Research On Monitoring Data Forecasting Method Of Mining-induced Overburden Deformation Based On LSTM

With the increasing depth of coal mining in China,dynamic disasters such as rock burst,coal and gas outburst are increasingly serious,which are closely related to overburden deformation in the process of deep mining.Physical simulation experiment is one of the frequently-used scientific research methods when investigate scientific problems in the process of coal mining,which is combined with the powerful function fitting ability of deep learning,it provides the possibility to explore the law of overburden deformation.Under this background,this paper utilizes the data generated by physical simulation experiment of mining-induced overburden deformation optical fiber monitoring,introduces the deep learning method,and focuses on the study of the forecasting method of mining-induced overburden deformation optical fiber monitoring data.The main contributions in this thesis are as follow.To solve the problem of missing data imputation of optical fiber monitoring data in physical simulation experiment of mining-induced overburden,a method of missing data imputation based on LSSVM is proposed.Three types of missing value imputation experiments named discrete missing value,continuous missing value and different missing ratio have been completed.Moreover,LSSVM was compared with back propagation nertual net,cubic spline interpolation and other missing value imputation methods,and the results showed that LSSVM interpolation method was superior to the other two methods.Using the optical fiber monitoring data in the physical simulation experiment,selecting the monitoring data obtained at the key layer monitoring points,and establishing the SMOTE-FDT-LSTM optical fiber monitoring data trend forecasting model,which expands sample data of monitored values of time series with synthetic minority oversampling technique sample enhancement technology.The monitoring data is tested for stationarity,and the differential stationary feature is extracted as the characteristic attribute of the input sample.The forecasting trend based on the Long short-term memory overburden deformation monitoring data is carried out.The validity of the forecasting model is verified by comparing with methods such as recurrent neural network and exponential smooth forecasting.Experimental results show that the SMOTE-FDT-LSTM forecasting model is superior to these two contrast methods,and has better forecasting accuracy.A distributed optical fiber monitoring physical simulation experiment data analysis system was designed and developed to effectively manage the corresponding physical simulation experiments and analyze the large amount of monitoring data generated in the simulation experiments.The system adopts B/S architecture,taking Django as the development framework and using python language to achieve the functions of the system.Visualization of learning algorithms such as experiment process management,experiment data management and other basic information management,missing value filling,and monitoring data trend prediction have been completed.The system test shows that the system function which is of stable operation meets the requirements.