In a recent visit to the Centre for Construction Work Health and Safety Research, Professor Matthew Hallowell of the University of Colorado presented research aimed at predicting injuries.
Over the past four years, Hallowell, a Beavers Endowed Professor of Construction Engineering, and his research team have used text-based injury reports to make reliable predictions of future injuries.
The reports were converted from unstructured text into structured quantitative data using natural language processing.
The system that Hallowell and his team developed can automatically detect causal factors from injury reports at a success rate of 97 per cent.
Using this system, the team extracted causal information from over 10,000 injury reports, resulting in a high-dimension dataset well suited for predictive analytics such as machine learning.
The team then used diagnostic statistics to identify unusual and sometimes unexpected clashes among certain attributes, which could not be identified through intuition and individual experience alone. For example, the team discovered that the combination of electricity and cold weather caused high rates of electrocution. This is because more clothing retains perspiration, which increases conductivity and the probability of arc flashes.
Machine learning was also used to create predictive models able to anticipate injury severity, type of injury and the body part injured, with a high degree of reliability.
The research suggests that it is now possible to predict the potential for injuries by using any new observation of construction attributes, such as the design of the facility, type of work performed, the characteristics of the environment and adjacent tasks, and the decisions of workers.