Full-waveform inversion detection method suitable for internal fractures of coal and rock mass

Abstract

The invention discloses a full-waveform inversion detection method suitable for internal fractures of coal and rock mass.The method comprises the steps that observation data of a seismic source are obtained, then an objective function is constructed, synthetic wave field data are obtained through calculation, and the gradient of velocity field model parameters is obtained through solving with a Lagrange adjoint method; the method is advantaged in that an accurate updating direction of velocity field model parameters can be acquired, finally, the velocity field model parameters are iteratively updated through the deep convolutional network until the maximum training frequency or the objective function value satisfies the precision requirement, and a characterization result of the deep convolutional network model is outputted, namely, a final inversion model parameter is acquired. Carrying out full-waveform inversion by adopting a velocity field model of the parameter; according to the method, accurate inversion modeling and internal fracture accurate detection can be performed on the coal and rock mass model through the steps, and the effective boundary strategy is adopted to replace the memory consumption with the calculated amount, so that the memory consumption is reduced, and data support is provided for subsequent mining.

Description

Technical Field

[0001] The invention relates to a full waveform inversion detection method suitable for internal cracks of coal and rock masses, belonging to the technical field of mineral resource exploration. Background Art

[0002] As an important energy and mineral resource, coal is of vital importance to the country's economic and social development. Coal rock structure model testing is used to understand the structure, mechanical properties and deformation laws of coal rock, and then reveal the dynamic evolution process of coal rock, so as to provide a basis for the prediction and early warning of disaster risks such as water inrush, roof collapse, and support failure in coal seams, and ultimately achieve safe and efficient production of coal resources. Therefore, the detection of internal cracks in coal rock is of great significance.

[0003] The following methods are currently available for internal crack detection tests on coal and rock samples in the laboratory: gamma or X-ray detection method, which is mature in technology, but the detection process is radioactive; electron spin resonance method, which has a small detection range; stress pick method, which has poor detection stability due to immature technology; infrared detection method, radar detection method, high-pressure water flow method, etc., all of which have problems with insufficient detection accuracy or detection range. In addition, there are also camera methods and vibration and acoustic wave analysis methods. Among them, the camera method cannot obtain internal information of coal and rock, while the vibration and ultrasonic wave analysis method is a method with the ability to finely detect coal and rock; coal and rock have multi-scale structural characteristics, and there are different structural changes from microcracks to fracture surfaces. Only by successfully obtaining the crack and structural interface information of coal and rock at different scales can its mechanical properties and deformation behavior be successfully simulated. In view of the multi-field coupled dynamic disaster problem in deep earth engineering, the coal and rock structure model faces the problem of large test range and high test accuracy requirements. The existing methods cannot effectively achieve high-precision characterization of microcracks and structural interfaces of such models.

[0004] Full-waveform inversion (FWI) is recognized as the most promising new generation of high-precision geophysical imaging technology. It makes full use of the full waveform information of seismic waves propagating underground, solves the accurate medium model of the underground medium through numerical simulation and inversion methods, and obtains high-resolution underground structure images. In recent years, classical full-waveform inversion has made considerable progress, especially in inversion accuracy and efficiency. Because the gradient obtained by automatic differentiation is equivalent to the gradient obtained by the adjoint method, deep learning FWI based on network automatic differentiation has been widely studied, which also makes it possible to carry out deep learning full-waveform inversion using physical constraints. However, when the gradient is calculated using the wave equation with automatic differentiation, it can be regarded as network training within the framework of recursive neural network. When inverting large-scale model parameters, complex wave equations or multi-parameters, the time recursive results need to occupy a large amount of memory and consume a lot of hardware resources; in addition, the existing deep learning tools lack typical objective functions designed specifically for full-waveform inversion, making it impossible to solve the local minimum problem in inversion. In response to the above problems, how to provide an improved full-waveform inversion detection method that can accurately invert and model the coal rock model and accurately detect internal cracks, while also effectively reducing the problem of excessive memory usage during the inversion process, thereby ensuring the effective use of hardware resources during the inversion process, is the research direction required by the present invention. Summary of the invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention provides a full waveform inversion detection method suitable for internal cracks in coal rock masses, which can accurately invert and model the coal rock mass model and accurately detect internal cracks, and the inversion process occupies less memory, thereby providing data support for subsequent mining.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a full waveform inversion detection method suitable for internal cracks in coal and rock masses, the specific steps are:

[0007] Step 1: Arrange multiple seismic sources in a row on one side of the coal and rock working face to be detected; arrange multiple detection points in a row on the side, place a detector at each detection point, and connect each detector to the acquisition base station and the acquisition host through the connecting main line, so that the detector, the acquisition base station and the acquisition host form an ultrasonic observation system; stimulate each seismic source in turn, and receive and obtain observation data through the ultrasonic observation system;

[0008] Step 2: First, construct a convolution-based objective function and establish a wave equation for generating synthetic wavefield data in the objective function;

[0009] Step 3: Based on the objective function and wave equation constructed in step 2, an augmented objective function of the objective function is constructed with the wave equation as a constraint, and the adjoint equation is derived using the Lagrange adjoint method;

[0010] Step 4: construct the gradient of the model parameters in the objective function using the wave equation of step 2 and the adjoint equation of step 3, wherein the gradient consists of the gradient of velocity and the gradient of density;

[0011] Step 5: Use the Lagrangian adjoint method to continuously calculate the gradient of the model parameters, and combine the deep neural network to update the model parameters, and finally obtain the model parameters required for inversion, and finally use the model parameters to perform full waveform inversion.

[0012] Further, the step 2 is specifically as follows:

[0013] Based on the convolutional objective function, the specific definition is as follows:

[0014]

[0015] In the formula, d(x r ,t) and d(x ref ,t) represents the observation data obtained by the detection point and the reference channel respectively, v(x ref ,t;m) and v(x r ,t;m) represents the synthetic wave field data of the reference channel and the detection point, x r and x ref Represents the position of the detection point and the reference channel, the spatial coordinate is x = [x, z], t represents time, represents the Euclidean norm; wherein the reference channel is a detection point arbitrarily selected from each detection point;

[0016] The wave equation for generating synthetic wavefield data in the objective function is:

[0017]

[0018] in, is the forward wave field variable, σ is the particle vibration stress corresponding to the forward wave field, v = [v x ,v z ] represents the vibration speed of the particle along the x and z directions, s is the source term, It means to find partial derivatives with respect to time t, and the link matrix E is:

[0019]

[0020] in, and It means partial derivatives with respect to x and z.

[0021] Further, the adjoint equation in step three is:

[0022]

[0023] Among them, the back propagation field is the particle vibration stress corresponding to the back-propagation wave field, and represents the vibration speed of the particle along the x and z directions, and the accompanying source is

[0024]

[0025] * indicates convolution operation, Represents the cross-correlation operation.

[0026] Further, the step 4 is specifically as follows:

[0027] Gradient of model parameters in the objective function Further transformed into:

[0028]

[0029] in, is the gradient of velocity, is the gradient of density;

[0030] When calculating the gradient of model parameters, the wave equation is discretized using the differential approximation of a high-order staggered grid, and the boundary absorption condition uses the convolution-type perfect absorbing boundary condition. The algorithm uses MPI to control the multi-node high-performance parallel computing of the GPU, and reduces memory consumption by adopting an effective boundary strategy to exchange computational effort for memory consumption.

[0031] Further, the step five is specifically as follows:

[0032] The gradient of the model parameters calculated by the Lagrangian adjoint method is assigned to the gradient of the network output:

[0033]

[0034] Use network automatic backpropagation to obtain the gradient of network parameters of each network layer The network parameters are updated as follows,

[0035]

[0036] Where k represents the k-th training number, αRepresents the learning rate; there are many methods to optimize network parameter updates, such as gradient descent, momentum-based optimization, adaptive learning rate optimization, and learning rate scheduler. Various types of optimizers can be selected and used in full waveform inversion.

[0037] The model parameters are updated by selecting the deep network to represent the model parameter information and training the network parameters. The deep neural network represents the model parameters in the following way:

[0038] m=CNN θ (λ)+m ini (9)

[0039] where m = [v, ρ] T represents the model parameters, m ini =[v ini ,ρ ini ] T represents the initial model parameters, v represents the velocity of the medium, ρ represents the density of the medium, and v ini represents the velocity of the medium, ρ ini represents the density of the medium, T represents the transpose of the matrix; CNN θ It is a deep convolutional neural network model used to represent model parameters, θ represents network parameters, λ represents random feature variables, and deep neural network CNN θ The characteristic variable λ is used as input and the model parameter m is used as output;

[0040] In order to carry out independent training and learning of model parameter information in the frequency band corresponding to each scale and avoid the problem that the features of certain scales are difficult to accurately obtain when multiple frequency bands are continuously learned, the following multi-scale deep learning strategy is adopted:

[0041]

[0042] Among them, m Ini represents the initial model for starting inversion of each frequency band; F i represents the frequency band of multi-scale inversion, which is F=F1,F2,...,F n , n represents the number of inverted frequency bands; Indicates that in the frequency band F i Carry out deep learning and characterization of model parameters; through the above calculation, in multi-scale inversion, except for the initial model provided by the first scale, the initial models of the inversion of the remaining scales are superimposed with the network characterization results of the inverted frequency bands to ensure the accuracy of the model parameter update.

[0043] Finally, the formulas in this step are iterated repeatedly to output new model parameters until the maximum number of training times is reached or the objective function value meets the accuracy requirements, and the characterization results of the deep convolutional neural network model are output, which are the final inversion model parameters, and the model parameters are used for full waveform inversion.

[0044] Compared with the prior art, the present invention first obtains the observation data of the earthquake source, then constructs the objective function, obtains the synthetic wave field data by calculation, and solves the gradient of the velocity field model parameters by the Lagrangian adjoint method, which is conducive to obtaining the accurate update direction of the velocity field model parameters, and finally iteratively updates the velocity field model parameters through the deep convolution network until the maximum number of training times or the objective function value meets the accuracy requirements, and outputs the characterization results of the deep convolution network model, which are the model parameters of the final inversion. At this time, the velocity field model of the parameters is used for full waveform inversion; the present invention can accurately invert and model the coal rock model and accurately detect internal fractures through the above steps, and adopts an effective boundary strategy to exchange the amount of calculation for memory consumption, thereby reducing memory consumption, thereby providing data support for subsequent mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a velocity profile diagram of the multi-scale inversion result of the present invention at the position of x=0.325m;

[0046] Figure 2 is the initial velocity model used in the multi-scale inversion of the present invention;

[0047] Figure 3 It is the speed model inverted by the deep network calling the Adagrad optimization algorithm in the present invention;

[0048] Figure 4 It is the real velocity model in the experimental proof. DETAILED DESCRIPTION

[0049] The present invention will be further described below.

[0050] like Figure 1 As shown, the specific steps of the present invention are:

[0051] Step 1: Arrange multiple seismic sources in a row on one side of the coal rock working face to be detected; arrange multiple detection points in a row on the side, place a detector at each detection point, and connect each detector to the acquisition base station and the acquisition host through the main line, so that the detector, the acquisition base station and the acquisition host form an ultrasonic observation system; stimulate each seismic source in turn, and receive and obtain observation data through the ultrasonic observation system; the observation data is obtained using 6×10 4 Hz Ricker wavelet is formed, and 620 observation data can be received after each source is excited;

[0052] Step 2: First, construct a convolution-based objective function and establish a wave equation for generating synthetic wavefield data in the objective function, specifically:

[0053] Based on the convolutional objective function, the specific definition is as follows:

[0054]

[0055] In the formula, d(x r ,t) and d(x ref ,t) represents the observation data obtained by the detection point and the reference channel respectively, v(x ref ,t;m) and v(x r ,t;m) represents the synthetic wave field data of the reference channel and the detection point, x r and x ref Represents the position of the detection point and the reference channel, the spatial coordinate is x = [x, z], t represents time, represents the Euclidean norm; wherein the reference channel is a detection point arbitrarily selected from each detection point;

[0056] The wave equation for generating synthetic wavefield data in the objective function is:

[0057]

[0058] in, is the forward wave field variable, σ is the particle vibration stress corresponding to the forward wave field, v = [v x ,v z ] represents the vibration speed of the particle along the x and z directions, s is the source term, It means to find partial derivatives with respect to time t, and the link matrix E is:

[0059]

[0060] in, and It means partial derivatives with respect to x and z.

[0061] Step 3: Based on the objective function and wave equation constructed in step 2, an augmented objective function of the objective function is constructed with the wave equation as a constraint, and the adjoint equation is derived using the Lagrange adjoint method. The adjoint equation is:

[0062]

[0063] Among them, the back propagation field is the particle vibration stress corresponding to the back-propagation wave field, and represents the vibration speed of the particle along the x and z directions, and the accompanying source is

[0064]

[0065] * indicates convolution operation, Represents the cross-correlation operation.

[0066] Step 4: Use the wave equation of step 2 and the adjoint equation of step 3 to construct the gradient of the model parameters in the objective function. The gradient consists of the gradient of velocity and the gradient of density, specifically:

[0067] Gradient of model parameters in the objective function Further transformed into:

[0068]

[0069] in, is the gradient of velocity, is the gradient of density;

[0070] When calculating the gradient of model parameters, the wave equation is discretized using the differential approximation of a high-order staggered grid, and the boundary absorption condition uses the convolution-type perfect absorbing boundary condition. The algorithm uses MPI to control the multi-node high-performance parallel computing of the GPU, and reduces memory consumption by adopting an effective boundary strategy to exchange computational effort for memory consumption.

[0071] Step 5: Use the Lagrangian adjoint method to continuously calculate the gradient of the model parameters, and combine the deep neural network to update the model parameters, and finally obtain the model parameters required for inversion. Finally, use the model parameters to perform full waveform inversion, specifically:

[0072] The gradient of the model parameters calculated by the Lagrangian adjoint method is assigned to the gradient of the network output:

[0073]

[0074] Use network automatic backpropagation to obtain the gradient of network parameters of each network layer The network parameters are updated as follows,

[0075]

[0076] Among them, k represents the k-th training number, α represents the learning rate; there are many methods to achieve network parameter update optimization, such as gradient descent, momentum-based optimization method, adaptive learning rate optimization method and learning rate scheduler, etc. Various types of optimizers can be selected and used in full waveform inversion.

[0077] The model parameters are updated by selecting the deep network to represent the model parameter information and training the network parameters. The deep neural network represents the model parameters in the following way:

[0078] m=CNN θ (λ)+m ini (9)

[0079] where m = [v, ρ] T represents the model parameters, m ini =[v ini ,ρ ini ] T represents the initial model parameters, v represents the velocity of the medium, ρ represents the density of the medium, and v ini represents the velocity of the medium, ρ ini represents the density of the medium, T represents the transpose of the matrix; CNN θ It is a deep convolutional neural network model used to represent model parameters, θ represents network parameters, λ represents random feature variables, and deep neural network CNN θ The characteristic variable λ is used as input and the model parameter m is used as output;

[0080] In order to carry out independent training and learning of model parameter information in the frequency band corresponding to each scale and avoid the problem that the features of certain scales are difficult to accurately obtain when multiple frequency bands are continuously learned, the following multi-scale deep learning strategy is adopted:

[0081]

[0082] Among them, m Ini represents the initial model for starting inversion of each frequency band; F i represents the frequency band of multi-scale inversion, which is F=F1,F2,...,F n , n represents the number of inverted frequency bands; Indicates that in the frequency band F i Carry out deep learning and characterization of model parameters; through the above calculation, in multi-scale inversion, except for the initial model provided by the first scale, the initial models of the inversion of the remaining scales are superimposed with the network characterization results of the inverted frequency bands to ensure the accuracy of the model parameter update.

[0083] Finally, the formulas in this step are iterated repeatedly to output new model parameters until the maximum number of training times is reached or the objective function value meets the accuracy requirements, and the characterization results of the deep convolutional neural network model are output, which are the final inversion model parameters, and the model parameters are used for full waveform inversion.

[0084] Tests prove:

[0085] Firstly, a two-dimensional fractured coal-rock model was used for multi-scale inversion test, and the Adagrad optimization method was selected for multi-scale inversion. Figure 1 The real velocity model is shown as Figure 4 As shown; the model size is 0.775m×0.61m, and the spatial step size of the model discretization is 1.25×10 -3 m. The staggered grid finite difference method is used in the numerical simulation to discretize the equations, and the absorbing boundary condition is introduced to reduce the artificial boundary reflection. The seismic observation data includes 34 seismic source data evenly distributed on one side of the working face of the two-dimensional fractured coal rock model. The observation data is obtained by using 6×10 4 Hz Ricker wavelet is generated, and 620 data can be received after each source excitation. The time sampling interval of the seismic record is 1.25×10 -7 s, track spacing 1.25×10 -3 m. The initial model used in the inversion is as follows Figure 2 As shown in Figure 1, the initial model is obtained by spatially averaging the true model using a window with a smoothing factor of 12. Based on the convolutional objective function, the inversion is carried out. The multiscale inversion is divided into five frequency bands, and the center frequencies are set to 3×10 4 Hz,5×10 4 Hz,8×10 4 Hz,12×10 4 Hz, 13×10 4 Each frequency band is iterated 30 times, and the inversion is iterated 150 times in total. The final inversion result is the fusion of the network's representation results in different frequency bands and the initial model; the final inversion result is as follows Figure 3 As shown. Figure 3 The inversion results and Figure 4 By comparing the inversion result with the real velocity model, it can be seen that the inversion result has a high similarity with the real velocity model, which indicates that the inversion accuracy of the present invention is high.

[0086] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A full waveform inversion detection method suitable for internal cracks in coal and rock masses, characterized in that: The specific steps are: Step 1: Arrange multiple seismic sources in a row on one side of the coal and rock working face to be detected; arrange multiple detection points in a row on the side, place a detector at each detection point, and connect each detector to the acquisition base station and the acquisition host through the connecting main line, so that the detector, the acquisition base station and the acquisition host form an ultrasonic observation system; stimulate each seismic source in turn, and receive and obtain observation data through the ultrasonic observation system; Step 2: First, construct a convolution-based objective function and establish a wave equation for generating synthetic wavefield data in the objective function; Step 3: Based on the objective function and wave equation constructed in step 2, an augmented objective function of the objective function is constructed with the wave equation as a constraint, and the adjoint equation is derived using the Lagrange adjoint method; Step 4: construct the gradient of the model parameters in the objective function using the wave equation of step 2 and the adjoint equation of step 3, wherein the gradient consists of the gradient of velocity and the gradient of density; Step 5: Use the Lagrangian adjoint method to continuously calculate the gradient of the model parameters, and combine the deep neural network to update the model parameters, and finally obtain the model parameters required for inversion, and finally use the model parameters to perform full waveform inversion.

2. The full waveform inversion detection method for internal cracks in coal and rock mass according to claim 1 is characterized in that: The step 2 is specifically as follows: Based on the convolutional objective function, the specific definition is as follows: In the formula, d(x r ,t) and d(x ref ,t) represents the observation data obtained by the detection point and the reference channel respectively, v(x ref ,t;m) and v(x r ,t;m) represents the synthetic wave field data of the reference channel and the detection point, x r and x ref Represents the position of the detection point and the reference channel, the spatial coordinate is x = [x, z], t represents time, represents the Euclidean norm; The wave equation for generating synthetic wavefield data in the objective function is: in, is the forward wave field variable, σ is the particle vibration stress corresponding to the forward wave field, v = [v x ,v z ] represents the vibration speed of the particle along the x and z directions, s is the source term, It means to find partial derivatives with respect to time t, and the link matrix E is: in, and It means partial derivatives with respect to x and z.

3. The full waveform inversion detection method for internal cracks in coal and rock mass according to claim 1 is characterized in that: The adjoint equation in step 3 is: Among them, the back propagation field is the particle vibration stress corresponding to the back-propagation wave field, and represents the vibration speed of the particle along the x and z directions, and the accompanying source is * indicates convolution operation, Represents the cross-correlation operation.

4. The full waveform inversion detection method for internal cracks in coal and rock mass according to claim 1 is characterized in that: The step 4 is specifically as follows: Gradient of model parameters in the objective function Further transformed into: in, is the gradient of velocity, is the gradient of density; When calculating the gradient of model parameters, the wave equation is discretized using high-order staggered grid differential approximation, the boundary absorption condition uses convolution-type perfect absorbing boundary conditions, and the algorithm uses MPI to control GPU multi-node high-performance parallel computing.

5. The full waveform inversion detection method for internal cracks in coal and rock mass according to claim 1 is characterized in that: The step five is specifically as follows: The gradient of the model parameters calculated by the Lagrangian adjoint method is assigned to the gradient of the network output: Use network automatic backpropagation to obtain the gradient of network parameters of each network layer The network parameters are updated as follows, Where k represents the k-th training number, and α represents the learning rate; The model parameters are updated by selecting the deep network to represent the model parameter information and training the network parameters. The deep neural network represents the model parameters in the following way: m=CNN θ (λ)+m ini (9) where m = [v, ρ] T represents the model parameters, m ini =[v ini ,ρ ini ] T represents the initial model parameters, v represents the velocity of the medium, ρ represents the density of the medium, and v ini represents the velocity of the medium, ρ ini represents the density of the medium, T represents the transpose of the matrix; CNN θ It is a deep convolutional neural network model used to represent model parameters, θ represents network parameters, λ represents random feature variables, and deep neural network CNN θ The characteristic variable λ is used as input and the model parameter m is used as output; In order to carry out independent training and learning of model parameter information in the frequency band corresponding to each scale and avoid the problem that the features of certain scales are difficult to accurately obtain when multiple frequency bands are continuously learned, the following multi-scale deep learning strategy is adopted: Among them, m Ini represents the initial model for starting inversion of each frequency band; F i represents the frequency band of multi-scale inversion, which is F=F1,F2,...,F n , n represents the number of inverted frequency bands; Indicates that in the frequency band F i Conduct deep learning and characterization of model parameters; Finally, the formulas in this step are iterated repeatedly to output new model parameters until the maximum number of training times is reached or the objective function value meets the accuracy requirements, and the characterization results of the deep convolutional neural network model are output, which are the final inversion model parameters, and the model parameters are used for full waveform inversion.

Images