Developing a model to predict unsafe behaviour of construction workers in Sri Lanka

Abstract

The construction industry is known to be one of the most accident-prone of work sectors around the globe. Although the construction output is less in Sri Lanka, compared to developed countries in general, the magnitude of the accident rate in the construction industry is significantly high. Most of the occupational accidents happen due to the unsafe behaviour of the workers. Along with this revelation, behaviour based safety has emerged as an effective approach to ensure occupational safety. The principal step of behaviour based safety approach involves the identification of the unsafe behaviour of the workers. The research, therefore, focused on investigating factors influencing construction workers’ unsafe behaviour and developing a model to predict unsafe behaviour originated from those factors. Quantitative research strategy was selected to carry out the study considering the nature of this investigation. The acts characterising the unsafe behaviour of construction workers, and the factors influencing those were identified through a literature survey. A pilot study was undertaken to validate and generalise the literature findings to the Sri Lankan construction industry. Fifteen unsafe acts those characterise the unsafe behaviour and fourteen factors those influence the unsafe behaviour were identified relevant to the local context. A survey approach was used to collect data. C1 grade building construction organisations were selected as the sampling framework. Twenty organisations were chosen within Colombo district to gather information from construction workers. The processed data were used to develop and train an Artificial Neural Network (ANN) predictive model that could predict unsafe behaviour of a construction worker with respect to a score. Backpropagation architecture using Neuroph Studio software was employed to develop the predictive model. 277 data points taken from the survey were used to train the network. The architecture of the trained model was demonstrated by conducting a sensitivity analysis. Mean Absolute Error was the technique used in this process. Sensitivity analysis showed that the model is highly sensitive to the neuron corresponding to “education”, while the lowest sensitivity was evident for the neuron corresponding to “employee involvement in safety”. The results suggests that educational level of a worker has the highest influence on his unsafe behaviour at work. Similarly, the co-workers’ involvement in safety on site has the lowest influence on unsafe behaviour of a worker. Furthermore, the predictive model was validated for generalisability using seven data points those were not used in training the network. The findings depict that the performance of the model is accurate due to high generalisation capabilities in the validation session. The model serve as a prototype tool to determine the unsafe behaviour level of construction workers and their safety training needs. This model can further be employed as a tool to proactively design interventions to avoid or minimise occupational accidents based on the unsafe behaviour levels of construction workers.