Study On Damage Identification Of Acoustic Emission Signals Of Prestressed Concrete Beams Based On Multi-channel CNN

Reinforced concrete structure is one of the most widely used structural forms in the current construction field.During its long service process,any small original defects may accumulate and develop into fatal structural damage,which indirectly contributed to nondestructive testing in recent decades The rapid development of technology in the world.Unlike most non-destructive testing technologies that emphasize the technical focus of result testing,acoustic emission technology emphasizes process detection.It has a stronger technical adaptability to the dynamic control of the entire process of structural damage evolution,and acoustic emission technology has gradually developed.Become an important technical branch in the field of non-destructive monitoring.Under the background of the era of big data,the empirical human identification has been unable to clarify the complex logic relationship between the massive structural state information,and the intelligent judgment level of structural damage has become an important indicator to examine the merits and demerits of non-destructive testing technology,and acoustic emission technology is no exception.From the point of view of the data source object,the acoustic emission takes the signal time history information or the characteristic parameters extracted from the time history information as the analysis object.After the clock accuracy of the acoustic emission instrument is improved to the nanosecond level,the big data characteristics of its data source become Especially prominent.This has greatly increased the need for intelligent judgment of acoustic emission technology,which is reflected in all technical details from signal denoising,feature extraction,damage identification,and safety warning.Convolutional neural network(CNN)is a type of feedforward neural network that includes convolution calculation and has a deep structure.CNN and its variant architecture,as the winning algorithm in the "Computer Information Feature Recognition Competition Held Worldwide in Recent Years",are better than other neural networks in terms of noise immunity,input diversification,network training difficulty,recognition timeliness and accuracy.There is a big improvement.The IMF-CNN proposed in this paper is a CNN variant algorithm specifically for acoustic emission signals.It emphasizes comprehensive analysis based on multi-channel input.The input IMF component is obtained by empirical mode decomposition(EMD).The analysis object can be the acoustic emission time history The IMF component can also be the acoustic emission characteristic parameter of the IMF component.In the process of IMF-CNN neural network analysis,three important sub-tasks need to be completed to realize the damage signal identification of concrete components.One is to collect the acoustic emission signals of component damage through the destruction test,and establish a reference classification database of component damage based on time series collation,test phenomenon inference,and data singular point recognition,that is,the second chapter of this article;second,use EMD decomposition to obtain acoustic emission signals The IMF component of the time history,because this step is located at the input layer of the neural network,the signal noise processing method and the process of denoising and decomposition should be considered while obtaining the IMF component,that is,the content of Chapter 3 of this article;the third is based on the "component The IMF-CNN network training,recognition efficiency and system development of the "Damaged Benchmark Classification Database" are the fourth chapter of this article.In the second chapter,the failure test is set as three-point bending test of prestressed reinforced concrete beams for two reasons.First,the bending failure process of reinforced concrete beams has obvious time sequence characteristics(including micro-crack initiation stage,crack forming stage,crack development stage and crack penetration stage),which is conducive to the establishment of benchmark classification database and the reference test of imf-cnn neural network analysis.Secondly,the addition of prestressed steel bars can effectively avoid the early cracking of components,and then distinguish the microcrack initiation stage from the crack forming stage from the data source.When the datum classification database is established,its time series boundary point can be obtained by combining the characteristic parameter analysis of acoustic emission signal and Kurtosis value analysis.In the third chapter,starting from the principle of EMD decomposition and PCA de-noising,the combined process of EMD decomposition and PCA denoising is discussed,and three algorithms of EMD-PCA,IMF-PCA and PCA-EMD are proposed.Secondly,the differences of the three algorithms are compared by the simulated acoustic emission signal.Finally,PCA-EMD algorithm was selected based on the comparison of the advantages and disadvantages of the algorithm,the acoustic emission signal of the test beam damage was denoised and decomposed,and the sensitive IMF component was extracted.In the fourth chapter,CNN multi-channel input data is generated by decomposing the extracted sensitive IMF components based on PCA-EMD algorithm.The neural network training corresponding to the benchmark classification database is completed,and the optimal network architecture is selected through the optimization and improvement of training parameters and structural parameters,so as to realize the damage phase or damage pattern recognition of prestressed reinforced concrete beams.Based on the Matlab language,a set of simple and easy acoustic emission damage pattern recognition system is developed,and the automatic and intelligent recognition for this problem is preliminarily realized.