Big Data Modeling Approaches for Engineering Applications

Abstract

Engineering is intrinsically a field in which the application of science and mathematics is utilized to solve problems in pursuit of the design, operation, maintenance, and other faculties of systems in complex systems. Many of these systems contain nonlinear interactions and as such, require tools of varying robustness and power to describe them. Forecasting of future states or designing such systems is very costly, time consuming, and computationally intensive, due to finite project timelines and technical constraints within industry. Increasing the calculation power will provide us with daily data production in modeling and analysis of complex dynamic systems to exceed 2.5 exabyte by 2020, which is a 44-fold increase from those seen in 2010, illustrating the rapid changes in this area. “Big data” is a relatively amorphous term used to describe the rise in data volumes that are difficult to capture, store, manage, process, and analyze, using traditional database methods and tools. The new reality of big data has and shall continue to have profound implications on modeling, as new and highly valuable information can be extracted for decision-making. Volume, often considered to be the primary characteristic of big data, refers to the absolute size of the dataset being considered. Variety in big datasets also provides additional challenges. Given the great diversity of data sources, including sensors, images, video feeds, financial transactions, location data, text documents, and others, reconciling these sources into unified modeling strategies is not straightforward. When considering so many different types of data, big data modeling strategies typically address three distinct types of data: structures data, semi-structured data, and unstructured data. In this chapter, a review of classical machine learning methods will be provided, including a selection of clustering, classification, and regression methods. Then it will detail the six approaches for applying scalable machine learning solutions to big data, specifically, representation learning methods for data reduction. Deep learning for capturing highly nonlinear behavior, distributed and parallel learning, transfer learning for cross-domain and cross-task learning activities, active learning, and kernel-based learning will address the challenges associated with big data.