A kind of identification of nonlinearity method of rock masses fracturing signal and blasting vibration signal
Technical field
The present invention relates to the recognition methods of a kind of rock masses fracturing signal and blasting vibration signal, especially relate to a kind of nonlinear rock masses fracturing signal and blasting vibration signal recognition methods.
Background technology
Micro seismic monitoring is at home and abroad widely applied as the effective ground pressure monitoring means of one, and rock masses fracturing signal is significant to micro seismic monitoring with the identification of blasting vibration signal.But rock masses fracturing signal and blasting vibration signal similarity are relatively big, and are disturbed by numerous noise signals, automatically identify comparatively difficulty.Currently mainly adopt artificial cognition rock mass microseismic signals, but artificial cognition is subject to individual factor impact, and identify limited amount, limit the real-time analysis of micro seismic monitoring.
Automatically know method for distinguishing currently for rock masses fracturing signal and blasting vibration signal can be divided into: multiparameter statistic law, machine learning method and waveform time-frequency method, these methods generally include feature extraction and two processes of feature identification.Conventional feature extraction includes earthquake magnitude, energy, apparent stress, apparent volume, static stress drops, dynamic stress drops and time domain waveform characteristic parameter (amplitude, frequency) etc., and characteristic recognition method includes fisher classification method, the Logistic Return Law, random forest method, neural network, support vector machine method and Bayes Method etc..Conventional characteristic parameter is difficult to automatically derive mostly, and waveform time domain characteristic parameter is analyzed on single yardstick and obtained, and its quantity of information is less, limits the automatic identification of rock mass microseismic signals.
Waveform time-frequency method can obtain the information of rock mass microseismic signals on multiple dimensioned, is widely used in signal analysis field.The multiscale analysis of waveform time-frequency method is mainly by wavelet analysis, wavelet packet analysis and frequency slice wavelet analysis.Tang Shoufeng etc. (2011) determine wavelet decomposition out to out according to sampling thheorem and Mallat algorithm, and propose to adopt wavelet character energy spectrum coefficient to break as coal petrography the quantitatively characterizing of microseismic signals identification;Zhu Quanjie etc. (2012a) use wavelet packet analysis that microseismic signals has carried out 5 layers of multi-resolution decomposition, and the Energy distribution of rock rupture signal and Blasting Vibration Signal has been carried out relative analysis;Blasting vibration, rock rupture and electromagnetic interference 3 class signal have been carried out 5 layers of WAVELET PACKET DECOMPOSITION in conjunction with wavelet analysis and fractal theory by Zhu Quanjie etc. (2012b), and using 23 small wave fractal box counting dimensions after screening as the characteristic vector of support vector machine identification;State of Zhao men of virtue and ability etc. (2015) adopt frequency slice wavelet transformation that rock body quality of mine destruction signals and blasting vibration signal special frequency band energy ratio and correlation coefficient have been studied.Wavelet analysis, wavelet packet analysis and frequency slice wavelet analysis have a good adaptivity, but the overlapping impact being highly susceptible in signal adjacent harmonic components, cause different frequency bands signal to there is aliasing.
There is bigger limitation with blasting vibration signal recognition methods in visible existing rock masses fracturing signal, it is necessary to studies the automatic identifying method that a kind of suitability is strong, accuracy is high.
Summary of the invention
The technical problem to be solved is to provide a kind of identification of nonlinearity method of rock masses fracturing signal and blasting vibration signal, and rock mass microseismic signals identification of nonlinearity method applicability is strong, accuracy is high for this.
The technical solution of invention is as follows:
A kind of rock masses fracturing signal and the identification of nonlinearity method of blasting vibration signal, comprise the following steps:
Step 1: import virgin rock microseismic signals time series x (n)
Import the time series x (n), n=1,2 of virgin rock microseismic signals ..., N, wherein N is total sampling number of rock mass microseismic signals, takes N=4000~7000, rock mass microseismic signals sample frequency f=4000~7000Hz;
Step 2:EMD decomposes the rock mass microseismic signals after normalization;
2.1 adopt formula x*(n)=x (n)/| xmax(n) | normalized virgin rock microseismic signals, wherein x*N () is the rock mass microseismic signals after normalization, | xmax(n) | for the absolute value of the virgin rock microseismic signals of peak swing;
2.2 adopt EMD to decompose (empirical mode decomposition) formula (1) decomposes the rock mass microseismic signals after normalization, obtains intrinsic modal components IMFj;
In formula, x*N () is the rock mass microseismic signals after normalization, IMFjThe jth intrinsic modal components obtained, IMF is decomposed for EMDjN () is IMFjThe value of nth point, rmN () decomposes the discrepance obtained for EMD, m is the number of intrinsic modal components;
Step 3:SVD decomposition obtains eigenmatrix singular value σi(i=1,2 ..., r)
3.1 adopt formula (2) to calculate each intrinsic modal components IMFjWith correlation coefficient co (j) of the rock mass microseismic signals after normalization, and then screen according to the size of correlation coefficient co (j) and obtain main intrinsic modal components;
Wherein,
The intrinsic modal components obtained after 3.2 employing SVD decomposition formula (3) (singular value decomposition) calculating siftings constitutes matrix X=[c1c2…cr]TSingular value σi(i=1,2 ..., r), wherein, c1,c2,…,crFor the intrinsic modal components obtained after screening, r is the intrinsic modal components number obtained after screening, and T is the transposition of matrix;
In formula, the orthogonal matrix on U, V respectively r × r and N × N rank;S is the clinodiagonal matrix on r × N rank, namely Wherein, σ=diag (σ1,σ2,…,σr), be diagonal entry it is σ1,σ2,…,σrDiagonal matrix;σi(i=1,2 ..., r) for the singular value of matrix X, and σ1≥σ2≥…≥σr;
Step 4: the Probability p (Z) using Logistic regression model signal calculated to be blasting vibration signal;
Choose the M group data training data as Logistic regression model of rock masses fracturing signal and blasting vibration signal, σ step 3.2 tried to achieve respectively1,σ2,…,σrIndependent variable as rock mass microseismic signals identification;Z, as the dependent variable of rock mass microseismic signals identification, sets Z=1 and represents that signal is as blasting vibration signal, and Z=0 represents that signal is rock masses fracturing signal, uses maximum likelihood estimate to try to achieve the parameter beta of Logistic regression model0,β1,β2,…,βr;
Logistic regression model is:
P (Z)=1/ (1+exp (-Z))=1/ (1+exp (-(β0+β1·σ1+β2·σ2+…+βr·σr)))(4)
Wherein, β0For constant term, β1,β2,…,βrFor with independent variable σ1,σ2,…,σrRelevant parameter;
Step 5: the size identification rock mass microseismic signals according to p (Z)
In described step 3, the intrinsic modal components of screening correlation coefficient co (j) >=0.03 is as main intrinsic modal components.
In described step 3, taking the p (Z)=0.5 cut off value as Logistic regression model identification rock mass microseismic signals, p (Z) > 0.5 is identified as blasting vibration signal;P (Z)≤0.5 is identified as rock masses fracturing signal.
In described step 1, total sampling number N of rock mass microseismic signals is taken as 5000, and sample frequency f is taken as 6000Hz.
In described step 4, M value is 100.
Beneficial effect:
A kind of rock masses fracturing signal of the present invention and the identification of nonlinearity method of blasting vibration signal, comprise the steps: to import the time series x (n), n=1,2 of rock mass microseismic signals ..., N.Setting rock masses fracturing signal identification classification as 0, blasting vibration signal identified category is 1;EMD decomposes the intrinsic modal components obtaining normalization rock mass microseismic signals, and screening obtains main intrinsic modal components c1,c2,…,cr.Wherein, r is the number of intrinsic modal components after screening;SVD split-matrix [c1c2…cr]T, obtain its singular value σi(i=1,2 ..., r);Logistic model calculates the Probability p (Z) of blasting vibration signal.Wherein, p (Z)=1/ (1+exp (-Z))=1/ (1+exp (-(β0+β1·σ1+β2·σ2+…+βr·σr)));Identify rock mass microseismic signals: p (Z) > 0.5 blasting vibration signal;P (Z)≤0.5 rock masses fracturing signal.The present invention obtains the intrinsic modal components of rock mass microseismic signals by EMD adaptive decomposition, rock mass microseismic signals is made to be provided with multi-scale information, the matrix that intrinsic modal components after screening is constituted by recycling SVD decomposes, obtain its singular value, realize Data Dimensionality Reduction and feature extraction, efficiently solve single time scale and analyze the less problem of quantity of information.Empirical mode decomposition carries out adaptive Time-frequency Decomposition from signal self local feature, is that a kind of more effective Time-Frequency Localization analyzes method, and is applicable to the analysis of nonlinear and non local boundary value problem.Additionally, Logistic model returns the non-linear relation that can effectively solve between multiparameter, easy to use, amount of calculation is little so that Classification and Identification has higher accuracy rate.The method has that the suitability is strong, accuracy high.
Accompanying drawing explanation
Fig. 1 is the method for the invention flow chart.
Fig. 2 is rock mass microseismic signals EMD catabolic process figure.Wherein, (a) is virgin rock microseismic signals, and (b) is normalization rock mass microseismic signals, and (c) decomposes, for EMD, the intrinsic modal components that normalization rock mass microseismic signals obtains.
Fig. 3 is the correlation coefficient figure of each intrinsic modal components and normalization rock mass microseismic signals.
Fig. 4 is that the intrinsic modal components obtained after SVD decomposes screening constitutes the singular value figure that matrix obtains.
Fig. 5 is the singular value box figure that EMD_SVD engineer applied obtains.
Fig. 6 is training and the prediction effect figure of Logistic model
Detailed description of the invention
Below in conjunction with accompanying drawing 1~6, a kind of rock masses fracturing signal propose the present invention and the identification of nonlinearity method of blasting vibration signal are described further.The description of inventive algorithm thought is as follows: the present invention obtains the intrinsic modal components of rock mass microseismic signals by EMD adaptive decomposition, rock mass microseismic signals is made to be provided with multi-scale information, the matrix that intrinsic modal components after screening is constituted by recycling SVD decomposes, obtain its singular value, realize Data Dimensionality Reduction and feature extraction, efficiently solve single time scale and analyze the less problem of quantity of information.Additionally, Logistic model returns the non-linear relation that can effectively solve between multiparameter so that Classification and Identification has higher accuracy rate.
Step 1: import rock mass microseismic signals time series x (n)
Import the time series x (n), n=1,2 of rock mass microseismic signals ..., N, wherein N is total sampling number of rock mass microseismic signals, takes N=4000~7000, rock mass microseismic signals sample frequency f=4000~7000Hz;
Step 2:EMD decomposes normalization rock mass microseismic signals x (n)
Adopt formula x*(n)=x (n)/| xmax(n) | normalized rock mass microseismic signals, wherein x*N () is rock mass microseismic signals after normalization, | xmax(j) | for the absolute value of original signal peak swing.Adopt formula (1) empirical mode decomposition (EMD) that normalization rock mass microseismic signals is decomposed again, obtain intrinsic modal components IMFj;
X in formula*N () is the rock mass microseismic signals after normalization, IMFjThe jth intrinsic modal components obtained, IMF is decomposed for EMDjN () is IMFjThe value of nth point, rmN () decomposes the discrepance obtained for EMD, m is the number of intrinsic modal components.
Step 3:SVD decomposition obtains eigenmatrix singular value σi(i=1,2 ..., r)
Employing formula (2) calculates intrinsic modal components IMFjWith correlation coefficient co (j) of normalized signal, and then screen according to the size of correlation coefficient and obtain main intrinsic modal components.
Wherein,
After employing formula (3) singular value decomposition (SVD) calculating sifting, intrinsic modal components constitutes matrix [c1c2…cr]TSingular value σi(i=1,2 ..., r).Wherein, c1,c2,…,crFor the intrinsic modal components obtained after screening, r is the intrinsic modal components number obtained after screening, and T is the transposition of matrix.
X=[c in formula1c2…cr]T;The orthogonal matrix on U, V respectively r × r and N × N rank;S is the clinodiagonal matrix on r × N rank, namely Wherein, σ=diag (σ1,σ2,…,σr), σi(i=1,2 ..., r) for the singular value of matrix, and σ1≥σ2≥L≥σr。
Step 4:Logistic model calculates the Probability p (Z) of blasting vibration signal
Choose the M group data training data as Logistic regression model of rock masses fracturing signal and blasting vibration signal, σ step 3.2 tried to achieve respectively1,σ2,…,σrIndependent variable as rock mass microseismic signals identification;Z, as the dependent variable of rock mass microseismic signals identification, sets Z=1 and represents that signal is as blasting vibration signal, and Z=0 represents that signal is rock masses fracturing signal, uses maximum likelihood estimate to try to achieve the parameter beta of Logistic regression model0,β1,β2,…,βr;
Logistic regression model is:
P (Z)=1/ (1+exp (-Z))=1/ (1+exp (-(β0+β1·σ1+β2·σ2+…+βr·σr)))(4)
Wherein, β0For constant term, β1,β2,…,βrFor with independent variable σ1,σ2,…,σrRelevant parameter;
Step 5: identify rock mass microseismic signals according to p (Z)
Identify that rock mass microseismic signals: p (Z) > 0.5 is blasting vibration signal;P (Z)≤0.5 is rock masses fracturing signal.
Embodiment 1:
Fig. 2 is rock mass microseismic signals EMD catabolic process figure.Wherein, (a) is virgin rock microseismic signals, and (b) is normalization rock mass microseismic signals, and (c) decomposes, for EMD, the intrinsic modal components that normalization rock mass microseismic signals obtains.Rock mass microseismic signals total sampling number N=5000, sample frequency f=6000Hz in figure.Known that normalization rock mass microseismic signals EMD decomposes by Fig. 2 (c) and obtain 7 intrinsic modal components IMFj(j=1,2 ..., 7), expand the analysis yardstick of signal, but add data volume, thereby increases and it is possible to there is chaff component.
Fig. 3 is the correlation coefficient figure of each intrinsic modal components and normalization rock mass microseismic signals.IMF is known by Fig. 31~IMF3Relatively big with the correlation coefficient of normalized signal, IMF4~IMF7Less with the correlation coefficient of normalized signal, and the correlation coefficient of IMF7 and normalized signal is only 0.02.It is chaff component, it is known that selecting IMF1~IMF6 is relatively reasonable as main intrinsic modal components by the achievement in research-correlation coefficient of (2014) such as the Xu Feng IMF component less than 0.03.
Fig. 4 is that the intrinsic modal components obtained after SVD decomposes screening constitutes the singular value figure that matrix obtains.Known that SVD decomposes the dimension reducing matrix by Fig. 4, and achieve the feature extraction of matrix.
Fig. 5 is the singular value box figure that EMD_SVD engineer applied obtains.Rock masses fracturing signal classification is designated 0, and blasting vibration signal classification logotype is 1.200 rock masses fracturing signals and 200 blasting vibration signal of engineer applied are randomly draw the signal (each microseismic event comprises multiple microseismic signals) triggered in each microseismic event obtained the earliest in the IMS Microseismic monitoring system of phosphorus ore sand bar ore deposit, Kaiyang.Each microseismic signals total sampling number N=5000, sample frequency f=6000Hz.The singular value σ of rock masses fracturing signal and blasting vibration signal is known by Fig. 51、σ2And σ3Obvious difference, and σ4、σ5And σ6There is some difference, but difference is inconspicuous.In view of being difficult to find one or more singular value cut off value identification rock masses fracturing signal and blasting vibration signal, this patent adopts a kind of nonlinear method to launch to identify the Logistic Return Law.
Fig. 6 is training and the prediction effect figure of Logistic model.Choose 1~100 group of data training data as Logistic model of rock masses fracturing signal and blasting vibration signal respectively, then choose 101~200 groups of data testing data as Logistic model respectively.Logistic model returns Probability p (Z)=1/ (1+exp (-Z))=1/ (1+exp (-(the 8.459-0.777 σ obtaining blasting vibration signal1-0.198·σ2-0.359·σ3+0.027·σ4-0.186·σ5+2.171·σ6))), and then obtain training and the test effect (Fig. 6 and Biao 1) of rock mass microseismic signals.For understanding this patent more intuitively, by the singular value initial data of 25 blasting vibration signal and rock masses fracturing signal with predict the outcome and be compiled in table 2 and table 3 respectively.
Being known by Fig. 6 and Biao 1: the training of Logistic model and prediction effect are all very good, training group rate of accuracy reached is to 92.5%, and check groups rate of accuracy reached is to 86.5%, and overall rate of accuracy reached is to 89.5%.Visible, feature extraction and recognition methods based on EMD_SVD and Logistic model are a kind of very effective rock mass microseism recognition methodss.
The training of table 1Logistic model and prediction effect statistical table
The singular value initial data of 225 groups of blasting vibration signal of table and predicting the outcome
The singular value initial data of 325 groups of rock masses fracturing signals of table and predicting the outcome
The foregoing is only embodiments of the invention, not in order to limit the present invention, all within present invention spirit and principle, changed, be equal to replacement, improvement etc., be should be included within protection scope of the present invention.