Deep-Learning-Based THz Wireless Channel Property Prediction in Motherboard Desktop Environment

Abstract

This paper proposes a residual network (ResNet) based feature concatenated neural network model to predict the type of scenario the channel is under and the attribute of the predicted scenario with power delay profile (PDP) as the inputs. The generalized model structure consists of three blocks for feature extraction, scenario prediction, and attribute prediction, respectively. The PDP data is collected from a motherboard desktop environment under five different physical arrangement scenarios. Within each scenario, data is collected several times while varying a different physical attribute for each scenario. Two steps of data augmentation are applied to expand the size and to improve the resolution (difference between the neighbouring attributes) of the measured dataset for the robust training and thorough evaluation of the proposed model. The proposed model is evaluated and compared with an multi-layer perceptron (MLP) based model on an expanded measured and averaged interpolated dataset. It is shown that both models perform very well on the expanded measured dataset with nearly 100% prediction accuracy on both scenarios and attributes. The MLP based model suffers performance degradation on the averaged interpolated dataset with up to a 9% drop of classification accuracy on attribute prediction tasks, while our ResNet based feature concatenated model performs equally in both scenarios. Feature activation map (FAM) and Grad-Class Activation Mapping (Grad-CAM) approaches are applied to provide visual explanations highlighting characteristics of the input PDP used for model decisions. FAM shows that the MLP based model focuses on the multipath generated peaks of the PDP where some interpolated neighboring data points cannot be distinguished. The Grad-CAM shows that the proposed ResNet based feature concatenated model performs better because it has strong attention not only on the multipath peaks, but also on the valleys between those peaks which hold distinguishing information.