| dc.description.abstract | The construction industry has been suffering from declining productivity since the 1950s.
To tackle this issue, our industry has developed numerous productivity-tracking methods and
systems. However, most of the existing tracking approaches focus on measuring work-level or
team-level productivity. Some researchers have tried visual tracking with in-situ complexities or
vibration data analysis for fall prediction. In this paper, we implemented vibration data analysis
methods to efficiently track and identify ongoing construction action using vibration data. These
data were collected from an accelerometer attached to power tools, representing 16 classes of
construction actions frequently needed in pipeline work. We trained a support vector machine
model and a decision tree model by feature matrixes and label matrixes generated from Y-axis
values of raw data. We applied data preprocessing, frequency-domain feature extraction,
training, 10-fold cross-validation, and parameter optimization. After cross-validation, results
showed the support vector machine to have a better average accuracy result compared with the
decision tree. Meanwhile, the support vector machine model successfully identified ongoing
construction action. Overall, this research makes a significant contribution to applying machine-learning
methods by vibration-data-processing techniques for tracking construction actions. In
the future, construction managers can use this system to track and identify ongoing action on the
site remotely, improving work efficiency and work-tracking robustness. | en |
