Neural Network-Assisted Distortion Representation for Small Sample Self-Interference Cancellation in ISAC Systems

In the process of realizing integrated sensing and communication, which radiates signals and receives echoes through the same station, digital-domain self-interference cancellation has always been a key issue. However, due to the fact that the system can obtain very few clean self-interference signals in this scenario, the reconstruction of entire self-interference becomes more challenging. Most traditional methods and neural network approaches are overly dependent on the amount of data available. Therefore, in this paper, a simple architecture combining deep learning with traditional algorithms is proposed. Parameters of the self-interference signal are extracted from the network to construct a nonlinear expression, and then the least squares estimation is applied to estimate and reconstruct the self-interference signal. On average, this method achieved approximately 55.35 dB of overall suppression across fifty sets of random data.