Research On Pipeline Leak Detection Based On Underwater Acoustic Signal Analysis

With the progress of industry and the development of population,the demand for water resources increases year after year.At the same time,the imbalance between supply and demand of water resources is emerging,partly because of the waste of water.For example,the underground water supply network in many cities has serious leakage problems due to aging,corrosion,geological settlement and other reasons.Underground pipeline network is the only way for urban water supply,and the leakage of water supply pipeline will cause serious consequences.In addition to the waste of resources and economic losses,it will even cause some secondary disasters,threatening the safety of drinking water.Therefore,the research of pipeline leak detection technology and development of related equipment have become hot topics.Pipeline detection equipment based on underwater acoustic signals has been widely used.The detection device collects underwater acoustic signals for analysis through sensors.Noise processing in complex environments is the primary problem of this method.Donoho wavelet threshold denoising method is a classic filtering method used to remove the Gaussian noise of the signal,but the main noise present in the underwater acoustic signal is non-stationary,for this reason,the threshold filtering method based on the modulus maximum value is adapted according to the characteristics of the leaked signal,retaining coefficients of weak signal and eliminating coefficients of high amplitude noise,which makes the wavelet denoising method achieve better RMSE indicators and improve Filter effect.In order to further improve the working performance of the detection equipment,a detection model was designed based on the SVM and CNN methods to realize automatic leakage identification.Firstly,a hardware platform with Zynq as the core is built,based on which stable underwater acoustic data is obtained,and then empirical mode decomposition(EMD)and Fourier transform(FFT)are used to extract the IMF energy of the signal and frequency spectrum to form different feature data sets for training and analysis.The results show that the CNN-based detection model can use its deep convolutional network to establish the implicit connection between the leaked signals,and it has stronger recognition ability than the SVM method in the experiment.After testing,the CNN model introduced with FFT frequency domain data for training has the best effect.The leak detection system deployed with this model can achieve good performance indicators and high reliability.