Machine Learning for Detecting Drift Fault of Sensors in Cyber-Physical Systems

Abstract

Cyber-Physical System (CPS) emerges as a potential direction to improve the applications relating to object-to-object, human-to-human and human-to-object communications in both the real world and virtual world. The examples of CPSs include Smart Girds (SG), Wireless Sensor Networks (WSNs), transportation networks etc. However, the Ubiquitiy of CPSs pose extra challenges to the systems, such as intrusions from external agent and failures originated internally in nodes of the networks. These internally appeared failures are generally referred to as faults. The Drift fault, that result in increasing linearly the output of sensor from normal value, is the most commonly occured fault in sensor nodes. In this paper, a fault detection model is proposed using machine learning to identify drift fault immediately after it appears in sensor. The proposed approach uses an Auto-Encoder (AE) to extract features from the raw signal of sensor. The vector of features is forwrded to Support Vector Machine (SVM) for fault identification. We compare the performance of four different architectures of AE: primary AE, Stacked AE (SAE), Denoising AE (DAE), and Stacked Denoising AE (SDAE). The number of features extracted by these network is varied from 2 to 15. The accuracy of the SVM is used as a parameter for comparison. The results shows that the DAE outperform the counter AE, SAE, and SDAE architectures.