Study On Characteristics Extraction And Damage Pattern Recognition Of Acoustic Emission Signals Of Refractory Based On Multi-fractal Theory

The composition of refractory is complicated, and it belongs to microheterogeneous material that porous and multiphase. The acoustic emission signals thatproduced during the damage process contains a lot of information about the source. Thekey of achieve the identification of damage pattern is to extract the damagecharacteristics of refractory effectively and select appropriate classifier. As therefractory acoustic emission signal has multi-fractal, non-line, non-stationarycharacteristics, used the method of combine multi-fractal theory and empirical modedecomposition to extract the characteristics of signals, and achieve the identification ofdamage pattern by classification methods of support vector machine and BP neural. Ithas positive significance for the study of microscopic damage of refractory. The mainresearch contents of this paper are as follows:(1) In this paper, magnesium carbon refractory for the study. And simulate the itsdamage situation under the compressive stress state by uniaxial compression, thencollected and analyzed the damage acoustic emission signals of the damage process.According to the relationship between the signal frequency of damaged components ofcomposite and its elastic modulus, density dependent, can selected the typical signals.(2) In order to selected the best damage characteristics from the multi-fractalparameters (Δα, Δf, K, MeanDq), a series of simulation acoustic emission signals withdifferent frequency structures was built according to the feature of refractory AE signal.Then selected the best characteristic from the simulation signals analysis. Finally, verifythe result by experimental signals. The analysis results show that the multi-fractalspectrum width a value Δα can characterize the acoustic emission signal characteristicswell, which best for damage characteristics.(3) As the acoustic emission signals was non-linear and non-stationary,decomposing the signal into several IMF components by the EMD method, then tookthe multi-fractal spectrum parameters of entire signal and the multi-fractal spectrumparameters of the IMF components as input feature vector of the classifier. Then usedthe SVM and BP neural network to classify the damage signal respectively, theclassification accuracy rate of two methods both reached above90%, which confirms the rationality of using the method based on EMD and multi-fractal spectrumparameters to extract damage feature of experimental signal.(4) According to analysis the pattern classification results with different trainingsamples showed that SVM can achieve higher classification accuracy with smallersample, which has more advantages than the BP neural network method.