Enhancing data-driven applications in construction

Digital transformation of the construction industry has been slow and challenging. With continuously improving information and communication technology, increasing amounts of construction data are automatically generated throughout the stages of construction for all construction management functions. However, due to the complex nature of construction, collected data are noisy, fragmented, and discordant, consisting of observational and subjective as well as structured and unstructured information. These types of data form natural barriers for use in any data-driven applications, limiting their ability to provide reliable, timely, and informed decision support. How to fully exploit the value of "big data"—specifically, learn as much as we can from the raw construction data that we collect—is a challenge the entire construction industry is facing. This research investigated this problem by addressing three specific challenges that hinder the digital transformation of the construction industry: 1) low automation for integrating and pre-processing fragmented construction data for project-level decision support; 2) lack of means for fusing information derived from various origins for data-driven simulation in real-time; and 3) slow implementation of machine learning, resulting in organizations 'drowning' in a flood of data. This research adopted methods from applied mathematics and statistics, data science, and computing science to develop methodologies capable of addressing these challenges. This research better exploits the value of construction data and improves its conversion into informed project decision support. Specifically, Through the development of an enhanced data-driven application framework with two embedded custom functions to automate key data preprocessing steps for data aggregation and merging, this research increases information flow between segmented data sets, thus enhancing data-driven simulation and analytics in general; Through the proposal of two methods for enabling real-time input model calibration for simulation, this research establishes a foundation of dynamic data-driven simulations to incorporate real-time data of diverse origins, extending their applications to all stages of a project's life cycle and potential connections with multiple project stakeholders; Through the development of a data solution to improve preliminary resource planning in industrial construction, this research not only provides vital decision support—a scientific and data-driven resource plan at the early planning stage—but also demonstrates the practicality of integrating unsupervised and supervised learning for large, unlabeled, and noisy construction data. This research has achieved the goal of bridging low-quality construction data to a real-time data solution and contributed to the academic literature and construction industry by: 1) proposing a novel framework for enhanced data-driven applications built upon fragmented construction data; 2) developing and generalizing functions to automate and streamline the otherwise manual data pre-processing steps; 3) proposing a numerical-based Bayesian inference method for systematically updating input models (any given univariate continuous probability distribution) of simulations as new observations become available; 4) proposing a Markov chain Monte Carlo-based weighted geometric average method to effectively fuse information generated from diverse sources (both subjective and objective) for stochastic simulation inputs; and 5) developing a data solution to scientifically plan project resources with incomplete engineering.