RobuT-Net: Dual-CNN-Based Robust Training Sequence Design for IoT Systems

Jae-Mo Kang

Abstract: This letter proposes a new methodology for training sequence design in Internet of Things (IoT) systems based on deep learning, called RobuT-Net. The proposed RobuT-Net is constructed via a dual convolutional neural network (CNN) architecture composed of two CNN modules to effectively and intelligently design a statistically robust training sequence for the minimum mean-square error (MMSE) channel estimator, against uncertainties in both channel and noise covariance matrices. Furthermore, we develop an effective learning strategy for the proposed RobuT-Net in an unsupervised manner, which leverages intentionally deformed samples for the channel and noise covariance matrices to mitigate the adverse impacts of the uncertainties. Simulation results substantiate the superiority and efficacy of the proposed scheme.

External IDs:dblp:journals/iotj/Kang24c