Estimation of the severity of occupational accidents in the building process with pre-informed artificial learning method

In this study, the relationship between accident severity and safety measures in occupational accidents occurring in the building construction process was investigated. An integrated model had been developed that can predict what precautions should be taken in future accidents and what the result might be if these precautions are not taken, with using historical accident data. This prediction model was developed by integrating AHP (Analytical Hierarchy Process) and ANN (Artificial Neural Networks) methods, which are frequently used by researchers today, to complement each other at the point where they are weak. The significance of this pre-informed artificial neural network model was tested with real data by conducting a field study. For the sample, four types of occupational accident that resulted in the most deaths in the national and international literature were selected and thirty-five past accident data were collected for each type in the same construction company. The binary comparison data for the AHP method, which defines the input weights of the ANN, were obtained by a professional survey company from the occupational health and safety experts working in the sector, by survey method. 120 of historical accident data were used for training networks and 20 for testing. For pre-informed neural networks, two alternative network configurations were developed using two different activation functions and a comparison was made with a neural network with the same configuration without pre-informing. As a result, it was seen that the pre- informing stage increased the learning rate of artificial neural networks by 5% in the training data set and 15% in the test data set. In addition, the alternative with parabolic activation function in pre-informed artificial neural networks showed a high learning performance, and it was observed that the relationship between risk reduction measures and accident severity was 90% significant, provided that it was limited to the collected accident data.