Approximate entropy-based leak detection using artificial neural network in water distribution pipelines

Correlation techniques are widely used to locate leaks in buried water pipes. However, a distinct peak in the cross-correlation of two spatially separately collected acoustic signals may result from a non-leak acoustic source outside the pipe. And the peak not related to a real leak will result in a false leak location. So it is necessary to determine whether or not a real leak exists beforehand. In this paper, a new leak detection method using approximate entropy is proposed to discriminate the leak acoustic signals from the non-leak signals. In this method, the autocorrelation function values for the delay τ larger than the correlation length of the signal, not the signal itself or its entire autocorrelation function values, are used to extract or evaluate the self-similarity degree of the signal by the approximate entropy. A neural-network approach has been developed as a classifier, which uses the identified self-similarity degrees as the network inputs. The proposed leak detection method has been employed to identify the leak in the buried water pipelines, and achieved a 92.5% correct detection rate.