Communication Scene Recognition Method Based on Multi Phone Sensors and Deep Learning

Wireless channel quality is determined by the user’s environment scene, such as indoor, corridor, elevator, subway station, walking outside, driving car and high-speed railway. Besides channel measurement at PHY layer, it is very important to recognize the communication scene for mobile user, to match the source coding scheme or transmitting resource allocation to dynamic channel characters. In this paper, we present a communication scene recognition framework based on deep learning methods, utilizing decision-level fusion of data from multiple Android phone sensors. Our approach utilizes the ResNet-34 model to process data from four sensor types: acceleration sensors, magnetic field intensity sensors, orientation sensors, and gyroscope sensors, resulting in high precision. The algorithm achieves an 8-scene recognition accuracy of 91.68% and operates at a recognition speed of 9.4 ms per recognition on a server platform equipped with a single RTX 3090 Ti graphics card.