine Learning VS Domain-specific Knowledge

Collision detection (CD) is a key capability of carrier sense multiple access (CSMA) based medium access control (MAC) protocol. Applying CD, the transmitter can abort transmission immediately so that the power can be saved. This technique does not need the peer receiver to give feedback on whether there is a packet collision, and hence, the overall overhead is significantly low. The challenge, however, is to operate in transmit time and instantly detect the week colliding signal in the presence of strong self-interference (SI). Using this dataset, we investigate two CD methods and compares them regarding the detection performance and the false alarm rate. The first method trains a convolutional neural network (CNN) model which operates on raw baseband samples, without the need for pre-decoding. The second method treats the SI as a normal signal and estimates the signal to noise ratio (SNR): low SNR implies there is a collision because the pure SI is expected to have high SNR. Both models are evaluated by IEEE 802.15.4-like measured and simulated signals, available in the dataset. The results show that collisions up to 30 dB below the SI signal can be detected precisely within 20 us, while the proposed models can deliver an acceptably low false alarm rate < 1.5%.