Symbolic Dynamic Filtering for Pattern Recognition in Distributed Sensor Networks
Proliferation of modern sensors and sensing applications provides us with large volumes of data. By extracting useful information from these data sets in real time, we enhance the ability to better comprehend and analyze the environment around us. Tools for data-driven feature extraction and pattern classification facilitate performance monitoring of distributed dynamical systems over a sensor network, especially if the physics-based models are either inadequate or unavailable . In this regard, a critical issue is real-time analysis of sensor time series for information compression into low-dimensional feature vectors that capture the relevant information of the underlying dynamics [3,9,10,22]. 308Time-series analysis is a challenging task if the data set is voluminous (e.g., collected at a fast sampling rate), high dimensional, and noise contaminated. Moreover, in a distributed sensor network, data collection occurs simultaneously at multiple nodes. Consequently, there is a need for low-complexity algorithms that could be executed locally at the nodes to generate compressed features and therefore reduce the communication overhead. In general, the success of data-driven pattern classification tools depends on the quality of feature extraction from the observed time series. To this end, several feature extraction tools, such as principal component analysis (PCA) , independent component analysis (ICA) , kernel PCA , dynamic time warping , derivative time-series segment approximation , artificial neural networks (ANN) , hidden Markov models (HMM) , and wavelet transforms [26,27,35] have been reported in technical literature. Wavelet packet decomposition (WPD)  and fast wavelet transform (FWT)  have been used for extracting rich problem-specific information from sensor signals. Feature extraction is followed by pattern classification (e.g., using support vector machines [SVM]) [3,9].