Intelligent identification method and system for surrounding rock structure surface based on while-drilling parameters and peeping calibration

Abstract

A surrounding rock structure plane intelligent identification method and system based on while-drilling parameters and peeping calibration, the method comprising: collecting while-drilling parameters and peeping data; establishing while-drilling-peeping depth corresponding point groups; constructing a segmented depth drift correction function to unify the depth coordinate system; determining the structure plane type, visual attribute and calculating the mechanical response characteristics; constructing a three-segment sample unit to extract differential response characteristics; converting hard boundary labels into probability distribution soft boundary labels; constructing a multi-task identification model; jointly optimizing model parameters; identifying on site to output the structure plane type, start and end depth and identification confidence. The system comprises: a while-drilling parameter acquisition module, a peeping data acquisition module, a while-drilling-peeping depth calibration module, a structure plane information extraction module, a three-segment sample construction module, a soft boundary label construction module, a peeping constraint and feature consistency module, a model training module, an on-site identification module, and an on-site feedback and model correction module to form an intelligent identification system. The identification precision is high.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering surrounding rock detection, drilling measurement data processing, borehole viewing image analysis and intelligent recognition technology, specifically involving an intelligent recognition method and system for surrounding rock structural surfaces based on drilling parameters and viewing calibration. Background Technology

[0002] In underground engineering projects such as coal mine roadways and tunnels, natural structural surfaces such as fissures, delamination, and rock strata interfaces are commonly found within the surrounding rock. The spatial distribution, scale, and extent of these structural surfaces directly determine the integrity and overall bearing capacity of the surrounding rock, serving as the core basis for evaluating roadway surrounding rock stability, predicting disaster risks, and optimizing support parameters. Therefore, achieving rapid and accurate identification of the type, location, and extent of surrounding rock structural surfaces is of significant engineering importance for ensuring the safe and efficient construction of underground projects.

[0003] Currently, the main methods for identifying surrounding rock structural surfaces in the field are traditional measurement methods such as manual logging and borehole inspection. However, these methods generally suffer from drawbacks such as low operational efficiency, poor real-time performance, strong human subjectivity, and limited identification accuracy. Among these methods, borehole inspection is the most widely used direct detection method in the field, which can directly obtain the morphology of the surrounding rock structure, fracture development, and rock strata boundaries of the borehole wall. However, the imaging results of this method are easily affected by factors such as mud obstruction in the borehole, image quality, field environment, and human interpretation experience. It is difficult to achieve accurate, real-time, and continuous quantitative identification of surrounding rock structural surfaces, and it cannot meet the construction requirements of intelligent and dynamic surrounding rock detection.

[0004] Detection while drilling (DWD) technology can continuously collect dynamic drilling response parameters such as drill pipe rotation speed, drilling speed, thrust, and slewing torque during borehole construction, enabling real-time characterization of the dynamic coupling process between the drill bit and the surrounding rock medium. When the drill bit traverses structural surfaces such as fractures, delaminations, fracture zones, and rock interfaces, drilling mechanical parameters will exhibit significant abrupt changes or abnormal fluctuations, which can be used to infer and determine the development characteristics of the surrounding rock structural surfaces within the borehole. This technology has advantages such as continuous data acquisition, real-time response, strong adaptability to operating conditions, and no need for secondary entry operations, making it suitable for rapid dynamic identification of downhole surrounding rock structural surfaces.

[0005] Current drilling-while-drilling (WSD) identification technologies still have certain technical shortcomings. Relying solely on WWD parameters for structural face identification is highly susceptible to interference from drilling condition fluctuations, drill bit wear, equipment vibration, sensor sampling delays, and inhomogeneous coal and rock media. This can easily lead to problems such as structural face boundary positioning errors, misjudgment of structural face types, and poor consistency of data from multiple boreholes, making it difficult to guarantee identification accuracy and robustness. Meanwhile, borehole inspection data can provide intuitive visual truth-value information on borehole wall structural face morphology, fracture edge characteristics, rock strata interfaces, and fracture zone ranges, accurately pinpointing the actual location and morphological range of structural faces. However, borehole inspection is an offline detection method after borehole formation and cannot meet the real-time dynamic identification requirements of the drilling process. In current engineering applications, dynamic WWD parameter identification and accurate borehole inspection identification are two independent operational processes and technical systems. The industry lacks a fusion calibration scheme that uses borehole inspection results as truth-value prior information to calibrate WWD data depth deviation, correct model training sample labels, and adjust structural face identification thresholds.

[0006] Furthermore, surrounding rock structural surfaces are not ideal point-like structures; most exhibit banded development characteristics with a certain thickness and stress-medium transition zones. In summary, existing technologies cannot achieve deep fusion, joint calibration, and collaborative identification of dynamic data while drilling and borehole observation data, making it difficult to balance real-time performance and accuracy. Therefore, there is an urgent need for an intelligent identification method and system for surrounding rock structural surfaces that integrates drilling response data and borehole observation data, possesses a soft-boundary supervised training mechanism, and can achieve accurate real-time output in the field, in order to address the technical limitations of existing technologies. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides an intelligent identification method and system for surrounding rock structural surfaces based on drilling parameters and borehole inspection calibration. This method can effectively improve the depth positioning accuracy, boundary identification stability, and real-time performance of surrounding rock structural surface identification. It can solve the problems of inconsistency between the depth coordinates of drilling parameters and borehole inspection data, insufficient structural surface boundary positioning accuracy, lack of visual prior constraints in the drilling identification model, and insufficient reliability of real-time on-site identification in the prior art. The system can deeply integrate the borehole inspection visual prior and the drilling mechanical response under a unified depth coordinate system, realizing fully automated processing of surrounding rock structural surfaces from data acquisition to intelligent identification and continuous optimization.

[0008] To achieve the above objectives, this invention provides an intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration, comprising the following steps: Step 1: Multi-source data acquisition; Drilling operations are carried out in the surrounding rock to be identified, and drilling parameters during the drilling process and inspection data after the drilling is completed are collected simultaneously; Step 2: Establish a set of corresponding points for drilling-visiting depths; using the borehole opening as a unified zero point of depth, establish a drilling parameter depth sequence and a borehole visiting depth sequence respectively, extract feature points from the two sequences and perform matching in the depth direction; Step 3: Segmented depth drift correction; construct a segmented depth drift correction function based on the depth offset of the corresponding point group, and correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain calibrated data under a unified borehole depth coordinate system; Step 4: Structural surface feature extraction; Determine the type and visual attributes of the structural surfaces based on the calibrated data, and calculate the mechanical response characteristics of the structural surfaces; Step 5: Construct a three-segment combined sample unit for drilling and inspection; with the structural surface depth interval as the center, construct a three-segment sample unit along the drilling depth direction, which includes a pre-stabilization segment, a structural surface influence segment, and a post-recovery segment, and extract the differential response features between adjacent segments; Step 6: Construct soft boundary labels; convert the hard boundary annotations of the structural surfaces into soft boundary labels in the form of probability distributions; Step 7: Construct a joint drilling-peeping structure recognition model; Construct a multi-task deep learning model that includes a drilling feature extraction unit, a peeping constraint unit, a feature consistency constraint unit, and multiple output units to achieve end-to-end recognition from drilling parameters to structure type and start and end depths; Step 8: Model training; Train the drilling-peeping joint structural surface recognition model using three-segment drilling-peeping joint sample units and soft boundary labels, and jointly optimize the model parameters through a multi-task loss function; Step 9: On-site deployment and real-time identification; Deploy the trained model on-site, form a flowing drilling window according to the depth step size, trigger candidate identification of structural surfaces based on mechanical response characteristics, and output the structural surface type, start and end depths, development length and identification confidence.

[0009] As a preferred option, the multi-source data acquisition process in step 1 is as follows: S11: Drilling operations are carried out in the surrounding rock to be identified in the underground engineering project; during the drilling operation, drilling parameters are collected in real time through a sensor system installed on the drilling rig; the drilling parameters include at least drilling depth, drilling time, drilling speed, rotational speed, thrust, and torque; S12: After drilling is completed, drilling observation data corresponding to the drill hole is collected using a drilling observation device; the drilling observation data includes at least one or more of drilling video, drilling image or drilling unfolded diagram; For situations where it is impossible to inspect the same borehole under certain engineering conditions, borehole inspection data from adjacent calibration boreholes are used. When the borehole inspection data comes from adjacent calibration boreholes, it is ensured that the distance between the adjacent calibration boreholes and the borehole to be identified does not exceed the preset borehole distance threshold, and the lithological consistency index, stratigraphic correspondence, or structural surface development similarity index of the borehole sections where the two are located meet the preset consistency conditions to ensure the transferability of calibration data.

[0010] As a preferred option, the process of establishing the corresponding point set for the drilling-visualization depth in step 2 is as follows: S21: Using the borehole opening as a unified depth zero point, the drilling parameters are arranged according to the drilling depth to form a drilling parameter depth sequence, and the inspection data are arranged according to the borehole depth to form a borehole inspection depth sequence. S22: Extract structural visual markers from the borehole inspection depth sequence; the structural visual markers are extracted by image segmentation, manual verification or color and texture abrupt change detection, including fracture edge points, rock layer interface points, bedding abrupt change points, delamination boundary points, fracture zone start and end points, weak interlayer interface points or borehole wall color band abrupt change points. Candidate abrupt change points of mechanical response are extracted from the depth sequence of drilling parameters. These candidate abrupt change points can be extracted by methods such as multi-parameter anomaly detection, including points of sudden drop in drilling speed, sudden increase in torque, sudden change in thrust, peak point of mechanical specific energy, sudden change point of energy per unit footage, change point of drilling stiffness, or joint anomaly point of multiple parameters. S23: Match the visual markers of the structural surface and the candidate mutation points of the mechanical response within the preset depth search radius to establish a set of corresponding points for the drilling-viewing depth; when the same visual marker of the structural surface corresponds to multiple candidate mutation points of the mechanical response, select the candidate mutation point of the mechanical response with the largest mechanical response anomaly intensity and a depth offset less than the preset offset threshold as the corresponding point.

[0011] As a preferred embodiment, the segmented depth drift correction process in step 3 is as follows: S31: Construct a piecewise depth drift correction function based on the depth offset of each corresponding point in the drilling-viewing depth corresponding point group. ,in To correct the depth before, For the corrected depth, For the first Each depth segment; For the first Depth correction within each depth segment; at the same time It is a strictly monotonically increasing function. Correction amount within each segment Determined by interpolation of the depth offset of the corresponding point within the corresponding depth interval; S32: Use the segmented depth drift correction function to correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain the calibrated drilling parameter sequence and calibrated borehole inspection data under a unified borehole depth coordinate system.

[0012] As a preferred option, the structural surface feature extraction process in step 4 is as follows: The structural surface type, starting depth, ending depth, and visual attributes of the surrounding rock structural surface are determined based on the calibrated borehole inspection data. The structural surface type includes one or more of the following: transverse fractures, longitudinal fractures, oblique fractures, bedding planes, rock layer interfaces, delamination, fracture zones, weak interlayer interfaces, and intact surrounding rock transition interfaces. The visual attributes include one or more of the following: visible width of the structural surface, continuity, filling state, opening state, degree of color change, and degree of texture change. Simultaneously, the mechanical response characteristics of the structural surface are calculated based on the calibrated drilling parameter sequence; the mechanical response characteristics of the structural surface include at least several of the following: energy per unit footage, mechanical specific energy, torque-to-thrust ratio, drilling rate residual, energy mutation coefficient, and drilling stiffness change rate.

[0013] As a preferred embodiment, the process of constructing the three-segment drilling-viewing combined sample unit in step 5 is as follows: S51: Taking the structural surface depth range defined by the structural surface initiation depth and the structural surface termination depth as the center, take a drilling interval of a preset length before the structural surface initiation depth as the pre-stabilization section; take the interval between the structural surface initiation depth and termination depth as the structural surface influence section; take a drilling interval of a preset length after the structural surface termination depth as the post-recovery section. S52: Extract the first differential response features between the pre-stabilization segment and the structural surface-affected segment, and the second differential response features between the structural surface-affected segment and the post-recovery segment; the first differential response features and the second differential response features include at least one or more of the following: mean difference, gradient difference, mechanical specific energy difference, and drilling stiffness change rate difference; S53: Based on the pre-stabilization section, the structural surface influence section, the post-recovery section, the first differential response characteristics, and the second differential response characteristics, a three-segment drilling-viewing joint sample unit is constructed that is sequentially distributed along the drilling depth direction.

[0014] As a preferred embodiment, the process of constructing soft boundary labels in step 6 is as follows: To address the fuzzy transition characteristics of the structure boundary during drilling caused by drill bit size effect, rock mass gradual change characteristics, and drilling disturbance transmission, soft boundary labels are used instead of hard boundary labels; the starting depth of the structure is converted into the probability distribution of the starting boundary, the ending depth of the structure is converted into the probability distribution of the ending boundary, and the influence segment of the structure is converted into a probability mask of the existence of the structure. The starting point boundary probability distribution and the ending point boundary probability distribution are constructed with Gaussian, triangular, or trapezoidal probability distribution soft tags centered on the starting depth and ending depth of the structural surface, respectively. The depth extension radius of the probability distribution soft tag is adaptively determined based on the drill bit diameter, current drilling speed, thrust fluctuation amplitude, torque fluctuation amplitude, or the visually visible width of the structural surface. The probability mask for the existence of the structural surface satisfies the following distribution characteristics: the existence probability in the central region of the structural surface is not less than 0.8, the existence probability in the transition region of the structural surface boundary varies continuously between 0.2 and 0.8, and the existence probability in the stable surrounding rock region outside the structural surface is not higher than 0.2. Through the aforementioned soft boundary labels, the three-segment drilling-peeping joint sample unit simultaneously possesses a structural surface type label, a starting boundary probability label, an ending boundary probability label, and a structural surface existence probability label.

[0015] As a preferred option, the process of constructing the drilling-viewing joint structural surface identification model in step 7 is as follows: A joint drilling-visual structural surface recognition model is constructed based on the drilling feature extraction unit, the peep constraint unit, the drilling-visual feature consistency constraint unit, the structural surface type output unit, and the structural surface start and end depth output unit. The feature extraction unit during drilling includes a one-dimensional multi-scale convolutional layer, a dilated convolutional layer, and a residual connection layer, which are used to extract multi-scale structural surface mechanical response features from the drilling parameter sequence. The observation constraint unit is used to encode the structural surface type, visible width, continuity, filling state or opening state obtained by borehole observation into a visual prior vector. The drilling-peeping feature consistency constraint unit is configured to: within the structural surface depth range, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features output by the peeping constraint unit is higher than a preset positive sample threshold; and in the non-structural surface segment, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features is lower than a preset negative sample threshold; the similarity measure uses cosine similarity, defined as... ,in The feature vectors during drilling, This is the visual prior vector.

[0016] As an alternative, step 10 is also included: on-site feedback and model adaptive correction; after completing the borehole inspection verification, the deviation between the on-site identification results and the verification results is calculated, and the depth drift correction function, abnormal trigger threshold and model parameters are corrected according to the type and degree of deviation, so as to achieve continuous optimization of the identification system.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention establishes a set of corresponding points for drilling-viewing depths by matching the depth direction of visual markers on structural surfaces with candidate abrupt change points in mechanical response, and uses a piecewise depth drift correction function to uniformly calibrate the drilling parameter depth sequence and the borehole viewing depth sequence. This correction function satisfies monotonically increasing and piecewise continuous constraints, and adaptively selects the interpolation method based on the density of corresponding points. It can effectively reduce depth deviations caused by factors such as drill pipe deformation, sensor sampling delay, viewing propulsion error, and fluctuations in drilling conditions, improve the depth correspondence accuracy between drilling parameters and borehole viewing data, and provide a unified coordinate benchmark for subsequent joint identification.

[0018] 2. This invention constructs a three-segment combined drilling-viewing sample unit, consisting of an initial stable segment, a segment influenced by the structural surface, and a subsequent recovery segment, distributed sequentially along the borehole depth direction. It also extracts differential response features between adjacent segments, enabling a complete characterization of the dynamic mechanical response changes during the drill bit's entry, traversal, and exit from the structural surface. Simultaneously, it replaces traditional hard boundary labels with probability distribution soft boundary labels centered on the structural surface's start and end depths. By jointly describing the spatial distribution of the structural surface through the probability distributions of the starting and ending boundaries and the structural surface existence probability mask, it better adapts to the fuzzy transition characteristics of geological structural surface boundaries, significantly improving the positioning accuracy of the structural surface's start and end depths.

[0019] 3. This invention encodes the borehole peeping visual prior into a visual prior vector through a peeping constraint unit, and applies positive and negative sample similarity constraints to the structured surface segment and the non-structured surface segment respectively through a drilling-peeping feature consistency constraint unit. Visual semantic guidance is integrated into the training of the drilling recognition model, which enhances the model's ability to distinguish different structured surface types and development ranges.

[0020] 4. This invention constructs a multi-task loss function that includes structural surface type classification loss, start and end point boundary localization loss, structural surface existence probability loss, drilling-viewing feature consistency loss, and physical response constraint loss. Through multi-objective joint optimization, the model learns structural surface type discrimination, precise boundary localization, and physical response consistency simultaneously. Furthermore, the Mahalanobis distance constraint ensures the separability of mechanical response features between the affected and stable segments of the structural surface, thereby improving the stability and recognition accuracy of model training.

[0021] 5. During on-site drilling, this invention calculates mechanical response characteristics such as energy mutation coefficient and drilling rate residual in real time based on the flow-while-drilling window. The abnormal triggering threshold determined adaptively by the stable drilling section drives the candidate identification of structural surfaces. The start and end depths of the structural surfaces are determined by continuous window-level fusion and weighted peak value calculation. The output includes the identification confidence level including type probability, boundary probability peak value, window consistency coefficient and physical response anomaly intensity, providing a real-time and reliable basis for surrounding rock logging, support parameter design and construction safety decision-making.

[0022] 6. After completing the borehole inspection verification, this invention triggers the segmented depth drift correction function, the abnormal trigger threshold correction, and the model parameter correction based on the initial depth deviation, the final depth deviation, the mechanical response characteristic threshold deviation, and the structural surface type deviation, respectively. It also sets a historical model retention strategy and update stability constraints. When the performance of the corrected model degrades, it automatically reverts to the parameters before the correction, thus realizing the continuous optimization and safety evolution of the identification system in field applications.

[0023] This method deeply integrates borehole inspection visual priors with drilling mechanical response in a unified depth coordinate system, realizing intelligent processing of the entire process of surrounding rock structure surface identification from data acquisition, model training, real-time on-site identification to adaptive continuous optimization. It can effectively improve the depth positioning accuracy, boundary identification stability, and real-time performance of surrounding rock structure surface identification. It can solve the problems of inconsistency between drilling parameters and borehole inspection data depth coordinates, insufficient structural surface boundary positioning accuracy, lack of visual prior constraints in the drilling identification model, and insufficient reliability of real-time on-site identification in existing technologies. It is suitable for surrounding rock detection in mine roadways, tunnels, and underground engineering.

[0024] This invention also provides an intelligent identification system for surrounding rock structural surfaces based on drilling parameters and inspection calibration, used to implement an intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration, comprising: The drilling parameter acquisition module is used to acquire drilling parameters in real time during the drilling process. The borehole observation data acquisition module is used to collect borehole observation data corresponding to the borehole after drilling is completed; The drilling-visual depth calibration module is used to extract structural surface visual markers from borehole visual data, extract mechanical response candidate abrupt change points from drilling parameters, match and establish corresponding point groups in the depth direction, and construct a segmented depth drift correction function to uniformly calibrate the two depth sequences. The structural surface information extraction module is used to determine the structural surface type, starting depth, ending depth and visual attributes of the surrounding rock structural surface based on the calibrated borehole inspection data, and to calculate the mechanical response characteristics of the structural surface based on the calibrated drilling parameter sequence. The three-segment sample construction module is used to construct a three-segment drilling-viewing joint sample unit along the borehole depth direction, with the structural surface depth range defined by the structural surface start depth and structural surface end depth as the center. The soft boundary label construction module is used to convert the starting depth of the structural surface into the probability distribution of the starting boundary, the ending depth of the structural surface into the probability distribution of the ending boundary, and the influence segment of the structural surface into the probability mask of the existence of the structural surface, so that the three-segment drilling-peeping joint sample unit simultaneously has structural surface type label, starting boundary probability label, ending boundary probability label and structural surface existence probability label. The peeping constraint and feature consistency module, built into the drilling-peeping joint structure surface recognition model, is used to encode the structure surface type, visible width, continuity, filling state, or open state obtained from borehole peeping into visual prior vectors, and to apply drilling-peeping feature consistency constraints during model training. The model training module is used to receive the three-segment drilling-peeping joint sample units output by the three-segment sample construction module and the soft boundary labels output by the soft boundary label construction module, and to train the drilling-peeping joint structural surface recognition model. The on-site identification module is used to deploy the drilling-viewing combined structural surface identification model on the on-site operating host. During the on-site drilling process, a flow-type drilling window is formed according to the preset depth step size, and the energy mutation coefficient, drilling speed residual, drilling stiffness change rate, mechanical specific energy peak value or torque-thrust ratio change within the window are calculated in real time. The on-site feedback and model correction module is used to calculate the deviation between the on-site identification results and the borehole inspection verification results after the borehole inspection verification is completed. Based on the deviation type and degree, it triggers the segmented depth drift correction function, the abnormal trigger threshold correction, and the model parameter correction respectively.

[0025] In this invention, the setup of the drilling parameter acquisition module facilitates real-time synchronous acquisition of core drilling parameters such as drilling depth, drilling speed, rotational speed, thrust, and torque. It can also expand to acquire auxiliary parameters such as hydraulic pressure, drilling vibration, and flushing fluid pressure, providing a multi-dimensional data foundation for subsequent extraction of candidate abrupt change points in mechanical response and calculation of structural surface mechanical response features. The setup of the peep data acquisition module facilitates the acquisition of borehole videos, images, or unfolded diagrams, and utilizes adjacent calibrated borehole data that meets consistency conditions, providing an intuitive source of visual information for extracting visual landmarks on the structural surface and constructing visual prior constraints. Together with the drilling parameter acquisition module, it forms a complementary dual-channel data acquisition system. The setup of the drilling-peep depth calibration module facilitates the extraction of visual landmarks and candidate abrupt change points in the structural surface, and the matching of corresponding point groups in the depth direction. This allows for the construction of a piecewise depth drift correction function that satisfies monotonically increasing and piecewise continuous constraints, effectively eliminating depth deviations in heterogeneous data caused by factors such as drill pipe deformation, sensor sampling delay, and peep propulsion errors, providing a unified depth coordinate benchmark for subsequent joint identification. The structural surface information extraction module facilitates the collaborative extraction of visual and mechanical features within the same depth coordinate system, providing complete multi-source feature input for the three-segment sample construction. The three-segment sample construction module comprehensively depicts the dynamic mechanical response changes before and after the drill bit traverses the structural surface, providing structured training samples for the model. The soft boundary label construction module ensures that labels better reflect the continuous transition characteristics of geological structural surface boundaries formed by drill bit size effects and gradual rock mass changes, improving the model's ability to identify blurred boundary regions. The peep constraint and feature consistency module integrates visual semantic guidance into the drilling feature learning process, enhancing the model's ability to distinguish different structural surface types and its robustness to noise interference. The model training module enables the model to simultaneously learn type discrimination, boundary localization, and physical consistency, and uses Mahalanobis distance constraints to ensure the separability of mechanical response features between the affected and stable segments of the structural surface, improving training stability and recognition accuracy. Based on the field identification module, real-time identification can be performed during drilling, outputting identification confidence scores including type probability, peak boundary probability, window consistency coefficient, and physical response anomaly intensity. This provides real-time, reliable, and intelligent decision-making support for surrounding rock logging and support design. Based on the field feedback and model correction module, depth correction function correction, anomaly trigger threshold correction, and model parameter correction can be triggered based on depth deviation, threshold deviation, and type deviation, respectively. Furthermore, historical model retention and updated stability constraints ensure that model performance does not degrade, achieving continuous adaptive optimization of the identification system in field applications.

[0026] This system can deeply integrate borehole inspection vision priors with drilling mechanical response in a unified depth coordinate system, realizing fully automated processing of the surrounding rock structure from data acquisition to intelligent identification and continuous optimization. Attached Figure Description

[0027] Figure 1 This is a flowchart of the intelligent recognition method in this invention.

[0028] Figure 2 The flowchart for constructing a three-segment drilling-viewing joint sample unit and soft boundary label is provided for this invention. Figure 3 This is a structural diagram of the drilling-viewing combined structural surface identification model in this invention.

[0029] Figure 4 This is a block diagram illustrating the principle of the intelligent recognition system in this invention. Detailed Implementation

[0030] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0031] Example 1:

[0032] like Figures 1 to 3 As shown, this embodiment provides a method for intelligent identification of surrounding rock structural surfaces based on drilling parameters and inspection calibration, characterized by the following steps: Step 1: Multi-source data acquisition; Drilling operations are carried out in the surrounding rock to be identified, and drilling parameters during the drilling process and inspection data after the drilling is completed are collected simultaneously to provide a multi-source data foundation for subsequent joint identification; As a preferred option, the multi-source data acquisition process is as follows: S11: As Figure 1 As shown, drilling operations are carried out in the surrounding rock to be identified in underground engineering. During the drilling operation, drilling parameters are collected in real time by a sensor system installed on the drilling rig. The drilling parameters include at least drilling depth, drilling time, drilling speed, rotation speed, thrust and torque, and may also include one or more of the following as extended parameters: hydraulic pressure, drill rod displacement, borehole inclination angle, drilling vibration, drilling current, drilling rig power, flushing fluid pressure or flushing fluid flow rate. As a preferred option, the drilling rig can be a bolt drilling rig, a water exploration drilling rig, a geological drilling rig, or other drilling equipment capable of collecting drilling parameters; S12: After drilling is completed, drilling observation data corresponding to the borehole is collected using drilling observation equipment (including drilling observation instrument, drilling television, borehole wall imaging equipment or borehole wall unfolding diagram generation equipment); the drilling observation data includes at least one or more of drilling video, drilling image or borehole unfolding diagram, and may also include one or more of the following: borehole wall panoramic unfolding diagram, structural surface manual verification result, structural surface image segmentation result or borehole wall color texture features. In cases where it is not possible to inspect the same borehole under certain engineering conditions, borehole inspection data from adjacent calibration boreholes can be used. When the borehole inspection data comes from adjacent calibration boreholes, ensure that the distance between the adjacent calibration boreholes and the borehole to be identified does not exceed the preset borehole distance threshold, and that the lithological consistency index, stratigraphic correspondence, or structural surface development similarity index of the borehole sections where the two are located meet the preset consistency conditions to ensure the transferability of calibration data. This technical solution utilizes the drilling rig's sensor system to synchronously collect core drilling parameters such as drilling depth, drilling speed, rotational speed, thrust, and torque, as well as extended parameters such as hydraulic pressure and drilling vibration, in real time during drilling operations. After drilling is completed, borehole inspection equipment, borehole television, or borehole wall imaging devices are used to acquire borehole video, images, or unfolded diagrams, thus constructing a dual-channel data acquisition system that complements the mechanical response during drilling and the visual information from the borehole wall. Furthermore, for engineering conditions where it is impossible to inspect the same borehole, data from adjacent calibration boreholes that meet preset spacing thresholds, lithological consistency, and stratigraphic correspondence are permitted, ensuring the transferability of calibration data under different working conditions and the engineering adaptability of the acquisition scheme.

[0033] Step 2: Establish a set of corresponding points for drilling-visiting depths; using the borehole opening as a unified zero point of depth, establish a drilling parameter depth sequence and a borehole visiting depth sequence respectively, extract feature points from the two sequences and perform matching in the depth direction; As a preferred option, the process of establishing the corresponding point set for drilling-visual depth is as follows: S21: Using the borehole opening as a unified depth zero point, the drilling parameters are arranged according to the drilling depth to form a drilling parameter depth sequence, and the inspection data are arranged according to the borehole depth to form a borehole inspection depth sequence. S22: Extract visual markers of structural surfaces from the borehole inspection depth sequence; the visual markers of structural surfaces can be extracted by image segmentation, manual verification or color and texture abrupt change detection, including fracture edge points, rock layer interface points, bedding abrupt change points, delamination boundary points, fracture zone start and end points, weak interlayer interface points or borehole wall color band abrupt change points, which are used to visually mark the position of the structural surfaces in the depth direction. Candidate abrupt change points of mechanical response are extracted from the depth sequence of drilling parameters. These candidate abrupt change points can be extracted by multi-parameter anomaly detection and other methods, including points of sudden drop in drilling speed, sudden increase in torque, sudden change in thrust, peak point of mechanical specific energy, sudden change point of energy per unit footage, point of change in drilling stiffness, or joint anomaly points of multiple parameters, which are used to identify possible locations of structural surfaces from the perspective of mechanical response. S23: Match the visual markers of the structural surface and the candidate abrupt change points of the mechanical response within the preset depth search radius to establish a set of corresponding points for the drilling-viewing depth; when the same visual marker of the structural surface corresponds to multiple candidate abrupt change points of the mechanical response, the candidate abrupt change point of the mechanical response with the largest mechanical response anomaly intensity and a depth offset less than the preset offset threshold is selected as the corresponding point to ensure the accuracy and uniqueness of the matching. In this technical solution, by using the borehole opening as a unified depth zero point, drilling parameters are arranged according to drilling depth, and inspection data is arranged according to borehole depth to form depth sequences, thus achieving ordered organization of heterogeneous data in the depth dimension. Based on this, visual markers of structural surfaces are extracted from the inspection sequence through image segmentation, manual verification, or color and texture abrupt change detection. Candidate abrupt change points of mechanical response are extracted from the drilling sequence through multi-parameter anomaly detection. These are then matched within a preset depth search radius, prioritizing the candidate abrupt change points with the largest anomaly intensity and offset less than a threshold as corresponding points, establishing a high-confidence depth correspondence between visual features and mechanical responses. This mechanism effectively solves the correspondence problem between drilling data and inspection data in the depth dimension, providing a reliable set of control points for subsequent segmented depth drift correction, and ensuring the accuracy of joint analysis of heterogeneous data from the source.

[0034] Step 3: Segmented depth drift correction; construct a segmented depth drift correction function based on the depth offset of the corresponding point group, and correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain calibrated data under a unified borehole depth coordinate system; As a preferred option, the process of segmented depth drift correction is as follows: S31: Construct a piecewise depth drift correction function based on the depth offset of each corresponding point in the drilling-viewing depth corresponding point group. ,in Depth before correction, in meters. This is the corrected depth, in meters (m). For the first Each depth segment is measured in meters. For the first The depth correction amount within each depth segment, in meters; Meanwhile, the segmented depth drift correction function satisfies the following constraint: the corrected depth monotonically increases with the original depth, i.e. For a strictly monotonically increasing function, adjacent segments are continuous at the connection points, i.e. Correction amount within each segment The depth correction amount is determined by interpolation of the depth offset of the corresponding point within the corresponding depth interval; The piecewise depth drift correction function is obtained by solving for the depth offset of corresponding points within the depth segment using piecewise linear interpolation, spline interpolation, or dynamic time warping constraints. The solution method is adaptively selected based on the density of the corresponding points: when the depth distance between adjacent corresponding points is less than the preset density threshold, piecewise linear interpolation is used; when the depth distance between adjacent corresponding points is greater than or equal to the preset density threshold, spline interpolation or dynamic time warping constraint is used. S32: Use the segmented depth drift correction function to correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain the calibrated drilling parameter sequence and calibrated borehole inspection data under a unified borehole depth coordinate system. In this technical solution, a piecewise depth drift correction function is constructed based on the depth offset of the corresponding point group during drilling and borehole inspection. Strict monotonically increasing constraints and piecewise continuity constraints are applied to ensure the physical consistency and spatial continuity of the corrected depth sequence. Simultaneously, piecewise linear interpolation, spline interpolation, or dynamic time warping constraints are adaptively selected to solve for the depth correction based on the density of the corresponding points, balancing correction accuracy and computational efficiency. Using this correction function to uniformly calibrate the drilling parameter depth sequence and the borehole inspection depth sequence effectively eliminates depth deviations in heterogeneous data caused by factors such as drill pipe deformation, sensor sampling delay, and inspection propulsion errors, providing a precise and unified depth coordinate benchmark for subsequent structural surface feature extraction and joint identification.

[0035] Step 4: Structural surface feature extraction; Based on the calibrated data, determine the type and visual attributes of the structural surfaces, and calculate the mechanical response features of the structural surfaces to provide a feature basis for subsequent sample construction and model training; As a preferred method, the process of extracting structural surface features is as follows: The structural surface type, starting depth, ending depth, and visual attributes of the surrounding rock structural surface are determined based on the calibrated borehole inspection data. The structural surface type includes one or more of the following: transverse fractures, longitudinal fractures, oblique fractures, bedding planes, rock layer interfaces, delamination, fracture zones, weak interlayer interfaces, and intact surrounding rock transition interfaces. The visual attributes include one or more of the following: visible width of the structural surface, continuity, filling state, opening state, degree of color change, and degree of texture change. Simultaneously, the mechanical response characteristics of the structural surface are calculated based on the calibrated drilling parameter sequence; these characteristics include at least several of the following: energy per unit footage, mechanical specific energy, torque-to-thrust ratio, drilling rate residual, energy mutation coefficient, and drilling stiffness variation rate; among which, energy per unit footage... ,in For thrust, For drilling speed, For torque, Rotational speed; drilling speed residual The difference between the measured drilling rate and the baseline drilling rate predicted from the current thrust, rotational speed, and torque; energy mutation coefficient. The normalized difference between the energy per unit footage within the current depth window and the energy per unit footage in adjacent stable drilling sections; the rate of change of drilling stiffness. It is the rate of change of the ratio of thrust change to drilling displacement along the borehole depth direction; In this technical solution, under a unified depth coordinate system, visual attribute information covering various structural surface types such as transverse fractures, bedding planes, fracture zones, and weak interlayer interfaces is extracted from calibrated borehole inspection data. At the same time, mechanical response features such as energy per unit footage, mechanical specific energy, torque-to-thrust ratio, drilling rate residual, energy mutation coefficient, and drilling stiffness change rate are calculated from the calibrated drilling parameter sequence. This achieves the collaborative extraction and correlation representation of visual features and mechanical response features of the same structural surface under multi-source data, providing a rich and highly corresponding feature foundation for subsequent three-segment sample construction and soft boundary label generation.

[0036] Step 5: Construct a three-segment combined sample unit for drilling and inspection; with the structural surface depth interval as the center, construct a three-segment sample unit along the drilling depth direction, which includes a pre-stabilization segment, a structural surface influence segment, and a post-recovery segment, and extract the differential response features between adjacent segments; As a preferred option, the process of constructing a three-segment drilling-viewing joint sample unit is as follows: S51: Centered on the structural surface depth range defined by the structural surface initiation depth and the structural surface termination depth, a drilling section of a predetermined length before the structural surface initiation depth is taken as the pre-stabilization section, used to characterize the complete surrounding rock drilling state before the structural surface influence; the interval between the structural surface initiation depth and termination depth is taken as the structural surface influence section, used to characterize the disturbance state during the process of the drill bit entering, crossing or exiting the structural surface; a drilling section of a predetermined length after the structural surface termination depth is taken as the post-recovery section, used to characterize the drilling recovery state after the drill bit crosses the structural surface. S52: Extract the first differential response features between the pre-stabilized section and the structural surface-affected section, and the second differential response features between the structural surface-affected section and the post-recovery section; the first differential response features and the second differential response features include at least one or more of the following: mean difference, gradient difference, mechanical specific energy difference and drilling stiffness change rate difference, used to quantify the magnitude and pattern of mechanical response changes before and after the drill bit crosses the structural surface. S53: Based on the pre-stabilization section, the structural surface influence section, the post-recovery section, the first differential response characteristics, and the second differential response characteristics, a three-segment drilling-viewing joint sample unit is constructed that is sequentially distributed along the drilling depth direction; In this technical solution, the drilling process is divided into three continuous spatial intervals centered on the depth range of the structural surface: a pre-stabilization segment, a structural surface-affected segment, and a post-recovery segment. This comprehensively depicts the dynamic mechanical response evolution process of the drill bit before and after crossing the structural surface. Simultaneously, differential response features such as the mean difference, gradient difference, mechanical specific energy difference, and drilling stiffness change rate difference between adjacent segments are extracted, quantifying the mechanical response changes caused by the structural surface into calculable feature parameters. This three-segment sample unit not only preserves complete drilling state information before and after the structural surface influence but also enhances the response abrupt change signal at the structural surface boundary through differential features, providing structured and highly discriminative sample input for subsequent soft boundary label construction and recognition model training.

[0037] Step 6: Construct soft boundary labels; convert the hard boundary annotations of the structural surfaces into soft boundary labels in the form of probability distribution to better adapt to the fuzzy transition characteristics of the geological structural surface boundaries; As a preferred option, the process of constructing soft boundary labels is as follows: Structural boundaries are typically not strictly abrupt change points in the drilling response, but rather are influenced by drill bit size, gradual rock mass characteristics, and drilling disturbance propagation, exhibiting as transition zones of a certain thickness. Therefore, to address the fuzzy transition characteristics of structural boundaries caused by drill bit size effects, gradual rock mass characteristics, and drilling disturbance propagation, soft boundary labels are used instead of hard boundary labels; the structural boundary initiation depth is converted into a probability distribution of the initiation boundary, the structural boundary termination depth is converted into a probability distribution of the termination boundary, and the structural boundary influence segment is converted into a probability mask of structural boundary presence. The starting point boundary probability distribution and the ending point boundary probability distribution are constructed with Gaussian, triangular, or trapezoidal probability distribution soft tags centered on the starting depth and ending depth of the structural surface, respectively. The depth extension radius of the probability distribution soft tag is adaptively determined based on the drill bit diameter, current drilling speed, thrust fluctuation amplitude, torque fluctuation amplitude, or the visually visible width of the structural surface. In a preferred embodiment, the probability mask for the existence of the structural surface satisfies the following distribution characteristics: the probability of existence in the central region of the structural surface is not less than 0.8, the probability of existence in the transition region of the structural surface boundary varies continuously between 0.2 and 0.8, and the probability of existence in the stable surrounding rock region outside the structural surface is not higher than 0.2. Through the aforementioned soft boundary labels, the three-segment drilling-peeping joint sample unit simultaneously possesses a structural surface type label, a starting boundary probability label, an ending boundary probability label, and a structural surface existence probability label; In this technical solution, to address the ambiguous transition characteristics of structural surface boundaries caused by drill bit size effects, gradual rock mass characteristics, and drilling disturbance transmission, soft boundary labels in the form of probability distributions are used instead of traditional hard boundary labels. By constructing a probability distribution with an adaptive depth extension radius centered on the start and end depths of the structural surface, and converting the affected segments of the structural surface into a probability mask satisfying high probability at the center, continuous transition at the edges, and low probability on the outer sides, the sample labels more accurately reflect the continuous transition characteristics of the structural surface boundaries in the drilling response. This provides the identification model with multi-probability dimensional learning objectives that are more consistent with geological realities, significantly improving the model's accuracy in locating the start and end boundaries and development range of structural surfaces.

[0038] Step 7: Construct a joint drilling-peeping structure recognition model; Construct a multi-task deep learning model that includes a drilling feature extraction unit, a peeping constraint unit, a feature consistency constraint unit, and multiple output units to achieve end-to-end recognition from drilling parameters to structure type and start and end depths; As a preferred option, the process of constructing the drilling-viewing joint structural surface identification model is as follows: like Figure 4 As shown, a joint drilling-visual structural surface recognition model is constructed based on the drilling feature extraction unit, the peep constraint unit, the drilling-visual feature consistency constraint unit, the structural surface type output unit, and the structural surface start and end depth output unit. The feature extraction unit during drilling includes a one-dimensional multi-scale convolutional layer, a dilated convolutional layer, and a residual connection layer, which are used to extract multi-scale structural surface mechanical response features (including abrupt change features, fluctuation features, and trend features) from the drilling parameter sequence. The observation constraint unit is used to encode the structural surface type, visible width, continuity, filling state or opening state obtained by borehole observation into a visual prior vector, providing visual semantic guidance for drilling features. The drilling-peeping feature consistency constraint unit is configured to: within the structural surface depth range, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features output by the peeping constraint unit is higher than a preset positive sample threshold; and in the non-structural surface segment, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features is lower than a preset negative sample threshold; the similarity measure uses cosine similarity, defined as... ,in The feature vectors during drilling, The visual prior vector is used; as another preferred approach, the feature consistency constraint can also be implemented using contrastive learning loss or mutual information maximization constraint. In this technical solution, an end-to-end mapping from drilling parameters to surface type and start / end depth is achieved by constructing a recognition model that includes a drilling feature extraction unit, a peep constraint unit, a drilling-peep feature consistency constraint unit, and multiple output units. The drilling feature extraction unit extracts multi-scale mechanical response features of the surface, such as abrupt changes, fluctuations, and trends, through one-dimensional multi-scale convolution, dilated convolution, and residual connections. The peep constraint unit encodes visual attributes such as surface type, continuity, and filling state into visual prior vectors, providing semantic guidance for the drilling features. The drilling-peep feature consistency constraint unit applies positive and negative sample similarity constraints to both surface and non-surface segments, enabling the model to automatically align mechanical response features with visual prior features during training, enhancing the model's ability to distinguish different surface types and its robustness to noise interference.

[0039] Step 8: Model training; Train the drilling-peeping joint structural surface recognition model using three-segment drilling-peeping joint sample units and soft boundary labels, and jointly optimize the model parameters through a multi-task loss function; As a preferred option, the model training process is as follows: S81: Constructing a Multi-Task Loss Function ,in For the classification loss of structural surface types, cross-entropy loss is used. Locate the loss at the starting boundary. For endpoint boundary positioning loss, either KL divergence loss or mean square error loss is used. To account for probabilistic losses in structural surfaces, a binary cross-entropy loss is employed. For the loss of consistency of features during drilling and inspection, For physical response constraint loss, the Mahalanobis distance between the mechanical response feature vector of the constraint structure surface affected segment and the mechanical response feature vectors of the pre-stabilized segment and the post-recovery segment is not less than the preset separability threshold. to The weight coefficients for each loss term can be dynamically adjusted based on the convergence speed of the loss of various labels in the training samples, or determined through grid search, Bayesian optimization, or a preset empirical range, so that the structural surface type recognition error and the boundary localization error simultaneously meet the preset training convergence conditions. S82: A joint drilling-visualization structure recognition model is trained using a three-segment joint sample unit for drilling-visualization, and a multi-task loss function is applied. The joint optimization of model parameters enables the drilling-viewing joint structure identification model to output the structure type probability, the probability distribution of the starting boundary, the probability distribution of the ending boundary, and the probability of the existence of the structure based on the drilling parameter sequence and the mechanical response characteristics of the structure. This technical solution achieves multi-dimensional joint optimization of the recognition model by constructing a multi-task loss function that includes structural surface type classification loss, start and end boundary localization loss, structural surface existence probability loss, drilling-visual feature consistency loss, and physical response constraint loss. Specifically, the classification loss and boundary localization loss drive the model to correctly identify the structural surface type and accurately calibrate the start and end depths, respectively; the existence probability loss guides the model to perceive the spatial distribution range of the structural surface; the feature consistency loss integrates the borehole visual vision prior into the drilling-while-drilling feature learning process; and the physical response constraint loss ensures the separability of the mechanical response features of the affected and stable sections of the structural surface through Mahalanobis distance constraints. The weight coefficients can be adaptively adjusted according to the training convergence, enabling coordinated optimization of various tasks during training. Ultimately, the model can simultaneously output complete recognition results, including structural surface type probability, start and end boundary probability distribution, and existence probability.

[0040] Step 9: On-site deployment and real-time identification; Deploy the trained model on-site, form a streaming drilling window according to the depth step size, trigger candidate identification of structural surfaces based on mechanical response characteristics, and output the structural surface type, start and end depths, development length and identification confidence. As a preferred option, the process of on-site deployment and real-time identification is as follows: S91: The triggering condition is when the energy mutation coefficient, drilling rate residual, drilling stiffness change rate, mechanical specific energy peak value, or torque-thrust ratio change exceeds the abnormal triggering threshold adaptively determined by the current stable drilling section of the borehole within the flow-while-drilling window. S92: Deploy the trained drilling-sight joint structural surface recognition model on the field operator; during the field drilling process, form a flow-type drilling window according to the preset depth step size, calculate the energy mutation coefficient, drilling speed residual, drilling stiffness change rate, mechanical specific energy peak value or torque-thrust ratio change in the window in real time, and execute S93 when the triggering condition is met. S93: When structural surface candidate identification is triggered, the boundary fine positioning window is started. The drilling-viewing joint structural surface identification model outputs the structural surface type, starting boundary probability distribution, ending boundary probability distribution, and structural surface existence probability within the current drilling depth range. The output results of multiple consecutive streaming drilling windows are fused at the window level: when the structural surface type probability of adjacent streaming drilling windows is greater than the preset type threshold, and the starting boundary probability distribution or the ending boundary probability distribution overlaps, the adjacent streaming drilling windows are merged into the same structural surface candidate interval, and the starting boundary probability distribution and the ending boundary probability distribution within the same structural surface candidate interval are weighted and the peak value is calculated to determine the starting depth and ending depth of the structural surface. S94: Calculate the identification confidence score; the identification confidence score is jointly determined by the probability of the structure type, the peak probability of the starting boundary, the peak probability of the ending boundary, the consistency coefficient of the continuous window, and the intensity of the physical response anomaly. The identification confidence score is based on the weighted geometric mean of the probability of the structure type, the peak probability of the starting boundary, and the peak probability of the ending boundary, with the consistency coefficient of the continuous window as the stability correction factor and the normalized value of the intensity of the physical response anomaly as the physical support correction factor. The three are weighted and fused to obtain the final identification confidence score; and the structure type, starting depth, ending depth, and development length are output according to the identification confidence score.

[0041] In this technical solution, online adaptive triggering for structural face candidate identification is achieved by using mechanical response characteristics such as energy mutation coefficient and drilling rate residual within the flow-while-drilling window as trigger indicators, and by employing an anomaly trigger threshold adaptively determined by the current stable drilling section of the borehole. During the identification process, the structural face type and start-end boundary probability distribution are output through a boundary precision positioning window, and the output results of multiple consecutive flow-while-drilling windows are fused at the window level—when the type probability and boundary probability distribution of adjacent windows meet the consistency condition, they are merged into the same structural face candidate interval, and the start-end depth of the structural face is accurately determined by weighted peak value calculation. Simultaneously, based on the weighted geometric mean of the structural face type probability, boundary probability peak value, and physical response anomaly intensity, the identification confidence level is calculated by fusing the window consistency coefficient and physical support correction factor, providing a quantifiable reliability evaluation basis for on-site surrounding rock logging and support decision-making.

[0042] Step 10: On-site Feedback and Model Adaptive Correction; After completing the borehole inspection verification, the deviation between the on-site identification results and the verification results is calculated. Based on the type and degree of deviation, the depth drift correction function, the anomaly trigger threshold, and the model parameters are adjusted to achieve continuous optimization of the identification system. As a preferred approach, the process of on-site feedback and model adaptive correction is as follows: S101: After completing the borehole inspection verification, the deviation between the field identification result output in step 9 and the borehole inspection verification result is calculated; the deviation calculation includes the starting depth deviation, the ending depth deviation, the structural surface mechanical response characteristic threshold deviation, and the structural surface type deviation; the starting depth deviation is the absolute value of the difference between the structural surface starting depth output in step 9 and the starting depth determined by the borehole inspection verification; the ending depth deviation is the absolute value of the difference between the structural surface ending depth output in step 9 and the ending depth determined by the borehole inspection verification. S102: When the initial depth deviation or the final depth deviation exceeds the preset depth deviation threshold, the visual marker points of the structural surface determined by the verification and the corresponding candidate mutation points of the mechanical response are added as new corresponding points and added to the drilling-viewing depth corresponding point group in step 2. Step 3 is re-executed to correct the segmented depth drift correction function, so as to improve the depth calibration accuracy of subsequent drilling. S103: When the deviation of the mechanical response characteristic threshold of the structural surface exceeds the preset threshold deviation ratio, the abnormal trigger threshold of the flow-while-drilling window in step 9 is corrected according to the mechanical response characteristic value of the structural surface confirmed by the review, so that the triggering condition matches the abnormal response level confirmed by the review. S104: When the deviation of the structural surface type exceeds the preset deviation ratio, the verified structural surface sample is added to the three-segment drilling-viewing joint sample unit in step 5, and the model calibration parameters or sample weights of the drilling-viewing joint structural surface identification model in step 8 are corrected to enhance the model's ability to identify specific types of structural surfaces. The correction process includes a historical model retention strategy and updated stability constraints: before correction, the current model parameters are saved as a historical version; after correction, the corrected model is evaluated using validation samples. When the boundary positioning error or structure type identification error of the corrected model on the validation samples is greater than that of the model before correction, the model parameters are automatically reverted to those of the model before correction to ensure that the model performance does not degrade during the correction process.

[0043] This technical solution calculates multi-dimensional deviations between the on-site identification results and the borehole inspection verification results, including deviations in initial depth, final depth, mechanical response characteristic thresholds, and structural surface type. For different types and degrees of deviation, it triggers segmented depth drift correction functions, abnormal trigger threshold corrections, and model parameter corrections, achieving comprehensive adaptive optimization of the identification system from depth calibration and triggering mechanisms to model capabilities. The correction process incorporates a historical model retention strategy and update stability constraints. Before correction, the current model parameters are saved as a historical version. After correction, model performance is evaluated using validation samples. When the model performance degrades after correction, it automatically reverts to the parameters before correction, ensuring that the model's performance does not degrade during continuous learning. This achieves safe, stable, and continuous evolution of the identification system in on-site applications.

[0044] Example 2:

[0045] like Figure 4 As shown, this embodiment provides an intelligent identification system for surrounding rock structural surfaces based on drilling parameters and inspection calibration, used to implement an intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration, including: The drilling parameter acquisition module is used to acquire drilling parameters in real time during the drilling process. Through the sensor system installed on the drilling rig, it can acquire at least the drilling parameters including drilling depth, drilling time, drilling speed, rotation speed, thrust and torque. It can also acquire one or more extended parameters among hydraulic pressure, drill pipe displacement, borehole inclination angle, drilling vibration, drilling current, drilling rig power, flushing fluid pressure or flushing fluid flow rate. The borehole observation data acquisition module is used to acquire borehole observation data corresponding to the borehole after drilling is completed. It can acquire at least one or more types of observation data, including borehole video, borehole image, or borehole unfolded diagram, through a borehole observation instrument, borehole television, borehole wall imaging device, or borehole wall unfolded diagram generation device. It can also acquire one or more of the following: borehole wall panoramic unfolded diagram, structural surface manual verification results, structural surface image segmentation results, or borehole wall color and texture features. When it is not possible to observe the same borehole, the observation data of adjacent calibrated boreholes that meet the preset spacing threshold, lithological consistency, or stratigraphic correspondence conditions can be used. The drilling-visual depth calibration module is used to form a drilling parameter depth sequence by arranging drilling parameters according to drilling depth, and a borehole visual depth sequence by arranging visual data according to borehole depth, with the borehole opening as a unified depth zero point. It extracts structural surface visual markers from the borehole visual depth sequence and mechanical response candidate abrupt change points from the drilling parameter depth sequence. Within a preset depth search radius, it matches the structural surface visual markers and mechanical response candidate abrupt change points to establish a drilling-visual depth corresponding point group. Based on the depth offset of the corresponding point group, it constructs a piecewise depth drift correction function that satisfies monotonically increasing constraints and piecewise continuity constraints, and corrects at least one of the drilling parameter depth sequence and the borehole visual depth sequence to obtain a calibrated drilling parameter sequence and calibrated borehole visual data under a unified borehole depth coordinate system. The drilling-visualization depth calibration module includes a visual marker extraction unit, a mechanical abrupt change point extraction unit, a corresponding point group construction unit, and a correction function generation unit. The visual marker extraction unit identifies fracture edge points, rock layer interface points, delamination boundary points, fracture zone start and end points, or borehole wall color band abrupt change points from borehole visualization data. The mechanical abrupt change point extraction unit detects drilling speed drop points, torque increase points, thrust abrupt change points, mechanical specific energy peak points, or energy abrupt change points per unit footage from the drilling parameter depth sequence. The corresponding point group construction unit matches structural surface visual marker points and candidate mechanical response abrupt change points within a preset depth search radius, prioritizing candidate abrupt change points with the largest mechanical response anomaly intensity and a depth offset less than a preset offset threshold as corresponding points. The correction function generation unit constructs a piecewise depth drift correction function based on the depth offset of the corresponding point group and adaptively selects piecewise linear interpolation, spline interpolation, or dynamic time warping constraint to solve for the depth correction amount based on the corresponding point density.

[0046] The structural surface information extraction module is used to determine the structural surface type, starting depth, ending depth, and visual attributes of the surrounding rock structural surface based on the calibrated borehole inspection data, and to calculate the mechanical response characteristics of the structural surface based on the calibrated drilling parameter sequence. The structural surface type includes one or more of the following: transverse fractures, longitudinal fractures, oblique fractures, bedding planes, rock layer interfaces, delamination, fracture zones, weak interlayer interfaces, and intact surrounding rock transition interfaces. The visual attributes include one or more of the following: visible width of the structural surface, continuity, filling state, opening state, degree of color change, and degree of texture change. The mechanical response characteristics of the structural surface include at least several of the following: energy per unit footage, mechanical specific energy, torque-to-thrust ratio, drilling rate residual, energy mutation coefficient, and drilling stiffness change rate. A three-segment sample construction module is used to construct a three-segment combined drilling-visualization sample unit along the borehole depth direction, centered on the structural surface depth range defined by the structural surface initiation depth and structural surface termination depth. The three-segment combined drilling-visualization sample unit includes a pre-stabilization segment, a structural surface influence segment, and a post-recovery segment, as well as a first differential response feature between the pre-stabilization segment and the structural surface influence segment, and a second differential response feature between the structural surface influence segment and the post-recovery segment. The pre-stabilization segment characterizes the complete surrounding rock drilling state before the structural surface influence; the structural surface influence segment characterizes the disturbance state during the drill bit's entry, crossing, or exit from the structural surface; and the post-recovery segment characterizes the drilling recovery state after the drill bit crosses the structural surface. The first and second differential response features include at least one or more of the following: mean difference, gradient difference, mechanical specific energy difference, and drilling stiffness change rate difference. The soft boundary label construction module is used to address the fuzzy transition characteristics of the structure boundary caused by drill bit size effects, rock mass gradation characteristics, and drilling disturbance transmission. It converts the starting depth of the structure boundary into a starting boundary probability distribution, the ending depth into an ending boundary probability distribution, and the affected segment into a structure existence probability mask. This allows the three-segment drilling-viewing joint sample unit to simultaneously possess structure type labels, starting boundary probability labels, ending boundary probability labels, and structure existence probability labels. The starting and ending boundary probability distributions are constructed with the starting and ending depths of the structure boundary as the center, respectively, and their depth extension radius is adaptively determined based on the drill bit diameter, current drilling speed, thrust fluctuation amplitude, torque fluctuation amplitude, or the visually visible width of the structure. The structure existence probability mask satisfies the following distribution characteristics: the existence probability in the central region of the structure boundary is not lower than a preset high threshold, the existence probability in the boundary transition region changes continuously, and the existence probability in the outer stable surrounding rock region is not higher than a preset low threshold. The peeping constraint and feature consistency module, built into the drilling-peeping joint structure surface recognition model, is used to encode the structure surface type, visible width, continuity, filling state, or open state obtained from drilling peeping into visual prior vectors. During model training, it applies drilling-peeping feature consistency constraints: within the structure surface depth range, the similarity between the output features of the drilling feature extraction unit and the visual prior vector is higher than a preset positive sample threshold; in non-structure surface segments, the similarity is lower than a preset negative sample threshold. The similarity metric uses cosine similarity, contrastive learning loss, or mutual information maximization constraints. The model training module receives the three-segment drilling-peeping joint sample unit output by the three-segment sample construction module and the soft boundary label output by the soft boundary label construction module. It trains a drilling-peeping joint structural surface recognition model, including a drilling feature extraction unit, a peeping constraint unit, a drilling-peeping feature consistency constraint unit, a structural surface type output unit, and a structural surface start-end depth output unit. The drilling feature extraction unit includes a one-dimensional multi-scale convolutional layer, a dilated convolutional layer, and a residual connection layer. During training, a multi-task loss function is used to jointly optimize the model parameters. This multi-task loss function includes structural surface type classification loss, start-point boundary localization loss, end-point boundary localization loss, structural surface existence probability loss, drilling-peeping feature consistency loss, and physical response constraint loss. The physical response constraint loss ensures that the Mahalanobis distance between the mechanical response feature vector of the structural surface influence segment and the mechanical response feature vectors of the preceding stable segment and the subsequent recovery segment is not less than a preset separability threshold. The model training module includes a label receiving unit, a loss function construction unit, and a weight coefficient determination unit. The label receiving unit receives the structure surface type label, starting boundary probability label, ending boundary probability label, and structure surface existence probability label output by the soft boundary label construction module. The loss function construction unit constructs a multi-task loss function based on the structure surface type label, starting boundary probability label, ending boundary probability label, and structure surface existence probability label. This function includes structure surface type classification loss, starting boundary localization loss, ending boundary localization loss, structure surface existence probability loss, drilling-peeping feature consistency loss, and physical response constraint loss. The weight coefficient determination unit dynamically adjusts the weight coefficients of each loss term based on the convergence speed of the loss for each type of label in the training samples, or determines them through grid search, Bayesian optimization, or a preset empirical range, ensuring that the structure surface type identification error and boundary localization error simultaneously meet the preset training convergence conditions.

[0047] The on-site identification module is used to deploy the trained drilling-viewing joint structural surface identification model on the on-site operating host. During on-site drilling, it forms a flowing drilling window at a preset depth step size, and calculates in real time the energy mutation coefficient, drilling rate residual, drilling stiffness change rate, peak mechanical energy, or torque-thrust ratio change within the window. When the above indicators exceed the abnormal trigger threshold adaptively determined by the current stable drilling section, it triggers structural surface candidate identification, starts the boundary fine positioning window, and outputs the structural surface type, starting boundary probability distribution, ending boundary probability distribution, and structural surface existence probability within the current drilling depth range. It performs window-level fusion of the output results of multiple consecutive flowing drilling windows to determine the structural surface starting depth and structural surface ending depth. It calculates and outputs the identification confidence score, which includes the structural surface type probability, the peak probability of the starting boundary, the peak probability of the ending boundary, the consistency coefficient of the continuous window, and the intensity of the physical response anomaly, as well as the structural surface type, starting depth, ending depth, and development length. The on-site identification module includes a streaming window management unit, an anomaly threshold adaptive unit, an inference execution unit, and a window fusion unit. The streaming window management unit forms streaming windows at preset depth steps and calculates in real time the energy mutation coefficient, drilling rate residual, drilling stiffness change rate, peak mechanical energy, or torque-to-thrust ratio change within the window. The anomaly threshold adaptive unit dynamically updates the anomaly trigger threshold based on the statistical values ​​of the mechanical response characteristics of the current stable drilling section, ensuring the triggering conditions match the current drilling conditions. The inference execution unit loads the trained drilling-viewing joint structural surface identification model, activates the boundary fine-positioning window, and outputs the identification results when triggering structural surface candidate identification. The window fusion unit performs depth adjacency judgment and boundary probability weighted fusion on the output results of multiple consecutive streaming windows. When the structural surface type probabilities of adjacent windows are all greater than the preset type threshold and the boundary probability distributions overlap, the adjacent windows are merged into the same structural surface candidate interval, and the starting and ending depths of the structural surface are determined by weighted peak value calculation.

[0048] The on-site feedback and model correction module is used to calculate the deviation between the on-site identification results and the borehole inspection verification results after the borehole inspection verification is completed. Based on the deviation type and degree, it triggers segmented depth drift correction, anomaly trigger threshold correction, and model parameter correction respectively. When the initial depth deviation or termination depth deviation exceeds the preset depth deviation threshold, the newly added corresponding point determined by the verification is added to the drilling-inspection depth corresponding point group and depth correction is re-executed. When the mechanical response characteristic threshold deviation exceeds the preset threshold deviation ratio, the anomaly trigger threshold is corrected. When the structural surface type deviation exceeds the preset type deviation ratio, the model calibration parameters or sample weights are corrected. The correction process includes a historical model retention strategy and updated stability constraints: before correction, the current model parameters are saved as a historical version; after correction, the model is evaluated on validation samples; when the model performance degrades after correction, it automatically reverts to the parameters before correction. This system can deeply integrate borehole inspection vision priors with drilling mechanical response in a unified depth coordinate system, realizing fully automated processing of the surrounding rock structure from data acquisition to intelligent identification and continuous optimization.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent identification of surrounding rock structural surfaces based on drilling parameters and inspection calibration, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition; Drilling operations are carried out in the surrounding rock to be identified, and drilling parameters during the drilling process and inspection data after the drilling is completed are collected simultaneously. Step 2: Establish a set of corresponding points for drilling-visiting depths; using the borehole opening as a unified zero point of depth, establish a drilling parameter depth sequence and a borehole visiting depth sequence respectively, extract feature points from the two sequences and perform matching in the depth direction; Step 3: Segmented depth drift correction; construct a segmented depth drift correction function based on the depth offset of the corresponding point group, and correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain calibrated data under a unified borehole depth coordinate system; Step 4: Structural surface feature extraction; Determine the type and visual attributes of the structural surfaces based on the calibrated data, and calculate the mechanical response characteristics of the structural surfaces; Step 5: Construct a three-segment combined sample unit for drilling and inspection; with the structural surface depth interval as the center, construct a three-segment sample unit along the drilling depth direction, which includes a pre-stabilization segment, a structural surface influence segment, and a post-recovery segment, and extract the differential response features between adjacent segments; Step 6: Construct soft boundary labels; convert the hard boundary annotations of the structural surfaces into soft boundary labels in the form of probability distributions; Step 7: Construct a joint drilling-peeping structure recognition model; Construct a multi-task deep learning model that includes a drilling feature extraction unit, a peeping constraint unit, a feature consistency constraint unit, and multiple output units to achieve end-to-end recognition from drilling parameters to structure type and start and end depths; Step 8: Model training; Train the drilling-peeping joint structural surface recognition model using three-segment drilling-peeping joint sample units and soft boundary labels, and jointly optimize the model parameters through a multi-task loss function; Step 9: On-site deployment and real-time identification; Deploy the trained model on-site, form a flowing drilling window according to the depth step size, trigger candidate identification of structural surfaces based on mechanical response characteristics, and output the structural surface type, start and end depths, development length and identification confidence.

2. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 1, the multi-source data acquisition process is as follows: S11: Drilling operations are carried out in the surrounding rock to be identified in the underground engineering project; during the drilling operation, drilling parameters are collected in real time through a sensor system installed on the drilling rig; the drilling parameters include at least drilling depth, drilling time, drilling speed, rotational speed, thrust, and torque; S12: After drilling is completed, drilling observation data corresponding to the drill hole is collected using a drilling observation device; the drilling observation data includes at least one or more of drilling video, drilling image or drilling unfolded diagram; For situations where it is impossible to inspect the same borehole under certain engineering conditions, borehole inspection data from adjacent calibration boreholes are used. When the borehole inspection data comes from adjacent calibration boreholes, it is ensured that the distance between the adjacent calibration boreholes and the borehole to be identified does not exceed the preset borehole distance threshold, and the lithological consistency index, stratigraphic correspondence, or structural surface development similarity index of the borehole sections where the two are located meet the preset consistency conditions to ensure the transferability of calibration data.

3. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 2, the process of establishing the corresponding point group for the drilling-viewing depth is as follows: S21: Using the borehole opening as a unified depth zero point, the drilling parameters are arranged according to the drilling depth to form a drilling parameter depth sequence, and the inspection data are arranged according to the borehole depth to form a borehole inspection depth sequence. S22: Extract structural visual markers from the borehole inspection depth sequence; the structural visual markers are extracted by image segmentation, manual verification or color and texture abrupt change detection, including fracture edge points, rock layer interface points, bedding abrupt change points, delamination boundary points, fracture zone start and end points, weak interlayer interface points or borehole wall color band abrupt change points. Candidate abrupt change points of mechanical response are extracted from the depth sequence of drilling parameters. These candidate abrupt change points can be extracted by methods such as multi-parameter anomaly detection, including points of sudden drop in drilling speed, sudden increase in torque, sudden change in thrust, peak point of mechanical specific energy, sudden change point of energy per unit footage, change point of drilling stiffness, or joint anomaly point of multiple parameters. S23: Match the visual markers of the structural surface and the candidate mutation points of the mechanical response within the preset depth search radius to establish a set of corresponding points for the drilling-viewing depth; when the same visual marker of the structural surface corresponds to multiple candidate mutation points of the mechanical response, select the candidate mutation point of the mechanical response with the largest mechanical response anomaly intensity and a depth offset less than the preset offset threshold as the corresponding point.

4. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 3, the process of segmented depth drift correction is as follows: S31: Construct a piecewise depth drift correction function based on the depth offset of each corresponding point in the drilling-viewing depth corresponding point group. ,in To correct the depth before, For the corrected depth, For the first Each depth segment; For the first Depth correction within each depth segment; at the same time It is a strictly monotonically increasing function. Correction amount within each segment Determined by interpolation of the depth offset of the corresponding point within the corresponding depth interval; S32: Use the segmented depth drift correction function to correct at least one of the drilling parameter depth sequence and the borehole inspection depth sequence to obtain the calibrated drilling parameter sequence and calibrated borehole inspection data under a unified borehole depth coordinate system.

5. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 4, the process of extracting structural surface features is as follows: The structural surface type, starting depth, ending depth, and visual attributes of the surrounding rock structural surface are determined based on the calibrated borehole inspection data. The structural surface type includes one or more of the following: transverse fractures, longitudinal fractures, oblique fractures, bedding planes, rock layer interfaces, delamination, fracture zones, weak interlayer interfaces, and intact surrounding rock transition interfaces. The visual attributes include one or more of the following: visible width of the structural surface, continuity, filling state, opening state, degree of color change, and degree of texture change. Simultaneously, the mechanical response characteristics of the structural surface are calculated based on the calibrated drilling parameter sequence; the mechanical response characteristics of the structural surface include at least several of the following: energy per unit footage, mechanical specific energy, torque-to-thrust ratio, drilling rate residual, energy mutation coefficient, and drilling stiffness change rate.

6. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 5, the process of constructing the three-segment drilling-viewing joint sample unit is as follows: S51: Taking the structural surface depth range defined by the structural surface initiation depth and the structural surface termination depth as the center, take a drilling interval of a preset length before the structural surface initiation depth as the pre-stabilization section; take the interval between the structural surface initiation depth and termination depth as the structural surface influence section; take a drilling interval of a preset length after the structural surface termination depth as the post-recovery section. S52: Extract the first differential response features between the pre-stabilization segment and the structural surface-affected segment, and the second differential response features between the structural surface-affected segment and the post-recovery segment; the first differential response features and the second differential response features include at least one or more of the following: mean difference, gradient difference, mechanical specific energy difference, and drilling stiffness change rate difference; S53: Based on the pre-stabilization section, the structural surface influence section, the post-recovery section, the first differential response characteristics, and the second differential response characteristics, a three-segment drilling-viewing joint sample unit is constructed that is sequentially distributed along the drilling depth direction.

7. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 6, the process of constructing soft boundary labels is as follows: To address the fuzzy transition characteristics of the structure boundary during drilling caused by drill bit size effect, rock mass gradual change characteristics, and drilling disturbance transmission, soft boundary labels are used instead of hard boundary labels; the starting depth of the structure is converted into the probability distribution of the starting boundary, the ending depth of the structure is converted into the probability distribution of the ending boundary, and the influence segment of the structure is converted into a probability mask of the existence of the structure. The starting point boundary probability distribution and the ending point boundary probability distribution are constructed with Gaussian, triangular, or trapezoidal probability distribution soft tags centered on the starting depth and ending depth of the structural surface, respectively. The depth extension radius of the probability distribution soft tag is adaptively determined based on the drill bit diameter, current drilling speed, thrust fluctuation amplitude, torque fluctuation amplitude, or the visually visible width of the structural surface. The probability mask for the existence of the structural surface satisfies the following distribution characteristics: the existence probability in the central region of the structural surface is not less than 0.8, the existence probability in the transition region of the structural surface boundary varies continuously between 0.2 and 0.8, and the existence probability in the stable surrounding rock region outside the structural surface is not higher than 0.

2. Through the aforementioned soft boundary labels, the three-segment drilling-peeping joint sample unit simultaneously possesses a structural surface type label, a starting boundary probability label, an ending boundary probability label, and a structural surface existence probability label.

8. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, In step 7, the process of constructing the drilling-viewing joint structural surface identification model is as follows: A joint drilling-visual structural surface recognition model is constructed based on the drilling feature extraction unit, the peep constraint unit, the drilling-visual feature consistency constraint unit, the structural surface type output unit, and the structural surface start and end depth output unit. The feature extraction unit during drilling includes a one-dimensional multi-scale convolutional layer, a dilated convolutional layer, and a residual connection layer, which are used to extract multi-scale structural surface mechanical response features from the drilling parameter sequence. The observation constraint unit is used to encode the structural surface type, visible width, continuity, filling state or opening state obtained by borehole observation into a visual prior vector. The drilling-peeping feature consistency constraint unit is configured to: within the structural surface depth range, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features output by the peeping constraint unit is higher than a preset positive sample threshold; and in the non-structural surface segment, ensure that the similarity between the output features of the drilling feature extraction unit and the borehole peeping visual prior features is lower than a preset negative sample threshold; the similarity measure uses cosine similarity, defined as... ,in The feature vectors during drilling, This is the visual prior vector.

9. The intelligent identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration according to claim 1, characterized in that, It also includes step 10: on-site feedback and model adaptive correction; after completing the borehole inspection verification, the deviation between the on-site identification results and the verification results is calculated, and the depth drift correction function, abnormal trigger threshold and model parameters are corrected according to the type and degree of deviation, so as to realize the continuous optimization of the identification system.

10. A smart identification system for surrounding rock structural surfaces based on drilling parameters and inspection calibration, used to implement the smart identification method for surrounding rock structural surfaces based on drilling parameters and inspection calibration as described in any one of claims 1 to 9, characterized in that, include: The drilling parameter acquisition module is used to acquire drilling parameters in real time during the drilling process. The borehole observation data acquisition module is used to collect borehole observation data corresponding to the borehole after drilling is completed; The drilling-visual depth calibration module is used to extract structural surface visual markers from borehole visual data, extract mechanical response candidate abrupt change points from drilling parameters, match and establish corresponding point groups in the depth direction, and construct a segmented depth drift correction function to uniformly calibrate the two depth sequences. The structural surface information extraction module is used to determine the structural surface type, starting depth, ending depth and visual attributes of the surrounding rock structural surface based on the calibrated borehole inspection data, and to calculate the mechanical response characteristics of the structural surface based on the calibrated drilling parameter sequence. The three-segment sample construction module is used to construct a three-segment drilling-viewing joint sample unit along the borehole depth direction, with the structural surface depth range defined by the structural surface start depth and structural surface end depth as the center. The soft boundary label construction module is used to convert the starting depth of the structural surface into the probability distribution of the starting boundary, the ending depth of the structural surface into the probability distribution of the ending boundary, and the influence segment of the structural surface into the probability mask of the existence of the structural surface, so that the three-segment drilling-peeping joint sample unit simultaneously has structural surface type label, starting boundary probability label, ending boundary probability label and structural surface existence probability label. The peeping constraint and feature consistency module, built into the drilling-peeping joint structure surface recognition model, is used to encode the structure surface type, visible width, continuity, filling state, or open state obtained from borehole peeping into visual prior vectors, and to apply drilling-peeping feature consistency constraints during model training. The model training module is used to receive the three-segment drilling-peeping joint sample units output by the three-segment sample construction module and the soft boundary labels output by the soft boundary label construction module, and to train the drilling-peeping joint structural surface recognition model. The on-site identification module is used to deploy the drilling-viewing combined structural surface identification model on the on-site operating host. During the on-site drilling process, a flow-type drilling window is formed according to the preset depth step size, and the energy mutation coefficient, drilling speed residual, drilling stiffness change rate, mechanical specific energy peak value or torque-thrust ratio change within the window are calculated in real time. The on-site feedback and model correction module is used to calculate the deviation between the on-site identification results and the borehole inspection verification results after the borehole inspection verification is completed. Based on the deviation type and degree, it triggers the segmented depth drift correction function, the abnormal trigger threshold correction, and the model parameter correction respectively.

Images