Deep Learning for 5G Localization: Hybrid CNN-Transformer Approach in Accuracy Enhancement

Accurate localization of user equipments (UEs) in 5G communication systems is critical for maintaining synchronization with the base stations and supporting emerging location-based services. However, hardware impairments such as, amplitude distortion, phase noise, I/Q imbalance and other nonlinearities degrades localization accuracy by deviating from ideal behaviour. This paper proposes a hybrid convolutional neural network (CNN)-transformer regression model to mitigate the adverse effects of hardware impairments and enhance 5G localization accuracy. In a proposed framework, the CNN extracts local spatial features from channel-related data, while the transformer captures long-range dependencies through a self-attention mechanism, wherein relationships between all input features are modelled simultaneously. Unlike conventional CNNs that operate with fixed receptive fields, transformers dynamically weigh input features based on their contextual relevance, enables the model to better adapt to multipath propagation, non-line-of-sight (NLoS) conditions and hardware-induced distortions. Compared to previous hybrid approaches such as, model-driven deep neural networks (MoD-DNNs) that have physical modelling into CNN architectures, but face challenges of computational complexity and limited generalization across environments. Proposed model offers reduced complexity and greater robustness to real-world variations. Experimental evaluations demonstrate that the proposed hybrid model achieves R-squared values of 0.994 on validation data and 0.995 on testing data. This indicates an excellent model fit. Results substantiate the model's capability to maintain high localization accuracy under severe hardware impairment conditions, thus validating its effectiveness for real-time 5G localization in practical deployment scenarios.