Sequential Machine Learning for Decision-Making in Mechanical Systems
Open Access
- Author:
- Ghalyan, Najah
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 23, 2019
- Committee Members:
- Asok Ray, Dissertation Advisor/Co-Advisor
Asok Ray, Committee Chair/Co-Chair
Eric Russell Marsh, Committee Member
Alok Sinha, Committee Member
David Jonathan Miller, Outside Member
William Kenneth Jenkins, Outside Member
Shashi Phoha, Outside Member
Daniel Connell Haworth, Program Head/Chair - Keywords:
- symbolic dynamics
Hidden Markov Models
ergodicity
measure-preserving transformations
image classification
fatigue damage
thermoacoustic instabilities
probabilistic finite state automata
time series
bag-of-words
support vector machines
generating partition
quantization
lossy source coding
sliding block source coding
clustering
dynamic programming
Viterbi algorithm
anomaly detection
change point detection
transient change detection
quickest change detection
data-driven modeling - Abstract:
- Sequential machine learning for anomaly detection is critical in applications where fast control action is required to avoid failure of the system. Example is thermoacoustic instabilities (TAI) in combustion systems, which may lead to damage in mechanical structures if the resulting pressure oscillations match one of the natural frequencies of the system. TAI typically develop on the order of milliseconds, which must be mitigated by sufficiently fast actuation of control signals. Likewise, fatigue damage is one of the most common source of failure in structural materials. Initiation and evolution of this type of damage are critically dependent on the microstructural initial defects that are usually distributed in a highly random fashion. Therefore, fatigue damage is a stochastic process, for which early detection is required for condition-based maintenance and life extension of the system. The current PhD dissertation considers the problem of sequential machine learning for anomaly detection and decision-making in mechanical systems with an emphasize on the aforementioned two applications. In particular, the dissertation develops several novel data-driven algorithms that utilize the theory of Symbolic Time Series Analysis (STSA) and Hidden Markov Models (HMMs) for anomaly detection by learning sequential patterns from observed time series. While standard partitions in STSA symbolize each observation in a time series individually, the dissertation proposes two novel algorithms that jointly symbolize the entire time series. The first algorithm amounts to a novel type of sliding block lossy source coding, which can be used to estimate a finite generator from observed time series. The second algorithm utilizes the Viterbi (dynamic programming) algorithm to jointly convert the time series to a symbol string with maximum posterior probability conditioned on the observed time series. Both algorithms induce sequence space partitions which are particularly important for data-drive modeling of dynamical systems using short-length time series measurements. Moreover, an HMM-based algorithm is developed for feature extraction. From the STSA perspective, this algorithm generates soft symbolization of the time series, retaining information associated with all possible symbol strings. Furthermore, the dissertation presents a novel framework in STSA for anomaly detection in dynamical systems using the ergodic theory of measure-preserving transformations (MPTs). Unlike a standard STSA that generates time-homogeneous Markov chains, the proposed MPT-based STSA generates time-inhomogeneous Markov chains that can greatly facilitate modeling of the dynamical system using short-length time series of measurements. The dissertation also introduces a novel detection criterion well-matched to low-delay, narrowly localized change point detection, and develops an HMM-based algorithm that can efficiently make change point detection and narrowly identify an interval within which the change point occurred using the joint likelihood of a sliding block conditioned on the block's entire past. All the algorithms developed in this work have been experimentally validated and compared with other standard detection techniques using experimental data generated from the two aforementioned applications. The results consistently show superior performance of the proposed detection algorithms. Moreover, from the perspectives of health monitoring and life extension of structural materials, the dissertation also addresses the problem of early detection of fatigue cracks in metallic alloys. To this end, optical images have been collected from an ensemble of test specimens to construct computationally efficient models of crack evolution; these images are segmented into two major categories. The first category comprises images of (structurally) healthy specimens, while the second category contains images of specimens with cracks, including those in early stages of crack evolution. Based on this information, algorithms for early detection of crack formation are formulated in the setting of image classification, where the bag-of-words (BoW) technique has been used to develop models of the sensed images from a microscope, resulting in computationally efficient crack detection algorithms. To evaluate the performance of these crack detection algorithms, experiments have been conducted on a special-purpose fatigue testing apparatus, equipped with a computer-controlled and computer-instrumented confocal microscope system. The results of experimentation with multiple test specimens show excellent crack detection capabilities when the proposed BoW-based feature extraction is combined with quadratic support vector machine (QSVM) for pattern classification. Comparative evaluation with other classification tools establishes superiority of the proposed BoW/QSVM technique.