Non-stationary signal feature extraction and analysis

Extraction of time-varying frequency characteristics of a non-stationary signal is important in understanding them better and has immense applications in a number of real-world problems. A joint time-frequency (TF) analysis is required for extracting TF features from non-stationary signals. Of the two well known approaches of performing TF analysis, the non-parametric TF energy distributions are commonly used for visualization purposes whereas the parametric and adaptive TF decomposition approaches are attractive for extracting TF features. While the adaptive TF decomposition techniques have several parametric advantages in extracting TF features, there exists no unique or automated methodology in utilizing them for better non-stationary signal analysis.;Keywords. Time-frequency, Non-stationary Signals, Signal Decompositions, Local Discriminant Bases, Feature Extraction, Wavelet Packets, Matching Pursuits;In this thesis, we propose a unique and efficient methodology for the analysis of non-stationary signals using adaptive TF techniques. The treatment of this work is general and applies to any non-stationary signal, however an emphasis is placed on addressing biomedical scenarios with small databases. A novel time-width versus frequency band (TWFB) mapping is introduced which facilitates the adaptation of a Local Discriminant Bases (LDB) like approach to any TF decomposition technique in automating the identification process of discriminatory TF subspaces. The strong TF features could then be extracted from these discriminatory TF subspaces for either characterization or classification of non-stationary signals. The theoretical properties of the TWFB mapping were investigated from the viewpoint of pattern recognition and synthetic signal examples are provided. Selective reconstruction is another important feature of TWFB mapping that finds utility in partial reconstruction and separation of overlapping signal structures. This aspect was also examined and examples are provided. The proposed work also presents the process of constructing target positive TF distributions (PTFD) using the discriminatory TF subspaces for extracting instantaneous TF features. Summaries of the results at various phases of the work using synthetic and real world signals are presented in the respective chapters. In conclusion, it is shown that the proposed TWFB mapping technique is a powerful pattern recognition tool with a wide range of real world applications in non-stationary signal analysis.