Wellbore inspection visual recognition method based on digital twinning

Abstract

This invention relates to the field of twin simulation technology, specifically to a visual recognition method for wellbore inspection based on digital twins. The invention acquires the real-time physical perception matrix of the inspection device and the expected simulation baseline matrix of the wellbore digital twin, extracts the local structural disturbance feature sequence through residual calculation, acquires the corresponding transient pose matrix and physical boundary deformation parameters, establishes a geometric mapping relationship, and combines it with the solid mechanics deformation compensation equation to perform in-situ physical quantity secondary compensation on the boundary transfer constraints, solving it into a local damage state update vector. The invention aggregates the structural response change rate, data delay fluctuation, and environmental mutation indicators to calculate the simulation deduction complexity. Based on this complexity, a hierarchical multi-scale simulation model reconstruction strategy is executed. This invention solves the technical contradiction between the decrease in model space fidelity and the loss of system real-time performance under complex dynamic working conditions, significantly improving the physical mapping accuracy and adaptive scheduling capability of digital twins.

Description

Technical Field

[0001] This invention relates to the field of twin simulation technology, specifically to a visual recognition method for wellbore inspection based on digital twins. Background Technology

[0002] In the process of modern deep resource extraction and underground space utilization, vertical or inclined shafts are the vital passage connecting the surface and deep underground space; digital twin technology, as a technical architecture that enables high-fidelity, low-latency two-way interaction between the physical and virtual worlds, is gradually being introduced into mine safety monitoring.

[0003] Existing mine digital twin intelligent monitoring systems, by constructing digital twin models and using deep learning algorithms to perform perception analysis and simulation of multi-source sensor data based on 3D visualized virtual scenes, have initially achieved virtual monitoring and data-driven operation of underground physical space. However, when directly transplanting the existing machine vision and digital twin collaborative architecture to high-speed, deep shaft inspections, serious systemic defects are exposed in the underlying software algorithm scheduling and spatial mapping mechanisms. First, existing technologies suffer from severe nonlinear distortion when processing the spatial mapping between 2D visual data and 3D curved surface twin models. The shaft is a typical ultra-deep, large-diameter cylindrical 3D confined structure. During the vertical traction operation of inspection equipment at depths of hundreds or even thousands of meters, complex dynamic pose fluctuations of six degrees of freedom, such as swaying, yaw, pitch, and high-frequency vibrations, inevitably occur. Currently, most 3D modeling and visualization systems cannot effectively map these 2D visual features to the digital twin model. In terms of modeling, static texture mapping or coarse coordinate linear conversion mechanisms are commonly used. When a 2D image is projected onto a 3D cylindrical surface, uncontrollable coordinate drift, lateral stretching deformation, and topological distortion will occur. Secondly, when facing high-frequency concurrency of massive heterogeneous data, the existing system is prone to computational power collapse and loss of real-time performance due to the use of full-processing mechanism. In the operation scenario of high-speed well inspection, visual sensors generate unstructured image data streams, while various monitoring sensors and hoists distributed inside the well wall are also synchronously generating massive amounts of structured time-series data. When facing such scenarios, existing digital twin collaborative algorithms usually adopt an inefficient coarse-grained synchronization strategy of "full data reception, full-band parsing, and global model update". This not only causes serious delays in instruction processing, but also causes the matching process between the visual representation of physical entities and 3D modeling data to freeze, completely losing the basic function of "real-time bidirectional mapping" that a digital twin system should have.

[0004] In summary, how to resolve the technical contradiction between the severe decrease in model spatial realism and the loss of system real-time performance when digital twin 3D simulation models synchronously process massive amounts of high-frequency heterogeneous sensing data under complex dynamic working conditions is a technical problem that urgently needs to be solved in this field.

[0005] To address this, a visual recognition method for wellbore inspection based on digital twins is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a visual recognition method for wellbore inspection based on digital twins. This invention obtains the real-time physical perception matrix of the inspection device and the expected simulation baseline matrix of the wellbore digital twin, extracts the local structural disturbance feature sequence through residual calculation, obtains the corresponding transient pose matrix and physical boundary deformation parameters, establishes a geometric mapping relationship, and combines it with the solid mechanics deformation compensation equation to perform in-situ physical quantity secondary compensation on the boundary transfer constraints, solving it into a local damage state update vector. It aggregates the structural response change rate, data delay fluctuation, and environmental mutation indicators to calculate the simulation deduction complexity. Based on this complexity, it executes a hierarchical multi-scale simulation model reconstruction strategy. This invention solves the technical contradiction between the decrease in model space fidelity and the loss of system real-time performance under complex dynamic working conditions, significantly improving the physical mapping accuracy and adaptive scheduling capability of digital twins.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A visual recognition method for wellbore inspection based on digital twins includes:

[0009] Step 1: Obtain the real-time physical perception matrix of the inspection equipment, and simultaneously extract the expected simulation reference matrix of the wellbore three-dimensional digital twin model based on the position coordinates of the inspection equipment; perform state-space residual calculation on the real-time physical perception matrix and the expected simulation reference matrix, and extract the local disturbance feature sequence.

[0010] Step 2: Obtain the transient pose matrix and radial deformation parameters of the well wall when acquiring the real-time physical sensing matrix; use the transient pose matrix to construct the geometric mapping relationship for the local disturbance feature sequence to solve the boundary transfer constraints, and combine the radial deformation parameters to construct the physical structure deformation compensation kernel function to perform in-situ deformation secondary compensation for the boundary transfer constraints, and solve the planar discrete data points into the local damage state update vector of the structural simulation model;

[0011] Step 3: Calculate the spatial change based on the local damage state update vector and the abnormal evolution topology of the model, and nonlinearly aggregate the spatial change, data delay fluctuation, and physical environment mutation index to calculate the model evolution complexity; according to the preset dynamic threshold range of the model evolution complexity, execute a hierarchical model reconstruction strategy on the wellbore 3D digital twin model.

[0012] Preferably, in step 1, extracting the local disturbance feature sequence includes: establishing a nonlinear mapping constraint relationship between the basic difference threshold and the physical indicators of environmental humidity and lighting intensity, and dynamically generating an adaptive difference threshold matrix; subtracting the data points in the same coordinate system in the real-time physical perception matrix and the expected simulation reference matrix to obtain an initial residual matrix; comparing each element in the initial residual matrix with the corresponding element in the environmental adaptive difference threshold matrix one by one; and selecting spatially connected regions with values ​​greater than the corresponding elements of the adaptive difference threshold matrix as local disturbance feature sequences.

[0013] Preferably, in step 2, acquiring the transient pose matrix during the acquisition of the real-time physical sensing matrix includes: synchronously receiving continuous triaxial acceleration and triaxial angular velocity data output by the inertial measurement unit inside the inspection equipment, and absolute distance data measured by the external positioning base station; constructing a Kalman filter state transition equation, filtering and solving the continuous triaxial acceleration, triaxial angular velocity data and absolute distance data, and outputting a time-continuous pose function; substituting the data timestamp of the real-time physical sensing matrix into the time-continuous pose function to calculate the transient pose matrix used for intra-frame motion distortion compensation.

[0014] Preferably, in step 2, the boundary transfer constraint is solved by constructing a geometric mapping relationship for the local disturbance feature sequence using the transient pose matrix, including: retrieving the standard cylindrical section equation and design inner diameter parameters of the wellbore 3D digital twin model under the current depth coordinates; constructing a set of geometric ray equations starting from the sensor optical center and passing through the discrete data points in the plane in the 3D virtual space; substituting the discrete data points in the plane in the local disturbance feature sequence as input variables into the set of geometric ray equations, solving for the set of 3D coordinates of the intersection of the set of geometric ray equations and the standard cylindrical section equation, and defining it as the boundary transfer constraint.

[0015] Preferably, in step 2, in-situ deformation secondary compensation is performed on the boundary transfer constraints, and the planar discrete data points are solved into a local damage state update vector of the structural simulation model. This includes: receiving real-time strain data sent by fiber optic strain gauges on the wellbore wall; converting the real-time strain data into radial deformation parameters for the corresponding elevation section according to the constitutive relation of elasticity; generating a three-dimensional structural deformation field function containing a normal displacement scalar and a tangential slip vector using the radial deformation parameters as independent variables; using the three-dimensional structural deformation field function as the physical structural deformation compensation kernel function, superimposing the three-dimensional coordinates of the intersection points in the boundary transfer constraints, synchronously correcting the spatial errors of the three-dimensional coordinates of the intersection points in the wellbore normal direction and tangential direction, and encapsulating them into the local damage state update vector.

[0016] Preferably, in step 3, the spatial variation, data latency fluctuation, and physical environment mutation index are nonlinearly aggregated to calculate the model evolution complexity, including: calculating the product of the three-dimensional intersection-union ratio and Hausdorff distance between the local damage state update vector and the reference damage vector at the same physical elevation in the historical period as the spatial variation; extracting the round-trip time delay variance of data packets between the edge computing node and the cloud server as the data latency fluctuation; reading the wellbore temperature gradient derivative, pore water pressure change rate, and microseismic energy release integral parameters, normalizing and summing them as the physical environment mutation index; applying a logarithmic function to the data latency fluctuation, a natural exponential function to the spatial variation, and a polynomial fitting operation to the physical environment mutation index; performing dimensionless normalization on the data latency fluctuation, spatial variation, and physical environment mutation index obtained after the above operations; assigning a first weight coefficient, a second weight coefficient, and a third weight coefficient to the three normalized values; summing the three values ​​after assigning weight coefficients and outputting the model evolution complexity.

[0017] Preferably, in step 3, a hierarchical model reconstruction strategy is executed on the well shaft 3D digital twin model according to the preset dynamic threshold range of the model evolution complexity, including: presetting a first threshold and a second threshold, wherein the value of the first threshold is less than the value of the second threshold; when the model evolution complexity is less than the first threshold, a first-level strategy is executed: intercepting the 3D mesh redrawing command and appending the local damage state update vector to the underlying time-series database to maintain the rendering state of the well shaft 3D digital twin model unchanged; when the model evolution complexity is not less than the first threshold and less than the second threshold, a second-level strategy is executed: generating a local material mask surrounding the local damage state update vector in the space of the well shaft 3D digital twin model, and only reconstructing the 3D mesh vertex coordinates and face normal vectors inside the local material mask; when the model evolution complexity is not less than the second threshold, a third-level strategy is executed: increasing the priority of the 3D rendering pipeline to the highest level of the system, and forcibly calling the global polygon mesh to perform overall structural deformation simulation rendering of the well shaft 3D digital twin model.

[0018] Preferably, in step 3, after executing the hierarchical model reconstruction strategy, a structural failure deduction step is also included: extracting the initial microcrack coordinates and orientation features from the local damage state update vector; mapping the initial microcrack coordinates and orientation features to a pre-constructed digital twin discrete element mechanical mesh containing the properties of the original rock mass joint surface; under the driving force of the input pore water pressure and principal stress coupled boundary conditions, calculating the stiffness degradation matrix of the anisotropic elements in the digital twin discrete element mechanical mesh using an anisotropic damage mechanics model; based on the stiffness degradation matrix, calculating and outputting the predicted structural failure probability density matrix and evolutionary morphology boundary; in the wellbore three-dimensional digital twin model, using a dynamic polygonal mesh to visualize and spatiotemporally overlay the evolutionary morphology boundary, and triggering an interruption alarm command when the peak value of the predicted structural failure probability density matrix is ​​greater than a preset failure probability safety threshold.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. This invention effectively improves the spatial mapping accuracy and structural fidelity of digital twin models under dynamic working conditions. By introducing a transient pose matrix to establish a geometric mapping relationship and combining physical boundary deformation parameters to construct a solid mechanics deformation compensation equation, this scheme performs in-situ secondary compensation of physical quantities for the system boundary transfer constraints. This mechanism can accurately solve discrete disturbance data under multi-field coupled sensing into local damage state update vectors, effectively ensuring the consistency of spatial coordinates and topological integrity between the physical entity and the virtual model, and providing a more reliable three-dimensional simulation benchmark for subsequent structural stress analysis.

[0021] 2. This invention optimizes the system computing power allocation mechanism when massive multi-source data is concurrent, effectively improving the real-time response performance of the digital twin system. By nonlinearly aggregating structural response change rate, data latency fluctuation, and multi-physics environment mutation indicators, this scheme proposes and quantifies the simulation inference complexity, and triggers a hierarchical multi-scale simulation model reconstruction strategy accordingly. This flexible adaptive scheduling logic can intelligently match computing resources according to the current physical and network environment, reasonably avoid redundant rendering of invalid background areas, and ensure the smooth execution of core simulation reconstruction tasks and the stable flow of the data bus.

[0022] 3. This invention achieves an intelligent functional upgrade from basic condition monitoring to forward-looking structural failure prediction. It deeply integrates the microscopic crack features captured purely by vision into discrete element mechanical calculations constrained by pore water pressure and principal stress. By introducing an anisotropic damage mechanics model, it strictly follows the objective laws of geotechnical engineering mechanics and can quantitatively output the predicted structural failure probability density matrix and evolutionary morphology boundary with strict physical meaning. This multi-physics coupled prediction mechanism effectively solves the problem of insufficient scientific rigor in conventional geometric extrapolation prediction, providing high-fidelity and quantifiable forward-looking decision support for the safety early warning and preventive maintenance of deep-earth engineering facilities. Attached Figure Description

[0023] Figure 1 A flowchart of a wellbore inspection visual recognition method based on digital twins provided in an embodiment of the present invention;

[0024] Figure 2 A flowchart of the simulation deduction complexity calculation and hierarchical reconstruction strategy provided in this embodiment of the invention;

[0025] Figure 3 A flowchart of secondary compensation for multi-field thermo-coupling deformation provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] Please see Figures 1 to 3 This invention provides a visual recognition method for wellbore inspection based on digital twins, and the technical solution is as follows:

[0028] A visual recognition method for wellbore inspection based on digital twins includes:

[0029] The real-time physical perception matrix of the inspection equipment is obtained, and the expected simulation reference matrix of the three-dimensional digital twin model of the wellbore is extracted synchronously based on the position coordinates of the inspection equipment; the state space residual of the real-time physical perception matrix and the expected simulation reference matrix are calculated, and the local disturbance feature sequence is extracted.

[0030] The transient pose matrix and radial deformation parameters of the well wall are obtained when the real-time physical sensing matrix is ​​acquired; the transient pose matrix is ​​used to construct a geometric mapping relationship for the local disturbance feature sequence, and the radial deformation parameters are combined to construct a physical structure deformation compensation kernel function to perform in-situ deformation secondary compensation on the boundary transfer constraints, and the planar discrete data points are solved into the local damage state update vector of the structural simulation model.

[0031] The spatial change is calculated based on the local damage state update vector and the abnormal evolution topology of the model. The spatial change, data delay fluctuation and physical environment mutation index are nonlinearly aggregated to calculate the model evolution complexity. According to the preset dynamic threshold range of the model evolution complexity, a hierarchical model reconstruction strategy is executed on the wellbore three-dimensional digital twin model.

[0032] Example 1:

[0033] This embodiment is applied to the safety inspection scenario of vertical shafts in deep underground coal or metal mines. In deep underground space, the inspection device mounted on the hoisting equipment faces extreme working conditions such as high concentrations of groundwater mist, uneven illumination from searchlights, high-frequency stick-slip vibrations of the hoist, and minor deformations of the shaft structure due to complex ground stress. When receiving such massive, high-frequency sensing data accompanied by severe noise, traditional digital twin monitoring systems typically use passive full-scale two-dimensional image mapping updates, resulting in severe spatial mapping coordinate drift, three-dimensional topology distortion, and severe congestion of the computing bus. This embodiment aims to solve the above-mentioned technical problems through multiphysics constraints and adaptive computing power scheduling algorithms.

[0034] As one embodiment of the present invention, refer to Figure 1 Flowchart of a visual recognition method for wellbore inspection based on digital twins, referencing Figure 2 The flowchart for simulation derivation complexity calculation and hierarchical reconstruction strategy is shown below. Figure 3 Flowchart of secondary compensation for multi-field thermo-coupling deformation.

[0035] Before extracting the expected simulation baseline matrix, a heterogeneous data spatiotemporal alignment step is included: establishing a dynamic circular buffer queue, pushing the real-time physical sensing matrix with a high sampling frequency and the physical boundary deformation parameters with a low sampling frequency into the dynamic circular buffer queue respectively, and attaching a system timestamp; using the system timestamp of the real-time physical sensing matrix as a time reference, performing a time series interpolation algorithm on the physical boundary deformation parameters with a low sampling frequency to reconstruct a high-dimensional physical state tensor that strictly corresponds to the exposure time of the real-time physical sensing matrix; using the high-dimensional physical state tensor to replace the original asynchronous data input in the subsequent compensation calculation process to eliminate the spatiotemporal simulation mapping misalignment caused by asynchronous sampling.

[0036] Specifically, in actual inspection buses, the image matrix output frequency of vision sensors is usually as high as sixty frames per second, while the sampling frequency of physical sensors such as well wall fiber optic strain gauges is often only ten times per second, and network transmission is subject to random jitter. A dynamic circular buffer queue is opened in the memory of the edge computing node to record the high-resolution timestamps of various data arrivals in real time. Before extracting the simulation reference field, the timestamp of a certain frame of real-time physical perception matrix being processed is locked, and the physical boundary deformation parameter nodes before and after that time point are retrieved from the buffer queue. The physical strain value at that moment is calculated and reconstructed using the Gaussian process regression algorithm.

[0037] This invention effectively resolves the timestamp misalignment between high-frequency visual perception and low-frequency physical sensing data at the software computation level by establishing a dynamic circular buffer queue and a time series interpolation algorithm. This solution does not rely on stringent and impractical hardware-level clock synchronization, and can reconstruct strictly aligned high-dimensional physical state tensors under complex dynamic conditions. It effectively compensates for simulation model spatial mapping errors caused by the asynchronous arrival of heterogeneous data. While ensuring the real-time performance of system bus data flow, it significantly improves the physical fidelity of multi-source sensing data fusion and reconstruction.

[0038] Further, the local disturbance feature sequence is extracted, including: establishing a nonlinear mapping constraint relationship between the basic difference threshold and the physical indicators of environmental humidity and lighting intensity, and dynamically generating an adaptive difference threshold matrix; subtracting the data points in the same coordinate system in the real-time physical perception matrix and the expected simulation reference matrix to obtain an initial residual matrix; comparing each element in the initial residual matrix with the corresponding element in the environmental adaptive difference threshold matrix one by one; and selecting spatially connected regions with values ​​greater than the corresponding elements of the adaptive difference threshold matrix as local disturbance feature sequences.

[0039] Specifically, during actual well inspections, the physical environment varies significantly at different elevations within the well. The system reads real-time data from the ambient humidity sensor and illuminance meter at the current elevation. For example, when the inspection device approaches the aquifer of the well, the ambient humidity may jump to between 85% and 95%, at which point a large amount of suspended water mist will cause scattering of light spots on the visual sensor. A mapping rule is established in the algorithm backend: when the ambient humidity exceeds a preset reference humidity threshold (e.g., 80%) and the local lighting intensity is lower than a preset reference illuminance (e.g., 50 lux), the value of the basic differential threshold is increased for that specific area.

[0040] The calculation process is as follows: The nonlinear mapping constraint relationship is realized through a pre-constructed two-dimensional lookup table: First, the environmental humidity range (e.g., 0 to 40%, 40% to 60%, 60% to 80%, 80% to 95%, 95% to 100%) and the lighting intensity range (e.g., 0 to 20, 20 to 50, 50 to 100, 100 to 200, 200 to 500 lux) are divided; Second, the threshold compensation coefficient in the lookup table is determined based on the calibration experiment. For example, the coefficient is 0.85 in a good environment with humidity below 40% and illuminance above 100 lux, while the coefficient nonlinearly jumps to 2.0 in an extremely harsh environment with humidity above 95% and illuminance below 20 lux.

[0041] The two-dimensional mapping lookup table is represented in system memory as a two-dimensional data matrix defined by row and column indices. Specifically, the row indexes divide the range according to ambient humidity, and the column indexes divide the range according to lighting intensity. Each cell within the matrix stores a preset threshold compensation coefficient. An example of the structure and data of the two-dimensional mapping lookup table is shown in Table 1.

[0042] Table 1 Adaptive Differential Threshold Compensation Coefficients under Different Environmental Conditions

[0043]

[0044] As shown in Table 1, the row index of the two-dimensional mapping lookup table is divided into 5 environmental humidity ranges, and the column index is divided into 5 lighting intensity ranges. These two categories intersect to define 25 independent physical environmental conditions in the wellbore. Each cell within the matrix stores a threshold compensation coefficient calibrated through offline grid search iterations. During real-time system operation, edge computing nodes instantly retrieve specific coefficients suitable for the current operating condition based on single-point physical indicators fed back by online sensors.

[0045] To prevent array index overflow in the lookup table algorithm caused by extreme environmental physical data or hardware failure data collected by multi-source sensors, the system incorporates a boundary clamping mechanism before performing the two-dimensional mapping lookup. When the real-time illumination intensity data exceeds 500 lux (e.g., when the inspection device is operating in a wellhead section with sufficient light, or under direct strong light from other equipment), the system uses software logic to force the input value to be clamped to the maximum boundary reference value allowed by the lookup table (i.e., 500 lux), and calls the "200 to 500 lux" end column index for matching. Similarly, if the ambient humidity sensor causes an overload of the underlying electrical signal due to a short circuit, electrical fault, or physical condensation on the probe surface, resulting in an abnormal value exceeding 100% of the measurement range, the system uses software logic to force the input value to be clamped to 100%, and matches the "95% to 100%" end row index.

[0046] The two-dimensional mapping lookup table is constructed through an offline calibration process, specifically including the following steps:

[0047] Step 1, Multi-condition benchmark sample collection: During the initial construction phase or routine maintenance cycle of the shaft to be inspected, simulated lowering operations are performed in the normal shaft wall section without structural defects using the inspection device; during this process, by manually adjusting the underground ventilation and humidification system and adjusting the searchlight power, 25 different physical environment conditions covering the above 5 humidity ranges and 5 illumination ranges are artificially simulated, and benchmark images of unstructured deformation under the corresponding conditions are collected and stored to form a benchmark sample set;

[0048] Step 2, Digital Damage Feature Injection: For each image in the benchmark sample set, computer-aided design is used to manually overlay simulated structural damage features with known real geometric dimensions and contrast at random pixel coordinates of the image; since the real spatial coordinates and edge pixel boundaries of the injected simulated damage are completely known quantities at the algorithm backend, this constitutes a standard test sample set.

[0049] Step 3, Threshold Compensation Coefficient Grid Search Iteration: For a specific environmental condition image in the standard test sample set, a threshold compensation coefficient is set between 0.5 and 2.5 with a step size of 0.05 for grid search. At each coefficient step size, the residual between the perception state matrix and the simulation reference field is calculated, and the damage feature recognition results under that coefficient are statistically analyzed. Two core indicators are calculated and evaluated: the first indicator is the false alarm rate, which is the proportion of artificially injected lesions that were not successfully detected out of the total injected pixels; the second indicator is the false alarm rate, which is the proportion of background pixels that were incorrectly identified as lesions due to water mist drift or uneven lighting out of the total background pixels.

[0050] Step 4, Optimal Solution Selection and Form Consolidation: A joint optimization objective function is set for the system. This objective function requires that the false alarm rate be minimized while maintaining a false alarm rate below a preset rigid constraint threshold (e.g., 2%, determined based on the maximum computing power tolerance of the digital twin base for redundant model reconstruction instructions). After traversing all step sizes, a specific threshold compensation coefficient that enables the joint optimization objective function to achieve its optimal solution is selected and used as the optimal calibration value for that specific environmental condition. The same iterative solution is then applied to 25 environmental conditions, and the 25 final optimal calibration values ​​are written into their corresponding cells and encapsulated and stored as the two-dimensional mapping lookup table.

[0051] Subsequently, a basic differential threshold of 20 grayscale units is set. The ambient humidity and lighting intensity at the current elevation are read in real time. The corresponding compensation coefficient is retrieved from the lookup table, and the basic differential threshold is multiplied by the compensation coefficient to obtain an adaptive differential threshold scalar. Finally, this scalar is iterated and assigned to a blank matrix to generate an adaptive differential threshold matrix. The mapping mechanism ensures, from a physical logic perspective, that the threshold can be dynamically increased to filter out environmental artifact noise in environments with strong water mist scattering and low signal-to-noise ratio. The preset basic differential threshold is preferably a grayscale value of 15 to 25. This setting range can effectively filter out the dark current noise of conventional visual sensors and retain the true physical boundary disturbances before environmental compensation. Subsequently, the matrix is ​​subtracted element by element. If the residual value in a connected region exceeds the dynamically adjusted threshold, it is determined to be a real structural defect (such as cracks or spalling). Otherwise, it is determined to be an artifact caused by water mist or light fluctuations and is removed.

[0052] This invention introduces ambient humidity and lighting intensity as dynamic adjustment variables, which changes the problem of poor environmental adaptability caused by the use of fixed differential thresholds in traditional algorithms. In the complex optical environment of high humidity and low signal-to-noise ratio in underground wells, this method can intelligently and dynamically shield high-frequency background noise caused by water mist drift and light and shadow changes, significantly reducing the false feature extraction rate caused by environmental interference, thereby improving the data purity and reliability of local perturbation feature sequences input into the digital twin simulation system.

[0053] Before acquiring the real-time physical sensing matrix of the inspection device and extracting the local structural disturbance feature sequence, the process includes a numerical perspective preprocessing step based on multispectral fusion: simultaneously acquiring the visible light band sensing matrix and the penetrating band intensity reflection matrix; introducing a defogging prior algorithm based on the physical optical scattering model to calculate the suspended water particle scattering distortion parameters in the visible light band sensing matrix; performing adaptive weight fusion calculation on the visible light band sensing matrix and the penetrating band intensity reflection matrix at the feature map level, and assigning dynamic gain weights to the penetrating band intensity reflection matrix according to the suspended water particle scattering distortion parameters; reconstructing the underlying structural representation matrix after filtering out optical scattering distortion through the adaptive weight fusion calculation, and using this underlying structural representation matrix as the real-time physical sensing matrix to perform subsequent system state residual calculation.

[0054] Specifically, in extreme sections of the wellbore with severe water seepage or high concentrations of water mist, a single visible light sensor can become "blinded" due to strong optical scattering, causing the digital twin system to completely lose its physical update source. Therefore, the inspection device simultaneously activates both the visible light sensing unit and the infrared thermal imaging sensing unit. After receiving the two sets of heterogeneous matrices, it does not directly perform spatial difference comparison. First, it runs a defogging prior algorithm based on an atmospheric physical scattering model in the background to quantitatively assess the optical scattering distortion parameters caused by suspended water particles in the current visible light matrix (i.e., the estimation of the water mist concentration distribution field). Subsequently, it performs nonlinear adaptive weight fusion calculation at the pixel feature map level: in areas with low water mist concentration... The visible light matrix is ​​given higher weights to preserve surface texture details; in the blind zone where the water mist scattering distortion parameter exceeds the standard, the gain weight of the penetrating band matrix is ​​automatically amplified; the weight allocation of the visible light matrix is ​​negatively correlated with the calculated scattering distortion parameter of the suspended water particles, and the weight allocation of the penetrating band matrix is ​​positively correlated with the scattering distortion parameter, and the sum of the normalized weights of the two at the same pixel coordinate is always 1; through this multimodal numerical calculation, a clear underlying structure representation matrix that is not affected by water mist is "seen through" and reconstructed in the algorithm memory; finally, this pure matrix that filters out physical environment artifacts is sent to the subsequent digital twin residual calculation stage.

[0055] This invention addresses the technical shortcomings of high-concentration water mist that frequently obstructs visible light in deep wells, leading to feature extraction failures. By fusing penetrating band data and combining it with a scattering model to assign adaptive weights, the optical distortion caused by suspended water vapor is directly filtered out at the software algorithm level. This supplementary mechanism eliminates the need for complex and energy-intensive physical defogging and dust removal hardware downhole, reconstructing a clear structural representation matrix through pure numerical simulation calculations. This greatly enhances data penetration and feature extraction robustness in extremely harsh optical environments.

[0056] Furthermore, acquiring the transient pose matrix during the acquisition of the real-time physical sensing matrix includes: synchronously receiving continuous triaxial acceleration and triaxial angular velocity data output by the inertial measurement unit inside the inspection equipment, as well as absolute distance data measured by the external positioning base station; constructing a Kalman filter state transition equation, wherein the state vector is defined to include the three-dimensional position, velocity, and attitude quaternions of the inspection equipment and the inertial sensor zero bias; using the continuous triaxial acceleration and triaxial angular velocity data as the system state prediction input, establishing a kinematic discrete transfer function; using the absolute distance data measured by the external positioning base station as the observation update input, constructing a measurement matrix; alternately filtering and solving the prediction and update to output a time-continuous pose function; substituting the data timestamp of the real-time physical sensing matrix into the time-continuous pose function to calculate the transient pose matrix used for intra-frame motion distortion compensation.

[0057] The specific mathematical calculation form of the Kalman filter state transition equation is as follows: A 16-dimensional state vector is defined, whose elements sequentially include three-dimensional position, three-dimensional velocity, attitude quaternions, and the three-axis zero bias of the accelerometer and the gyroscope; the predicted state vector of the system at the current moment is equal to the optimal estimated state vector of the previous moment multiplied by a 16th-order state transition matrix, plus a 16x6 control input matrix multiplied by the 6-dimensional acceleration and angular velocity vectors currently output by the inertial measurement unit, and superimposed with a 16-dimensional process noise covariance vector; wherein, the main diagonal block matrix of the state transition matrix is ​​based on the kinematic differential equation. The process is discretized and includes the sampling period time difference variable; the control input matrix is ​​constructed based on the current attitude rotation matrix; the process noise covariance is a diagonal matrix, whose diagonal elements consist of the noise variances of position, velocity, attitude, and zero bias; at the same time, the observation update equation is constructed: the scalar absolute ranging observation value in the case of a single base station is equal to the observation mapping matrix of 1 row and 16 columns (whose non-zero elements are the partial derivatives of the Euclidean distance with respect to the three-dimensional position components) multiplied by the current predicted state vector, and the observation noise determined by the ultra-wideband ranging standard deviation is added; by alternately executing the state prediction and observation update solution, a smooth time continuous pose function is output.

[0058] Specifically, the inspection device operates at a speed of 3 to 5 meters per second under the traction of a flexible steel wire rope, accompanied by high-frequency vibration. In this embodiment, the inertial measurement unit outputs the acceleration and angular velocity of the device at a high sampling rate of 200 Hz, while the ultra-wideband positioning base station deployed on the well wall provides absolute ranging values ​​at a frequency of 50 Hz. A Kalman filter state prediction and observation update module is constructed to correct the high-frequency integral drift error of the inertial measurement unit with ultra-wideband low-frequency absolute position data, generating a smooth time-continuous pose function. For the visual perception matrix (e.g., using a rolling shutter sensor to acquire data at 30 frames per second), the high-resolution timestamp of the exposure of each row of pixels in the matrix is ​​accurately read and substituted into the above pose function to calculate the independent six-degree-of-freedom translation vector and rotation quaternion at the moment of exposure of that row. Based on the transient pose matrix, the original perception matrix is ​​subjected to row-by-row reverse geometric compensation calculation to restore its true physical geometry.

[0059] This invention effectively overcomes the distortion of observation data caused by high-speed movement and nonlinear vibration of the inspection device in the confined space of a deep well. By fusing high-frequency inertial data and low-frequency absolute positioning data through Kalman filtering, and performing row-by-row pose calculation on the sensing matrix in the high-resolution timestamp dimension, this technique eliminates intra-frame motion distortion caused by equipment yaw, pitch and sway from the kinematic level, providing a high-fidelity reference coordinate system with global consistency constraints for the subsequent three-dimensional spatial mapping of the digital twin model.

[0060] After outputting the time-continuous pose function containing translation vectors and rotation quaternions, the process further includes a second-order fine calibration step based on the prior topology of the digital twin: extracting the rigid engineering structure model in the three-dimensional digital twin model of the wellbore located at the elevation segment corresponding to the absolute distance data, and using it as a prior three-dimensional spatial landmark; extracting planar feature points corresponding to the prior three-dimensional spatial landmarks from the real-time physical perception matrix, and calculating the deviation between the actual observed feature points and the projection points of the theoretical model to construct a reprojection error objective function; constructing a visual-inertial tightly coupled nonlinear optimization framework by combining the reprojection error objective function with the pre-integration error of the inertial measurement unit; and iteratively solving the nonlinear optimization framework using a factor graph optimization algorithm to secondarily correct the yaw angle and translation drift parameters in the time-continuous pose function.

[0061] Specifically, basic ultra-wideband positioning inevitably generates electromagnetic multipath effects in well shafts filled with metal conduits and high-voltage cables, resulting in local spur errors at the centimeter or even decimeter level in the initial pose function output by the Kalman filter. To meet the requirements of high-fidelity digital twin mapping, after obtaining the basic pose, a search is initiated on the cloud model to extract the 3D model data of rigid engineering components (such as standard flange joints or ladder brackets) with known absolute coordinates at the current basic elevation, serving as reliable prior 3D spatial landmarks. The edge corner features of these components are extracted in the real-time physical perception matrix, and the geometric distance between the actual visual observation pixels and the theoretical projection points of the model is calculated to form the reprojection error. Subsequently, a factor graph network is constructed in the background, using the visual reprojection error as the observation factor and the high-frequency pre-integration error of the inertial measurement unit as the motion factor, and performing joint nonlinear least squares optimization. Through continuous iteration within the time sliding window, the algorithm accurately "squeezes out" the residual yaw drift and small horizontal offset in the basic pose, outputting a limit-precision pose tensor after secondary fine calibration.

[0062] This invention, building upon the basic ultra-wideband and inertial filtering positioning mechanism, supplements it with a secondary pose fine calibration mechanism based on digital twin prior topology. Addressing the technical shortcomings of ultra-wideband signals in the complex electromagnetic environment of deep wells, which are prone to multipath interference and struggle to meet millimeter-level spatial mapping requirements, this solution cleverly utilizes the rigid structure within the digital twin as a priori landmark to construct an optimized framework with tight visual-inertial coupling. This progressive strategy of "global filtering coarse positioning plus local twin tight coupling fine calibration" not only eliminates the long-duration cumulative drift of inertial navigation but also provides a high-precision pose reference with highly stable spatial reference for subsequent digital twin spatial mapping.

[0063] Furthermore, the boundary transfer constraint is solved by constructing a geometric mapping relationship for the local disturbance feature sequence using the transient pose matrix, including: retrieving the standard cylindrical section equation and design inner diameter parameters of the wellbore 3D digital twin model under the current depth coordinates; constructing a set of geometric ray equations starting from the sensor optical center and passing through the discrete data points in the plane in the 3D virtual space; substituting the discrete data points in the plane in the local disturbance feature sequence as input variables into the set of geometric ray equations, solving for the set of 3D coordinates of the intersection points of the set of geometric ray equations and the standard cylindrical section equation, and defining them as the boundary transfer constraint.

[0064] The specific construction and solution process is as follows: Extract the 3rd-order rotation component matrix and the 3D translation component vector representing the transformation from the camera coordinate system to the world coordinate system from the transient pose matrix; for each 2D pixel coordinate in the local perturbation feature sequence, use the camera intrinsic parameter matrix containing the focal length and principal point coordinates to back-project it onto the normalized plane to obtain the local direction vector; use the rotation component matrix to transform the local direction vector to the world coordinate system, and use the camera optical center corresponding to the translation component vector as the starting point to construct a spatial geometric ray equation in the 3D virtual space; subsequently, retrieve the standard cylindrical section equation for the current depth of the digital twin, substitute the above geometric ray equation into the cylindrical section equation, and solve the quadratic equation to calculate the set of 3D intersection points of the ray and the cylindrical surface; repeat this spatial projection and nonlinear intersection operation for all discrete pixels in the feature sequence, and the generated coordinate point set is the boundary transfer constraint data.

[0065] Specifically, assuming the wellbore design inner diameter at the current inspection depth is 4.0 meters, the mathematical equation of the standard cylindrical section corresponding to this elevation is retrieved from memory. Using the calculated transient pose matrix as the rotation and translation transformation operator between the world coordinate system and the local camera coordinate system, for each discrete data point (i.e., defect edge pixel) in the local disturbance feature sequence, a geometric ray equation is constructed in the three-dimensional virtual space, starting from the sensor optical center, passing through the data point, and extending outward. Subsequently, the simultaneous equation system is solved to obtain the set of three-dimensional intersection points of all geometric rays corresponding to the local disturbance feature sequence and the theoretical cylindrical section equation with a radius of 4.0 meters in space. These point sets containing absolute X, Y, and Z coordinate values ​​constitute the boundary transfer constraint data for updating the constraint structure simulation model.

[0066] This invention changes the conventional approach of directly and crudely attaching two-dimensional features as orthogonal textures to the surface of a three-dimensional mesh. By constructing a rigorous set of geometric projection ray equations and performing nonlinear intersection operations between each point of the two-dimensional feature and the global cylindrical surface equation of the digital twin, this method effectively overcomes the problems of tangential stretching, scaling imbalance, and topological destruction caused by projecting planar features onto a restricted, highly curved cylindrical surface at the computer simulation level, and significantly improves the spatial mapping fidelity of structural damage features.

[0067] Furthermore, in-situ deformation secondary compensation is performed on the boundary transfer constraints, and the discrete planar data points are solved into a local damage state update vector of the structural simulation model. This includes: receiving real-time strain data transmitted by fiber optic strain gauges on the wellbore wall; converting the real-time strain data into radial deformation parameters for the corresponding elevation section according to the constitutive relation of elasticity; generating a three-dimensional structural deformation field function containing a normal displacement scalar and a tangential slip vector using the radial deformation parameters as independent variables; using the three-dimensional structural deformation field function as the physical structural deformation compensation kernel function, superimposing the three-dimensional coordinates of the intersection points in the boundary transfer constraints, simultaneously correcting the spatial errors of the three-dimensional coordinates of the intersection points in the wellbore normal direction and tangential direction, and encapsulating them into the local damage state update vector.

[0068] The specific calculation process for the three-dimensional structural deformation field function and deformation compensation is as follows: Dimensionless circumferential strain data measured by fiber optic strain gauges arranged circumferentially along the well wall is read. Based on the generalized Hooke's law of elasticity and cylinder theory, this data is multiplied by the designed inner diameter of the well barrel to calculate the corresponding radial displacement. A cubic spline interpolation algorithm is applied to the radial displacement data of a finite number of discrete elevation points to generate a continuous radial deformation distribution function along the elevation direction. For any point in three-dimensional space, the radial deformation at its elevation is calculated and multiplied by a radial unit vector pointing towards the cylinder axis to obtain the result. The normal displacement vector is calculated. Simultaneously, considering the tangential slip due to gravity settlement in the deep well, a coupling relationship between vertical and radial strain is established based on Poisson's ratio. The radial deformation at the corresponding elevation is multiplied by an empirical settlement coefficient (e.g., 0.1) to obtain the tangential slip vector downwards along the Z-axis. The normal displacement vector and the tangential slip vector are added together to synthesize the total deformation vector function. For each three-dimensional intersection point coordinate in the boundary transfer constraints, the corresponding spatial displacement correction value is substituted into the above total deformation vector function to obtain the corresponding value. This value is then superimposed onto the original coordinates using vector addition in situ to complete the two-dimensional deformation compensation.

[0069] Specifically, the wellbore wall undergoes creep due to in-situ stress. Micro-strain values ​​in the current area are read using embedded fiber optic strain gauges, and converted into physical radial deformation parameters (e.g., a 3.5 mm inward contraction displacement of the wellbore wall in a certain area) according to the rock mechanics constitutive equations. A three-dimensional structural deformation field function is generated, converting this contraction displacement into a displacement scalar inward along the normal direction of the wellbore cross-section, and a tangential slip vector affected by gravity settlement. After obtaining the set of three-dimensional coordinates of the intersection points calculated in the aforementioned steps, the deformation field function is used as a compensation kernel function to perform vector addition on each intersection point coordinate, i.e., superimposing the aforementioned 3.5 mm normal displacement and the corresponding tangential displacement onto the original geometrically calculated coordinates. The three-dimensional point cloud data formed after secondary displacement correction is formatted and encapsulated as a local damage state update vector, and input into the next stage.

[0070] This invention introduces the hidden engineering mechanical response into the geometric mapping process of the digital twin system. Since the physical wellbore is not an absolutely rigid body, simple geometric ray projection cannot reflect the actual deformation of the well wall under multi-field stress. This method uses a three-dimensional structural deformation field function to perform in-situ superposition correction on the initial projection points, which makes up for the dynamic micro-difference between the virtual design model and the physical reality, so that the final output local damage state update vector has high fidelity accuracy at the mechanical constraint level in the spatial coordinate system.

[0071] The secondary compensation process, which involves constructing a solid mechanical deformation compensation equation, also includes a thermodynamic deformation compensation step based on multi-field coupling: Real-time surface temperature data of the wellbore's inner wall is read using a surface temperature sensor, and combined with the wellbore ventilation velocity and pre-stored historical geothermal gradient data in the digital twin of the wellbore structure, a transient heat conduction inversion differential equation is constructed; the transient heat conduction inversion differential equation is solved to estimate the transient thermal stress field inside the wellbore structure, and the transient thermal stress field is converted into a thermal expansion displacement vector; the thermal expansion displacement vector is tensor-superimposed with the three-dimensional structural deformation field to generate a thermodynamically coupled comprehensive deformation field; the thermodynamically coupled comprehensive deformation field is used to replace the original three-dimensional structural deformation field, and spatial displacement superposition and dense correction are performed on the three-dimensional coordinates of the intersection points in the boundary transfer constraints.

[0072] Specifically, in the actual operation of kilometer-deep wells, the temperature of the surrounding primary rock mass often reaches over 40 degrees Celsius, while the temperature of fresh convective ventilation injected into the wellbore from the surface is usually around ten degrees Celsius. This long-term, huge temperature gradient causes extremely complex thermal expansion or contraction stresses to form inside the concrete well wall, which is tens of centimeters thick. Fiber optic strain gauges, which are only attached to the surface, can only capture the local mechanical strain and cannot accurately isolate and quantify the deep thermodynamic deformation components. Therefore, while performing basic fiber optic strain compensation, the temperature of the well wall obtained by infrared thermometers must also be considered. Real-time surface temperature and current ventilation velocity are combined with the constant geothermal data of the original rock mass at this elevation pre-stored in the digital twin model; a system of partial differential equations is established based on Fourier's law of heat conduction and numerically solved in the background to deduce the transient thermal stress field of the temperature gradient inside the well wall, and then the three-dimensional thermal expansion displacement vector is calculated; subsequently, tensor addition is performed to superimpose the thermal expansion displacement vector with the basic three-dimensional structural deformation field generated by the aforementioned fiber optic strain gauge to generate a thermo-mechanical coupled comprehensive deformation field; finally, this comprehensive deformation field is used to perform a rigorous three-dimensional coordinate secondary correction on the intersection of the boundary geometric projection.

[0073] This invention, building upon the existing reliance on fiber optic strain gauges for mechanical deformation compensation, further introduces a thermodynamic deformation compensation mechanism based on transient thermal conduction inversion. Deep wellbores are subjected to severe temperature differences due to high geothermal temperatures and strong convection ventilation, leading to hidden thermal expansion and contraction deep within the well walls—a phenomenon difficult to quantify comprehensively using conventional surface monitoring. This solution strictly adheres to thermodynamic conduction laws, inverting the deep thermal stress field using multimodal data and superimposing it with the mechanical deformation field to achieve comprehensive deformation compensation under multi-physics coupling. This supplementary mechanism effectively eliminates the system blind spots of single mechanical compensation under extreme geothermal conditions, significantly improving the spatial mapping of structural damage characteristics in terms of mechanical rigor and physical fidelity.

[0074] Furthermore, the spatial variation, data latency fluctuation, and physical environment abrupt change index are nonlinearly aggregated to calculate the model evolution complexity. This includes: calculating the product of the three-dimensional intersection-union ratio (IUU) and Hausdorff distance between the local damage state update vector and the reference damage vector at the same physical elevation in the historical period as the spatial variation; extracting the round-trip time delay variance of data packets between the edge computing node and the cloud server as the data latency fluctuation; reading the wellbore temperature gradient derivative, pore water pressure change rate, and microseismic energy release integral parameters, normalizing them, and summing them as the physical environment abrupt change index; applying a logarithmic function to the data latency fluctuation, a natural exponential function to the spatial variation, and a polynomial fitting operation to the physical environment abrupt change index; performing dimensionless normalization on the data latency fluctuation, spatial variation, and physical environment abrupt change index obtained after the above operations; assigning a first weight coefficient, a second weight coefficient, and a third weight coefficient to the three normalized values; summing the three values ​​after assigning weight coefficients to output the model evolution complexity.

[0075] Specifically, nonlinear aggregation calculations of varying complexity are performed in the background. First, the three-dimensional intersection-union ratio and Hausdorff distance of the old and new damage vectors are calculated, and their multiplication yields the spatial change at the structural geometry level. Simultaneously, the network monitoring module calculates the round-trip delay variance of the most recent 100 data packets (e.g., stabilizing within a 15-millisecond fluctuation range) to assess communication congestion. The physical sensing module reads the temperature change rate, pressure change rate, and microseismic integral (e.g., microseismic energy jumps to a higher level), and performs interval mapping to 0-1 normalization on these three sets of heterogeneous data before summing them. Subsequently, the delay... The logarithm of the late variance is used to smooth the peak of abrupt changes, and the natural exponential operation is performed on the spatial variation to amplify the sudden propagation characteristics of micro-cracks. The environmental indicators are subjected to a polynomial operation of a preset number of times; the preset number of times is preferably 2 to 3 times, so as to moderately amplify the abrupt characteristics of the physical environment nonlinearly, while effectively avoiding the risk of computing power overflow caused by high-order calculations. Finally, according to the business focus, the results of these three operations are assigned weight coefficients of 0.3, 0.4 and 0.3 respectively, and the three are weighted and added together to obtain a scalar value, which is the model evolution complexity score at the current moment.

[0076] This invention proposes a nonlinear evaluation mechanism for multimodal data fusion. By scientifically integrating the spatial variation representing structural degradation, the time delay fluctuation representing computational transmission pressure, and the environmental mutation index representing physical state stability, this calculation method provides a quantifiable and unified standard for measuring the urgency and computational overhead of digital twin system state updates under complex dynamic conditions, laying a reliable data foundation for the subsequent realization of intelligent computing power scheduling.

[0077] Furthermore, based on the preset dynamic threshold range of the model evolution complexity, a hierarchical model reconstruction strategy is executed on the well shaft 3D digital twin model, including: pre-setting a first threshold and a second threshold, where the value of the first threshold is less than the value of the second threshold; when the model evolution complexity is less than the first threshold, a first-level strategy is executed: intercepting the 3D mesh redrawing command and appending the local damage state update vector to the underlying temporal database to maintain the rendering state of the well shaft 3D digital twin model unchanged; when the model evolution complexity is not less than the first threshold and less than the second threshold, a second-level strategy is executed: generating a local material mask surrounding the local damage state update vector in the space of the well shaft 3D digital twin model, and only reconstructing the 3D mesh vertex coordinates and face normal vectors inside the local material mask; when the model evolution complexity is not less than the second threshold, a third-level strategy is executed: increasing the priority of the 3D rendering pipeline to the highest level of the system, and forcibly calling the global polygon mesh to perform overall structural deformation simulation rendering of the well shaft 3D digital twin model.

[0078] Specifically, the first and second thresholds are not fixed constants, but adaptive parameters calibrated based on historical data of specific shaft operating conditions. The specific acquisition method is as follows: The arithmetic mean and standard deviation of the system simulation complexity of the inspection shaft within a continuous stable operating cycle (e.g., the past 30 days) are statistically analyzed; the sum of the arithmetic mean and one standard deviation is set as the first threshold (e.g., calibrated to 25 for a specific mine) to filter out normal system fluctuations; the lower bound of the extreme value of the simulation complexity when the shaft experienced abnormal events such as minor structural displacements in its history is extracted, and converted using an engineering early warning safety factor; the converted result is set as the second threshold (e.g., calibrated to 75 for this mine). This calibration method, combining statistics and historical operating conditions, ensures the reproducibility and accuracy of the dynamic threshold range under different geological conditions and communication environments. When the calculated model evolution complexity score is 18 (less than 25), it is judged as redundant data of extremely minor defects or environmental noise. The first-level strategy is executed, blocking the GPU rendering operation and only adding a log of "minor disturbance at coordinates X, Y" to the database to save system computing power. When the score is 45 (between 25 and 75), it is judged as a normal, slowly evolving structural crack. The second-level strategy is executed, defining a local material mask boundary of, for example, 2 meters × 2 meters, in the model, and only allocating computing power to redraw the vertex offsets of the 3D mesh within this small block. When the score soars to 85 (greater than 75) due to sudden water inrush or displacement, the third-level strategy is executed. The operating system kernel suspends all low-priority processes and allocates all video memory and computing resources to the main rendering program, instantly completing the macroscopic deformation reconstruction and alarm display of the well shaft 3D model.

[0079] This invention breaks away from the rigid process of traditional digital twin systems that require "full updates for every data item." By implementing a three-level elastic model reconstruction strategy, this method can automatically identify and intercept a large number of low-value invalid rendering updates when faced with massive concurrent sensor surges. It precisely allocates valuable computing and rendering resources to specific spatiotemporal nodes experiencing local high-frequency evolution or global catastrophes. This alleviates the congestion of the data flow bus and ensures the extremely low latency and real-time response capability of the digital twin system.

[0080] Furthermore, after implementing the hierarchical model reconstruction strategy, the structural failure deduction step is also included: the underlying mechanical evolution calculation of the anisotropic damage mechanics model is implemented as follows: First, a third-order diagonal matrix is ​​defined as the second-order damage tensor, whose diagonal elements represent the damage degree in the three principal directions X, Y, and Z (values ​​ranging from 0 to 1); based on the strain equivalence assumption, the nominal stress tensor is divided by the difference between the corresponding direction's one minus the damage variable to calculate the effective stress tensor; then, the stress state is judged element by element: if the principal stress is positive, it is determined to be under tension; if the equivalent shear stress is greater than the preset threshold of 5 MPa, it is determined to be under shear; for tensile damage, its evolution equation is set as the nonlinear equation after the equivalent strain minus the initial tensile threshold. Power function; for shear damage, the evolution equation is a nonlinear power function of shear strain minus the initial shear threshold; the calculated tensile and shear damage variables are mapped and distributed to the diagonal elements of the damage tensor according to the principal stress directions; each stress-strain component in the initial 6th order elastic constitutive matrix considering the isotropic assumption is multiplied by the product of the corresponding direction minus the damage variable, and the anisotropic stiffness degradation matrix containing the stiffness attenuation coefficient is calculated and output; the above boundary conditions, damage tensor and stiffness matrix are iteratively updated within a preset period of 30 to 180 days with a time step of 1 day; if any diagonal element of damage in a certain element is greater than 0.8, it is determined to be a failure, and finally the predicted structural failure probability density matrix is ​​constructed statistically.

[0081] Driven by the coupled boundary conditions of pore water pressure and principal stress, the stiffness degradation matrix of anisotropic elements in the digital twin discrete element mechanical mesh is calculated using an anisotropic damage mechanics model. Specifically, this includes: calculating the effective stress tensor of the element by superimposing the principal stress and pore water pressure according to the effective stress principle; constructing anisotropic damage evolution equation based on the strain equivalence assumption, using the effective stress tensor as the driving force and combining the properties of rock mass joint surfaces, to solve for the orthogonal anisotropic damage variables under tension and shear states; substituting the damage variables into the initial elastic constitutive matrix to calculate and output the stiffness degradation matrix of the anisotropic element containing the stiffness attenuation coefficient.

[0082] Specifically, the surrounding rock of a real wellbore is not a homogeneous medium, but contains a large number of complex primary joints and fractures. After completing the spatial simulation mapping, the three-dimensional starting coordinates, ending coordinates, and extension dip angle of newly formed microcracks are accurately analyzed from the local damage state update vector. A digital twin discrete element mechanical mesh, pre-established using geological exploration data, is loaded in the background. This mesh is already assigned the friction angle and cohesion constant of the joint surfaces of the primary rock mass. Subsequently, the pore water pressure values ​​fed back from the hydrological monitoring well at the current elevation and the triaxial principal stress values ​​fed back from the geostress sensor are read. Driven by the above-mentioned real physical boundary conditions, the prediction engine runs an anisotropic damage mechanics algorithm to calculate the stiffness degradation matrix of each discrete element mesh over time. The algorithm calculates the future preset values ​​according to the set time step. The probability that these initial microcracks will continuously develop, interconnect, and eventually form a large-area sliding surface within a period, wherein the preset period ranges from 30 to 180 days. This time span matches the creep evolution cycle of conventional mine structures and engineering preventive maintenance plans. The output predicts the structural failure probability density matrix and evolutionary morphology boundary. Based on this boundary, the graphics engine overlays a semi-transparent dynamic mesh with a warning color on the current real well wall surface of the 3D twin model to visualize the future development morphology. If the predicted instability probability value of a node reaches the safety set threshold, wherein the safety set threshold ranges from 80% to 90%, it indicates that the micro-damage evolution of the engineering node has approached the critical state of macro-physical fracture, and then sends an audible and visual interruption alarm with coordinate information to the central control room.

[0083] This invention deeply integrates high-fidelity spatial geometric evolution characteristics with multi-physics boundary conditions such as pore water pressure and geostress at the underlying level, and relies on an anisotropic damage mechanics model and discrete element mesh for rigorous physical simulation reasoning. This deduction process can objectively, quantitatively, and completely conform to the laws of geotechnical engineering mechanics to calculate the future failure probability and evolution mode of the structure. This provides a highly rigorous decision-making basis based on physical science for the forward-looking safety early warning, life assessment, and precise maintenance decision-making of complex engineering facilities.

[0084] This embodiment significantly improves the robustness and spatial fidelity of the system through refined algorithm design for each key aspect of wellbore inspection; the introduction of multispectral fusion and adaptive threshold effectively filters out optical distortion caused by high-concentration water mist; relying on the visual-inertial tight coupling mechanism of twin prior topology, it overcomes long-term positioning drift in deep wells; and the use of multi-field thermo-mechanical coupling deformation compensation corrects hidden deep stress deformation. These targeted technical means achieve stepwise reduction of errors in the physical perception and kinematic dimensions, providing a data source that closely approximates the real physical state for the digital twin model.

[0085] Example 2:

[0086] This embodiment provides a complete end-to-end implementation plan for a wellbore inspection visual recognition method based on digital twins. It elaborates in detail the complete logical data flow process of multi-source heterogeneous data from low-level perception, spatial alignment, deformation compensation, computing power scheduling to disaster prediction. The specific steps are as follows:

[0087] As one embodiment of the present invention, refer to Figure 1 Flowchart of a visual recognition method for wellbore inspection based on digital twins, referencing Figure 2 The flowchart for simulation derivation complexity calculation and hierarchical reconstruction strategy is shown below. Figure 3 Flowchart of secondary compensation for multi-field thermo-coupling deformation.

[0088] The first stage involves multi-source state perception, numerical perspective, and strict alignment with spatiotemporal context:

[0089] First, the underlying data acquisition bus is activated to acquire multi-field coupled sensing data from the inspection device. The real-time physical sensing matrix is ​​a three-dimensional point cloud data matrix generated by a three-dimensional laser scanner mounted on the inspection device. This matrix is ​​a three-dimensional tensor with M rows, N columns, and 3 channels, where M and N are the number of sampling points of the scanner in the horizontal and vertical directions, respectively. For example, M equals 1024 and N equals 768. The three channels of the third dimension store the X, Y, and Z three-dimensional spatial coordinates of each sampling point in the world coordinate system, in meters. The generation process of this matrix is ​​as follows: the three-dimensional laser scanner... The wellbore inner wall is scanned using a 0-degree circumferential scan method. The distance value of the spatial point is calculated based on the time-of-flight ranging principle. Combined with the transient pose matrix of the scanner, the distance value in the local polar coordinate system is converted into Cartesian coordinates in the world coordinate system. The coordinates are arranged according to the scanning sequence to form the tensor matrix mentioned above. When a visible light camera is used for assisted acquisition, the generated perception matrix is ​​a two-dimensional grayscale image matrix with M rows, N columns, and 1 channel. The value range is 0 to 255, representing the surface reflectivity of the corresponding spatial point. This data structure and format limit provides a clear dimensional benchmark for subsequent state space matching and residual calculation.

[0090] To address water mist interference in deep wells, not only is a visible light band sensing matrix acquired, but also penetrating band intensity reflection matrices from infrared or lidar are simultaneously acquired. A defogging prior algorithm based on a physical optical scattering model is introduced to calculate the scattering distortion parameters of suspended water particles in the visible light band sensing matrix, and dynamic gain weights are assigned to the penetrating band intensity reflection matrix at the feature map level. Through adaptive weight fusion calculation, a low-level structural representation matrix that filters out optical scattering distortion is reconstructed at the front end, serving as the core real-time physical sensing matrix. Simultaneously, to address the spatial mapping misalignment problem caused by inconsistent sampling frequencies of different sensors, a dynamic circular buffer queue is established in memory. The high-sampling-frequency real-time physical sensing matrix and the low-sampling-frequency physical boundary deformation parameters (such as fiber optic strain data), temperature data, and ultra-wideband positioning data are pushed into the buffer queue with timestamps. Using the timestamp of the real-time physical sensing matrix as the time reference, a time-series interpolation algorithm is performed on the low-sampling-frequency physical data to reconstruct a high-dimensional physical state tensor that strictly corresponds to the exposure time of the real-time physical sensing matrix, eliminating asynchronous sampling spatiotemporal errors.

[0091] The second stage involves environmental adaptive state residual calculation and feature extraction.

[0092] After completing spatiotemporal alignment, the expected simulation baseline matrix of the internal structure of the well shaft structure is extracted synchronously based on physical coordinates. To further accurately remove background noise, the current environmental humidity and lighting intensity physical indicators are extracted, and a nonlinear mapping constraint relationship between the basic difference threshold and the above physical indicators is established to dynamically generate a pixel-by-pixel adaptive difference threshold matrix. The reconstructed real-time physical perception matrix and the expected simulation baseline matrix are subtracted pixel-by-pixel in the same coordinate system to generate an initial residual matrix. The initial residual matrix is ​​compared element-by-element with the adaptive difference threshold matrix to select spatially connected regions whose values ​​are greater than the corresponding elements of the adaptive difference threshold matrix, and the data in the connected regions are extracted as local structural perturbation feature sequences.

[0093] The specific steps for calculating the state space residual are as follows: For each three-dimensional spatial sampling point in the real-time physical perception matrix, the corresponding point with the closest spatial Euclidean distance is retrieved in the expected simulation reference matrix; when the spatial Euclidean distance between two points is less than a preset matching threshold (e.g., 0.05 meters), the two points are considered to correspond to the same physical location, and the X, Y, and Z coordinates of the two points are subtracted respectively to obtain the three-dimensional spatial residual vector; after traversing all sampling points, an initial residual matrix with the same dimension as the original matrix is ​​generated, and each element of the matrix stores the three-dimensional residual vector of the corresponding position; if a point cannot find a matching point with a distance less than the above threshold in the expected simulation reference matrix, its residual vector is marked as an outlier; in addition, when high concentration of water mist causes local missing laser point cloud data, the bilinear interpolation method is applied to the missing area, and the spatial interpolation is performed using the residual values ​​of the surrounding effective sampling points to ensure the spatial alignment and topological integrity of the initial residual matrix.

[0094] The third stage involves multi-source fusion coarse localization and twin-prior-guided secondary fine calibration:

[0095] The algorithm executes a progressive pose calculation logic. First, global coarse localization is performed: a Kalman filter state transition equation is constructed, and the continuous acceleration and angular velocity data output by the inertial measurement unit (IMU) are fused and filtered with the absolute distance data sent by the ultra-wideband positioning base station to output a time-continuous pose function containing translation vectors and rotation quaternions. Then, a secondary fine calibration based on twin prior topology is performed: rigid engineering structure models (such as metal tank connection parts) in the corresponding elevation section of the well shaft's 3D digital twin model are extracted as prior 3D spatial landmarks; corresponding planar feature points are extracted from the real-time physical perception matrix, and the deviation between the actual observed feature points and the theoretical model projection points is calculated to construct a reprojection error objective function. This reprojection error objective function and the pre-integration error of the IMU are used to construct a visual-inertial tightly coupled nonlinear optimization framework, and the factor graph optimization algorithm is called for iterative solution to correct the yaw angle and translation drift parameters in the time-continuous pose function for the second time, outputting a high-precision transient pose matrix with high-stability spatial reference.

[0096] The fourth stage involves the transfer of simulated boundary conditions and secondary compensation for thermo-mechanical coupling deformation.

[0097] After obtaining the high-precision pose, the transient pose matrix is ​​used as the coordinate transformation operator to construct a set of geometric ray equations pointing from the local observation coordinate system to the global standard cylindrical section of the digital twin. The three-dimensional coordinate set of the intersection points is solved as the boundary transfer constraint. In the physical quantity compensation stage, multi-field coupling compensation of mechanical and thermodynamics is performed. First, based on fiber optic strain gauge data and elastic constitutive relations, the deformation field of the basic three-dimensional structure, including normal displacement and tangential slip, is calculated. Second, the real-time surface temperature of the well wall is read using a surface temperature sensor, and a transient heat conduction inversion differential equation is constructed by combining ventilation wind speed and historical geothermal gradient data to estimate the transient thermal stress field in the deep layer of the well wall and convert it into a thermal expansion displacement vector. Third, the thermal expansion displacement vector and the deformation field of the basic three-dimensional structure are tensor superimposed to generate a thermo-mechanically coupled comprehensive deformation field. Finally, the comprehensive deformation field is used to perform spatial displacement superposition and dense correction on the three-dimensional coordinates of the intersection points in the boundary transfer constraint, and the planar discrete data is accurately solved into the local damage state update vector of the structural simulation model.

[0098] In the fifth stage, the nonlinear aggregation calculation of the simulation derivation complexity is as follows:

[0099] First, the spatial variation is calculated. The complement of the 3D intersection-union ratio (representing overlap) is added to the Hausdorff distance (representing shape deviation) normalized to the reference radius to obtain a comprehensive similarity index. Then, a nonlinear transformation is performed: the logarithm (base-1) of the data delay fluctuation represented by the round-trip variance of the most recent 100 data packets is extracted to smooth abrupt peaks; the natural exponential function of the natural constant e (e.g., with an exponential scaling factor of 2) is extracted from the spatial variation to nonlinearly amplify the propagation weight of micro-cracks; the derivatives of the temperature gradient and pressure are then... The physical environment mutation index, after normalization and summation of the rate of change, is transformed using a quadratic polynomial. Since the output ranges of the three functions mentioned above are different, the calculation results need to be divided by the maximum value of their respective theoretical intervals to perform max-min normalization and uniformly map them to the interval between 0 and 1. Finally, considering that spatial geometric changes have the most profound impact on computing power scheduling, weight coefficients of 0.3, 0.4, and 0.3 are assigned to the normalized data latency fluctuation, spatial change, and physical environment mutation index, respectively. The weighted summation outputs a scalar of model evolution complexity between 0 and 1.

[0100] The sixth stage involves the adaptive scheduling of hierarchical multi-scale simulation model reconstruction strategies.

[0101] The calculated simulation complexity triggers a preset dynamic threshold comparison mechanism in memory. When the complexity is less than the first threshold, the first-level silent strategy is executed, intercepting the 3D mesh redrawing instruction and only writing the local damage state update vector to the underlying time-series database. When the complexity is between the first and second thresholds, the second-level minimally invasive update strategy is executed, generating a local material mask in the digital twin space and only scheduling computing power to reconstruct the mesh vertex coordinates and normals inside the mask. When the complexity is greater than or equal to the second threshold, the third-level global preemption strategy is executed, suspending idle background processes and forcibly calling the global polygon mesh to perform overall structural deformation simulation rendering of the digital twin model.

[0102] The seventh stage involves forward extrapolation of the structural failure probability in the evolution of anisotropic damage:

[0103] After the model update is completed, the underlying physics and mechanics engine is launched to perform disaster simulation. The initial microcrack coordinates and orientation features in the local damage state update vector are extracted and mapped to a pre-constructed digital twin discrete element mechanical mesh containing the properties of the original rock mass joint surface. Under the driving force of the coupled boundary conditions of pore water pressure and principal stress in real-time monitoring, the anisotropic damage mechanics model is run to calculate the stiffness degradation matrix of the anisotropic element over time. Based on this mechanical degradation process, the predicted probability density cloud map of microcrack initiation and expansion to form macroscopic sliding surfaces in the future preset period is calculated and output. The graphics rendering engine visualizes the cloud map and triggers the disaster early warning window output command when the probability density peak exceeds the structural design limit state equation, completing the entire business closed loop from perception correction, computing power scheduling to physical forward defense.

[0104] This embodiment provides a closed-loop solution for multi-source data flow and simulation, effectively resolving the technical contradiction between high-fidelity mapping and real-time computing power congestion. By nonlinearly aggregating and calculating the complexity and triggering a hierarchical multi-scale reconstruction strategy, this solution breaks through the rendering bus bottleneck caused by high-frequency concurrency of massive heterogeneous data, ensuring stable operation with low latency. At the same time, by deeply integrating full-volume perception data into the anisotropic mechanics prediction engine, it achieves a leap from passive status quo display to proactive structural failure early warning, providing a complete intelligent system solution for deep-earth engineering safety.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A visual recognition method for wellbore inspection based on digital twins, characterized in that, include: Step 1: Obtain the real-time physical perception matrix of the inspection equipment, and simultaneously extract the expected simulation reference matrix of the wellbore three-dimensional digital twin model based on the position coordinates of the inspection equipment; perform state-space residual calculation on the real-time physical perception matrix and the expected simulation reference matrix, and extract the local disturbance feature sequence. Step 2: Obtain the transient pose matrix and radial deformation parameters of the well wall when acquiring the real-time physical sensing matrix; use the transient pose matrix to construct the geometric mapping relationship for the local disturbance feature sequence to solve the boundary transfer constraints, and combine the radial deformation parameters to construct the physical structure deformation compensation kernel function to perform in-situ deformation secondary compensation for the boundary transfer constraints, and solve the planar discrete data points into the local damage state update vector of the structural simulation model; Step 3: Calculate the spatial change based on the local damage state update vector and the abnormal evolution topology of the model, and nonlinearly aggregate the spatial change, data delay fluctuation, and physical environment mutation index to calculate the model evolution complexity; according to the preset dynamic threshold range of the model evolution complexity, execute a hierarchical model reconstruction strategy on the wellbore 3D digital twin model.

2. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, Step 1 involves extracting local disturbance feature sequences, including: establishing a nonlinear mapping constraint relationship between the basic difference threshold and physical indicators such as environmental humidity and lighting intensity, and dynamically generating an adaptive difference threshold matrix; subtracting data points in the same coordinate system from the real-time physical perception matrix to obtain an initial residual matrix; comparing each element in the initial residual matrix with the corresponding element in the environmental adaptive difference threshold matrix; and selecting spatially connected regions with values ​​greater than the corresponding elements in the adaptive difference threshold matrix as local disturbance feature sequences.

3. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, In step 2, acquiring the transient pose matrix during the acquisition of the real-time physical sensing matrix includes: synchronously receiving continuous triaxial acceleration and triaxial angular velocity data output by the inertial measurement unit inside the inspection equipment, and absolute distance data measured by the external positioning base station; constructing a Kalman filter state transition equation, filtering and solving the continuous triaxial acceleration, triaxial angular velocity data and absolute distance data, and outputting a time-continuous pose function; substituting the data timestamp of the real-time physical sensing matrix into the time-continuous pose function to calculate the transient pose matrix used for intra-frame motion distortion compensation.

4. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, In step 2, the geometric mapping relationship for the local disturbance feature sequence is constructed using the transient pose matrix to solve the boundary transfer constraint. This includes: retrieving the standard cylindrical section equation and design inner diameter parameters of the wellbore 3D digital twin model at the current depth coordinates; constructing a set of geometric ray equations starting from the sensor optical center and passing through the discrete data points in the plane; substituting the discrete data points in the plane in the local disturbance feature sequence as input variables into the set of geometric ray equations, solving for the set of 3D coordinates of the intersection of the geometric ray equations and the standard cylindrical section equation, and defining it as the boundary transfer constraint.

5. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, In step 2, in-situ deformation secondary compensation is performed on the boundary transfer constraints, and the discrete planar data points are solved into a local damage state update vector of the structural simulation model. This includes: receiving real-time strain data transmitted by fiber optic strain gauges on the wellbore wall; converting the real-time strain data into radial deformation parameters for the corresponding elevation section according to the constitutive relation of elasticity; generating a three-dimensional structural deformation field function containing a normal displacement scalar and a tangential slip vector using the radial deformation parameters as independent variables; using the three-dimensional structural deformation field function as the physical structural deformation compensation kernel function, superimposing the three-dimensional coordinates of the intersection points in the boundary transfer constraints, synchronously correcting the spatial errors of the three-dimensional coordinates of the intersection points in the wellbore normal direction and tangential direction, and encapsulating them into the local damage state update vector.

6. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, In step 3, the spatial variation, data latency fluctuation, and physical environment mutation index are nonlinearly aggregated to calculate the model evolution complexity. This includes: calculating the product of the three-dimensional intersection-union ratio and Hausdorff distance between the local damage state update vector and the reference damage vector at the same physical elevation in the historical period as the spatial variation; extracting the round-trip time delay variance of data packets between the edge computing node and the cloud server as the data latency fluctuation; reading the wellbore temperature gradient derivative, pore water pressure change rate, and microseismic energy release integral parameters, normalizing them, and summing them as the physical environment mutation index; applying a logarithmic function to the data latency fluctuation, a natural exponential function to the spatial variation, and a polynomial fitting operation to the physical environment mutation index; performing dimensionless normalization on the data latency fluctuation, spatial variation, and physical environment mutation index obtained after the above operations; assigning a first weight coefficient, a second weight coefficient, and a third weight coefficient to the three normalized values; summing the three values ​​after assigning weight coefficients and outputting the model evolution complexity.

7. The wellbore inspection visual recognition method based on digital twins according to claim 1, characterized in that, In step 3, a hierarchical model reconstruction strategy is executed on the well shaft 3D digital twin model according to the preset dynamic threshold range of the model evolution complexity. This includes: pre-setting a first threshold and a second threshold, where the value of the first threshold is less than the value of the second threshold; when the model evolution complexity is less than the first threshold, a first-level strategy is executed: intercepting the 3D mesh redrawing command and appending the local damage state update vector to the underlying temporal database to maintain the rendering state of the well shaft 3D digital twin model unchanged; when the model evolution complexity is not less than the first threshold and less than the second threshold, a second-level strategy is executed: generating a local material mask surrounding the local damage state update vector in the space of the well shaft 3D digital twin model, and only reconstructing the 3D mesh vertex coordinates and face normal vectors inside the local material mask; when the model evolution complexity is not less than the second threshold, a third-level strategy is executed: increasing the priority of the 3D rendering pipeline to the highest level of the system, and forcibly calling the global polygon mesh to perform overall structural deformation simulation rendering of the well shaft 3D digital twin model.

8. The wellbore inspection visual recognition method based on digital twin according to claim 1, characterized in that, In step 3, after executing the hierarchical model reconstruction strategy, the following structural failure inference steps are also included: extracting the initial microcrack coordinates and orientation features from the local damage state update vector; mapping the initial microcrack coordinates and orientation features to a pre-constructed digital twin discrete element mechanical mesh containing the properties of the original rock mass joint surface; and calculating the stiffness degradation matrix of the anisotropic elements in the digital twin discrete element mechanical mesh using an anisotropic damage mechanics model under the input pore water pressure and principal stress coupled boundary conditions. Based on the stiffness degradation matrix, the predicted structural failure probability density matrix and evolutionary morphology boundary are calculated and output. In the three-dimensional digital twin model of the wellbore, the evolutionary morphology boundary is visualized and spatiotemporally superimposed using a dynamic polygonal mesh. An interrupt alarm command is triggered when the peak value of the predicted structural failure probability density matrix is ​​greater than a preset failure probability safety threshold.

Images