Hierarchical Symbolic Perception in Dynamic Data Driven Application Systems

Open Access
Sarkar, Soumalya
Graduate Program:
Mechanical Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
June 12, 2015
Committee Members:
  • Asok Ray, Dissertation Advisor
  • Christopher Rahn, Committee Member
  • Daniel Connell Haworth, Committee Member
  • Shashi Phoha, Committee Member
  • Thomas Wettergren, Committee Member
  • Karen Ann Thole, Committee Member
  • Symbolic Dynamics
  • Data Driven
  • Hierarchical approach
  • Multi-scale analysis
  • Combustion Instability
  • Fault detection
A dynamic data-driven application system (DDDAS) is a recently formalized architecture that integrates simulation with dynamically assimilated data, multiscale modeling and computation via completion of a two-way feedback loop between model execution and the data acquisition modules. The perception layer of DDDAS requires precise feature extraction from dynamic data, multi-scale (both spatial and temporal) data analysis and hierarchical sensor fusion, which are capable of real-time execution. This dissertation proposes hierarchical approaches that tackle different aspects of the perception layer of DDDAS, based on the concept of symbolic time series analysis (STSA). STSA is a nonlinear feature-extraction tool that is applied on dynamic data by constructing probabilistic finite state automata (PFSA) via symbolization based on an alphabet size. Along with proposing a supervised alphabet size selection algorithm, the dissertation develops optimal STSA algorithms that capture the texture of individual temporal data (by Generalized D-Markov machines) and that of the causal cross-dependence (by ×D-Markov machines) from one temporal data set to another with reduced dimensionality. The aforementioned patterns are aggregated to construct a spatio-temporal pattern network (STPN) from a sensor network and it is pruned via an information-theoretic approach for fault diagnosis in distributed systems. Finally, a hierarchical architecture named multiscale symbolic time series analysis (MSTSA) is formalized, where a semantic finite length learning-and-inferencing mechanism is stacked to capture the multi-timescale nature of DDDAS events in real time. In addition to extensive validation by complex simulation experiments, the proposed techniques have been applied on experimental data for early detection of Lean Blow out (LBO), and thermo-acoustic instability in combustors and the performance is compared with state-of-the-art results. A hierarchical framework is developed combining deep learning and STSA with the objective of autonomously tracking the temporal evolution of coherent structures that are distributed in the flame during combustion instability. MSTSA has been successfully implemented in real-time activity recognition from noisy sensor data for border surveillance scenarios. It is envisioned that the general contributions, made in this dissertation, will be useful for many other potential application areas in DDDAS paradigm, such as weather system causality understanding, smart grid and buildings, distributed energy systems, situation awareness in ISR (intelligence, surveillance and reconnaissance) missions, material damage prediction and future ground and air transportation systems.