Reduced Complexity Deep Learning Approach for Bluetooth Ranging in Multipath Environments

The multicarrier phase difference (MCPD) wireless ranging technique, which uses the two-way channel frequency response (CFR) measurements, helps mitigate time and phase offsets. However, these measurements are challenged by an increased number of multipath components (MPCs) due to squaring the one-way CFR, a low number of snapshots, and model deficiencies in practical scenarios. Deep learning methods have produced remarkable improvements in different fields of research by learning complex systems and mitigating the system model imperfections using collected data. Although data-driven deep learning models produce good ranging accuracy for the environment it is trained on, generalizing the model to new unseen environments is challenging. Furthermore, deep learning algorithms are expensive in terms of memory and computational resources on embedded devices. To mitigate these problems, we provide system-level information to our low-complexity deep learning model using novel randomized low-rank approximation (RLA) of the signal subspace from the two-way CFR covariance matrix for Bluetooth ranging. The estimated signal subspace is used as an input to our separable convolution neural network (CNN) deep learning model, used for the first time for ranging. Our model greatly reduces the amount of parameters of the deep network, reducing memory requirements and being computationally more efficient than traditional approaches such as multiple signal classification (MUSIC). Overall, our approach outperforms MUSIC in terms of ranging accuracy by achieving overall root mean square error (RMSE) of 0.5 m (range measurements from 1 to 7 m) with 44% improvement over MUSIC in new unseen environments, and without the need for fine-tuning or additional data collection.