Date of Award

9-10-2019

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Industrial and Manufacturing Systems Engineering

Keywords

Artificial Intelligence AI, Machine Learning ML, Deep Learning, Convolutional Neural Networks CNN, Big Data Analytics, Breast Cancer, Gene Biomarkers, Treatment Therapies, Survivability, Cyber Physical Systems (CPS), Cloud Computing, Internet of Things (IoT), Smart Manufacturing Systems Industry 4.0

Supervisor

El Maraghy, W.

Rights

info:eu-repo/semantics/openAccess

Abstract

The evolution of technology in all fields led to the generation of vast amounts of data by modern systems. Using data to extract information, make predictions, and make decisions is the current trend in artificial intelligence. The advancement of big data analytics tools made accessing and storing data easier and faster than ever, and machine learning algorithms help to identify patterns in and extract information from data. The current tools and machines in health, computer technologies, and manufacturing can generate massive raw data about their products or samples. The author of this work proposes a modern integrative system that can utilize big data analytics, machine learning, super-computer resources, and industrial health machines’ measurements to build a smart system that can mimic the human intelligence skills of observations, detection, prediction, and decision-making. The applications of the proposed smart systems are included as case studies to highlight the contributions of each system. The first contribution is the ability to utilize big data revolutionary and deep learning technologies on production lines to diagnose incidents and take proper action. In the current digital transformational industrial era, Industry 4.0 has been receiving researcher attention because it can be used to automate production-line decisions. Reconfigurable manufacturing systems (RMS) have been widely used to reduce the setup cost of restructuring production lines. However, the current RMS modules are not linked to the cloud for online decision-making to take the proper decision; these modules must connect to an online server (super-computer) that has big data analytics and machine learning capabilities. The online means that data is centralized on cloud (supercomputer) and accessible in real-time. In this study, deep neural networks are utilized to detect the decisive features of a product and build a prediction model in which the iFactory will make the necessary decision for the defective products. The Spark ecosystem is used to manage the access, processing, and storing of the big data streaming. This contribution is implemented as a closed cycle, which for the best of our knowledge, no one in the literature has introduced big data analysis using deep learning on real-time applications in the manufacturing system. The code shows a high accuracy of 97% for classifying the normal versus defective items. The second contribution, which is in Bioinformatics, is the ability to build supervised machine learning approaches based on the gene expression of patients to predict proper treatment for breast cancer. In the trial, to personalize treatment, the machine learns the genes that are active in the patient cohort with a five-year survival period. The initial condition here is that each group must only undergo one specific treatment. After learning about each group (or class), the machine can personalize the treatment of a new patient by diagnosing the patients’ gene expression. The proposed model will help in the diagnosis and treatment of the patient. The future work in this area involves building a protein-protein interaction network with the selected genes for each treatment to first analyze the motives of the genes and target them with the proper drug molecules. In the learning phase, a couple of feature-selection techniques and supervised standard classifiers are used to build the prediction model. Most of the nodes show a high-performance measurement where accuracy, sensitivity, specificity, and F-measure ranges around 100%. The third contribution is the ability to build semi-supervised learning for the breast cancer survival treatment that advances the second contribution. By understanding the relations between the classes, we can design the machine learning phase based on the similarities between classes. In the proposed research, the researcher used the Euclidean matrix distance among each survival treatment class to build the hierarchical learning model. The distance information that is learned through a non-supervised approach can help the prediction model to select the classes that are away from each other to maximize the distance between classes and gain wider class groups. The performance measurement of this approach shows a slight improvement from the second model. However, this model reduced the number of discriminative genes from 47 to 37. The model in the second contribution studies each class individually while this model focuses on the relationships between the classes and uses this information in the learning phase. Hierarchical clustering is completed to draw the borders between groups of classes before building the classification models. Several distance measurements are tested to identify the best linkages between classes. Most of the nodes show a high-performance measurement where accuracy, sensitivity, specificity, and F-measure ranges from 90% to 100%. All the case study models showed high-performance measurements in the prediction phase. These modern models can be replicated for different problems within different domains. The comprehensive models of the newer technologies are reconfigurable and modular; any newer learning phase can be plugged-in at both ends of the learning phase. Therefore, the output of the system can be an input for another learning system, and a newer feature can be added to the input to be considered for the learning phase.

Share

COinS