Characterization of multiple time-varying transient sources from multivariate data sequences
Wachowski, Neil, author
Azimi-Sadjadi, Mahmood R., advisor
Breidt, F. Jay, committee member
Fristrup, Kurt, committee member
Pezeshki, Ali, committee member
Characterization of multiple time-varying transient sources using sequential multivariate data is a broad and complex signal processing problem. In general, this process involves analyzing new observation vectors in a data stream of unknown length to determine if they contain the signatures of a source of interest (i.e., a signal), in which case the source's type and interference-free signatures may be estimated. This process may continue indefinitely to detect and classify several events of interest thereby yielding an aggregate description of the data's contents. Such capabilities are useful in numerous applications that involve continuously observing an environment containing complicated and erratic signals, e.g., habitat monitoring using acoustical data, medical diagnosis via magnetic resonance imaging, and underwater mine hunting using sonar imagery. The challenges associated with successful transient source characterization are as numerous as the application areas, and include 1) significant variations among signatures emitted by a given source type, 2) the presence of multiple types of random yet structured interference sources whose signatures are superimposed with those of signals, 3) a data representation that is not necessarily optimized for the task at hand, 4) variable environmental and operating conditions, and many others. These challenges are compounded by the inherent difficulties associated with processing sequential multivariate data, namely the inability to exploit the statistics or structure of the entire data stream. On the other hand, the complications that must be addressed often vary significantly when considering different types of data, leading to an abundance of existing solutions that are each specialized for a particular application. In other words, most existing work only simultaneously considers a subset of these complications, making them difficult to generalize. The work in this thesis was motivated by an application involving characterization of national park soundscapes in terms of commonly occurring man-made and natural acoustical sources, using streams of "1/3 octave vector'' sequences. Naturally, this application involves developing solutions that consider all of the mentioned challenges, among others. Two comprehensive solutions to this problem were developed, each with unique strengths and weaknesses relative to one another. A sequential random coefficient tracking (SRCT) method was developed first, that hierarchically applies a set of likelihood ratio tests to each incoming vector observation to detect and classify up to one signal and one interference source that may be simultaneously present. Since the signatures of each acoustical event typically span several adjacent observations, a Kalman filter is used to generate the parameters necessary for computing the likelihood values. The SRCT method is also capable of using the coefficient estimates produced by the Kalman filter to generate estimates of both the signal and interference components of the observation, thus performing separation in a dual source scenario. The main benefits of this method are its computational efficiency and its ability to characterize both components of an observation (signal and interference). To address some of the main deficiencies of the SRCT method, a sparse coefficient state tracking (SCST) approach was also developed. This method was designed to detect and classify signals when multiple types of interference are simultaneously present, while avoiding restrictive assumptions concerning the distribution of observation components. This SCST method uses generalized likelihood ratios tests to perform signal detection and classification during quiescent periods, and quiescent detection whenever a signal is present. To form these tests, the likelihood of each signal model is found given a sparse approximation of an incoming observation, which makes the temporal evolution of source signatures more tractable. Robustness to structured interference is incorporated by virtue of the inherent separation capabilities of sparse coding. Each signal model is characterized by a Bayesian network, which captures the dependencies between different coefficients in the sparse approximation under the associated hypothesis. In addition to developing two complete transient source characterization systems, this thesis also introduces several concepts and tools that may be used to aid in the development of new systems designed for similar tasks, or supplement existing ones. Of particular note are a comprehensive overview of existing general approaches for detecting changes in the parameters of sequential data streams, a new method for performing fusion of sequential classification decisions based on a hidden Markov model framework, and a detailed analysis of the 1/3 octave data format mentioned above. The latter is especially helpful since this data format is commonly used in audio analysis applications. A comprehensive study is carried out to evaluate the performance of the developed methods for detecting, classifying, and estimating the signatures of signals using 1/3 octave soundscape data that is corrupted with multiple types of structured interference. The systems are benchmarked against a Gaussian mixture model approach that was adapted to handle the complexities of the soundscape data, as such approaches are frequently used in acoustical source recognition applications. Performance is mainly measured in terms of the receiver operator characteristics (ROC) of the test statistics implemented by each method, the improvement in signal-to-noise ratio they offer when estimating signatures, and their overall ability to accurately detect and classify signals of interest. It was observed that both the SRCT and SCST methods perform exceptionally on the national park soundscape data, though the latter performs best in the presence of heavy interference and is more flexible in new environmental and operating conditions.