A Neural Network System for Fault Prediction in Pipelines by Acoustic Emission Techniques

ABSTRACT The problem of evaluating the risk of failure associated with the propagation of a crack in a pipe under pressure has great practical relevance, and it may be tackled with acoustic-emission techniques. Artificial neural networks may be trained to classify the acoustic emissions generated by the crack according to the phase of propagation, and such a classification permits to evaluate the risk of mantaining a system in operation. In order to train the network, a human specialist has to estimate the transition times between any two consecutive phases by inspecting the results of a previous hydrostatic test, and such determination of the transition times has a high degree of subjectivity and uncertainty, affecting the classification performance of the network. In this paper, we propose a human-independent method for the estimation of the transition times, and we show successful applications to the data from two hydrostatic tests. For a test on a 2 m-long pipe, the method exhibited 98% of correct-classification rate, an improvement of 8% over results obtained with human-determined transition times. For a 40 m-long pipe, under experimental conditions comparable to those found in industrial applications, the method exhibited 91% of correct-classification rate. The proposed method provides a fully automated framework for the evaluation of the state of a crack.