Joint Interference Cancellation with Imperfect CSI

Precoding and Modulation techniques are widely investigated to mitigate interference in the space, time, and frequency domains, respectively. However, next-generation (xG) channels are increasingly dense and mobile, making these domains highly correlated and resulting in high-dimensional channel tensors. Separate interference cancellation methods treating the channel as independent 2-D matrices fail to cancel the joint interference across multiple Degrees of Freedom (DoF). High Order Generalized Mercer’s Theorem (HOGMT) based precoding and Multidimensional Eigenwave Multiplexing (MEM) modulation have been proposed to cancel the joint interference by leveraging the jointly orthogonal eigencomponents of the channel. However, these methods rely heavily on perfect Channel State Information (CSI), and their performance degrades with imperfect CSI. In this paper, we propose a joint interference cancellation method based on the Convolutional Neural Network (CNN) with a single linear kernel (JIC-CLK). We show that a CNN with a single linear kernel is equivalent to a wireless system, where the linear kernel itself is identical to the transmitted signal with the CSI as input. By training the output to approach the desired signal at the receiver with MSE loss, the kernel converges to the optimal transmitted signal in the MMSE sense. The proposed method adapts to both perfect and imperfect CSI and can be extended to high-dimensional channels. The accuracy and generality of the proposed method are validated by simulations in three cases: 1) Perfect CSI, 2) Imperfect CSI, and 3) 4-D spatio-temporal channels with imperfect CSI.