Neural network identification of water pipe blockage from smart embedded passive acoustic measurements

This study presents a new neural network approach to identify the presence and type of obstruction in pipes from measurements of passive acoustic emissions. Inserts were used in a fluid re-circulation loop to simulate different types of blockage at various flow rates within the turbulent regime, generating patterns of acoustic emissions. The data were pre-processed using Fourier analysis, and two candidate sets of statistical descriptors were extracted for each measurement. The first set used average and spread of the Fourier transform amplitudes, the second used data binning to obtain a concise representation of the spectrum of amplitudes. Experimental evidence showed the second set of descriptors was the most suitable to train the neural network to recognize with accuracy the presence and type of blockage. The obtained results compare favourably with the literature, indicating that the approach provides a tool to enhance process monitoring in water supply systems, in particular early detection of upstream blockages.