Study On Recognition Methods Of Tensile Shear Failure Of Red Sandstone Based On Artificial Intelligence

In this thesis,red sandstone is used as test material for direct shear test and Brazilian split test.The vertical force of direct shear test is 11 k N,the loading rate is500N/s,and the loading rate of Brazilian split test is 100N/s.Three groups of specimens are tested for each type to obtain acoustic emission signals of rock tensile failure and shear failure.Five AE parameters including rise time,peak frequency,counting,absolute energy and amplitude are selected to analyze the distribution of AE parameters in the whole test process.The failure type discrimination method based on acoustic emission parameters is used to screen the acoustic emission data obtained from the test,and 200 groups of acoustic emission data are selected for each specimen.Draw waveform diagram for selected acoustic emission data,and summarize the characteristics of acoustic emission signal waveform under different failure modes.Wavelet analysis is used to decompose the waveform of acoustic emission signal,reduce noise,nalyze time-frequency,alculate the sample entropy of signal waveform,and analyze the waveform characteristics of acoustic emission signal with different damage modes.Three convolutional neural networks VGG19,Res Net50 and Inception V3 are selected to input the filtered acoustic emission signal waveform,noise reduction waveform and time-frequency phase plane.The training results show that VGG19 convolutional neural network has the best training effect when the waveform is used as the input;Res Net50 convolution neural network has the best effect when the noise reduction waveform is used as the input;When the frequency phase plane is used as the input,VGG19 convolutional neural network has the best training effect.When VGG19 convolutional neural network is used as the training network,the waveform graph as the input model has the best training effect;When Res Net50 convolutional neural network is used as the training network,the noise reduction waveform as the input model training effect is the best;When Inception V3 convolution neural network is used,the model training effect with time frequency phase plane as input is the best.The statistical analysis of the five selected parameters,namely rise time,peak frequency,count,absolute energy and amplitude,shows that the five parameters are concentrated in a range.The rise time,peak frequency and count of rock shear failure are greater than the splitting failure,and the amplitude and absolute energy are less than the splitting failure.The five acoustic emission parameters are input into the decision tree for parameter identification,one for each group,two for each group,three for each group,four for each group,and five for each group.The results show that the combination of peak frequency,count,amplitude and rise time parameters has the highest recognition rate of 88.33%.The recognition effect of decision tree parameters is lower than that of neural network.However,the two methods have their own advantages.Convolutional neural network model has high recognition rate,but needs more time to train the model;Although the recognition rate of the decision tree model is not high,the recognition accuracy of specific parameters is also above 80%,and the training time required by the decision tree model is extremely short.