STAR methodology-based learning about construction accidents and their prevention

Abstract

This paper presents the results of a feasibility study concerning the application of STAR methodology-based machine learning to construction accidents and their prevention. A ten-stage knowledge acquisition process is presented and its individual stages described. Knowledge about construction accidents was acquired using a collection of 225 examples, based on actual accidents records. Inductive learning with a system based on the STAR methodology was employed. This system was used in both the generalization and specialization modes of operation. The decision rules obtained are complex, but their interpretation is clear and they seem to be consistent with the present understanding of causal relationships between accident results and various factors affecting them. Also, the rules were verified using average overall and omission empirical error rates, which were calculated as average for three randomly determined sequences of examples. These error rates were calculated for all seven steps in the machine learning process, and were used to construct learning curves for both error rates. The relationships between error rates and the number of examples used for learning are analyzed, and coefficients of linear regression given and discussed. The 225 examples used were found to be grossly insufficient to produce reliable knowledge about accidents and therefore a large study is postulated which would involve the collection of a larger number of construction accident records. In general, our study demonstrated the feasibility of machine learning in acquiring knowledge about construction accidents.