An Optimized Machine Learning Approach to Classify Incidents in the Canadian Construction Industry

Learning from previous construction incidents is essential to avoid coming ones. Compared to other industries, the construction industry is argued to be the most hazardous. Construction incidents do not only negatively impact humans in workplaces but also may cause substantial financial losses. To this end, an adequate interpretation and prediction of construction incidents are vital. This study contributes to the body of knowledge by providing a holistic incident evaluation model that aims to automatically classify current incidents and predict future ones. A data set of 13,153 records of incidents reported by a construction company located in Alberta, Canada, is employed in this study. The study adopts text-mining techniques to represent word vectors (attributes) that machine learning (ML) models can utilize. Moreover, the study employs shallow ML models, such as decision tree (DT), K-nearest neighbors (KNN), Random Forest (RF), and Naive Bayes. It also adopts deep ML models, such as deep neural network (DNN). Further, the study discusses the attributes that can be used to predict the severity of possible incidents, i.e., weather conditions and work packages. Consequently, it utilizes the same ML models to predict the severity of possible future incidents based on correlation analysis of the identified attributes. All of the parameters of the ML models are optimized by the grid optimization technique. Further, the results reveal that the DNN model is the best-performing ML model. The study provides the participants with a tool to classify safety records automatically and theoretically discusses the potential of forecasting possible incidents.