Research On Metal Sheet Damage Identification Based On Modal Decomposition And Machine Learning

Metal thin plates are widely used in the fields of infrastructure construction and aerospace industry.These thin plates must undergo strict damage detection before and during use.The complex on-site environment and large-scale detection range are more suitable for the technical means such as ultrasonic Lamb wave,which is simple and has strong flaw detection capability.The high-frequency Lamb wave has multi-modal characteristics,which will seriously interfere with the test output in actual testing,and the characteristic parameters cannot be effectively identified.In this paper,for the complicated environment of the thin plate detection,the direct sound injection method is used to transmit the ultrasonic waves excited by the stacked piezoelectric ceramics to the detection thin plate,and the excitation frequency of the Lamb wave is reduced to reduce the modal interference from the Lamb wave itself.In this paper,the principle of ultrasonic Lamb wave detection is described,and the theoretical dispersion curve of 1.5mm thick metal plate is drawn by MATLAB program.The experimental platform was built to collect 90 sample data under four types of no damage,pass defect,groove defect and solder joint.At the same time,the four types of waveform coding were distinguished and the relationship between input and output was established.Due to the complexity of the ultrasonic Lamb wave body and the interference of the collection environment and the reflection wave of the metal plate,it is necessary to effectively reduce the noise of the signal preprocessing.Among them,the traditional empirical mode decomposition,the improved full noise assisted aggregate empirical mode decomposition and the variational mode decomposition have certain effect.Through the Hilbert envelope spectrum,the advantages of the three methods are compared.Icemdan method is selected in this paper as a pretreatment method.Then,the time-domain feature parameters of each IMF component are extracted,and the feature information of FFT spectrum and Hilbert marginal spectrum is extracted to construct the feature vector for the subsequent defect classification.Finally,on this basis,two models of BP neural network and limit learning machine are established.48 sets of training sets obtained in this paper are input into the model,and the remaining 24 sets of test sets are used to verify the effectiveness of these two models for the feature parameter extraction method in this paper.The feature parameters extracted in this paper can be effectively identified and matched in the two models of BP neural network and extreme learning machine.The judgment of the defect type and the accuracy of recognition are at high values.The remaining 18 sets of data sets are used in the design experiment Comparing the recognition capabilities of the two models,the results show that the extreme learning machine is superior to the BP neural network in identifying the accuracy rate and resource consumption rate of the four defects,and is more suitable for engineering applications.