Advancing regression based analytics for steel fabrication productivity modeling

Abstract: Accounting for seven percent of the Gross Domestic Product (GDP), the construction industry is the fifth largest contributor to the Canadian economy (Statistics Canada 2019). Structural steel is one of the primary materials used in the construction industry for providing structural stability in residential and commercial buildings, as well as critical infrastructure and industrial facilities such as bridges, oil and gas pipe racks. With emerging economic diversification efforts in Canada, it is expected that the construction industry, and the utilization of steel products therein, will continue to grow over the coming years. To respond to increasing demand, the construction industry is relying more on off-site prefabrication to shorten project delivery time, reduce construction cost, and improve overall quality. Off-site fabrication shops provide a safer work environment for labourers, conducive to higher productivity, while also removing uncertainties and risks associated with site conditions and environmental factors.Despite this pressing need, the construction industry faces several challenges when shifting to more data-driven productivity modelling systems. First, because of the complexity, variability, and uncertainty of construction conditions and activities, productivity-influencing factors cannot be exhaustively identified and quantified, making it practically impossible to account for every relevant detail. As a response to this challenge, prefabrication facilities isolate environmental factors and implement manufacturing-like methodologies that minimize productivity-influencing factors. Second, in order for practitioners to trust and apply developed productivity models, the generated models need to be transparent, easy to use and adaptable. Third, with limited resources available in the construction industry, implementing and maintaining data-driven models need to minimize overhead costs and take advantage of readily available information as much as practically possible to optimize data collection efforts. Therefore, a systematic, transparent, and quantitative approach to determine labour productivity, based on historical project data, is optimal to support project cost estimating, resource scheduling and productivity analysis. Furthermore, an innovative approach is highly desirable to account for sufficient project details describing product uniqueness, complexity, and uncertainty involved in steel fabrication processes. This research proposes a new methodology that correlates labour productivity data with project design features. This methodology essentially utilizes efficient data-driven methods to capture implicit patterns in historical data and steel structure design details to produce labour productivity models. The novelty of the present research lies in its simple-to-understand and easy-to-implement analytical approach in selecting model input parameters and classifying steel fabrication projects based on work content and design features. The focus of this research is on applications of Multiple Linear Regression (MLR) and proposes enhanced methods to cater to application needs, first by selecting a proper set of input variables through a proposed method called Modified Stepwise Regression (MSR), then by splitting the feature domain by Model Trees (MT) into different branches to predict non-linearity using piecewise linear models. Compared to other predictive methods, this approach would satisfy the construction industry application need for transparency and ease of use in modelling productivity, while maintaining minimal data collection efforts and achieving high prediction accuracy.The contributions of this study include: (1) proposing an application framework based on Modified Stepwise Regression (MSR) for selecting relevant input variables and streamlining a predictive model without losing the model's predictive power; the MSR method leverage a simple but different method to select input variables while also verifying MLR underlying assumptions; (2) developing and validating a stee fabrication labour productivity model and identifying effects of work content factors; (3) developing an analytical methodology to generate a system of Multiple Linear Regression (MLR) equations by coupling the power of MSR and Model Tree (MT); (4) formalizing a quantitative approach to analyze the trade-off between model fit quality, prediction accuracy, and model complexity; and (5) providing an analytical means to elucidate productivity data structure and influencing factors by classifying the data and identifying significant variables for each class.Although this research focusses on applying the proposed methodologies and framework to steel fabrication productivity modelling, the proposed data-driven methodologies and application framework can be implemented wherever there is a need for a transparent, accurate and generalized predictive model to quantify input-output relationships. Concrete slump and viaduct installation time-predictive models are just a few examples of the generic applicability of the methodologies proposed and demonstrated in this thesis.