Machinery detection by impulsive noise recognition using WiFi sensing

Abstract

Engines and electrical devices in operation generate electromagnetic pulses, also called impulsive noise (IN), that interfere with wireless signals. The IN shall affect the channel estimation process and is, therefore, present in the channel state information (CSI) provided by wireless receivers. In this paper, impulsive noise (IN) is used as a fingerprint of electrical devices to identify the IN sources that interfere with a WiFi signal, taking into account that each individual machine has a unique pattern of impulsive noise. In this sense, the WiFi CSI provides valuable information to recognize the IN sources through deep learning (DL) strategies. Two DL models have been proposed and tested on two experimental data sets for multiclass and multilabel analysis; in multiclass, devices can operate alone during the measurement, and in multilabel, multiple devices can work simultaneously in a more realistic scenario. The model transferability between location and days has also been evaluated by analyzing two different IN feature sets for device classification with the Few-shot-learning (FSL) model. Results show that the proposed DL models can recognize multiple devices working simultaneously through the IN and also offer an acceptable transferability performance ($\sim$ 80% accuracy for a five-class problem).