Developing a Low-Order Statistical Feature Set Based on Received Samples for Signal Classification in Wireless Sensor Networks and Edge Devices

Abstract

Classifying fluctuating operating wireless environments can be crucial for successfully delivering authentic and confidential packets and for identifying legitimate signals. This study utilizes raw in-phase (I) and quadrature-phase (Q) samples, exclusively, to develop a low-order statistical feature set for wireless signal classification. Edge devices making decentralized decisions from I/Q sample analysis is beneficial. Implementing appropriate security and transmitting mechanisms, reducing retransmissions and increasing energy efficiency are examples. Wireless sensor networks (WSNs) and their Internet of Things (IoT) utilization emphasize the significance of this time series classification problem. Here, I/Q samples of typical WSN and industrial, scientific and medical band transmissions are collected in a live operating environment. Analog Pluto software-defined radios and Raspberry Pi devices are utilized to achieve a low-cost yet high-performance testbed. Features are extracted from Matlab-based statistical analysis of the I/Q samples across time, frequency (fast Fourier transform) and space (probability density function). Noise, ZigBee, continuous wave jamming, WiFi and Bluetooth signal data are examined. Supervised machine learning approaches, including support vector machines, Random Forest, XGBoost, k nearest neighbors and a deep neural network (DNN), evaluate the developed feature set. The optimal approach is determined as an XGBoost/SVM classifier. This classifier achieves similar accuracy and generalization results, on unseen data, to the DNN, but for a fraction of time and computation requirements. Compared to existing approaches, this study’s principal contribution is the developed low-order feature set that achieves signal classification without prior network knowledge or channel assumptions and is validated in a real-world wireless operating environment. The feature set can extend the development of resource-constrained edge devices as it is widely deployable due to only requiring received I/Q samples and these features are warranted as IoT devices become widely used in various modern applications.

Over-the-Air Modulation Classification using Deep Learning in Fading Channels for Cognitive Radio

“…Clock offset occurs as a result of the imperfections of transmitters and receivers' internal clock sources (20) . Clock offset deviates from the optimum values for the center frequency, which is used to down-convert the signal to baseband, and the sampling rate of the digital-to-analog converter.…”