무선 센서 네트워크에서 센서 고장 진단을 위한 지능형 기계 학습 기법

Abstract

Wireless Sensor Network (WSN) being highly diversified Cyber-Physical System makes it vulnerable to numerous failures. These failures due to abnormal behaviors in the network can cause serious threat towards safety, economy, and reliability of systems. Abnormal behaviors of sensors are primarily triggered by low-quality production, electromagnetic interference, and complex environments. The precise detection and diagnosis of abnormal behaviors in WSN is a challenging issue due to the diversity of deployment and limitations in the resources.

In this dissertation, a data-driven supervised machine learning-based techniques are considered to scrutinize the behavior of sensors through their data for the timely detection and diagnosis of abnormal behaviors (faults or anomaly). In this study, most of the faults that commonly occur in WSN are considered such as drift, hard-over, spike, erratic, data-loss, stuck, and random fault.

A trusted dataset published by the researchers at the University of North Carolina composed of temperature and humidity sensor healthy measurements of multi-hop scenario was acquired and the aforementioned faults were injected in the non-faulty (healthy) sensor measurements. This practice is common among researchers due to the lack in availability of defective datasets.

Events from fault occurrences were generated to replicate realistic scenarios of WSN. For instance, fault may occur in WSN for a short length as well as long, or it may occur in the combination of both. To detect and diagnose the faults in timely manner, an ensemble learning-based lightweight machine learning classification technique is adopted, which is known as Extremely Randomized Trees or Extra-Trees.

Furthermore, multiple data labelling approaches such as multi-label/multi-class were utilized in order to get the best performance out of machine learning classifiers. In this study, the proposed Extra-Trees-based detection and diagnosis scheme has shown the ability of robustness towards signal noise and strong reduction of bias and variance error.

The performance of the proposed scheme was compared with those of the state-of-the-art machine learning algorithms such as support vector machine, neural network, random forest, and decision tree. Performance evaluation shows the efficiency of the proposed scheme in terms of lightweightness and detection/diagnosis accuracy, precision, F1-score, and area value under the ROC curve. To achieve the lightweight measure, the proposed scheme training time was compared to the aforementioned state-of-the-art machine learning classifiers.