Research On Statistical Characteristics Of Water Distribution System Leakage Signals And Intelligent Identification Technology

At present,the leakage detection of the water distribution system mainly relies on artificial listening.This method is highly dependent on the experience of the leak detection workers,and it takes several years to train a qualified leak detection personnel.In addition,artificial listening needs a very quiet environment,and the workers often work in the dead of night.Long-term day-night reversal work will cause great harm to the staff.In order to solve the technical difficulty of leak detection in the water distribution system effectively,the theoretical and technical research on the characteristics of the leaked acoustic emission signals is carried out in this paper.In this paper,tens of thousands of leak detection signals have been collected,and more than 3,000 signals have been marked manually.A variety of spectrum analysis methods are used to study the signals,the individuality and common law of the spectral characteristics of the leaked acoustic signals are summarized,and some statistical models for the spectrum of the leaked acoustic signals are established.The noise preprocessing method strengthens the spectral statistical characteristics;According to the spectral characteristics of the leaked signals,some leakage identification models based on machine learning are established,which can realize large-scale automatic leakage identification.The main conclusions are as follows:(1)The leaked acoustic emission signal is called "noise" from the sense of hearing,but it is a random signal with "weak periodicity",which is obviously different from white noise in the general sense.There are seldom the same two leaked signals because of the randomness,so the idea of probability and statistics is used in the study of this paper.9 basic characteristics are analyzed,for example marginal spectrum,spectral centroid,MFCC feature,spectral flatness,spectral centroid bandwidth,spectral drop rate,time spectrum and higher-order statistics.The results show that the dominant frequency center,spectral flatness,and spectral drop rate of the leaked signals have significant statistical distribution characteristics,and it can be used to distinguish whether there is a leak or not.MFCC characteristics,time spectrum and high-order spectrum are also significant,which can be used alone or in combination to judge the leak.(2)The leakage signals of the water distribution system often contain uncertain interference,and the noise reduction preprocess helps to strengthen the features of the leakage signals.The leaked acoustic emission signal is a random signal with "weak periodicity".The purpose to strengthen the statistical characteristics of the signals is aimed at.Based on wavelet threshold noise reduction,EEMD noise reduction and singular value decomposition noise reduction,an adaptive joint noise reduction method is proposed for water distribution system leakage signals.This method can effectively improve the statistical distribution concentration of leaky signal features,and helps to strengthen the discrimination between leaky signal and non-leakage signal spectral features.(3)Based on feature engineering and noise reduction.The decision tree model,SVM model,BP neural network,RBF neural network and ensemble learning models are used for training,and the classification effects of each model are ranked according to comprehensive assessment system including accuracy rate F1 value,recall rate,negative positive rate,of which the negative positive rate is the economic decision indicator.The BP neural network model and the Boosting integrated learning algorithm can achieve the best classification effect,reaching a recall rate close to 100%,a negative positive rate of 7.35%,and a F1 value above 0.95,which can meet the accuracy in practical projects.In practice,the number of non-leakage signals is much more than that of leaked signals,which has a great impact on the model effect.By adjusting the proportion of leakage and non-leakage signals in training samples,it proposed that the proportion of leakage and non-leakage samples used for training should be within the range of [1,4/3].