Modelling serviceability deterioration of concrete stormwater pipes using genetic algorithm trained neural network

The deterioration and aging of stormwater pipe systems in Australia has become a major concern in recent years. The inspection and evaluation of pipe condition for the whole systems is not economically feasible due to the conceivable system size and inherent difficulties of buried pipes. One frequently used strategy is to develop a deterioration model that can predict the change of pipe condition over time for individual pipes based on pipe attributes. Furthermore, the significant factors influencing the deteriorating process can be identified through model analysis. Based on the outcomes of the deterioration model, a proactive maintenance and rehabilitation can be operated efficiently. This paper used a feed forward neural networks (NN) approach to model serviceability deterioration of concrete stormwater pipes, which make up the bulk of the stormwater network in Australia. System condition data was collected using CCTV images. For efficient and quick training over conventional back-propagation training algorithms, the weights for NN model were obtained by using a genetic algorithm (GA). The proposed approach is tested and compared with multiple discriminant analysis, a traditionally statistical method. The robustness and generalization capability of the NN model was demonstrated. They were further enhanced by the use of GA. However, further improvements in data collection and condition grading schemes should be carried out to increase the predictive performance of the NN model.