Diagnosis of horizontal pipe leaks using neural networks

Abstract

This document presents an experimental study that supports probabilistic decisions based on neural networks to detect the presence of leaks in pipeline transport systems, since such leaks can cause serious consequences. In addition to the economic losses presented by the lost product, process stoppage and repair of the damage, there can be insurmountable environmental and social losses such as the death of human beings. The probabilistic model correlates measurements of inlet and outlet pressures and flow to the state of leakage. The study and experimentation presented in this work are based on information acquired by simulating the behavior of the fluid in a pilot tube installed in the Fluid Mechanics laboratory of the Universidad Francisco de Paula Santander, Seccional Ocaña. Finally, experimental tests were carried out to obtain the data of the physical variables of the flow sensors at the entrance and exit, with these data a multilayer neural network of perception was trained. The results obtained from the test equipment, with intentionally caused leaks, showed that the structure of the multilayer Perceptron neural network was capable of detecting leaks in the pilot tubes. With the accomplishment of this work it was found that the neural network presents favorable results at the time of detecting the leak compared with other methods of detection since this one presents immunity to the noise, a parallel structure and, consequently, capacities of fast processing and classification.