Classification of Acoustic Plastic Pipe Water Leak Signals with Gaussian Mixture Models

Project Type: Master/Diploma Thesis
Student: Priewald Robin


 Distributing water has become a standard in industrialised countries like the UK nowadays, but the pipes have to be maintained to reduce water losses through ever-present leaks. To optimise the operation of water supply grids occurring leaks must be detected as fast and reliably as possible. Acoustic leak detection in metal pipes is already common sense, but the newly introduced plastic pipes hitherto remain problematic due to the high acoustic attenuation which causes reductions in the operating range. The present study assesses leak detection in plastic pipes using acoustic signals which were recorded with special ground microphones. The available database comprises recordings from three different sites which were used to evaluate the detection range and to optimise the systems detection accuracy. A statistical classifier based on Gaussian mixture models (GMMs) was used. In exhaustive searches Line Spectral Frequency (LSF) yielded the best results which in a 4-fold cross-validation experiment using data separated by site and date achieved an estimated leak detection accuracy of p=97.2% (90.3%, 99.1%) at a significance level of 5%. This involved 71 tests of which each was based on a recording with a length of approx. one to ten minutes. The maximum detection range was found to be around 25 m, depending on the individual conditions which vary from site to site.