Sequential Anomaly Detection Using Wireless Sensor Networks in Unknown Environment

Abstract

Anomaly detection is an important problem for environment, fault diagnosis and intruder detection in Wireless Sensor Networks (WSNs). A key challenge is to minimize the communication overhead and energy consumption in the network when identifying these abnormal events. We present a machine learning (ML) framework that is suitable for WSNs to sequentially detect sensory level anomalies and time-related anomalies in an unknown environment. Our system consists of a set of modular, unsupervised, machine learning algorithms that are adaptive. The modularity of the ML algorithms to maximize the use of resource constrains sensor nodes in different environmental monitoring tasks without reprogramming. The developed ML framework consists of the following modular components. First, an unsupervised neural network is used to map multi-dimensional sensor data into discrete environmental states/classes and detect sensor level anomalies. Over time, the labeled classes form a sequence of environmental states. Next, we use a variable length Markov model in the form of a Probabilistic Suffix Tree (PST) to model the relationship between temporal events. Depending on the types of applications, high order Markov models can be expensive. We use a symbol compression technique to bring down the cost of PST models by extracting the semantic meaning out of temporal sequences. Lastly, we use a likelihood-ratio test to verify whether there are anomalous events. We demonstrate the efficiency our approach by applying it in two real-world applications: volcano monitoring and traffic monitoring applications. Our experimental results show that the developed approach yields high performances based on different benchmarks compared to traditional fixed length Markov models.