Research On Feature Extraction And Intelligent Recognition Method Of Pipeline Acoustic Signal

As one of the important fluid medium transportation tools,pipelines play a very important role in national defense industry and urban construction.In order to ensure the safe operation of the pipeline and effectively reduce the risk of pipeline failure,it is very necessary to conduct rapid and accurate state inspection of the pipeline on a regular basis.At present,among many pipeline state detection methods,the detection method based on active acoustics has obvious advantages for the state detection of buried pipelines.This thesis aims at detecting the state of buried pipelines,collecting pipeline acoustic signals under different working conditions based on the active detection method of pipeline acoustics,and combining machine learning and deep learning methods to perform feature extraction and pattern recognition on the collected pipeline acoustic signals.The purpose is to solve the problem of status identification of buried pipelines.The main research contents of the thesis are as follows:(1)Starting from the theory of active detection based on pipeline acoustics,systematically expound the theories involved in the detection method,demonstrate the feasibility of the method in pipeline detection,and provide theoretical support for the core research of this page.At the same time,the pipeline acoustic wave detection experimental platform based on acoustic active detection conducts preliminary time-domain analysis and frequencydomain analysis of pipeline acoustic signals under different working conditions.The research results indicate the necessity of in-depth subsequent analysis and processing of pipeline acoustic signals.(2)Facing the key point of pipeline state feature extraction based on acoustic feature analysis,a method of pipeline feature extraction and state recognition based on PCA-BP neural network model is proposed.Starting from the multi-condition pipeline acoustic signal acquisition based on the acoustic active detection method,the signal framing method is used to frame the pipeline acoustic signal and extract the sound pressure level of each frame signal as a feature vector.Then,the feature vector is optimized based on the principal component analysis method,and the feature parameters that can best characterize the pipeline state are screened out.Finally,the pipeline state recognition is realized base on the BP neural network.The results show that the use of the principal component analysis method to optimize the acoustic features extracted from different signal frames—the sound pressure level further improves the identification effect of the BP neural network classifier,and also illustrates the effectiveness of the feature selection results.(3)Aiming at the problem that the degree of pipeline blockage is difficult to distinguish under different conditions,a method for identifying characteristics of pipeline acoustic signal based on the ICEEMDAN-SVM model is proposed.Firstly,the acoustic active detection method is used to collect the sound signals of the pipeline under different blocking conditions,and the sound signal of the pipeline is decomposed based on the improved CEEMDAN method to obtain the IMF component series.Then,according to the correlation coefficient and the variance contribution rate,the effective components that characterize the original signal are screened out,and the information entropy of the selected components(including time domain energy entropy,power spectrum entropy and singular spectrum entropy)is calculated as the feature vector set.Finally,based on the SVM classification model,the blockage state pattern recognition is realized.The research results show the effectiveness of this method in identifying different pipeline states and different degrees of blockage.(4)Aiming at the disadvantages of traditional pipeline fault pattern recognition methods such as relying on manual selection of features and complicated and changeable diagnosis process,a pipeline state acoustic signal intelligent recognition method based on 1D CNN-LSTM hybrid neural network model is proposed.A 1D CNN-LSTM model is established based on onedimensional convolutional neural network and long-short-term memory network theory,and the relevant parameters of the model are optimized.Finally,the collected pipeline acoustic signals from multiple operating conditions(normal,blocked,leaking)are standardized and then input into the 1D CNNLSTM model to adaptively complete feature extraction,dimensionality reduction and classification.Experimental research results show that this method combines the advantages of one-dimensional convolutional neural network and long-and short-term memory network,not only can effectively reduce the uncertainty caused by artificial feature extraction,but also realizes the intelligent recognition of pipeline acoustic signal features in different states.The accuracy rate has reached more than 99%.This paper has carried out research on the signal feature extraction and pattern recognition methods of pipeline state,enriched the theoretical methods and means of pipeline fault diagnosis,and has good theoretical and engineering value for the state detection and pattern recognition of buried pipelines.