Coal mining digital twinning method based on multi-level dynamic deviation correction and AI optimization

Abstract

The invention provides a coal mining digital twinning method based on multi-level dynamic deviation correction and AI optimization. The method comprises the steps of monitoring data acquisition and dynamic preprocessing, and performing denoising and space consistency detection on data; constructing and pre-processing the coal mine digital twinborn body; generating a nonlinear deviation field by adopting a high nonlinear variable interpolation and model mapping method; performing Kalman filtering error analysis and correction under the continuous analysis step, and marking an optimization model state in combination with a space mutation step; performing staged sliding window long-time-sequence fluctuation analysis, and classifying and marking regions according to an absolute value of an error and fluctuation; performing multi-level weighted error optimization based on a genetic algorithm, and adjusting material parameters and boundary conditions; carrying out material constitutive rectification based on dynamic Bayesian, and gradually enhancing the adaptability and accuracy of the model; and constructing a hybrid network structure AI optimization prediction model based on deep data driving, capturing the dynamic change of the material constitutive and the characteristic difference of the model constitutive, and realizing the precise digital twinborn simulation of the whole coal mining process.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mining engineering, and in particular to a coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization. Background Art

[0002] During coal mining, complex geological conditions, rock movement and changes in stress distribution bring huge safety risks. Therefore, simulation and simulation technology in the mining process play an important role in predicting potential geological disasters and guiding mining operations. Current coal mining simulation mainly relies on static processes, which are constructed based on previous geological surveys and exploration data to perform unsupervised simulation of complex mining environments. However, it has the following significant limitations:

[0003] 1. Lack of deep integration of dynamic correction and real-time monitoring data: Existing simulation technologies usually build geological models based on static data before mining to simulate stress distribution and material response. However, during the mining process, geological conditions and stress fields will continue to change with the progress of the operation, and the existing static models cannot capture and update these changes in time. Due to the lack of deep integration with real-time monitoring data, the mining process and stress distribution in the model are usually difficult to reflect the actual situation, resulting in low prediction accuracy.

[0004] 2. The application of digital twin technology is not yet in-depth: Although digital twin technology has been widely used in industrial manufacturing, urban management and other fields, its application in the field of coal mining is still in its early stages, especially in the dynamic correction and nonlinear stress change processing. Existing coal mine simulation technology is limited to limited numerical simulation, lacks real-time feedback on the mining process and refined digital twin process construction, and cannot provide accurate virtual-real mapping to adapt to changes in geological conditions.

[0005] 3. Insufficient dynamic parameter adjustment and multi-level data optimization capabilities: In the coal mining environment, geological mutations at different levels, material soft-hardening effects, and stress concentration phenomena gradually emerge during the mining process. Existing static simulation models are difficult to adjust these parameters in real time, and react slowly to abnormal signals at monitoring points and sensitive areas of stress concentration. At the same time, the calculation results of static models are often not optimized in real time for dynamic monitoring data, and lack multi-level parameter correction and anomaly detection mechanisms, which makes them insufficient in long-term time series analysis and dynamic prediction.

[0006] In summary, there is little research on the real in-depth use of digital twin technology for supervised, interactive dynamic twin simulation of coal mining. Summary of the invention

[0007] The applicant of the present invention has found through research that the application of digital twin technology in the process of coal mining can realize real-time data correction, high-precision simulation prediction, etc., which significantly improves the transparency and accuracy of the mining process. However, the existing simulation methods of coal mining are mostly static simulations, lacking dynamic fusion with real-time monitoring data, and less optimization of material parameters and boundary conditions with complex nonlinear changes. In response to this technical demand, the present invention provides a digital twin method for coal mining based on multi-level dynamic correction and AI optimization. The method first performs dynamic collection and denoising of monitoring data, constructs a geological digital twin model, generates a nonlinear deviation field and performs accurate data mapping; in the simulation process, a Kalman filter is used to dynamically correct errors, and a sliding window analysis is used to identify mutation areas; for high-volatility areas, genetic algorithms are further used to optimize material parameters and boundary conditions, and dynamic Bayesian methods are combined to gradually adjust material constitutive properties to achieve high-precision fitting under complex geological conditions; finally, a deep data-driven AI network model is used to dynamically optimize and predict material constitutive and model constitutive, complete real-time update of virtual-real mapping, and significantly improve the digital twin accuracy and decision support capabilities of coal mining.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] The digital twin method of coal mining based on multi-level dynamic correction and AI optimization includes the following steps:

[0010] S1. Monitoring data acquisition and dynamic preprocessing: Determine the acquisition frequency of the sensor and calibrate the timestamp, perform structured settings on the data collected before processing, set the upper and lower noise thresholds of mean ±3 times the standard deviation for the monitoring points, and perform spatial consistency detection, analyze the correlation between the relative positions of multiple sensors and stress values, adjust the threshold strength according to the fluctuation adaptively, and perform smooth interpolation repair on the data of abnormal points;

[0011] S2. Construction and pre-processing of digital twins of coal mines: Through the rapid modeling method of drilling, exploration, and geophysical unstructured data, the true three-dimensional state of coal mine geology is constructed, and the model is adaptively meshed, the initial ground stress field is balanced, the static boundary conditions are set, and the material parameters are assigned to provide a digital foundation for twin simulation;

[0012] S3, high nonlinear variable interpolation and model mapping: for the monitoring data points in step S1, based on the position of the model grid node, a continuous nonlinear deviation field is generated by radial basis function interpolation, the model coordinate system is spatially aligned, the coordinates of each grid node are matched to complete the numerical mapping, and the bilinear interpolation method is used to further smooth the transition;

[0013] S4, Kalman filter error analysis and correction under continuous analysis steps: In each calculation step, the simulation calculation results and monitoring data are subjected to Kalman filter analysis, the Kalman gain is calculated and the state value and covariance matrix are updated, the state value and error of the model are dynamically adjusted through the prediction-correction process, the weight is adaptively adjusted according to the error fluctuation, and the spatial analysis of all monitoring point data is performed to calculate the spatial coefficient of variation in the entire monitoring point set, that is, the ratio of the standard deviation to the mean, and mark the spatial mutation step that exceeds the threshold, providing a hierarchical processing basis for the subsequent optimization algorithm;

[0014] S5. Staged sliding window long time series fluctuation analysis: After a working face is mined, a dynamic sliding window is set in the geological structure and stress concentration area where error fluctuations are prone to occur, and the time series error data in the window is processed by rolling average. The results are classified as abnormal according to the large absolute value of the error and the large error fluctuation. The area with continuous large error or frequent fluctuations in the window is detected by the increasing-decreasing threshold method. If the error fluctuation area contains the spatial mutation step marked in step S4, it will be given a higher weight value in the subsequent optimization;

[0015] S6. Multi-level weighted error optimization based on genetic algorithm: taking error convergence degree as fitness target, calculating mean square error between each monitoring point and interpolation point, forming initial population with different material parameters and boundary condition combinations, setting weights for spatial mutation step in step S4, large absolute error value and large error fluctuation in step S5, analyzing long-term dynamic characteristics of error according to error time series trend after rolling average, adjusting local material parameters for large error fluctuation and mutation marked areas based on weights, fine-tuning boundary conditions for areas with large absolute error values, maintaining the amplitude at 5% to 10% to maintain convergence stability, and obtaining initial boundary state and material properties for the next step after optimization;

[0016] S7. Correction of material constitutive parameters based on dynamic Bayesian: The optimized results of step S6, geological exploration, historical monitoring data and known material constitutive model library are used as the initial prior distribution of material constitutive parameters. A likelihood function is constructed based on the difference between simulated stress-strain and monitoring data in multiple steps to reflect the soft and hardened nonlinear response characteristics of materials under different stress states. Based on the combination of the two, the posterior distribution is updated in the next mining stage of the working face. As the mining of the working face deepens, the new material constitutive characteristics are input into the model as the prior distribution of the next stage, gradually enhancing the adaptability and accuracy of the model to complex nonlinear and high-stress geological conditions, forming a dynamic feedback-multiple iterations of progressive optimization effect;

[0017] S8. Construct an AI optimization prediction model based on a hybrid network structure driven by deep data: use the optimization and monitoring data generated in steps S4 to S7 as input, and use the recurrent neural network based on the nonlinear relationship between stress and strain to capture the progressive changes of material parameters on the time axis; use a convolutional neural network to extract the boundary and material spatial distribution characteristics, and integrate their time series and spatial characteristics; based on supervised learning and dynamic Bayesian updating, construct a dynamic material constitutive and model constitutive property update mechanism to accurately describe the dynamic property changes of geology under mining, and continuously improve the ability to accurately map the virtual and real states of dynamic twin mining in coal mining.

[0018] Furthermore, in step S1, the potential time drift error inside the device is eliminated by comparing and correcting the clocks inside and outside the sensor, the collected raw data is structured and normalized to the minimum and maximum values, and the data of each sensor is correlated by combining the Pearson correlation coefficient with the mutual information. When the stress value monitored by the sensor fluctuates frequently, the threshold is automatically relaxed to adapt to higher dynamic changes, and vice versa. The regional correlation interpolation method is used for multi-point anomalies, and the stress values ​​of adjacent time and space nodes are fully considered to ensure the accuracy of abnormal point repair. The end time and initial time of each step are set to ensure that the end time of the calculation step is accurately aligned with the acquisition time, and a smaller time step is selected by adjusting the time integration step to better capture nonlinear behavior while meeting the acquisition interval frequency.

[0019] Furthermore, in step S3, radial basis function interpolation is used to generate a nonlinear deviation field from the monitoring point to the model node, so that it has global continuity and local accuracy. The radial basis function is expressed as follows:

[0020]

[0021] Where r represents the coordinate of the target grid node; r i Represents the coordinates of the monitoring point; ||rr i || is the Euclidean distance between the grid node and the monitoring point; σ is the parameter that controls the smoothness of the radial basis function, which is dynamically adjusted according to the density and spacing of the monitoring points to match the distribution of nonlinear features;

[0022] Then, the deviation field is spatially aligned according to the model's three-dimensional coordinate system. For each grid node x j , map the interpolated nonlinear deviation field to the three-dimensional coordinate space of the model, assuming that the value of the monitoring data is f(r i ), then node x j The mapping value F(x j ) is calculated by radial basis function interpolation as:

[0023]

[0024] Among them, λ i The monitoring point r i The weight of the impact on interpolation is determined by the monitoring data and model calibration parameters; N is the total number of monitoring points;

[0025] After the mapping is completed, bilinear interpolation is applied to the adjacent grid nodes to eliminate the value mutation. For the four adjacent nodes Q in the two-dimensional plane 11 , Q 12 , Q 21 , Q 22 Interpolate as follows:

[0026]

[0027] Among them, (x1, y1), (x2, y1), (x1, y2), (x2, y2) are the coordinates of four adjacent grid nodes, respectively. 11 )、f(Q 21 )、f(Q 12 )、f(Q 22 ) are the values ​​corresponding to the four grid nodes respectively;

[0028] This bilinear interpolation formula can be generalized in three-dimensional grids to achieve a more accurate smooth transition of the nonlinear deviation field by calculating eight adjacent nodes in the three-dimensional solid.

[0029] Furthermore, in step S4, the state transition equation is used to predict the next state. and error covariance P k|k-1 :

[0030]

[0031] Among them, F k is the state transfer matrix; B k is the control input matrix; u k is the external input quantity; Q k is the process noise covariance;

[0032] The calibration phase uses monitoring data k Calculate the residual y k , and through the Kalman gain K k Balance observations and observed values:

[0033]

[0034] Among them, H k is the observation matrix; R k is the observation noise covariance matrix;

[0035] State and error update Final updated state estimate and error covariance P k|k for:

[0036]

[0037] P k|k =(IK k H K ) k|k-1

[0038] Where I is the identity matrix.

[0039] Furthermore, in step S5, the dynamic sliding window size W is defined. If a continuous large error is detected, the detection threshold is dynamically increased to focus on the change area, otherwise it is reduced to avoid repeated optimization. The error values ​​that significantly deviate from the relatively stable trend have a large absolute value, the error fluctuates significantly in a short time, and the error fluctuations with high sensitivity under dynamic response are marked respectively, and the error value e in the sliding window is marked. i Average to get the average error value within the sliding window:

[0040]

[0041] Where W is the size of the sliding window; i is the current position, indicating the center of the sliding window; j is the specific position within the sliding window, traversing each data point in the window; e j is the error value of the jth position;

[0042] Judgment criteria: When the average error When a certain threshold α is exceeded, the window is marked as an area with a large absolute error value:

[0043]

[0044] Among them, α is based on historical monitoring data or a preset empirical threshold, which can be set according to three times the standard deviation of the monitoring data;

[0045] Calculate the standard deviation σ of the error value within the sliding window e,i , which is used to quantify the volatility of the error value:

[0046]

[0047] Judgment criteria: When the standard deviation σ within the sliding window e,i When the preset fluctuation threshold β is exceeded, the window is marked as an area with large error fluctuations:

[0048] σ e,i >β

[0049] Among them, β is a dynamically adjusted threshold, which is set by monitoring the historical fluctuation characteristics of the data.

[0050] Furthermore, in step S6, different levels of weights are set for fitness evaluation during the optimization process. The fitness function is defined by the weighted mean square error. The lower the fitness value, the closer the individual parameter combination is to the optimal one. The focus is continuously placed on the error-significant region to achieve multi-level optimization of boundary conditions and material parameters. The fitness function is calculated as follows:

[0051]

[0052] Where, N is the total number of monitoring points; w space is the weight coefficient of spatial error; ∈ space,i is the spatial error of the ith monitoring point; w absolute is the weight coefficient of absolute error; ∈ absolute,i is the absolute error of the ith monitoring point; w variance is the weight coefficient of the variance error; ∈ variance,i is the variance error of the ith monitoring point; w space 、w absolute 、w variance In each iteration of the genetic algorithm, adjustments are made based on the error characteristics to ensure that the model's fitness is improved in areas with high error incidence. During the crossover and mutation process, the individual parameter update formula is:

[0053] Child param =(1-α)·Parent1 param +α·Parent2 param

[0054] Among them, α∈[0,1] is the crossover coefficient, which ensures that the new individual is reasonably adjusted based on the parent parameters; Parent1 param is the parameter of parent individual 1; Parent2 param is the parameter of parent individual 2.

[0055] Furthermore, in step S7, the optimization parameters, original data, and known nonlinear response multidimensional parameters are used to stabilize the prior stability, and the likelihood function is constructed based on the difference of stress-strain data. The deviation field distribution form is defined by the Gamma distribution. After the constitutive parameters are updated, a progressive optimization mechanism based on dynamic feedback is constructed. After each working surface is completed, the posterior distribution is updated to the prior distribution of the next stage, and gradually converges to the true mechanical properties of the material. The core dynamic Bayesian update formula is as follows:

[0056] p(θ t+1 |D 1:t+1 )∝p(D t+1 |θ t)·p(θ t |D 1:t )

[0057] Among them, θ t+1 represents the updated parameter, p(D t+1 |θ t ) is the likelihood value of the current step observation data, p(θ t |D 1:t ) represents the posterior distribution of the previous stage.

[0058] Furthermore, in step S8, the dynamic material constitutive update constructs the initial material constitutive distribution based on historical data, known material constitutive model libraries, and dynamic Bayesian results, uses a recurrent neural network to capture the soft hardening trend of the material over time, and a convolutional neural network to obtain spatial features, and gradually updates the stress-strain curve of the material. The model performs nonlinear strengthening or attenuation on the high stress area to form an adaptive adjustment of the constitutive structure; the dynamic material constitutive structure is described by a nonlinear stress-strain relationship, and the AI ​​model constructs a material nonlinear constitutive update function through supervised learning and dynamic Bayesian updating:

[0059] σ t+1 =f(σ t ,ε;θ AI )

[0060] Among them, σ t+1 is the stress prediction value at the current moment, ε is the strain, θ AI are model parameters, which are dynamically updated through the learning process;

[0061] In the dynamic model constitutive model, the traditional constitutive model parameters are used as the initial parameters of the model, and the AI ​​algorithm is combined to realize data-driven adaptive correction. The material change trend is captured by learning the historical sequence of stress and strain data, and the calculation residual is fed back as the adjustment factor of φ and c parameters through Bayesian optimization to realize the a posteriori update of the original model parameters. Taking the common Moore-Coulomb criterion as an example, the update formula formed by the deep learning model is:

[0062] τ=(c+Δc)+σ·tan(φ+Δφ)

[0063] Among them, Δc and Δφ are the adjustment values ​​at the current moment, predicted by the AI ​​model; Adaptive error correction: Combined with Kalman filtering or Bayesian update method, a new posterior distribution is generated:

[0064] P(c,φ|D 监测 )≈P(c,φ)·L(D 监测 |c,φ)

[0065] Where P(c,φ) represents the joint probability distribution of material constitutive parameters; L(D 监测|c,φ) represents the difference between the actual data and the constitutive model, and the optimal (c,φ) value is updated through this distribution;

[0066] Starting from the computational essence of material and model constitutive elements, strong nonlinear deviations are gradually overcome through AI deep data drive to achieve self-optimized virtual-reality dynamic correction and precise twin mapping.

[0067] The digital twin method for coal mining based on multi-level dynamic correction and AI optimization provided by the present invention innovatively applies digital twin technology to real-time monitoring and dynamic simulation of the coal mining process, and uses high-precision data fusion and adaptive correction strategies to achieve synchronous mapping and multi-level optimization of virtual and real dual spaces. Through the numerical simulation method optimized by intelligent algorithms, the present invention realizes high-precision dynamic mapping and real-time feedback of the coal mining process, making up for the shortcomings of traditional static numerical simulation. Compared with the prior art, the present invention has the following advantages:

[0068] 1. Interactive supervised twin simulation: The present invention realizes virtual-real synchronization through interactive and supervised twin simulation by dynamically sensing real-time data in the coal mining process. The monitoring data and simulation results are constantly compared and corrected through a loop feedback mechanism, and a flexible real-time model correction mechanism is constructed, which can dynamically capture the changing characteristics of the mining environment and generate high-precision simulation results. Compared with traditional static simulation methods, the present invention has obvious advantages in virtual-real interaction and dynamic response.

[0069] 2. Multi-level optimization and correction mechanism: The present invention uses a multi-level dynamic correction process to perform adaptive weighted correction on simulation deviations at key monitoring points and mutation areas. Under the multi-level optimization framework, based on Kalman filtering and sliding window analysis, genetic algorithms and dynamic Bayesian methods are used to optimize and adjust material parameters and boundary conditions, and correction is performed on nonlinear mutation points and high error fluctuation areas to improve the overall convergence speed and accuracy of the simulation model.

[0070] 3. Deep data-driven AI optimization and prediction: Combining deep learning technology with a complex data-driven algorithm framework, the present invention uses recurrent neural networks and convolutional neural networks to accurately describe the boundary state and material changes of coal mining in time and space. Based on dynamic Bayesian updating, an adaptive multi-level feature learning mechanism is constructed, which can effectively identify and predict key stress-strain trends, material constitutive and model constitutive evolution in the mining process. Unlike traditional limited data prediction methods, the present invention breaks the limitations of traditional static simulations and constructs a flexible real-time feedback and dynamic adjustment framework, which provides important support for the accurate prediction and optimization of complex coal mining processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1It is a schematic flow chart of the digital twin method for coal mining based on multi-level dynamic correction and AI optimization provided by the present invention. DETAILED DESCRIPTION

[0072] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below with reference to specific diagrams.

[0073] Please refer to Figure 1 As shown, the present invention provides a coal mining digital twin method based on multi-level dynamic correction and AI optimization, comprising the following steps:

[0074] S1. Monitoring data acquisition and dynamic preprocessing: Determine the acquisition frequency of the sensor and calibrate the timestamp, perform structured settings on the data collected before processing, set the upper and lower noise thresholds of mean ±3 times the standard deviation for the monitoring points, and perform spatial consistency detection, analyze the correlation between the relative positions of multiple sensors and stress values, adjust the threshold strength according to the fluctuation adaptively, and perform smooth interpolation repair on the data of abnormal points;

[0075] S2. Construction and pre-processing of digital twins of coal mines: Through the rapid modeling method of unstructured data such as drilling, exploration, and geophysical exploration, the true three-dimensional state of coal mine geology is constructed, and the model is adaptively meshed, the initial ground stress field is balanced, the static boundary conditions are set, and the material parameters are assigned, etc., to provide a digital foundation for twin simulation;

[0076] S3, high nonlinear variable interpolation and model mapping: for the monitoring data points in step S1, based on the position of the model grid node, a continuous nonlinear deviation field is generated by radial basis function (RBF) interpolation, which is spatially aligned according to the model coordinate system, matching the coordinates of each grid node to complete the numerical mapping, and further smoothing the transition by bilinear interpolation;

[0077] S4, Kalman filter error analysis and correction under continuous analysis steps: In each calculation step, the simulation calculation results and monitoring data are subjected to Kalman filter analysis, the Kalman gain is calculated and the state value and covariance matrix are updated, the state value and error of the model are dynamically adjusted through the prediction-correction process, the weight is adaptively adjusted according to the error fluctuation, and the spatial analysis of all monitoring point data is performed to calculate the spatial coefficient of variation in the entire monitoring point set, that is, the ratio of the standard deviation to the mean, and mark the spatial mutation step that exceeds the threshold, providing a hierarchical processing basis for the subsequent optimization algorithm;

[0078] S5. Staged sliding window long time series fluctuation analysis: After a working face is mined, a dynamic sliding window is set in the geological structure (such as faults, folds) and stress concentration area where error fluctuations are prone to occur, and the time series error data in the window is processed by rolling average. The results are classified as abnormal according to the large absolute value of the error and the large error fluctuation. The area with continuous large error or frequent fluctuations in the window is detected by the increasing-decreasing threshold method. If the error fluctuation area contains the spatial mutation step marked in step S4, it will be given a higher weight value in the subsequent optimization;

[0079] S6. Multi-level weighted error optimization based on genetic algorithm (GA): taking error convergence degree as fitness target, calculating mean square error (MSE) between each monitoring point and interpolation point, forming an initial population with different material parameters (elastic modulus, Poisson's ratio, etc.) and boundary condition combinations, setting weights for the spatial mutation step in step S4, the large absolute error value and the large error fluctuation in step S5, analyzing the long-term dynamic characteristics of the error according to the error time series trend after rolling average, adjusting local material parameters for the error fluctuation and mutation mark areas based on weights, fine-tuning boundary conditions for the areas with large absolute error values, maintaining the amplitude at 5% to 10% to maintain convergence stability, and obtaining the initial boundary state and material properties of the next step after optimization;

[0080] S7. Correction of material constitutive parameters based on dynamic Bayesian method: the optimized results of step S6, geological exploration, historical monitoring data and known material constitutive model library are used as the initial prior distribution of material constitutive parameters. A likelihood function is constructed based on the difference between simulated stress-strain and monitoring data in multiple steps to reflect the nonlinear response characteristics of materials such as softness and hardening under different stress states. The posterior distribution is updated in the next mining stage of the working face based on the combination of the two. As the mining of the working face deepens (for example, mining of 3 or more working faces is completed), the new material constitutive characteristics are input into the model as the prior distribution of the next stage, and the adaptability and accuracy of the model to complex nonlinear and high-stress geological conditions are gradually enhanced, forming a progressive optimization effect of dynamic feedback-multiple iterations.

[0081] S8. Construct an AI optimization prediction model based on a hybrid network structure driven by deep data: use the optimization and monitoring data generated in steps S4 to S7 as input, and use the recurrent neural network (RNN) based on the nonlinear relationship between stress and strain to capture the progressive changes of material parameters on the time axis; use a convolutional neural network (CNN) to extract the boundary and material spatial distribution characteristics, and integrate their time series and spatial characteristics; based on supervised learning and dynamic Bayesian updating, construct a dynamic material constitutive and model constitutive property update mechanism to accurately describe the dynamic property changes of geology under mining, and continuously improve the ability to accurately map the virtual and real states of dynamic twin mining in coal mining.

[0082] As a specific embodiment, in step S1, the potential time drift error inside the device is eliminated by comparing and correcting the internal and external clocks of the sensor, the collected raw data is structured and normalized to the minimum and maximum values, and the data of each sensor is correlated by combining the Pearson correlation coefficient with the mutual information. When the stress value monitored by the sensor fluctuates frequently, the threshold is automatically relaxed to adapt to higher dynamic changes, and vice versa. The regional correlation interpolation method is used for multi-point anomalies, and the stress values ​​of adjacent time and space nodes are fully considered to ensure the accuracy of abnormal point repair; the end time (end time) and initial time (initial time) of each step (Step) are set to ensure that the end time of the calculation step (Step) is accurately aligned with the acquisition time, and a smaller time step is selected by adjusting the time integral (increment size) step to better capture nonlinear behavior while meeting the acquisition interval frequency.

[0083] As a specific embodiment, in step S3, radial basis function (RBF) interpolation is used to generate a nonlinear deviation field from the monitoring point to the model node, so that it has global continuity and local accuracy. The radial basis function is expressed as follows:

[0084]

[0085] Where r represents the coordinate of the target grid node; r i Represents the coordinates of the monitoring point; ||rr i || is the Euclidean distance between the grid node and the monitoring point; σ is the parameter that controls the smoothness of the radial basis function, which is dynamically adjusted according to the density and spacing of the monitoring points to match the distribution of nonlinear features;

[0086] Then, the deviation field is spatially aligned according to the model's three-dimensional coordinate system. For each grid node x j , map the interpolated nonlinear deviation field to the three-dimensional coordinate space of the model, assuming that the value of the monitoring data is f(r i ), then node x j The mapping value F(x j ) is calculated by radial basis function interpolation as:

[0087]

[0088] Among them, λ i The monitoring point r i The weight of the impact on interpolation is determined by the monitoring data and model calibration parameters; N is the total number of monitoring points;

[0089] After the mapping is completed, bilinear interpolation is applied to the adjacent grid nodes to eliminate the value mutation. For the four adjacent nodes Q in the two-dimensional plane 11 , Q 12 , Q 21 , Q 22 Interpolate as follows:

[0090]

[0091] Among them, (x1, y1), (x2, y1), (x1, y2), (x2, y2) are the coordinates of four adjacent grid nodes, respectively. 11 )、f(Q 21 )、f(Q 12 )、f(Q 22 ) are the values ​​corresponding to the four grid nodes respectively;

[0092] This bilinear interpolation formula can be generalized in three-dimensional grids to achieve a more accurate smooth transition of the nonlinear deviation field by calculating eight adjacent nodes in the three-dimensional solid.

[0093] As a specific embodiment, the state transition equation is used in step S4 to predict the next state. and error covariance P k|k-1 :

[0094]

[0095] Among them, F k is the state transfer matrix; B k is the control input matrix; u k is the external input quantity; Q k is the process noise covariance;

[0096] The calibration phase uses monitoring data k Calculate the residual y k , and through the Kalman gain K k Balance observations and observed values:

[0097]

[0098] Among them, H k is the observation matrix; R k is the observation noise covariance matrix;

[0099] State and error update Final updated state estimate and error covariance P k|k for:

[0100]

[0101] P k|k=(IK k H K ) k|k-1

[0102] Where I is the identity matrix.

[0103] As a specific embodiment, the dynamic sliding window size W is defined in step S5. If a continuous large error is detected, the detection threshold is dynamically increased to focus on the change area, otherwise it is reduced to avoid repeated optimization. The error values ​​that significantly deviate from the relatively stable trend have a large absolute value, the error fluctuates significantly in a short time, and the error fluctuations with high sensitivity under dynamic response are marked respectively, and the error value e in the sliding window is marked. i Average to get the average error value within the sliding window:

[0104]

[0105] Where W is the size of the sliding window; i is the current position, indicating the center of the sliding window; j is the specific position within the sliding window, traversing each data point in the window; e j is the error value of the jth position;

[0106] Judgment criteria: When the average error When a certain threshold α is exceeded, the window is marked as an area with a large absolute error value:

[0107]

[0108] Among them, α is based on historical monitoring data or a preset empirical threshold, which can be set according to three times the standard deviation of the monitoring data;

[0109] Calculate the standard deviation σ of the error value within the sliding window e,i , which is used to quantify the volatility of the error value:

[0110]

[0111] Judgment criteria: When the standard deviation σ within the sliding window e,i When the preset fluctuation threshold β is exceeded, the window is marked as an area with large error fluctuations:

[0112] σ e,i >β

[0113] Among them, β is a dynamically adjusted threshold, which is set by monitoring the historical fluctuation characteristics of the data.

[0114] As a specific embodiment, different levels of weights are set in step S6 for evaluating fitness during the optimization process. The fitness function is defined by the weighted mean square error (MSE). The lower the fitness value, the closer the individual parameter combination is to the optimal one. The focus is continuously placed on the error-significant area to achieve multi-level optimization of boundary conditions and material parameters. The fitness function is calculated as follows:

[0115]

[0116] Where, N is the total number of monitoring points; w space is the weight coefficient of spatial error; ∈ space,i is the spatial error of the ith monitoring point; w absolute is the weight coefficient of absolute error; ∈ absolute,i is the absolute error of the ith monitoring point; w variance is the weight coefficient of the variance error; ∈ variance,i is the variance error of the ith monitoring point; w space 、w absolute 、w variance In each iteration of the genetic algorithm, adjustments are made based on the error characteristics to ensure that the model is more responsive in areas with high error rates. During the crossover and mutation process, the individual parameter update formula is:

[0117] Child param =(1-α)·Parent1 param +α·Parent2 param

[0118] Among them, α∈[0,1] is the crossover coefficient, which ensures that the new individual is reasonably adjusted based on the parent parameters; Parent1 param is the parameter of parent individual 1; Parent2 param is the parameter of parent individual 2.

[0119] As a specific embodiment, in step S7, the optimization parameters, original data, known nonlinear responses and other multidimensional parameters are used to stabilize the prior stability, the likelihood function is constructed based on the difference of stress-strain data, and the deviation field distribution form is defined by Gamma distribution. After the constitutive parameters are updated, a progressive optimization mechanism based on dynamic feedback is constructed. After each working surface is completed, the posterior distribution is updated to the prior distribution of the next stage, and gradually converges to the true mechanical properties of the material. The core dynamic Bayesian update formula is as follows:

[0120] p(θ t+1 |D 1:t+1 )∝p(D t+1 |θ t )·p(θ t |D 1:t )

[0121] Among them, θ t+1 represents the updated parameter, p(D t+1 |θ t ) is the likelihood value of the current step observation data, p(θ t |D 1:t ) represents the posterior distribution of the previous stage.

[0122] As a specific embodiment, the dynamic material constitutive update in step S8 constructs an initial material constitutive distribution based on historical data, a known material constitutive model library, and a dynamic Bayesian result, uses a recurrent neural network (RNN) to capture the soft hardening trend of the material over time, and a convolutional neural network (CNN) to obtain spatial features, and gradually updates the stress-strain curve of the material. The model performs nonlinear strengthening or attenuation on high stress areas (such as faults) to form an adaptive adjustment of the constitutive structure; the dynamic material constitutive structure is described by a nonlinear stress-strain relationship, and the AI ​​model constructs a material nonlinear constitutive update function through supervised learning and dynamic Bayesian updating:

[0123] σ t+1 =f(σ t ,ε;θ AI )

[0124] Among them, σ t+1 is the stress prediction value at the current moment, ε is the strain, θ AI are model parameters, which are dynamically updated through the learning process;

[0125] In the dynamic model constitutive model, the traditional constitutive model parameters are used as the initial parameters of the model, and the AI ​​algorithm is combined to realize data-driven adaptive correction. The material change trend is captured by learning the historical sequence of stress and strain data, and the calculation residual is fed back as the adjustment factor of φ and c parameters through Bayesian optimization to realize the a posteriori update of the original model parameters. Taking the common Moore-Coulomb criterion as an example, the update formula formed by the deep learning model is:

[0126] τ=(c+Δc)+σ·tan(φ+Δφ)

[0127] Among them, Δc and Δφ are the adjustment values ​​at the current moment, predicted by the AI ​​model; Adaptive error correction: Combined with Kalman filtering or Bayesian update method, a new posterior distribution is generated:

[0128] P(c,φ|D 监测 )≈P(c,φ)·L(D 监测 |c,φ)

[0129] Where P(c,φ) represents the joint probability distribution of material constitutive parameters; L(D 监测|c,φ) represents the difference between the actual data and the constitutive model, and the optimal (c,φ) value is updated through this distribution;

[0130] Starting from the computational essence of material and model constitutive elements, strong nonlinear deviations are gradually overcome through AI deep data drive to achieve self-optimized virtual-reality dynamic correction and precise twin mapping.

[0131] In order to better understand the coal mining digital twin method based on multi-level dynamic correction and AI optimization provided by the present invention, the following will be described in detail in conjunction with a specific embodiment. This embodiment conducts a field test on the Shaanxi Coal Mengcun Coal Mine, and the specific steps are as follows:

[0132] S1. Use Python to write scripts to obtain real-time monitoring data using sensors such as strain gauges and accelerometers installed on site in coal mines. After collecting data, use the Pandas library to clean the data, including removing missing values ​​and noise. Calculate the Pearson correlation coefficient and mutual information between multiple sensors to evaluate the correlation between relative position and stress value. Use scipy to smooth and interpolate data to ensure that there are no abnormal points in the collected monitoring data.

[0133] S2. Use UMAT (user-defined material subroutine) in Abaqus to define the material constitutive model, and perform initial static boundary condition setting, stress field balance and meshing.

[0134] S3. Based on the real-time monitoring data collected in step S1, the radial basis function (RBF) interpolation method is called by Python to perform nonlinear deviation field interpolation to generate a continuous deviation field based on the model grid. Fortran is used to realize spatial alignment of the deviation field, and the interpolation results of the monitoring points are mapped to the Abaqus grid model, and data smoothing is achieved by the bilinear interpolation method.

[0135] S4. Implement the Kalman filter algorithm through Python and Fortran interface, calibrate the simulated data and monitoring data obtained each time, calculate the gain and update the model state. After each iteration, adjust the weight adaptively according to the error fluctuation. Use Python to write a sliding window analysis program, perform rolling average on the time series data in the sliding window, and classify them according to "large absolute error value" and "large error fluctuation".

[0136] S5. Update each working face after mining is completed, combine the dynamic Bayesian method, build the likelihood function based on simulation data, monitoring data and historical data, and update the material model. Use Fortran and Python interfaces to calculate the posterior distribution and use the new material constitutive properties as the prior input for the next stage.

[0137] S6, Python and TensorFlow are used to build a deep learning model, and RNN is used to capture the softening and hardening trends of materials and establish a nonlinear time series relationship between stress and strain. Convolutional neural network CNN is used to extract the boundary of the mine and the spatial distribution characteristics of materials. RNN and CNN are combined to form a hybrid network structure. The network weights are continuously updated through supervised learning to accurately predict the geological stress changes and material behaviors under different working faces. Based on the dynamic Bayesian method and network output, the material constitutive model is gradually optimized, and the new dynamic characteristics are fed back to the next working face. Finally, the whole process from the collection and processing of monitoring data to the dynamic optimization of the digital twin and then to the construction of the AI ​​model realizes the closed-loop optimization from model construction to dynamic feedback in coal mining.

[0138] The digital twin method for coal mining based on multi-level dynamic correction and AI optimization provided by the present invention innovatively applies digital twin technology to real-time monitoring and dynamic simulation of the coal mining process, and uses high-precision data fusion and adaptive correction strategies to achieve synchronous mapping and multi-level optimization of virtual and real dual spaces. Through the numerical simulation method optimized by intelligent algorithms, the present invention realizes high-precision dynamic mapping and real-time feedback of the coal mining process, making up for the shortcomings of traditional static numerical simulation. Compared with the prior art, the present invention has the following advantages:

[0139] 1. Interactive supervised twin simulation: The present invention realizes virtual-real synchronization through interactive and supervised twin simulation by dynamically sensing real-time data in the coal mining process. The monitoring data and simulation results are constantly compared and corrected through a loop feedback mechanism, and a flexible real-time model correction mechanism is constructed, which can dynamically capture the changing characteristics of the mining environment and generate high-precision simulation results. Compared with traditional static simulation methods, the present invention has obvious advantages in virtual-real interaction and dynamic response.

[0140] 2. Multi-level optimization and correction mechanism: The present invention uses a multi-level dynamic correction process to perform adaptive weighted correction on simulation deviations at key monitoring points and mutation areas. Under the multi-level optimization framework, based on Kalman filtering and sliding window analysis, genetic algorithms and dynamic Bayesian methods are used to optimize and adjust material parameters and boundary conditions, and correction is performed on nonlinear mutation points and high error fluctuation areas to improve the overall convergence speed and accuracy of the simulation model.

[0141] 3. Deep data-driven AI optimization and prediction: Combining deep learning technology with a complex data-driven algorithm framework, the present invention uses recurrent neural networks and convolutional neural networks to accurately describe the boundary state and material changes of coal mining in time and space. Based on dynamic Bayesian updating, an adaptive multi-level feature learning mechanism is constructed, which can effectively identify and predict key stress-strain trends, material constitutive and model constitutive evolution in the mining process. Unlike traditional limited data prediction methods, the present invention breaks the limitations of traditional static simulations and constructs a flexible real-time feedback and dynamic adjustment framework, which provides important support for the accurate prediction and optimization of complex coal mining processes.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A digital twin method for coal mining based on multi-level dynamic deviation correction and AI optimization, characterized in that: The following steps are involved: S1. Monitoring data acquisition and dynamic preprocessing: Determine the acquisition frequency of the sensor and calibrate the timestamp, perform structured settings on the data collected before processing, set the upper and lower noise thresholds of mean ±3 times the standard deviation for the monitoring points, and perform spatial consistency detection, analyze the correlation between the relative positions of multiple sensors and stress values, adjust the threshold strength according to the fluctuation adaptively, and perform smooth interpolation repair on the data of abnormal points; S2. Construction and pre-processing of digital twins of coal mines: Through the rapid modeling method of drilling, exploration, and geophysical unstructured data, the true three-dimensional state of coal mine geology is constructed, and the model is adaptively meshed, the initial ground stress field is balanced, the static boundary conditions are set, and the material parameters are assigned to provide a digital foundation for twin simulation; S3, high nonlinear variable interpolation and model mapping: for the monitoring data points in step S1, based on the position of the model grid node, a continuous nonlinear deviation field is generated by radial basis function interpolation, the model coordinate system is spatially aligned, the coordinates of each grid node are matched to complete the numerical mapping, and the bilinear interpolation method is used to further smooth the transition; S4, Kalman filter error analysis and correction under continuous analysis steps: In each calculation step, the simulation calculation results and monitoring data are subjected to Kalman filter analysis, the Kalman gain is calculated and the state value and covariance matrix are updated, the state value and error of the model are dynamically adjusted through the prediction-correction process, the weight is adaptively adjusted according to the error fluctuation, and the spatial analysis of all monitoring point data is performed to calculate the spatial coefficient of variation in the entire monitoring point set, that is, the ratio of the standard deviation to the mean, and mark the spatial mutation step that exceeds the threshold, providing a hierarchical processing basis for the subsequent optimization algorithm; S5. Staged sliding window long time series fluctuation analysis: After a working face is mined, a dynamic sliding window is set in the geological structure and stress concentration area where error fluctuations are prone to occur, and the time series error data in the window is processed by rolling average. The results are classified as abnormal according to the large absolute value of the error and the large error fluctuation. The area with continuous large error or frequent fluctuations in the window is detected by the increasing-decreasing threshold method. If the error fluctuation area contains the spatial mutation step marked in step S4, it will be given a higher weight value in the subsequent optimization; S6. Multi-level weighted error optimization based on genetic algorithm: taking error convergence degree as fitness target, calculating mean square error between each monitoring point and interpolation point, forming initial population with different material parameters and boundary condition combinations, setting weights for spatial mutation step in step S4, large absolute error value and large error fluctuation in step S5, analyzing long-term dynamic characteristics of error according to error time series trend after rolling average, adjusting local material parameters for large error fluctuation and mutation marked areas based on weights, fine-tuning boundary conditions for areas with large absolute error values, maintaining the amplitude at 5% to 10% to maintain convergence stability, and obtaining initial boundary state and material properties for the next step after optimization; S7. Correction of material constitutive parameters based on dynamic Bayesian: The optimized results of step S6, geological exploration, historical monitoring data and known material constitutive model library are used as the initial prior distribution of material constitutive parameters. A likelihood function is constructed based on the difference between simulated stress-strain and monitoring data in multiple steps to reflect the soft and hardened nonlinear response characteristics of materials under different stress states. Based on the combination of the two, the posterior distribution is updated in the next mining stage of the working face. As the mining of the working face deepens, the new material constitutive characteristics are input into the model as the prior distribution of the next stage, gradually enhancing the adaptability and accuracy of the model to complex nonlinear and high-stress geological conditions, forming a dynamic feedback-multiple iterations of progressive optimization effect; S8. Construct an AI optimization prediction model based on a hybrid network structure driven by deep data: take the optimization and monitoring data generated in steps S4 to S7 as input, and use the recurrent neural network based on the nonlinear relationship between stress and strain to capture the progressive changes of material parameters on the time axis; use a convolutional neural network to extract the boundary and material spatial distribution characteristics, and integrate their time series and spatial characteristics; Based on supervised learning and dynamic Bayesian updating, a dynamic material constitutive and model constitutive property updating mechanism is constructed to accurately describe the dynamic property changes of geology under mining, and continuously improve the ability to accurately map the virtual and real states of dynamic twin coal mining.

2. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In the step S1, firstly, the potential time drift error inside the device is eliminated by comparing and correcting the clocks inside and outside the sensor, and the collected raw data is structured and normalized to the minimum and maximum values. The Pearson correlation coefficient is combined with the mutual information to perform correlation analysis on the data of each sensor. When the stress value monitored by the sensor shows frequent fluctuations, the threshold is automatically relaxed to adapt to higher dynamic changes, and vice versa. The regional correlation interpolation method is used for multi-point anomalies, and the stress values ​​of adjacent time and space nodes are fully considered to ensure the accuracy of abnormal point repair. The end time and initial time of each step are set to ensure that the end time of the calculation step is accurately aligned with the acquisition time, and a smaller time step is selected by adjusting the time integration step to better capture nonlinear behavior while meeting the acquisition interval frequency.

3. The digital twin method for coal mining based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S3, radial basis function interpolation is used to generate a nonlinear deviation field from the monitoring point to the model node, so that it has global continuity and local accuracy. The radial basis function is expressed as follows: Where r represents the coordinate of the target grid node; r i Represents the coordinates of the monitoring point; rr i is the Euclidean distance between the grid node and the monitoring point; σ is the parameter that controls the smoothness of the radial basis function, which is dynamically adjusted according to the density and spacing of the monitoring points to match the distribution of nonlinear features; Then, the deviation field is spatially aligned according to the model's three-dimensional coordinate system. For each grid node x j , map the interpolated nonlinear deviation field to the three-dimensional coordinate space of the model, assuming that the value of the monitoring data is f(r i ), then node x j The mapping value F(x j ) is calculated by radial basis function interpolation as: Among them, λ i The monitoring point r i The weight of the impact on interpolation is determined by the monitoring data and model calibration parameters; N is the total number of monitoring points; After the mapping is completed, bilinear interpolation is applied to the adjacent grid nodes to eliminate the value mutation. For the four adjacent nodes Q in the two-dimensional plane 11 , Q 12 , Q 21 , Q 22 Interpolate as follows: Among them, (x1, y1), (x2, y1), (x1, y2), (x2, y2) are the coordinates of four adjacent grid nodes, respectively. 11 )、f(Q 21 )、f(Q 12 )、f(Q 22 ) are the values ​​corresponding to the four grid nodes respectively; This bilinear interpolation formula can be generalized in three-dimensional grids to achieve a more accurate smooth transition of the nonlinear deviation field by calculating eight adjacent nodes in the three-dimensional solid.

4. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S4, the state transition equation is used to predict the next state and error covariance P kk-1 : Among them, F k is the state transfer matrix; B k is the control input matrix; u k is the external input quantity; Q k is the process noise covariance; The calibration phase uses monitoring data k Calculate the residual y k , and through the Kalman gain K k Balance observations and observed values: Among them, H k is the observation matrix; R k is the observation noise covariance matrix; State and error update Final updated state estimate and error covariance P kk for: P k|k =(I-K k H K )P k|k-1 Where I is the identity matrix.

5. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S5, the dynamic sliding window size W is defined. If a continuous large error is detected, the detection threshold is dynamically increased to focus on the change area, otherwise it is reduced to avoid repeated optimization. The error values ​​that significantly deviate from the relatively stable trend have a large absolute value, the error fluctuates significantly in a short time, and the error fluctuations with high sensitivity under dynamic response are marked separately, and the error value e in the sliding window is marked. i Average to get the average error value within the sliding window: Where W is the size of the sliding window; i is the current position, indicating the center of the sliding window; j is the specific position within the sliding window, traversing each data point in the window; e j is the error value of the jth position; Judgment criteria: When the average error When a certain threshold α is exceeded, the window is marked as an area with a large absolute error value: Among them, α is based on historical monitoring data or a preset empirical threshold, which can be set according to three times the standard deviation of the monitoring data; Calculate the standard deviation σ of the error value within the sliding window e,i , which is used to quantify the volatility of the error value: Judgment criteria: When the standard deviation σ within the sliding window e,i When the preset fluctuation threshold β is exceeded, the window is marked as an area with large error fluctuations: s e,i >b Among them, β is a dynamically adjusted threshold, which is set by monitoring the historical fluctuation characteristics of the data.

6. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S6, different levels of weights are set for fitness evaluation during the optimization process. The fitness function is defined by the weighted mean square error. The lower the fitness value, the closer the individual parameter combination is to the optimal one. The focus is continuously placed on the error-significant region to achieve multi-level optimization of boundary conditions and material parameters. The fitness function is calculated as follows: Where, N is the total number of monitoring points; w space is the weight coefficient of spatial error; ∈ space，i is the spatial error of the ith monitoring point; w absolute is the weight coefficient of absolute error; ∈ absolute，i is the absolute error of the ith monitoring point; w variance is the weight coefficient of the variance error; ∈ variance，i is the variance error of the ith monitoring point; W space 、w absolute 、w variance In each iteration of the genetic algorithm, adjustments are made based on the error characteristics to ensure that the model's fitness is improved in areas with high error incidence. During the crossover and mutation process, the individual parameter update formula is: Child param =(1-α)·Parentl param +α·Parent2 param Among them, α∈[0,1] is the crossover coefficient, which ensures that the new individual is reasonably adjusted based on the parent parameters; Parentl param is the parameter of parent individual 1; Parent2 param is the parameter of parent individual 2.

7. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S7, the optimization parameters, original data, and known nonlinear response multidimensional parameters are used to stabilize the prior stability, and the likelihood function is constructed based on the difference of stress-strain data. The deviation field distribution form is defined by Gamma distribution. After the constitutive parameters are updated, a progressive optimization mechanism based on dynamic feedback is constructed. After each working surface is completed, the posterior distribution is updated to the prior distribution of the next stage, and gradually converges to the true mechanical properties of the material. The core dynamic Bayesian update formula is as follows: p(θ t+1 |D 1：t+1 )∝p(D t+1 |θ t )·p(θ t |D 1：t ) Among them, θ t+1 represents the updated parameter, p(D t+1 |θ t ) is the likelihood value of the current step observation data, p(θ t |D 1：t ) represents the posterior distribution of the previous stage.

8. The coal mining digital twin method based on multi-level dynamic deviation correction and AI optimization according to claim 1 is characterized in that: In step S8, the dynamic material constitutive update constructs the initial material constitutive distribution based on historical data, known material constitutive model library, and dynamic Bayesian results, uses a recurrent neural network to capture the soft hardening trend of the material over time, and a convolutional neural network to obtain spatial characteristics, and gradually updates the stress-strain curve of the material. The model performs nonlinear strengthening or attenuation on the high stress area to form an adaptive adjustment of the constitutive structure; the dynamic material constitutive structure is described by a nonlinear stress-strain relationship, and the AI ​​model constructs a material nonlinear constitutive update function through supervised learning and dynamic Bayesian updating: s t+1 =f(σ t ,e;i AI ) Among them, σ t+1 is the stress prediction value at the current moment, ε is the strain, θ AI are model parameters, which are dynamically updated through the learning process; In the dynamic model constitutive model, the traditional constitutive model parameters are used as the initial parameters of the model, and the AI ​​algorithm is combined to realize data-driven adaptive correction. The material change trend is captured by learning the historical sequence of stress and strain data, and the calculation residual is fed back as the adjustment factor of φ and c parameters through Bayesian optimization to realize the a posteriori update of the original model parameters. Taking the common Moore-Coulomb criterion as an example, the update formula formed by the deep learning model is: τ=(c+Δc)+σ·tan(φ+Δφ) Among them, Δc and Δφ are the adjustment values ​​at the current moment, predicted by the AI ​​model; Adaptive error correction: Combined with Kalman filtering or Bayesian update method, a new posterior distribution is generated: P(c,φ|D 监测 )≈P(c,φ)·L(D 监测 |c,φ) Where P(c, φ) represents the joint probability distribution of material constitutive parameters; L(D 监测 |c, φ) represents the difference between the actual data and the constitutive model, and the optimal (c, φ) value is updated through this distribution; Starting from the computational essence of material and model constitutive elements, strong nonlinear deviations are gradually overcome through AI deep data drive to achieve self-optimized virtual-reality dynamic correction and precise twin mapping.

Images