The construction industry and industries that support construction (such as manufacturing and supply) are characterized by the high risk associated with projects. In 2020 alone, according to the Association of Worker’s Compensation Boards of Canada, the construction industry had the highest number of fatalities across the industries tracked; and trades, transport, and equipment operators and related occupations saw the highest total lost time claims in Canada at over seventy thousand. The construction industry is complex, hazardous, and dynamic in nature, which contributes to the frequency of safety incidents.
In response to the inherent risk present in construction projects, many companies have implemented a series of policies, plans, and procedures in a systematic approach commonly known as a Safety Management System (SMS). As part of their SMS, many companies maintain large databases of incident reports which have been collected over many years. According to Dr. Lianne Lefsrud, the principal investigator (PI) of the CIC project titled Enhancing Safety Management Systems on Construction Projects: A data-driven approach, “These reports often include some descriptive analysis, but don’t allow for an in-depth examination of trends or leading indicators. If incidents can be reported in greater detail, collected, and analyzed across the industry, then the information gathered will help build an understanding of how to prevent and mitigate risks.”
Dr. Lefsrud and Dr. Yasser Mohammed (Co-PI) and their teams of researchers seek to use artificial intelligence and machine learning (AI/ML) analysis to help identify trends and leading indicators, better design prevention and mitigation strategies, and leverage the information gathered from organizations across the Canadian construction industry.
Implementing AI/ML into companies’ legacy management systems has historically been a slow process for several reasons. Firstly, operators tend to only analyze incidents with severe consequences with the aim of preventing their recurrence, while minor incidents are usually logged without being evaluated in depth. The issue with this is that high-frequency, low-consequence incidents are often the precursors of more significant incidents. Secondly, the records of these low-impact incidents are often incomplete, scattered across many sources of data, or retained as private records that are not publicized or shared. It can also be difficult to ascertain ownership of such records, making them even more difficult to properly circulate and maintain. Finally, operators who incorporate AI/ML services into their business model (whether as part of an internal department or as an outsourced service) are often surprised to learn that AI/ML researchers and suppliers require large datasets in order to train their algorithms, which results in operators seeing AI/ML as a high-effort, low-payoff venture because of the amount of time and investment of information that needs to happen before seeing meaningful results.
Through this project’s collaboration with industry partners, CIC researchers are working to create an easier way to share incident and near-miss reports to facilitate easier identification of leading indicators that can help predict high-consequence incidents, identify the variables that are the most insightful and determine what information should be prioritized for data governance and data quality initiatives, and embed graduate students in the operations that occur to facilitate the students’ learning from leading operators while also developing AI/ML skills to support long-term capabilities and learning.
“Technological, organizational, and human factors are often interacting in complex, multi-directional, and uncertain ways, which can be difficult to disentangle in the aggregate,” Dr. Lefsrud told us. “AI/ML methods—like Bayesian network analysis—can display both horizontal and vertical dependencies, data and knowledge uncertainty, and practical applications. The findings are displayed in such a way that the variables and their relationships are intuitively understandable.
“The risk management and AI/ML methods that we are developing are context-independent. While these methods originate from the petrochemical and energy industries, they can be pivoted to construction, transportation, mining, and other industries to glean new insights.”
According to Dr. Lefsrud, the methods developed in this research can be used in reliability evaluation, quantitative risk assessment, diagnosis, prediction, automated reasoning, and developing a reaction network or model when there is a large dataset with limited prior knowledge of the system. The resulting framework is useful for ensuring a reduced incident rate, increased productivity, and enhanced project performance. It also enhances companies’ environmental, social, and governance (ESG) performance.
Looking for more details? Visit this project’s research page.