Georgia Institute of Technology
Automated Prediction of Behavior and Productivity of Construction Workers through Wearable Motion Sensors and Deep-Learning Approaches
Abstract
dc:description.abstractConstruction tasks involve numerous activities composed of one or more body motions. Because construction projects are labor-intensive and rely heavily on manual tasks, understanding the ever-changing behaviors and activities is essential to effectively manage construction workers regarding their safety and productivity. Although several research efforts have shown promising results in automated motion and activity recognition of the workers using motion sensors, there remains a lack of understanding of how the number of motion sensors and their locations affect the performance of the recognition and contribute to improving recognition performance and reducing implementation costs. Further research is needed to seek a motion recognition model that accurately identifies various motions using motion sensors attached to the workers’ bodies. Furthermore, although detection-based approaches have performed well in identifying posture-related safety hazards, researchers have made few efforts to directly obtain labor productivity from workers’ behavior because this would require the discretization of continuous movements to estimate a production amount. To address these challenges, this study proposes artificial intelligence (AI)-based methods for effectively and efficiently recognizing the various motions of individual construction workers and predicting their labor productivity using wearable motion sensors and deep-learning (DL) approaches. The development of the motion recognition method using a DL approach includes evaluating the effectiveness of the number of motion sensors and their locations to maximize the AI-based motion recognition methods and sensor implementation efficiency. The evaluation is conducted by generating different datasets containing motion sensor data collected from the wearable motion sensors located on different body parts. Using the performance of five machine learning models trained using the datasets, the evaluation formula is provided to obtain quantitative contribution scores for each node. These scores indicate the relative impact of using different nodes on the performance of AI-based motion recognition algorithms. With the formula, the desired number and locations of motion sensors are identified. Based on those findings, the long short-term memory (LSTM) network for recognizing construction workers’ motions is developed. The LSTM network classifies various motions of workers that can be utilized as primitive elements for monitoring safety and productivity. Multiple case studies are conducted to validate the technical and practical feasibility of the developed network. Furthermore, this study presents an automated labor productivity prediction method that estimates the production amount of individual workers and predicts their labor productivity. The production amount is derived by discretizing activities classified using an LSTM network and a dynamic time warping (DTW) technique. With the proposed methods, it is expected that the individual workers’ behavior and working conditions can be automatically monitored and managed without excessive manual observation. Thus, this study contributes to building the theoretical and practical foundation for automating jobsite monitoring practices and generating essential information regarding workers’ safety and productivity.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Civil and Environmental Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kim, Kinam
- Advisor dc:contributor.advisor
-
- Cho, Yong Kwon
- Committee members dc:contributor.committeemember
-
- Wang, Yang
- Marks, Eric
- Ueda, Jun
- Jeong, Inbae
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/72582