| For the past many decades, several concepts and measures for studying nonlinear sensor data have been proposed and investigated. There have been many attempts to understand behavior, reliability, and performance of sensor data. This dissertation presents novel methodologies for analyzing, classifying, and recognizing patterns of nonlinear sensor data based on recurrence network analysis.;First, a comprehensive overview of recurrence theory and their quantification possibilities is presented. New measures of recurrence networks are defined by using the complex network properties. These measures are intended to recognize and classify patterns of sensor data using the feature extraction method. In this dissertation, we introduce new methodologies to classify sensor data based on recurrence quantification analysis with artificial neural network classifier. In addition, we present methods to classify and recognize patterns of sensor data based on a recurrence network. The goal of the proposed method is to compute all symmetric patterns and super families of any similar structure, even if they are non-periodic or multivariate time series.;We introduce and apply the recurrence based feature extraction method to complex and nonstationary sensor data such as physiological signals and machining sensor signals. In the lung sounds classification problem, our results show that the proposed method gives 100% classification performance. In the classification of electromyogram signals, our results show that the proposed method classifies these signals with 98.28% accuracy. Lastly, the proposed method was applied to estimate surface roughness in turning process. Acoustic emission signals are transformed into recurrence plots and a set of recurrence statistics are computed using the recurrence quantification analysis. The surface roughness parameters are estimated using a multilayer neural network, taking the recurrence statistics of acoustic signals as inputs. The estimation accuracy of the proposed method is in the range of 90.13% to 91.26%. Furthermore, these accurate results indicate that the proposed method is very effective and amenable for practical implementation.;Using the phase synchronization method, we construct the stock correlation network. It is used for observing, analyzing, and predicting the stock market dynamics. The proposed method captures the dynamic behavior of the time series of stocks and mitigates the information loss. It provides valuable insights into the behavior of highly correlated stocks which can be useful for making trading decisions. The network exhibits a scale free degree distribution for both chaotic and non-chaotic periods.;We illustrate similarities and dissimilarities with respect to the classification by considering large-scale sensor data. The selected applications of the introduced techniques to data from different applications demonstrate the ability of these techniques. We derive new methods to examine massive, multi-dimensional, multi-source, time-varying information stream of data. |
