Enhancing Wireless Sensing Using Simulated Datasets: A CNN-Based Approach to mmWave Angle of Arrival Estimation

Wireless sensing, with applications such as crowd monitoring and localization, extracts valuable location-related information from communication signals. Deep learning models are widely used in this field to identify relationships between sensing data and critical communication metrics, such as channel state information and received signal strength. However, the process of data collection remains resource-intensive, involving the installation of communication infrastructure, the setup of multiple sensing points, and participant recruitment. To address these challenges, this paper proposes a simulation-based approach for generating wireless sensing datasets. Channel data was generated using ray-tracing techniques, preserving essential location information while incorporating various noise power levels. The generated dataset was utilized to train a CNN-based model for angle-of-arrival estimation, which was subsequently validated against DeepSense, a real-world mmWave sensing and communication dataset. Results demonstrate that the simulated dataset enables effective model performance in real-world scenarios. This approach mitigates the cost and time constraints of traditional data collection, offering researchers greater flexibility in training and exploration of model capabilities.