Machine learning predictions for lost time injuries in power transmission and distribution projects

Although advanced machine learning algorithms are predominantly used for predicting outcomes in many fields, their utilisation in predicting incident outcome in construction safety is still relatively new. This study harnesses Big Data with Deep Learning to develop a robust safety management system by analysing unstructured incident datasets consisting of 168,574 data points from power transmission and distribution projects delivered across the UK from 2004 to 2016. This study compared Deep Learning performance with popular machine learning algorithms (support vector machine, random forests, multivariate adaptive regression splines, generalised linear model, and their ensembles) concerning lost time injury and risk assessment in power utility projects. Deep Learning gave the best prediction for safety outcomes with high skills (AUC = 0.95, R2 = 0.88, and multi-class ROC = 0.93), thus outperforming the other algorithms. The results from this study also highlight the significance of quantitative analysis of empirical data in safety science and contribute to an enhanced understanding of injury patterns using predictive analytics in conjunction with safety experts’ perspectives. Additionally, the results will enhance the skills of safety managers in the power utility domain to advance safety intervention efforts. •Deep learning predictive techniques for injury detection.•Addressing the count variable problem using deep learning.•Performance comparisons of deep learning models with conventional machine learning techniques.•Interpreting deep learning models using local and global interpretation techniques.