APPLICATION OF DATA MINING TO IMPROVE CONSTRUCTION SAFETY

Abstract

Past research and industry applications have highlighted and demonstrated the usefulness of data mining and analytics approaches. However, data mining and analytics are not widely utilised in the construction industry. Until now, construction industry remains one of the most dangerous industry in Singapore and internationally. Therefore, there is a strong need to alleviate this perennial problem by providing leading indicators of safety performance through data mining (DM), which is defined herein as a combination of machine learning, analytics and database technology. This study aims to identify a set of leading indicators that predicts safety risk in construction projects using data mining techniques. The leading indicators can then be used to guide managers and workplace safety and health practitioners to focus their efforts on high risk sites so that accidents can be prevented proactively. This study also demonstrate how different data mining processes and techniques can be utilised to effectively learn from construction dataset. The data was obtained from a BCA-grade A1 Construction Company and the data were accumulated from year 2010 to 2015. The key types of data include safety inspection records, accident cases and project-related data (e.g. project manpower, project delay and project progress). This study follows the industry-recognised CRISP-DM (Cross-Industry Standard Process for Data Mining) framework to systematically apply data mining processes and techniques. Out of thirty-four x-attributes (also known as independent variables or input variables), 10 leading indicators were identified. It was found that relying on safety-related data alone cannot effectively predict safety risk in the construction. Of the 10 leading indicators identified, six of them are project-related (contract sum, manpower at project and organisation level, project progress, days delayed and magnitude of delay) and four of them are safety-related (scaffold, overhead protection, safe means of access and inspection score). Five popular machine learning techniques were used to train models for prediction of accident occurrence and severity. During testing, it was found that Random Forest (RF) provided the best prediction performance with an overall score of 0.83. Comparing with similar works found in the literature, this result is promising. It also suggest that the severity and occurrence of accidents do not happen by chance. It is recommended that such predictive DM model be deployed in construction organisations (e.g. LTA, HDB, large private developers and construction companies and industry associations) to improve safety by forecasting monthly safety outlook of construction worksites, so that pre-emptive inspections and interventions can be implemented. Key findings were presented to experienced safety practitioners, the Executive Director of the Workplace Safety and Health Institute (WSHI) and the Construction and Landscape Committee of the Workplace Safety and Health Council and there was consensus that this research could bring practical benefits to the construction industry. Future work to further validate and implement this study will potentially be funded by the WSHI. Although this research is focused on a single construction company, it serves as a strong foundation for future DM applications in construction safety in Singapore and globally.