Research On Microseismic Signal Processing Of Tunnel And Intelligent Early Warning Of Rockburst Based On Deep Learning

The outline of China’s 13th five year plan for scientific and technological innovation proposes to strengthen the development and utilization of deep resources,including exploration and development of mineral and energy resources,utilization of urban underground space,disaster reduction and prevention,etc.However,the exploration of deep resources especially in the construction of deep underground engineering and tunnel engineering is often faced with various risks and geological hazards,such as rockburst,large deformation,etc.As a new technology for rock mass micro-fracture monitoring,microseismic monitoring has developed rapidly and become one of the important means of underground disaster early warning.It has the characteristics of 7×24-hour continuous monitoring,which leads to a large number of monitoring data collection and accumulation,and brings great challenges to the timely,rapid and effective processing of data.Now,most of the data processing work is completed by the staff with rich practical experience and solid seismology background.The long time of processing,the low efficiency and accuracy have large effect on the timeliness of geological disaster prediction and early warning.Combined with the development of microseismic activity or the evolution law of focal parameters,there are more subjective factors for disaster warning,and its effective early warning method and stability need to be further improved.In this paper,taking the rock burst disaster of deep tunnel as the research object,combined with microseismic monitoring technology,artificial intelligence algorithm,deep learning and Internet of things technology,the research on tunnel microseismic signal processing and intelligent early warning of rock burst based on deep learning is carried out.Based on sufficient microseismic monitoring data,the intelligent classification model of microseismic waveform,and denoising and detection model of surrounding rock mass are established.The microseismic source location method is optimized and improved.Combined with the evolution trend of microseismic information in the whole process of rockburst disaster,the microseismic prediction and rockburst early warning model are constructed,and the comprehensive early warning process of rockburst microseismic is proposed.On this basis,the platform of tunnel microseismic automatic monitoring and intelligent early warning of the rockburst is establishend,which improves the timeliness and accuracy of dynamic early warning of rockburst disaster.Through the research,this paper obtains the following main results and understanding:(1)The intelligent classification model of surrounding rock microseismic signals is constructed.The time-frequency analysis of on-site monitoring signals can preliminarily distinguish and identify the micro-fracture signals,which have the characteristics of relatively low strength and frequency,single waveform components and faster attenuation.Two kinds of sample databases of micro fracture waveform and noise waveform(blasting,mechanical and unknown waveform)are established.The intelligent classification model of surrounding rock microseismic waveform is constructed based on deep convolution neural network.The good performance of the method is proved by training,validation,testing and comparative analysis of existing methods based on relevant indicators,and it can also be good used for micro-fracture signals with different signal-to-noise ratio levels.The model also has good generalization ability,and it can also maintain high accuracy for the classification of microseismic waveform of surrounding rock under different background geological structure regions,which can better detect micro-fracture events with M_W≥0.5.The trained model can ensure the accuracy without adjustment,which has a good application prospect in real-time monitoring,intelligent detection and classification.(2)A integrated task model of microseismic waveform denoising and detection based on deep convolution encoding-decoding neural network is constructed.The model integrates two convolutional encoding-decoding networks with similar structures,which can solve the problems of denoising and duration picking of surrounding rock micro-fracture signal at one time.The micro-fracture signal is usually polluted by different types and intensities of noise(non Gaussian noise).Although the model was trained by semisynthetic data,the micro-fracture signal and noise components can be correctly distinguished and separated,even if the frequency band of noise overlaps with that of micro-fracture signal.The results show that the leakage of the micro-fracture signal after denoising is very small,and its shape and amplitude characteristics are well preserved.These characteristics are also applicable to the prediction noise(non-Gaussian noise and Gaussian noise)obtained from noisy signal and denoised signal.The model also shows high accuracy in signal duration picking(including onset time picking).Although the training data of the model comes from the semisynthetic data,the application effect of the model in the actual noisy signals also keeps good performance in terms of the improvement of SNR,the good shape recovery,and the high accuracy of duration picking of micro-fracture signal.In addition,the method also has a good ability to pick up the duration of microseismic signals that are not manual picking due to serious noise pollution,which can improve and correct the picking results of waveform duration compared with the manually picked.Compared with the high pass filter and STA/LTA method,this method significantly improves the SNRs,which cause less waveform distortion,better waveform recovery,and high accuracy of picking in low SNR.It also has higher hit-rate and lower average deviation on onset time picking,which can meet the engineering requirements of picking accuracy.(3)Optimization and evaluation of tunnel microseismic array and source location method.The residual criterion and hyperbolic density are introduced to evaluate and analyze the accuracy and effectiveness of source location of three types of tunnel"non surrounded"micro seismic arrays,including axial extension,lateral extension and twin-tube array.The results are verified by manual percussion experiment and field application,which shows that the source location effect of twin-tube array.The weighted coefficient is introduced to optimize the microseismic location objective function based on L1 norm criterion.Combined with drilling-blasting method and initial judgment of source location,the propagation velocity model of tunnel surrounding rock is constructed,which improves the accuracy of source location.The chaos initialization strategy,adaptive learning factor,weight coefficient improvement and population diversity are introduced to optimize the PSO algorithm and improve the effect of microseismic source location.Compared with different source location methods,the optimized method has the characteristics of strong reliability and high stability,which can jump out of local optimum and improve the convergence accuracy,finally find a better solution than other algorithms.(4)The intelligent prediction model of rock burst microseismic parameters time series based on multi-variable and multi-objective is constructed.The multi-variable parameters energy,apparent volume,event number,and their cumulative value,and energy index in the whole process of rockburst development are selected to construct the rockburst microseismic index database.The multi-objective time series prediction model of rockburst microseismic index based on convolution neural network is established.The prediction performance of different neural network models is compared and analyzed by combining with various evaluation indexes.The prediction method of microseismic index time series is proposed,which realizes the accurate prediction of future microseismic parameters,and provides data basis and supports for subsequent dynamic early warning of rock burst disaster.(5)A comprehensive intelligent early warning model of rockburst microseismic based on the precursor information of disaster and its evolution trend is established.Based on the change trend of cumulative apparent volume and energy index,the whole process of rockburst disaster is divided into three stages:rockburst initiation stage,rockburst catastrophic stage and rockburst forming stage.Combined with sliding time window method,the sample database corresponding to different rockburst stages is established,and the rockburst early warning model is constructed based on high-resolution convolutional neural network.The performance of the model is compared with various evaluation indexes and different methods.The accuracy and good performance of the model for different rock burst stages are demonstrated,and the robustness of the model for different rockburst data is verified.The rockburst catastrophic stage is determined as one of the thresholds of rockburst warning.The rockburst risk degree and its growth trend in the whole process of different rockburst disasters are explored and studied.The fitting curve of the value of the growth rate of rockburst risk degree gerater than 0 apperas no less than 6 times in continuous increasing is considered as another threshold of rockburst warning.Combined with the microseismic monitoring technology,the precursor information of evolution trend of multi-parameter disaster is deduced based on the rockburst microseismic prediction model,and the comprehensive early warning process of rockburst microseismic is finally established.(6)Based on various intelligent algorithms and monitoring technology such as microseismic monitoring technology,microseismic signal processing(intelligent classification,denoising and duration picking of surrounding rock microseismic waveform),source location,parameter calculation,microseismic prediction and rockburst early warning,combined with Java and python programming language and B/S framework system,a platform of tunnel microseismic automatic monitoring and intelligent rockburst early warning is established.The platform realizes the automation,high efficiency and intelligence of the whole microseismic monitoring workflow,which greatly improves the data quality and processing speed,and ensures the timeliness and accuracy of rock burst microseismic warning.More,the platform has been well applied in actual engineering.