Leak Detection in Water Distribution Pipes Based on CNN with Mel Frequency Cepstral Coefficients

Abstract

The acoustic signals collected by a listening device on pipelines are often utilized for detecting leakage in buried water pipelines. In practice, the operator carries with a listening device to collect the acoustic signal on the road along the direction of pipelines and justifies the acoustic signal whether it is the leakage or not. Based on operator's experiences and skills, such a detection method has successfully applied for detecting leak in the absence of a fixed non-leak acoustic source occurring in or outside the detected pipelines. However, the leak signals are always inevitably corrupted with non-leak acoustic sources. It affects not only the accuracy of leakage detection but also the time for locating the leakage. In this paper, an acoustic leak detection approach based on convolutional neural network with Mel frequency cepstrum coefficients for underground water pipelines is proposed. Mel frequency cepstrum coefficients of acoustic signals are first extracted as features. Then, a convolutional neural network model has been established as a classifier for detecting whether the pipes are leaky or not, which uses Mel frequency cepstral coefficients as the network inputs. As the result, the convolutional neural network model accuracy of classification is above 98%. The proposed method can provide highly detection efficiency than that of the traditional approach which depends on operator's experiences. The proposed leak detection method based on convolutional neural network can be implemented in the listing devices for real-time leak detection in water pipes.