Non-Productive Time Reduction in Off-Site Construction: A Predictive Analytics Approach Using Deep Learning

Abstract

This study proposes a data-driven framework to reduce Non-Productive Time (NPT) in Off-Site Construction (OSC) by integrating predictive modeling with lean principles. Through comprehensive computer vision (CV) techniques, large-scale video data are analyzed to identify key operational states, worker positions, and material flows. Two primary classification approaches-a Random Forest (RF) model and a Deep Learning (DL) network-are employed to detect NPT episodes. Experimental results reveal that the DL model achieves substantially higher accuracy and F1 scores, demonstrating its capability to handle complex spatial-temporal interactions. Feature-weight analysis further indicates that task transitions, assembly processes, and coordinate-based congestion zones are critical drivers of NPT. Building on these insights, targeted lean interventions-such as just-in-time material delivery, proactive maintenance, and optimized workstation layouts-can streamline operations, minimize idle intervals, and enhance overall productivity. The conclusions underscore the effectiveness of merging data analytics with lean strategies in uncovering hidden inefficiencies within OSC, providing a clear direction for future research and industry adoption.