Research On Feature Extraction And Recognition Of Metal Defects Based On Ultrasonic Testing

Ultrasonic testing is a widely used method in nondestructive testing,and it is also a common method for the detection of defects in metallic materials.The feature extraction and recognition method of the defect signal plays an important role in the qualitative identification of the internal defects of the material.Feature extraction ensures the number and resolution of features after reduction,and the recognition method ensures the accuracy of signal classification.The paper mainly focuses on the feature extraction and recognition of defect signals of coal mining machinery,it is of great significance to prevent accidents and reduce accidents in coal mining.This paper mainly focuses on the detection of casting defects in the shearer,from the feature extraction of defect signals,attribute reduction and recognition of the defect signal.The main research contents include the following aspects:(1)Research on wavelet de-noising and wavelet packet energy feature extraction of experimental data.Experiments were carried out using artificial simulated defect signals collected by an ultrasonic flaw detector,the wavelet threshold method is used to de-noise the signal and compare the de-noising results.This paper analyzes the feasibility of using defect signal energy as feature information,and studies the energy feature extraction using wavelet packet.Experiments show that the energy information of the defect signal has good resolution to distinguish different defects.(2)In order to solve the problem of the high dimensionality of the wavelet packet energy feature extraction,the neighborhood rough set algorithm is used to reduce the redundant attributes.By studying the importance of neighborhood value,the idea of non-uniform neighborhood value is put forward for the problem that the uniform neighborhood method is not ideal,and the calculation method of non-uniform neighborhood and dependency is given,as well as the process of attribute reduction algorithm.The method of neighborhood value granulation with non-uniform standard deviation and non-uniform slope is realized.The experimental results show that the two methods retain the most resolute features while reducing the dimension,and the method of non-uniform slope is better.(3)Research the defect signal intelligent recognition algorithm,the BP neural network and the support vector machine are used to classify and identify the defect signals.Genetic algorithm is used to optimize the training parameters of support vector machine.Using different methods to classify the reduced characteristic attributes.The experimental results show that the support vector machine is more accurate than BP neural network in the case of small sample.According to the classification accuracy of non-uniform slope,BP neural network reached 79%,support vector machine reached 93%,the latter effect is obviously better.At the same time,it shows that the method of non-uniform neighborhood value is superior to the uniform neighborhood value,and verifies the feasibility of the non-uniform neighborhood rough set.