In this paper, we propose a novel complex convolutional neural network (CNN) CSI enhancer for integrated sensing and communications (ISAC), which exploits the correlation between the sensing parameters (such as angle-of-arrival and range) and the channel state information (CSI) to significantly improve the CSI estimation accuracy and further enhance the sensing accuracy. Within the CNN CSI enhancer, we use the complex-valued computation layers to form the CNN, which maintains the phase information of CSI. We also transform the CSI into the sparse angle-delay domain, leading to heatmap images with prominent peaks that can be efficiently processed by CNN. Based on the enhanced CSI outputs, we further propose a novel biased fast Fourier transform (FFT)-based sensing scheme for improving the range sensing accuracy, by artificially introducing phase biasing terms. Extensive simulation results show that the ISAC complex CNN CSI enhancer can converge within 30 training epochs. The normalized mean square error (NMSE) of its CSI estimates is about 17 dB lower than that of the linear minimum mean square error (LMMSE) estimator, and the bit error rate (BER) of demodulation using the enhanced CSI estimation approaches that with perfect CSI. Finally, the range estimation MSE of the proposed biased FFT-based sensing method approaches that of the subspace-based sensing method, at a much lower complexity.

翻译：本文提出了一种面向集成感知与通信（ISAC）的新型复数卷积神经网络（CNN）信道状态信息（CSI）增强器。该增强器利用感知参数（如到达角和距离）与CSI之间的相关性，显著提升CSI估计精度，进而增强感知准确性。在CNN CSI增强器中，我们采用复数计算层构建CNN，以保留CSI的相位信息。同时，我们将CSI变换至稀疏角度-时延域，生成具有显著峰值的热力图图像，可被CNN高效处理。基于增强后的CSI输出，我们进一步提出一种新颖的有偏快速傅里叶变换（FFT）感知方案，通过人为引入相位偏置项来提升距离感知精度。大量仿真结果表明，该ISAC复数CNN CSI增强器可在30个训练周期内收敛。其CSI估计的归一化均方误差（NMSE）比线性最小均方误差（LMMSE）估计器低约17 dB，且利用增强CSI估计进行解调的误码率（BER）趋近于理想CSI下的性能。最后，所提出的有偏FFT感知方法的距离估计均方误差（MSE）在显著降低复杂度的情况下，趋近于子空间感知方法的性能。