Data-driven production planning and control using work density for on-site building construction

The construction industry faces challenges due to escalating project complexity, demands for accelerated project delivery, and dwindling contractor margins. Traditional planning and control methods, characterized by deterministic and transformation-focused approaches, struggle to manage the dynamic interactions within production systems, including workflows and variabilities. Despite the potential of takt production and other location-based methods to streamline workflows and enhance project outcomes, their practical adoption remains limited due to several inherent challenges.This dissertation addresses the challenges in adopting location-based production planning and control methods by identifying key obstacles such as the non-repetitive nature of work, dynamic production rates, and the lack of objective and data-driven systems necessary for effective production planning and control. To mitigate these issues, I propose a comprehensive Data-Driven Production Framework (DDPF), which integrates planning and control components using work density to measure, store, and utilize actual production data. While this framework is developed using takt production principles, its final objectives are left to the users, making it relevant for other location-based methods.Employing a qualitative case study approach based on Design Science Research (DSR) principles, this research develops and validates artifacts including a comprehensive suite of tools and methods under the DDPF. I validated the application of these artifacts in real-world settings on two case studies, highlighting their practical applications and the benefits of using work density for planning and control.The first case study involves the application of the Work Density Method (WDM) through the ViWoLZo tool (Visual Workload Leveling and Zoning), demonstrating the utility of work density in enhancing decision-making during production planning by enabling an objective way to reduce variability and improve process flow, among several other production metrics proposed to support planning. Furthermore, the role of cost modeling is explored, reimagining cost from a production perspective and integrating cost considerations directly into the planning process to ensure cost-effective production decisions.In ViWoLZo, I utilize work density maps for each process step with non-uniformly spaced grid lines to calculate work density per cell and aggregate the cells into zones to calculate the workload per step per zone. However, its application highlighted challenges with manual and subjective data collection and analysis required to create the work density maps, leading to the development of the DDPF with a novel data model and data schema, leveraging real-time data collection and advanced data-driven methods. This data model moves away from the original definition of work density that is cell-based and is calculated using estimates, to a new definition that is location-based, where a location is a room or zone, and is measured empirically. Comparing it to the original definition, this work density can be considered cumulative work density for the cells in a room or a zone, hence referred to as measured workload.The second case study focuses on the DDPF’s application in production control, utilizing advanced data collection and analysis methods, including a 360° camera-based tracking, and a production performance dashboard for supporting production control. Algorithms for work detection, such as the computer vision-based algorithm used in this research, are constantly improving. Current implementations are limited as they only work with objects in line-of-sight and with a certain range and accuracy, limited by the nature of the work to be detected, the resolution of the sensor, and the types of detection models used. I used the actual performance data to measure the duration of steps by location, representing the step’s measured workload. A combined dataset with the workload for a location along with the contextual data describing what work was done, by whom, where, and ow, are stored as historical work density. I used this dataset to fit machine learning (ML) models that can predict work density for future planning efforts.In this research, unsupervised learning methods were employed to understand the data and enhance the supervised learning-based prediction of work density. The investigation of the model’s prediction performance demonstrated the DDPF’s potential utility in production planning, achieving an expected error of 3.7 days compared to the historical data. The model’s performance was constrained due to the limited amount of data collected and higher variability in the higher range of values of measured workloads. The distribution of measured workloads in the dataset has a mean of 20.44 days with a standard deviation of 19.03. To use the model on takt plans requiring smaller workloads, the ML model will have to be exposed to empirical data from processes with similar workloads. Although on separate occasions, the ML models have been trained and tested on real-world data and work density has been used for planning with ViWoLZo, and they are connected using the DDPF, the prediction model and ViWoLZo have not yet been integrated for use in real-world scenarios to plan future projects using predicted work density.The use of work density enabled ViWoLZo and the production performance dashboard to translate production system dynamics into a data-driven model. Interviews with field personnel confirmed the DDPF’s effectiveness in implementing takt-like production to reduce process duration through increased concurrency, enhance stakeholder communication through transparency and visual management, and improve overall project performance through reduced variability and increased predictability of the production performance. The ML models and data analyses demonstrate that contextual data contain valuable insights that can streamline the implementation of location-based production planning and control methods. However, due to the small size of data collected from a single case study, the complex relationships within the data, and many features left to be collected and tested, these results are just the building blocks for further research.The dissertation concludes with recommendations for broadening the adoption of the DDPF, emphasizing the need for advancements in data collection technologies and analysis methods, and the need to standardize historical work density libraries. Improved performance combined with simpler implementation can potentially increase the adoption of these methods in the industry, and in turn, lead to further advancements.