Theses and Dissertations from UMD

Permanent URI for this communityhttp://hdl.handle.net/1903/2

New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a give thesis/dissertation in DRUM

More information is available at Theses and Dissertations at University of Maryland Libraries.

Browse

Subject: search.filters.subject.Convolutional Neural Network

Search Results

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    INTERPRETABLE AND SPEED ADAPTIVE CONVOLUTIONAL NEURAL NETWORK FOR PROGNOSTICS AND HEALTH MANAGEMENT OF ROTATING MACHINERY
    (2023) Lee, Nam Kyoung; Pecht, Michael; Azarian, Michael H; Mechanical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    Faulty rotating machines exhibit vibrational characteristics that can be distinguished from healthy machines using prognostics and health management methods. These characteristics can be extracted using signal processing techniques. However, these techniques require certain inputs, or parameters, before the desired characteristics can be extracted. Setting the parameters requires skill and knowledge, as they should reflect the component geometries and the operational conditions. Using convolutional neural networks for diagnosing faults on a rotating machine eliminates the need for parameter setting by replacing signal processing with mathematical operations in the networks. The parameters that affect the outcomes of the operations are learned from data during the training of the neural networks. The networks can capture characteristics that are related to the health state of a machine, but their operations are not interpretable. Unlike signal processing, the internal operations of the networks have no constraints that guide the networks to transform vibrations into certain information, that is, vibrational characteristics. Without the constraints, there is no basis for understanding the characteristics in terms that can be associated with the physics of failure. The lack of interpretability impedes the physical validation of vibrational characteristics captured by the networks.This dissertation presents a method for changing the internal operations of a convolutional neural network to emulate a specific type of signal processing known as envelope analysis. Envelope analysis demodulates vibrations to extract vibrational signatures associated with mechanical impact on a defective rolling component. An understanding of envelope analysis, along with knowledge of the geometries of machine components and operational speeds, allows for a physical interpretation of the signatures. The dissertation develops speed adaptive convolutional layers and a rotational speed estimation algorithm to identify defect signatures whose frequency components change as the speed changes. The characteristics that are captured by the developed convolutional neural network are verified through a feature selection process that is designed to filter out physically implausible features. Case studies on three different systems demonstrate the feasibility of using the developed convolutional neural network for the diagnosis.
  • Thumbnail Image
    Item
    SMART STRUCTURAL CONDITION ASSESSMENT METHODS FOR CIVIL INFRASTRUCTURES USING DEEP LEARNING ALGORITHM
    (2018) Liu, Heng; Zhang, Yunfeng; Civil Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    Smart Structural Health Monitoring (SHM) technique capable of automated and accurate structural health condition assessment is appealing since civil infrastructural resilience can be enhanced by reducing the uncertainty involved in the process of assessing the condition state of civil infrastructures and carrying out subsequent retrofit work. Over the last decade, deep learning has seen impressive success in traditional pattern recognition applications historically faced with long-time challenges, which motivates the current research in integrating the advancement of deep learning into SHM applications. This dissertation research aims to accomplish the overall goal of establishing a smart SHM technique based on deep learning algorithm, which will achieve automated structural health condition assessment and condition rating prediction for civil infrastructures. A literate review on structural health condition assessment technologies commonly used for civil infrastructures was first conducted to identify the special need of the proposed method. Deep learning algorithms were also reviewed, with a focus on pattern recognition application, especially in the computer vision field in which deep learning algorithms have reported great success in traditionally challenging tasks. Subsequently, a technical procedure is established to incorporate a particular type of deep learning algorithm, termed Convolutional Neural Network which was found behind the many success seen in computer vision applications, into smart SHM technologies. The proposed method was first demonstrated and validated on an SHM application problem that uses image data for structural steel condition assessment. Further study was performed on time series data including vibration data and guided Lamb wave signals for two types of SHM applications - brace damage detection in concentrically braced frame structures or nondestructive evaluation (NDE) of thin plate structures. Additionally, discrete data (neither images nor time series data), such as the bridge condition rating data from National Bridge Inventory (NBI) data repository, was also investigated for the application of bridge condition forecasting. The study results indicated that the proposed method is very promising as a data-driven structural health condition assessment technique for civil infrastructures, based on research findings in the four distinct SHM case studies in this dissertation.
  • Thumbnail Image
    Item
    Multimodal Biomedical Data Visualization: Enhancing Network, Clinical, and Image Data Depiction
    (2017) Cheng, Hsueh-Chien; Varshney, Amitabh; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    In this dissertation, we present visual analytics tools for several biomedical applications. Our research spans three types of biomedical data: reaction networks, longitudinal multidimensional clinical data, and biomedical images. For each data type, we present intuitive visual representations and efficient data exploration methods to facilitate visual knowledge discovery. Rule-based simulation has been used for studying complex protein interactions. In a rule-based model, the relationships of interacting proteins can be represented as a network. Nevertheless, understanding and validating the intended behaviors in large network models are ineffective and error prone. We have developed a tool that first shows a network overview with concise visual representations and then shows relevant rule-specific details on demand. This strategy significantly improves visualization comprehensibility and disentangles the complex protein-protein relationships by showing them selectively alongside the global context of the network. Next, we present a tool for analyzing longitudinal multidimensional clinical datasets, that we developed for understanding Parkinson's disease progression. Detecting patterns involving multiple time-varying variables is especially challenging for clinical data. Conventional computational techniques, such as cluster analysis and dimension reduction, do not always generate interpretable, actionable results. Using our tool, users can select and compare patient subgroups by filtering patients with multiple symptoms simultaneously and interactively. Unlike conventional visualizations that use local features, many targets in biomedical images are characterized by high-level features. We present our research characterizing such high-level features through multiscale texture segmentation and deep-learning strategies. First, we present an efficient hierarchical texture segmentation approach that scales up well to gigapixel images to colorize electron microscopy (EM) images. This enhances visual comprehensibility of gigapixel EM images across a wide range of scales. Second, we use convolutional neural networks (CNNs) to automatically derive high-level features that distinguish cell states in live-cell imagery and voxel types in 3D EM volumes. In addition, we present a CNN-based 3D segmentation method for biomedical volume datasets with limited training samples. We use factorized convolutions and feature-level augmentations to improve model generalization and avoid overfitting.