Chaos Theory-Inspired Evolutionary Method to Refine Imperfect Sensor Data for Data-Driven Construction Simulation
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
2018 American Society of Civil Engineers. In construction planning, inherent uncertainties in activity sequences and variations of work packages can often cause deviations from project plans. In order to effectively study and preempt scenarios that may lead to undesired project time and cost overruns, several construction domain-specific simulation platforms have been designed and introduced in the past. Although transitioning to simulation-based decision-making has great potential to streamline project conceptualization and early planning, the inability of simulations to evolve with the real system can significantly limit their applicability and render them unreliable for construction-phase decision-making. This issue has been identified as one of the grand challenges to industry-wide adoption of simulation models throughout the lifecycle of construction and infrastructure projects. A potential solution to this problem is to equip simulations with sensing systems that interact with and collect project data in runtime. This approach, however, requires meticulous effort to procure, set up, operate, synchronize, calibrate, and maintain sensing devices over a large project area. This practical challenge can potentially hinder the ability of simulation systems to adapt and remain relevant for decision-making. Moreover, sensor readings are often noisy and imperfect. If used as inputs to a simulation model, this noise in sensor data can create volatility in simulation outputs. Chaos theory describes how small variations in input can cause high output errors even in simple systems. To this end, this paper presents an evolutionary algorithm to process and significantly reduce noise in imperfect sensor data captured by low-cost consumer-grade sensors. The main contribution of this work to the body of knowledge is a scientific methodology of refining imperfect (noisy) sensor data and producing clean datasets that can be used to generate more stable simulation input models. This methodology was validated in a field experiment in which simulation models were created using noisy (imperfect) as well as refined sensor data. The output of each simulation model was compared with ground truth values. Analysis of results shows that using refined sensor data to generate simulation models significantly improves the accuracy and reliability of the simulation output.
