A method for inverting geostress of deep rock mass based on blasting seismic wave velocity

Abstract

The application provides a deep rock mass ground stress inversion method based on blasting seismic wave velocity. According to the deep rock mass blasting seismic wave velocity obtained by monitoring and measuring the deep rock mass on site, a network with a structure of pix2pix is referenced to perform parameter inversion. The network takes seismic wave field wave velocity propagation data as input, and trains corresponding seismic wave velocity model true values, so as to establish an inversion model and inversely calculate some initial parameters of the rock mass. The purpose is to establish a suitable model, make the prediction results consistent with the results of the on-site test, correctly reflect or predict the mechanical properties of the deep rock mass, set a data set model in line with the actual geological conditions for the deep rock mass blasting seismic wave velocity inversion problem, and successfully perform preliminary supervised learning network pre-training on the tunnel seismic wave velocity inversion network through training, verification and testing on the parameter inversion network, and then complete network parameter initialization, thereby proving the effectiveness of the parameter inversion network.

Description

Technical Field

[0001] This invention relates to the field of geostress inversion in deep-buried tunnels and underground engineering rock masses, specifically to a geostress inversion method for deep rock masses based on blasting seismic wave velocity. Background Technology

[0002] In recent years, due to the increasing depletion of shallow resources, both domestic and international efforts have shifted towards the exploitation of deep resources. The failure of deep rock masses is a result of the combined effects of high ground stress and explosive impact loads. Particularly in the study of the impact of ground stress on the seismic wave velocity generated after deep rock blasting, the design and construction of deep-buried tunnels and underground engineering projects must consider the influence of ground stress state. Ground stress inversion can obtain information on the ground stress state of the rock mass, thereby helping to optimize engineering design and construction schemes and improve engineering quality and safety. Therefore, fully considering the ground stress factors affecting blasting seismic waves has significant practical engineering implications for related blasting projects.

[0003] Currently, there are very few methods for geostress inversion based on seismic wave velocity, and existing seismic inversion methods require a large amount of labeled data, i.e., real wave velocity models, for training. In this case, wave velocity models are difficult to obtain in practice, and it is necessary to rely on relatively accurate prior wave velocity data. Secondly, data from on-site project conditions, borehole information, and excavation exposure are not effectively utilized, and they cannot be fully matched with the actual detection situation, resulting in inaccurate information on the geostress state of the rock mass.

[0004] Therefore, to address the aforementioned technical issues, a method for inverting the geostress of deep rock masses based on blasting seismic wave velocity is needed. Summary of the Invention

[0005] Purpose of the invention:

[0006] This invention provides a method for inverting the geostress of deep rock masses based on blasting seismic wave velocities. The purpose is to solve the technical problems of existing seismic inversion methods, which require a large number of real wave velocity models for training, but wave velocity models are difficult to obtain in actual situations; secondly, they fail to effectively utilize knowledge of on-site project conditions, borehole information, and excavation exposure, and cannot fully match the actual detection situation, resulting in inaccurate information on the geostress state of the rock mass.

[0007] This invention provides a method for inverting the geostress of deep rock masses based on blasting seismic wave velocity, the specific steps of which include:

[0008] Step 1: Collect seismic wave velocity information by deploying geophones, and encode geological environment information and combine it with prior borehole information to establish and collect seismic wave velocity data for tunnels, providing a seismic wave velocity database for subsequent inversion.

[0009] Step 2: Supervised pre-training is performed using the seismic wave velocity inversion database to establish a tunnel wave velocity model inversion method and a pix2pix deep neural network model. The PD-FEM numerical calculation method and system are used to simulate the specific blasting damage throughout the entire process, and a constitutive model of surrounding rock damage under high strain rate conditions is established to simulate the blasting unloading damage effect of deep surrounding rock.

[0010] Step 3: Through the simulation of the damage effect of deep surrounding rock blasting unloading, the rock mass model data is brought into the P-wave instantaneous energy density model and the P-wave average energy density model to obtain the stress and strain state of the rock mass. This provides the PD program running conditions for the stress and strain changes under blasting loading, and then establishes the inversion optimization method for synchronous tunnel excavation and updates the inversion database.

[0011] Step 4: By updating the inversion database and synchronizing the inversion optimization method of tunnel excavation, the specific damaged parts are divided, and the relationship between the mechanical model and stress waves is linked. By controlling the blasting stress wave velocity and using the deep neural network model to pre-train the geological environment vector, the parameter-optimized tunnel inversion deep neural network model is obtained, thereby realizing the inversion of rock mass in deep buried tunnel based on blasting seismic wave velocity.

[0012] The beneficial effects of this invention are:

[0013] This invention provides a method for inverting the geostress of deep rock masses based on blasting seismic wave velocities. Based on the blasting seismic wave velocities of deep rock masses obtained from on-site monitoring, a pix2pix network is used for parameter inversion. The network takes seismic wave velocity propagation data as input and trains the corresponding true values ​​of the seismic wave velocity model, thereby establishing an inversion model and calculating certain initial parameters of the rock mass. The aim is to establish a suitable model whose prediction results are consistent with those of on-site experiments, accurately reflecting or predicting the mechanical properties of deep rock masses. For the problem of inverting blasting seismic wave velocities in deep rock masses, a dataset model conforming to actual geological conditions is set up. Through training, verification, and testing on the parameter inversion network, preliminary supervised learning network pre-training of the tunnel seismic wave velocity inversion network is successfully performed, thereby completing network parameter initialization and proving the effectiveness of the parameter inversion network, achieving geostress inversion of deep rock masses based on blasting seismic wave velocities.

[0014] Furthermore, the geostress inversion method of this invention utilizes blasting experiments to obtain seismic wave velocity and waveform data, directly inverting the geostress distribution in deep rock masses. This avoids the problems of existing geostress inversion methods that require estimation of underground medium parameters and the influence of seismic wave sources. Simultaneously, this method is applicable to geostress inversion in relatively deep rock masses, providing valuable information and support for geotechnical engineering and earthquake disaster prediction. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of the geostress inversion method for deep rock masses based on blasting seismic wave velocity, as described in this invention.

[0016] Figure 2 This is a schematic diagram of the deep rock blasting field observation and system layout of the present invention;

[0017] Figure 3 This is a diagram showing the bond breakage damage within the PD region of the present invention.

[0018] Figure 4 The present invention employs a near-field dynamic inversion model of the rock mass in a two-dimensional form.

[0019] Figure 5 The following is a flowchart illustrating the PD-FEM process used in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be described in more detail below with reference to the accompanying drawings.

[0021] This embodiment provides a geostress inversion method for deep rock masses based on blasting seismic wave velocity, which is obtained by measuring the blasting seismic wave velocity of deep rock masses on-site based on on-site monitoring.

[0022] By using a pix2pix network for parameter inversion, the network takes seismic wave velocity propagation data as input and trains the corresponding true values ​​of the seismic wave velocity model, thereby establishing an inversion model and calculating some initial parameters of the rock mass.

[0023] To address the problem of seismic wave velocity inversion in deep rock blasting, a geologically accurate dataset model was established. Through training, validation, and testing on the parameter inversion network, the network pre-training of the tunnel seismic wave velocity inversion network was successfully completed, network parameter initialization was achieved, and inversion was realized.

[0024] Using a pix2pix network for parameter inversion, input stress wave data can be transformed into an output seismic wave velocity model. In stress wave inversion, pix2pix is ​​used to convert input seismic data into a deep rock blasting seismic wave velocity model.

[0025] Specifically, pix2pix is ​​used to convert seismic data into a seismic wave velocity model for deep rock blasting.

[0026] The pix2pix tunnel inversion deep neural network model was pre-trained in a supervised manner using a pre-built tunnel inversion database, incorporating wave velocity model labels, to preliminarily determine the network model parameters.

[0027] The construction process of the pix2pix tunnel inversion database includes: using existing geological exploration reports to build a wave velocity model ahead of the tunnel, obtaining corresponding seismic observation data through numerical simulation, and combining the on-site tunnel empty mining noise signal to obtain noisy seismic data and wave velocity model that conform to the on-site characteristics, thus forming the tunnel inversion database.

[0028] Seismic wave data was used as input, and deep rock blasting seismic waves were used as output. A generator model was then trained using pix2pix to convert the input data into the output data. During training, a real deep rock blasting seismic wave velocity model could be used as the target output, allowing the generator to gradually learn how to convert seismic wave data into a deep rock blasting seismic wave velocity model.

[0029] In stress wave inversion, this method can be used to convert seismic wave data into a seismic wave velocity model for deep rock blasting, thereby more accurately predicting the properties of the underground medium.

[0030] The training set is based on the on-site geological wave velocity model of the rock mass. By using fixed seismic sources for successive blasting, geophones are fixed on each working face, and on-site seismic wave velocity signals are collected by observation. The collected on-site seismic wave velocity signals are added to the rock mass blasting record as data for the deep rock mass inversion database.

[0031] Using the obtained deep rock mass inversion database and in conjunction with actual engineering projects, a deep rock mass model for blasting that conforms to geological significance was set up. Through training and testing of the "data-model", the size, location and distribution of the deep rock mass model were reconstructed. The tunnel inversion deep neural network model was pre-trained in a supervised manner with wave velocity model labels, and the network model parameters were initially determined.

[0032] The geological environment, noise information, and observation system layout of the deep rock mass blasted are used as two additional channels to input the tunnel inversion deep neural network model together with the seismic observation data, forming a tunnel engineering geological environment vector, an environmental noise matrix, and an observation method matrix; with prior borehole information, a borehole wave velocity matrix containing borehole location and wave velocity information is generated.

[0033] The construction environment matrix and observation method matrix are used as two channels to input the tunnel pix2pix inversion deep neural network model together with the seismic observation data. The tunnel engineering geological environment vector is another input of the tunnel inversion deep neural network model. The predicted wave velocity model output by the tunnel inversion deep neural network model is used to perform parameter inversion of the measured data using a parameter inversion network. The results show that the parameter inversion network is suitable for real data and has good inversion performance, so the network parameters can be optimized and updated.

[0034] In summary, seismic wave velocity inversion using a pix2pix network can predict the properties of the subsurface medium. Combined with geological environment and observational information, accurate inversion of seismic wave velocities ahead of tunnels can be achieved. A generator model is trained, taking seismic wave data as input and generating a corresponding deep rock blasting seismic wave velocity model as output. When constructing the tunnel inversion database, a tunnel-ahead wave velocity model can be built using on-site geological exploration reports, and corresponding seismic observation data can be obtained through numerical simulation. Simultaneously, combining on-site tunnel empty mining noise signals can obtain noisy seismic data and wave velocity models. These data constitute the tunnel inversion database, used for training and validating the pix2pix network model. During training, a real deep rock blasting seismic wave velocity model is used as the target output, allowing the generator model to gradually learn how to convert seismic wave data into a deep rock blasting seismic wave velocity model. By optimizing network parameters, the inversion of tunnel seismic wave velocities is achieved. Furthermore, geological environment, noise information, and observation system deployment methods are added as additional channels along with seismic observation data to the network model, providing more constraints and contextual information, thus improving inversion accuracy.

[0035] like Figure 1 and Figure 2 As shown, a method for geostress inversion in deep rock masses based on blasting seismic wave velocity includes the following steps:

[0036] Step 1: Collect seismic wave velocity information by deploying geophones, and encode geological environment information and combine it with prior borehole information to establish and collect seismic wave velocity data for tunnels, providing a seismic wave velocity database for subsequent inversion.

[0037] In deep rock mass construction, seismic sources and geophones are deployed on the sidewalls and tunnel face to collect seismic wave signals carrying information about adverse geological structures ahead of the tunnel, enabling the detection and identification of such structures. To select the geophone installation points, five geophones are installed on each side of the sidewalls, with a 10-meter interval between each pair, and the closest geophone is 30 meters from the tunnel face. Simultaneously, numerous wave velocity models ahead of the tunnel are generated, and corresponding seismic observation data is obtained through numerical simulation, serving as label data for the tunnel inversion deep neural network model. Furthermore, a tunnel engineering geological environment vector needs to be encoded, including information such as stratum lithology, tunnel depth, geological structure, and surrounding rock grade, as well as environmental noise and observation method matrices. With prior borehole information, a borehole wave velocity matrix containing borehole location and wave velocity information is generated. Through these efforts, accurate detection and identification of adverse geological structures ahead of the tunnel can be achieved, optimizing seismic wave energy propagation and improving construction safety and efficiency.

[0038] Based on the actual conditions at the blasting site, the locations for the geophones were selected, and seismic wave data was collected using the geophones. Five geophones were distributed on each of the left and right sides of the tunnel face, with a 10-meter interval between every two geophones. The nearest geophone was located 30 meters from the tunnel face. The overall equipment and blasting sequence were arranged and illustrated.

[0039] Based on the on-site geological exploration report, a large number of tunnel front wave velocity models are automatically designed. The corresponding seismic observation data is obtained through numerical simulation, and the noise signal collected on-site is acquired. Finally, noisy seismic data and wave velocity models that conform to the on-site characteristics are obtained, forming a tunnel pix2pix inversion database. These wave velocity models serve as label data for the pre-training of the tunnel inversion deep neural network model.

[0040] The geological lithology, tunnel depth, geological structure, and surrounding rock grade of the current construction section of the tunnel are encoded to form a tunnel engineering geological environment vector; the noise from uncontrolled mining is obtained in the current construction environment, and the signal-to-noise ratio is estimated by collecting seismic data; noise information of different magnitudes is encoded to form an environmental noise matrix; the layout information of different sidewall observation systems is encoded to form an observation method matrix; and, with prior borehole information, a borehole wave velocity matrix containing borehole location and wave velocity information is generated.

[0041] In deep rock mass construction, seismic sources and geophones deployed on the sidewalls and tunnel face are used to collect seismic wave signals carrying information about adverse geological formations ahead. Data processing is then used to detect and identify these adverse geological structures ahead of the tunnel, determining the optimal placement of seismic sources and geophones. Seismic sources are also deployed on the tunnel face alongside the tunnel construction site to enhance the energy of seismic waves propagating forward, resulting in better detection performance.

[0042] Seismic wave data is collected using a rock detector inside the rock mass, including the following steps:

[0043] (1) During the propagation of blasting stress waves, the deformation of the rock mass is related to the magnitude of the stress wave amplitude. The rock mass is treated as an elastic medium within the area affected by blasting seismic waves.

[0044] (2) During the propagation of blasting seismic waves in the rock mass, the propagation speed, vibration amplitude, and waveform frequency of the blasting seismic waves are affected by the rock mechanical properties, and they carry information related to the physical and mechanical properties of the rock during the propagation process.

[0045] (3) By analyzing the measured blasting seismic wave signals, the rock mass characteristics are further obtained, including the rock mass's elastic modulus, Poisson's ratio, and rock mass integrity.

[0046] Step 2: Supervised pre-training is performed using the seismic wave velocity inversion database to establish a tunnel wave velocity model inversion method and a pix2pix deep neural network model. The PD-FEM numerical calculation method and system are used to simulate the specific blasting damage throughout the entire process, and a constitutive model of surrounding rock damage under high strain rate conditions is established to simulate the blasting unloading damage effect of deep surrounding rock.

[0047] PD stands for peridynamics, FEM for finite element method, and PD-FEM is a numerical calculation method combining peridynamics and finite element method. This embodiment optimizes network parameters through supervised pre-training on a seismic wave velocity inversion database. An inversion method for a tunnel wave velocity model is established, and a pix2pix inversion deep neural network model is used. This model utilizes seismic observation data, actual working condition information, and geological environment vectors as inputs, and improves model accuracy through additional channels of the construction environment matrix and observation method matrix. Furthermore, a full-process simulation is performed using the PD-FEM numerical calculation method and system, dividing the rock mass model into a peridynamic zone, a finite element zone, and a coupling zone to simulate force and displacement transmission and rock mass fracture. A constitutive model of surrounding rock damage under high strain rate conditions is also established, and finite element simulation is performed using PD-FEM to simulate the damage effect of deep surrounding rock blasting unloading.

[0048] Based on the seismic wave data obtained in step 1, a deep neural network model for tunnel pix2pix inversion is constructed. The input of the network is the seismic observation data and actual working condition information from step 1, and the output is the predicted wave velocity model.

[0049] Based on the seismic wave data obtained in step 1, a tunnel pix-to-pixel inversion deep neural network model is constructed. This model is computed using the loss function of CGAN and optimized using an optimization function based on wave velocity data. The loss function includes the CGAN loss function L... CGAN (G, D) and a loss function L1 representing image differences. From the perspective of the loss function, D is the CGAN discriminator, and its role remains unchanged. G is the CGAN generator, and its role is to sharpen the wave velocity data. The loss function of CGAN is:

[0050] L cGAN (G,D)=E x,y [logD(x,y)]+E x,z [log(1-D(x,G(x,z)))]

[0051] L L1 (G)=E x，y，z [||yG(x,y)||1]

[0052] The optimization function for wave velocity data is:

[0053]

[0054] Where: λ is the balance discriminant and L L1 The impact of distance on generator training. In the formula: x is the wave velocity data input to the generator; y is the label wave velocity data corresponding to x. In the network structure, x serves as the input wave velocity data to G, G(x) is the generated wave velocity data corresponding to x, and y is the label wave velocity data corresponding to x. G transforms the input wave velocity data to its corresponding output wave velocity data. G is a U-shaped network, and the network depth is adjusted by the resolution of x and y. x, G(x), and y are concatenated and input into D. x concatenated with G(x) is judged as false by D, and x concatenated with y is judged as true by D.

[0055] The tunnel engineering geological environment vector is embedded into the network structure as another input through a fully connected structure. The construction environment matrix and observation method matrix are input into the network as two additional channels along with the input seismic observation data. For borehole information, the network output and borehole wave velocity matrix are used to calculate the loss function to update the network parameters. The tunnel inversion deep neural network model is pre-trained in a supervised manner with wave velocity model labels through the tunnel inversion database to preliminarily determine the network parameters.

[0056] The PD-FEM numerical calculation method and system for full-process simulation divides the rock mass calculation model into a near-field dynamic zone, a finite element zone, and a coupling zone based on whether damage occurs. A model of appropriate size is established according to actual engineering conditions, and the model is divided into three regions: the region where rock mass fracture may occur is the near-field dynamic zone, the region far from the crack is the finite element zone, and the junction of the two regions is the coupling zone. Force and displacement are transferred in the coupling zone. The gravity and tectonic stress of the overlying rock strata are transformed into force boundary conditions in the finite element region, and constraints are transformed into displacement boundary conditions. The nodal displacements of the finite element region are calculated and used as the boundary conditions of the near-field dynamic zone. Under the above boundary conditions, the displacement of material points and damage are iteratively calculated to solve the deformation characteristics of the rock mass, achieving engineering-scale simulation of rock mass fracture.

[0057] The entire process is numerically calculated using PD-FEM. By dividing the crack damage state of each part of the rock mass model into different regions, the FEM part models the undamaged part, and the PD part models the failed area. They are directly coupled to simulate the propagation failure of deep rock mass cracks.

[0058] The approach involves direct coupling of two parts, and the steps are as follows:

[0059] First, initialize the rock mass model and parameters, and determine the total number of particles n. p Total number of time steps n tSecondly, determine the bonds between each particle; the total number of bonds around the i-th particle is n. b (i). After applying initial conditions and determining whether to add dynamic relaxation, if relaxation is applied, the time step Δt is equal to 1 and the t-th iteration begins and boundary conditions are applied, thereby calculating the total PD force F(i) on the i particles within the near field range.

[0060] Using the known displacement and velocity fields at time step i (i < n), the displacement of the points within the overlapping region is calculated using the displacement of the points within the overlapping region. The force density at the points within the overlapping region is calculated. The force density is applied to the finite elements of the overlapping region in the form of body forces. The displacement at the (n+1)th time step is obtained by integration. The numerical method of the near-field dynamics equation is used. The force generated by the interaction between the embedded near-field dynamics material points and the material points outside the connecting elements is called the coupling force. It is distributed to the nodes of the connecting elements located at the interface through shape functions as part of the internal forces acting on the nodes.

[0061] The domain is discretized into many material points, each associated with its volume, and the union of all volumes forms the entire volume of the object. Each discretized material point x... i Constitutive equation:

[0062]

[0063] Among them, u i u is the displacement in the i-direction at time t; j x represents the displacement in the j direction at time t; i Let x be a matter node; x j For the neighborhood H of x xj Any node within; V' j Let b represent the volume of a point of matter in the j-direction; j Let the force be in the j-direction.

[0064] In the key displacement after rock mass blasting, the external force acting on the material point can be written as:

[0065]

[0066] Where F is the external force acting on the particle; c is the micromodulus; V' j V is the volume of the material point in the j-direction; i Let η be the volume of the material point in the i-direction; η be the relative displacement; and ζ be the bond vector.

[0067] The relative displacement and relative position of the material point are expressed as:

[0068]

[0069] Among them, u i u represents the displacement of point i in the direction of matter;j The displacement of the material point j in the direction of l; ij The point bond length of the substance.

[0070] Since point forces are equal in magnitude and opposite in direction in bond-based PD theory, the stiffness matrix between material points is:

[0071]

[0072] Where k is the stiffness between material points; b is the body force of the material points.

[0073] To achieve direct-projection coupling, the discrete PD motion equations can be rewritten as:

[0074]

[0075] Where U is the displacement vector of the PD particle; vector F is the sum of internal and external forces; subscript p indicates a variable related to the PD region; single and double underscores indicate variables located outside and inside the overlapping region, respectively; parameter c n Let represent the damping coefficient at the nth time step. The coefficients of the virtual diagonal density matrix D can be determined using Greschgorin's theorem.

[0076] By directly assembling the finite element equations without constructing a global stiffness matrix, the coupling between the FEM and PD is achieved. The FEM model is then:

[0077]

[0078] Where U is the displacement vector of the PD particle; vector F is the sum of internal and external forces; subscript p indicates a variable related to the PD region; single and double underscores indicate variables located outside and inside the overlapping region, respectively; parameter c n Let M represent the damping coefficient at the nth time step; the coefficients of the virtual diagonal density matrix M are determined by Greschgorin's theorem.

[0079] For surrounding rock damage under high strain rate conditions, the dynamic damage constitutive models of rock materials include: static loading state model (linear elastic stage), blast loading state model (plastic stage), blast unloading state model (plastic stage), yield function model, damage evolution model, volumetric strain model, and equivalent stress model, as detailed below:

[0080] (1) Loading the static state model (linear elastic stage):

[0081]

[0082] Where K is the bulk modulus, p c The hydrostatic pressure of the rock under uniaxial compression is expressed in N and μ. cFor volumetric strain.

[0083] (2) Explosive loading state model (plastic stage):

[0084]

[0085] Explosive unloading state model (plastic stage):

[0086]

[0087] Where, p max The maximum hydrostatic pressure reached before unloading, N; μ max K is the volumetric strain, and K1 is the bulk modulus of the plastic stage.

[0088] (3) Yield function model:

[0089]

[0090] Where A is the viscosity coefficient; D is the damage variable; B is the pressure intensification coefficient; C is the strain rate sensitivity coefficient; f is the yield coefficient; σ eq The critical stress for crack propagation in the model is given in MPa and K. s P is the fracture strength factor of the model. * ε is the gas pressure in the model crack, N; r is the borehole radius, cm; ε * This is the equivalent plastic strain.

[0091] The equation consists of the equivalent stress defined by the difference between the dimensionless equivalent stress and the compressive and tensile strengths of the rock material.

[0092] (4) Damage evolution model:

[0093]

[0094] Where D is the damage variable, For the equivalent plastic strain increment, Δμ p For the plastic volumetric strain increment, The equivalent plastic strain at material failure. For volumetric plastic strain.

[0095] (5) Satisfies volumetric strain, where V and V0 are the volumes before and after deformation, respectively, and ε x ε y ε z Let be the strain components. The volumetric strain model is:

[0096]

[0097] Where, ε x ε represents the strain component in the x-direction of the model.y ε represents the strain component in the y-direction of the model. z Let z be the strain component in the z-direction of the model.

[0098] (6) The equivalent stress model is:

[0099]

[0100] Where, σ x For the stress components in the x-direction of the model; σ y σ represents the stress component in the y-direction of the model. z Let z be the stress component in the z-direction of the model.

[0101] After completing step 2, a constitutive model of surrounding rock damage under high strain rate conditions is established. The constitutive subroutine of the material is written in Fortran language, and the finite element simulation of the material constitutive model is realized through PD-FEM. Finally, the finite element simulation of the blasting unloading damage effect of deep surrounding rock is carried out based on the newly established constitutive model.

[0102] Step 2 enables automated wave velocity model updates and calculations of rock mass stress and strain states as tunnel excavation progresses, thus providing a more accurate basis for simulation and optimization design of tunnel engineering.

[0103] Steps 1 and 2 provide the conditions for the PD program to run in step 3.

[0104] Step 3: Through simulation of the damage effect of deep surrounding rock blasting unloading, the rock mass model data is input into the P-wave instantaneous energy density model and the P-wave average energy density model to obtain the stress-strain state of the rock mass. This provides the PD program running conditions for stress-strain changes under blasting loading, helps to establish an inversion optimization method for synchronous tunnel excavation, and updates the inversion database.

[0105] After simulating the damage effects of deep surrounding rock blasting and unloading, a network-updated inversion optimization method for synchronous tunnel excavation is established. During tunnel excavation, the tunnel wave velocity model is automatically redesigned using the new excavation findings, and corresponding observation data is generated. This can replace the previously designed wave velocity model in the tunnel pix2pix inversion database, enabling the updating of labeled data in the database. Furthermore, the established rock mass model data is incorporated into the energy density formula. First, under constant force, the stress-strain state of the rock mass is obtained. Then, under blasting loading, the changes in stress and strain of the rock mass over time are analyzed.

[0106] A network-updated synchronous tunnel excavation inversion optimization method is established. As tunnel excavation progresses, the new excavation reveal results are used to automatically redesign the tunnel wave velocity model and generate corresponding observation data to replace the previously designed wave velocity model in the tunnel pix2pix inversion database, thereby updating the label data in the tunnel inversion database.

[0107] Stress waves generate circumferential tensile and shear stresses in rock. Under the action of these circumferential tensile and shear stresses, combined with cracks formed after rock blasting, the circumferential stress caused by stress wave propagation exceeds the dynamic tensile strength of the rock. When this stress exceeds the dynamic tensile strength of the rock, cracks in the rock are activated. After cracks form, the number of activated cracks under the action of stress waves exhibits an exponential distribution. The relationship between the number of rock cracks and the volumetric strain of the cracked rock under the action of stress waves is as follows:

[0108]

[0109] Where n is the number of cracks generated; ε v It is volumetric strain; A and m are material-related coefficients.

[0110] Radial cracks do not propagate uniformly; the dominant crack will emerge from the crack cluster. Under stress wave action, crack propagation is modeled using a planar crack model. During rock blasting, the energy release rate is obtained, and the energy release rate during rock blasting under stress wave action is expressed as follows:

[0111]

[0112] Where K1 is the stress intensity factor; E m G represents the dynamic elastic modulus of rock, expressed in GPa. I denoted as energy release rate; μ is the Poisson's ratio of the rock mass.

[0113] Crack propagation Δ α When the stress intensity factor at the crack tip is equal, the crack propagation Δ is obtained. α Shear stress σ during the process θ The work done is:

[0114]

[0115] Where n is the number of radial cracks; r is the distance between the rock mass particles and the blast center, in cm; Δa is the change in r, in cm. During the blasting process, as the distance between the rock mass particles and the blast center gradually increases, the fracture characteristics of the rock mass will also change accordingly.

[0116] In the plane, since the strength is σ i Plane longitudinal waves with c p Since the velocity propagates in any direction within the plane, the instantaneous energy density e of the longitudinal wave is... p The model is:

[0117]

[0118] Among them, e p The instantaneous energy density is expressed in kJ / m³. 3 ;σ i c represents the intensity of a plane longitudinal wave, in km / s. p ρ is the longitudinal wave velocity in km / s; ρ is the density of the rock mass in g / cm³.

[0119] Longitudinal wave average energy density The model is:

[0120]

[0121] in, The average energy density of the longitudinal wave is kJ / m 3 ρ is the density of the rock mass, g / cm³; c p σ is the longitudinal wave velocity in km / s; T is the longitudinal wave propagation time; i x represents the intensity of a plane longitudinal wave, in km / s; i Displacement during the propagation of a plane longitudinal wave.

[0122] Step 4: By updating the inversion database and synchronizing the inversion optimization method of tunnel excavation, specific damage areas are divided by PD-FEM numerical calculation, and the mechanical model is associated with stress waves. By controlling the blasting stress wave velocity and using the deep neural network model to pre-train the geological environment vector, the parameter-optimized tunnel inversion deep neural network model is obtained, thereby realizing the inversion of rock mass in deep-buried tunnels based on blasting seismic wave velocity.

[0123] By updating the inversion optimization method for synchronous tunnel excavation via a network, the following steps are taken: First, a dynamic relaxation algorithm is used to calculate the static stress field under initial stress and output displacement data. The displacement data is then used to initialize the model, followed by dynamic calculations. The damaged area is then divided into PD (peripheral dynamics) and FEM (finite element method) sub-regions, considering bond interactions and damage value updates. The mechanical model is then correlated with stress waves, and the blasting influence range is controlled by adjusting the blasting stress wave velocity. Simultaneously, the tunnel's engineering geological environment vector is used as network input, and a deep neural network model is pre-trained. As tunnel excavation progresses, the tunnel wave velocity model is redesigned using the new excavation findings, and corresponding observation data is generated to update the tunnel inversion database. Finally, the optimized deep neural network model is used to optimize the seismic wave velocity inversion in the area to be constructed ahead of the tunnel.

[0124] like Figure 3 As shown, write a numerical calculation program. The algorithm flow of the program design is as follows:

[0125] First, the dynamic relaxation algorithm is used to calculate the static stress field of the model under the initial stress. Then, the displacement data of each mass point in the model under the static stress field is output. The model is then initialized with the mass point displacement data obtained under the static stress field, and then dynamic calculation is performed.

[0126] like Figure 4 As shown, the specific damage region is divided using the PD-FEM numerical calculation method, dividing the research object into PD sub-regions and FEM sub-regions. In the non-local interaction region, the research object is further divided into PD and FEM sub-regions. Peri-field dynamics modeling is used in the non-local interaction region, while four-node modeling is used in other regions, without setting an overlap region. Peri-field dynamics modeling is used, while four-node isoparametric element modeling is used in other regions. An overlap region is set between the two types of sub-regions. In bond-type PD theory, the interaction between two material points only considers the bond force between the two material points. In state-type PD theory, when analyzing the interaction between two material points, the influence of other material points in the neighborhood on the bond force is considered.

[0127] Specifically, the key uses four nodes to achieve hybrid modeling:

[0128] The element stiffness can be taken as:

[0129]

[0130] Where ξ′ is the relative distance between the finite element node and the material point; k p denoted as , where is the overall element stiffness; k is the element stiffness; l is the point bond length; and m is the point mass.

[0131] The force vector state of PD is uniformly expressed as:

[0132] TXT]<X'-X> =t<ξ>M(Y)

[0133] ξ = x' - x, η = u' - u

[0134] Among them, PD theory is a nonlocal continuous medium mechanics theory that uses integral expressions to represent the interaction forces between material points x and x′ within a neighborhood radius δ.

[0135] The kinetic equation of matter point x is expressed as:

[0136]

[0137] Where ρ is the density of node x; u is the displacement at time t; b(x,t) is the body force; x′ is any node in the neighborhood Hx of x; ζ is the bond vector; η is the relative displacement; T[x,t] and T[x',t] are the force vector density states of nodes x and x', respectively.

[0138] like Figure 5 As shown, in PD theory, the interaction between two material points only considers the bond force between the two material points. However, when analyzing the explosive interaction between two material points, PD theory considers the influence of other material points in the neighborhood on the bond force. If it is determined that the bond is not broken, the point-to-point force between the particles is calculated, and the damage value is updated. A second check is performed to determine whether dynamic relaxation should be applied. If any bond is broken, the dynamic relaxation algorithm is used to update the particle displacement, thus completing the algorithm for PD as a nonlocal mechanical model.

[0139] By linking the mechanical model with the stress wave, and through the analysis of a single parameter of the stress wave, the influence of stress wave energy on crack propagation is clarified. In blasting, the blasting influence range can be increased by extending the action time, and the development of stress waves can be controlled by selecting appropriate types and proportions of blasting methods, thus making the blasting stress wave velocity controllable.

[0140] As data acquisition depth extends into deeper areas, the tunnel pix-to-pix inversion database is expanded using on-site noise data obtained from blasting vibration meters and seismic observation data. Dedicated blasting vibration testing instruments are used to test vibrations caused by blasting, monitoring indicators such as peak vibration, dominant frequency, and vibration duration. Through blasting vibration monitoring, the attenuation law of blasting seismic waves and the relationship between seismic wave parameters and blasting methods can be measured, allowing for targeted optimization of problems in blasting design. This transforms the tunnel inversion database from an initial database entirely labeled with wave velocity models to a partially unlabeled semi-supervised database, gradually approaching an unsupervised database. The expanded inversion database is used to conduct several rounds of additional training on the tunnel inversion neural network using a seismic wave forward modeling physics-driven approach to adjust network parameters. This training process gradually becomes unsupervised as exploration progresses and the database expands.

[0141] A rock failure process analysis program based on the finite element method was used to simulate the fracture process of rock with pre-existing cracks under the impact load of a borehole explosion. Different explosion stress waveforms were generated by using different types and proportions of explosives and different charging schemes. For the simulation of explosion stress waves, a semi-theoretical and semi-empirical simplification method was used to simplify the explosion stress time history curve, and the results were compared with typical simplified models.

[0142] Specifically, during the calculation, since the solid part of the finite element region only undergoes elastic deformation without damage or fracture, the permeability of the fluid mesh nodes corresponding to the finite element region remains unchanged, which is the permeability of the rock matrix. Material points in the solid domain, transition domain, and fracture domain of the near-field dynamic region of the solid part need to be considered.

[0143] After the calculation is completed, the displacement and damage data of the finite element nodes and material points in the model can be obtained. Based on the data, deformation and damage diagrams during the blasting process can be plotted, thereby obtaining the laws and states of seismic waves in deep rock blasting. A tunnel inversion deep neural network model is constructed. The input of the network is seismic observation data and actual working condition information, and the output is a predicted wave velocity model.

[0144] The tunnel engineering geological environment vector is embedded into the network structure as another input through a fully connected structure. The construction environment matrix and observation method matrix are input into the network as two additional channels along with the input seismic observation data. For borehole information, the network output and borehole wave velocity matrix are used to calculate the loss function to update the network parameters. The tunnel inversion deep neural network model is pre-trained in a supervised manner with wave velocity model labels through the tunnel pix2pix inversion database to preliminarily determine the network parameters.

[0145] A network-updated synchronous tunnel excavation inversion optimization method is established. As tunnel excavation progresses, the new excavation reveal results are used to automatically redesign the tunnel wave velocity model and generate corresponding observation data to replace the previously designed wave velocity model in the tunnel inversion database, thereby updating the label data in the tunnel inversion database.

[0146] This approach allows for ongoing exploration during excavation, expanding the tunnel inversion database with newly acquired field noise data and seismic observation data. This transforms the database from an initial fully labeled database with wave velocity models to a partially unlabeled, semi-supervised database, gradually approaching an unsupervised database. The expanded inversion database is then used to perform several rounds of additional training on the tunnel pix2pix inversion neural network using a seismic wave forward modeling physics-driven approach to adjust the network parameters. This training process gradually becomes unsupervised as exploration progresses and the database expands.

[0147] The basic network parameters for the current excavation stage, obtained after the above optimization, are backed up: Individual iterative optimization is performed using observation data from the current tunnel excavation location to determine the forward seismic velocity conditions reflected in the current tunnel construction location's detection data. This single data point is used to iteratively update the network parameters, generating the velocity distribution of the detection area ahead of the tunnel at the current location. Before the next excavation and detection, the backed-up basic network parameters are restored, and the optimized tunnel inversion deep neural network model is used to achieve the seismic wave velocity inversion task for the area to be constructed ahead of the tunnel. This effectively utilizes data from on-site project conditions, borehole information, and excavation exposure, matching it with actual detection conditions to obtain accurate rock mass in-situ stress state information, thereby realizing deep rock mass inversion based on blasting seismic wave velocity.

[0148] The geostress inversion method for deep rock masses based on blasting seismic wave velocity in this embodiment is integrated into a computer program system that combines experimental and inversion methods, and is implemented in software form.

[0149] This invention is based on the aforementioned method for inverting the geostress of deep rock masses based on blasting seismic wave velocity. By controlling the entire simulated blasting experiment, the simulation process is divided into three parts, each simulated through blasting. During this process, the strain and stress changes of the deep rock mass are recorded, providing data for subsequent finite element analysis and creating conditions for subsequent pix2pix inversion, thus completing the entire experimental inversion process. This process is summarized and integrated into a computer program system that combines experimental and inversion methods.

[0150] Specifically, the system is manifested as a computer-readable storage medium for storing computer instructions. When these instructions are executed by a processor, they complete the steps in the PD-FEM numerical calculation method for simulating the entire process of deep rock mass fracturing. Its usage and storage are implemented in software form and can be stored in a computer-readable storage medium. Therefore, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, is essentially embodied in the form of a software product.

Claims

1. A method for inverting in-situ stress in deep rock masses based on blasting seismic wave velocity, characterized in that, The specific steps of this method include: Step 1: Collect seismic wave velocity information by deploying geophones, encode geological environment information and combine it with prior borehole information to establish a database of seismic wave velocity data for tunnels, and obtain a seismic wave velocity database for inversion. Step 2: Supervised pre-training is performed using the seismic wave velocity inversion database to establish a tunnel wave velocity model inversion method and a pix2pix deep neural network model. The PD-FEM numerical calculation method and system are used to simulate the full-process specific blasting damage, and a constitutive model of surrounding rock damage under high strain rate conditions is established to simulate the blasting unloading damage effect of deep surrounding rock. Step 3: Through the simulation of the damage effect of deep surrounding rock blasting unloading, the rock mass model data is brought into the P-wave instantaneous energy density model and the P-wave average energy density model to obtain the stress and strain state of the rock mass. This provides the PD program running conditions for the stress and strain changes under blasting loading, and then establishes the inversion optimization method for synchronous tunnel excavation and updates the inversion database. Step 4: By updating the inversion database and synchronizing the inversion optimization method of tunnel excavation, specific damage areas are divided by PD-FEM numerical calculation, and the mechanical model is associated with stress waves. By controlling the blasting stress wave velocity and using the deep neural network model to pre-train the geological environment vector, the parameter-optimized tunnel inversion deep neural network model is obtained, thereby realizing the inversion of rock mass in deep buried tunnel based on blasting seismic wave velocity. In step 2, the pix2pix inversion deep neural network model is computed using the loss function of CGAN and optimized using the optimization function of wave velocity data; Loss functions include the loss function L of CGAN. CGAN (G,D) and the loss function L1 representing image differences, where D is the CGAN discriminator and G is the CGAN generator; The loss function of CGAN is: ； ； The optimization function for wave velocity data is: ； Where: λ is the balance discriminant and L L1 The effect of distance on generator training; x represents the wave velocity data input to the generator; y represents... The corresponding label wave velocity data; in the network structure, x serves as the input wave velocity data for G, and G(x) is related to... The corresponding generated wave velocity data, y is related to The corresponding label wave velocity data, G transforms the input wave velocity data into its corresponding output wave velocity data. G is a U-shaped network, and the depth of the network is adjusted by the resolution of x and y. In step 3, the instantaneous energy density model of the longitudinal wave is: ； Among them, e p The instantaneous energy density is expressed in kJ / m³. 3 ;σ i c represents the intensity of a plane longitudinal wave, in km / s. p ρ is the longitudinal wave velocity in km / s; r is the density of the rock mass in g / cm³. Longitudinal wave average energy density The model is: ； in, The average energy density of the longitudinal wave is kJ / m 3 ; r is the density of the rock mass, g / cm³; c p σ is the longitudinal wave velocity in km / s; T is the longitudinal wave propagation time; i x represents the intensity of a plane longitudinal wave, in km / s; i Displacement during the propagation of a plane longitudinal wave.

2. The in-situ stress inversion method for deep rock masses based on blasting seismic wave velocity according to claim 1, characterized in that, In step 1, collecting seismic wave velocity information by deploying detectors includes: (1) During the propagation of blasting stress waves, the deformation of the rock mass is related to the magnitude of the stress wave amplitude. The rock mass is treated as an elastic medium within the area affected by blasting seismic waves. (2) During the propagation of blasting seismic waves in the rock mass, the propagation speed, vibration amplitude and waveform frequency of the blasting seismic waves are affected by the rock mechanical properties, and they carry information related to the physical and mechanical properties of the rock during the propagation process; (3) By analyzing the measured blasting seismic wave signals, the rock mass characteristics are further obtained, including the rock mass's elastic modulus, Poisson's ratio, and rock mass integrity.

3. The in-situ stress inversion method for deep rock masses based on blasting seismic wave velocity according to claim 1, characterized in that, In step 2, the constitutive model of surrounding rock damage under high strain rate conditions includes: a static loading state model, a blasting loading state model, a blasting unloading state model, a yield function model, a damage evolution model, a volumetric strain model, and an equivalent stress model.

4. The in-situ stress inversion method for deep rock masses based on blasting seismic wave velocity according to claim 3, characterized in that, The static state model under load is as follows: ； Where K is the bulk modulus, p c The hydrostatic pressure of the rock under uniaxial compression is expressed in N and μ. c For volumetric strain; Explosive loading state model: ； Explosive unloading state model: ； Where, p max The maximum hydrostatic pressure reached before unloading, N; μ max K is the volumetric strain; K1 is the bulk modulus of the plastic stage; Yield function model: ； Where A is the viscosity coefficient; D is the damage variable; B is the pressure intensification coefficient; C is the strain rate sensitivity coefficient; f is the yield coefficient; σ eq The critical stress for crack propagation in the model is given in MPa and K. s P is the fracture strength factor of the model. * ε is the gas pressure in the model crack, N; r is the borehole radius, cm; ε * Equivalent plastic strain; Damage evolution model: ； Where D is the damage variable; This represents the equivalent plastic strain increment. This represents the plastic volumetric strain increment; The equivalent plastic strain at material failure; For volumetric plastic strain; Satisfies volumetric strain, where V and V0 are the volumes before and after deformation, respectively, and ε x ε y ε z For the strain components; the volumetric strain model is: ； Where, ε x ε represents the strain component in the x-direction of the model. y ε represents the strain component in the y-direction of the model. z The strain component in the z-direction of the model; The equivalent stress model is as follows: ； Where, σ x For the stress components in the x-direction of the model; σ y σ represents the stress component in the y-direction of the model. z Let z be the stress component in the z-direction of the model.

5. The in-situ stress inversion method for deep rock masses based on blasting seismic wave velocity according to claim 1, characterized in that, In step 4, the mechanical model is associated with stress waves specifically as follows: The key uses four nodes to achieve hybrid modeling: The element stiffness can be taken as: ； Where ξ′ is the relative distance between the finite element node and the material point; k p ρ is the overall element stiffness; k is the element stiffness; l is the point bond length; m is the point mass. The force vector state of PD is uniformly expressed as: ； ； The kinetic equation for a point x is expressed as: ； Where ρ is the density of node x; u is the displacement at time t; b(x, t) is the body force; x′ is any node in the neighborhood Hx of x; ζ is the bond vector; η is the relative displacement; T[x, t] and T[x′, t] are the force vector density states of nodes x and x′, respectively.

6. A computer device, characterized in that: The computer device includes a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the geostress inversion method for deep rock masses based on blasting seismic wave velocity as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program containing a geostress inversion method for deep rock masses based on blasting seismic wave velocity. When the geostress inversion method for deep rock masses based on blasting seismic wave velocity is executed by a processor, it implements the geostress inversion method for deep rock masses based on blasting seismic wave velocity as described in any one of claims 1 to 5.

Images