Neural network approach for failure rate prediction

• The aim of this work is predict failure rate using artificial neural network.
• The correlation between experimental and predicted data was acceptable.
• In the problem of failure rate prediction BFGS algorithm was used to learn the network.

The aim of this paper was to present the possibility of artificial neural networks application to the failure rate modeling. Operating data from one Polish water utility were used to forecast output value of failure frequency. The prediction results indicate that artificial networks may be used to model the damages frequency in the water supply systems. It was found that the artificial neural network (multilayer perceptron) trained by quasi-Newton approach gave acceptable, from engineering point of view, convergence. The network was learnt using 173 and 147 data (house connections and distribution pipes, respectively). 50% of all data was chosen for training, 25% for testing and 25% for validation. In prognosis phase, the best created network used 100% of 133 and 114 values for testing. The correlation between experimental and predicted data (relating to house connections and distribution pipes, respectively) was characterized by indicator R2 = 0.9510 and R2 = 0.9268 (learning phase). Worse results were obtained in prognosis phase. In this step of modeling once created network predicted failure rate using not known input signals. The coefficient R2 was equal to 0.4142 for house connections. For the distribution pipes the significant relation between experimental and modeled data was not found. The created model could be used by water utility in the future to establish the level of failure frequency and to plan the renovation of the most deteriorated pipes.