Efficient Machine Learning Classifier for Fault Detection in Wireless Sensor Networks

Abstract

The deployment of wireless sensor networks in unpredictable and dangerous conditions makes them prone to software, hardware, and communication errors. Sensors are physical devices that are deployed in inaccessible environment which makes them malicious. The Fault occurs in the sensed data and its detection should be precise and rapid to limit the loss. The status of sensed data should be explicitly determined to guarantee the normal function of the Sensor Networks. For the purpose of fault detection machine learning classifiers are employed because they are effective and used to classify sensed data into faulty and non-faulty data. The faults due to Dos, Probe, R2L, and U2R are considered for implementation. KDD CUP 99 dataset is chosen for training and test purpose, and the dataset contains 41 features which are categorized as content, basic, TCP features. The required feature for each fault category is selected through recursive feature elimination technique. The performance of the classifier is measured and evaluated in terms of Accuracy, precision, recall, and F-measures. From experimental results, it is observed that Random Forest classifier is best suited for Wireless Sensor Networks fault detection. The simulation result shows that Multi-layer perceptron outperforms the other classifier with 92% of accuracy.