Research On Identification Of Quiet Period Of Yellow Sandstone Acoustic Emission Based On Machine Learning

With the deterioration of coal mining conditions,the advent of mine disasters is becoming more and more sudden.Traditional rock mechanics methods such as mine pressure testing have certain limitations during testing.With the popularization of acoustic emission(AE)systems,To detect the internal fracture of the rock so as to prevent coal and rock dynamic disasters has become the mainstream.Acoustic emission is caused by rock breaking.The quiet period of acoustic emission is a special stage in the phenomenon of acoustic emission.The quiet period of acoustic emission is before the rock fracture stage,which is unique.Therefore,the study of the quiet period of acoustic emission can predict rocks.? rupture.In order to explore the precursor information of rock failure and improve the accuracy of the identification of the quiet period,this paper takes the uniaxial compression experiment of yellow sandstone as an example,based on machine learning to identify the rock failure stage,and analyze the results to explore the influence of the acoustic emission quiet period The following work is mainly done for the parameters.Select and clean the acoustic emission time-domain data and frequency-domain data.The time-domain data is the time-domain characteristic parameters of acoustic emission during the rock fracture process,and the frequency-domain data is the waveform data of each impact point at a sampling rate of 1Mhz.Time domain data is subjected to various pre-processing methods such as data de-averaging and box graph processing to achieve the purpose of cleaning abnormal data and dimensionality reduction;frequency domain data is subjected to wavelet transform to denoise,and the denoised data uses wavelet In packet decomposition,the energy ratio is calculated according to the respective frequency bands,and a group of 8 wavelet packet energies are used as its frequency domain characteristic parameters;the time domain data and frequency domain data are combined as the feature vector of machine learning.The initial loading stage,linear elastic stage,elastoplastic stage,and peak failure stage are the four major stages of rock fracture.The latter stage of the elastoplastic stage is the quiet period of acoustic emission,which is divided into the fifth stage,consisting of five stages.The target value of machine learning.Based on acoustic emission time-domain data,frequency-domain data,and composite data(time-frequency domain data set),four machine learning models are combined to evaluate the accuracy of the experimental data.Then use grid search to traverse the parameters to find the optimal parameters of each of the four machine learning methods of k-nearest neighbor,random forest,gradient boosting regression tree,and kernel support vector machine.And analyze its precision rate,recall rate,f1 value and other parameters to evaluate its ability to recognize the quiet period of acoustic emission.Analyze the factors that affect the quiet period of acoustic emission according to the model learning results,combine parameter correlation,acoustic emission definition classification,and machine learning contribution rate to comprehensively select the most effective parameters for the acoustic transmitter learning model,and establish a new model after selecting the parameters Compare with the old model in terms of computing time and data volume.The machine learning model that is most effective for the acoustic emission quiet period obtained by the accuracy rate evaluation is used as a template,and the acoustic emission quiet period recognition system is designed and developed based on this template.The main research conclusions are as follows:(1)The machine learning model can be directly applied to a single time domain data set or frequency domain data set,but it can also be used in combination.By comparing the accuracy of the machine learning model,it can be concluded that the acoustic emission data set is Adapting to the four machine learning models of k-nearest neighbor,random forest,gradient boosting regression tree,and kernel support vector machine,the accuracy of different methods and data sets is between0.690 and 0.91,and the composite(time-frequency domain)data set is better than time The data set in the frequency domain is better than the data set in the frequency domain.The accuracy of the frequency domain data set is poor overall,the distribution is 0.7-0.82,and the average value of the data set is 0.78.The better one is the time domain data set,the distribution is 0.81-0.88,and the average value of the data set is 0.84.The best is the composite(time-frequency domain)data set.The gap between the machine learning models in this data set is small,and the average data accuracy rate is 0.9.From the point of view of machine learning methods,the kernel support vector machine performs best,with an average accuracy of 0.87,followed by a gradient boosting regression tree,with an average accuracy of 0.84,followed by k nearest neighbors,with an accuracy of 0.83,and finally Random forest,its average accuracy rate is 0.81.(2)Through the cross-grid search of four machine learning methods and three acoustic emission data sets,the optimal parameters of each machine learning model are obtained.Through comparison,it is concluded that the accuracy of the four machine learning models is 0.68-0.71 Among them,the kernel support vector machine KVSM has the lowest value of 0.68,the k-nearest neighbor and gradient boosting random tree RGBT are both close to 0.7,and the random forest has the highest value of 0.71.The comprehensive accuracy,recall,f1 score and other parameters are the best.The model of is RF(random forest),followed by KNN(k nearest neighbor)and RGBT(gradient boosting regression tree),and finally KVSM(kernel support vector machine).(3)By comparing the definition,classification,correlation and characteristic contribution rate of acoustic emission,the acoustic emission parameters established by the model are optimized.After optimization,the amount of redundant data is reduced by 60%,and the calculation time of model establishment is compared with the original The model has been significantly improved,the random forest is reduced by 10%,and the gradient boosting regression tree is reduced by 25%.Although the accuracy rate is reduced,1.5% and 1.1% respectively,its overall efficiency has been greatly improved,and with the data The increase in the amount becomes more and more significant.The final acoustic emission characteristic parameters selected are ASL,center frequency,peak frequency,absolute energy,duration,ringing count,and wavelet packet energy accounted for 1-8.