Autonomous navigation operation method and system of a micro drilling robot

Abstract

The application discloses a kind of autonomous navigation operation method and system of micro drilling robot, executes illumination abnormal response detection and feature sparse area positioning to borehole inner wall image sequence, structure anchor point sequence is generated jointly;According to structure anchor point sequence, initial pose sequence is generated by executing cross-frame pose transmission, and borehole point cloud map is constructed by combining hole wall ellipsoidal deformation analysis and geometric constraint fusion;The cumulative drift of initial pose sequence is implemented directional decomposition, and the axial correction amount and radial constraint mark are respectively obtained by absolute depth anchoring and wall structure frequency domain matching, and generate directional correction instruction;According to directional correction instruction, frame-by-frame re-correction and global consistency optimization are executed to initial pose sequence, and navigation positioning map is generated by point cloud registration and geographic coordinate system unification, realizes the stable pose estimation and high-precision map construction of micro drilling robot under the condition of hole wall texture deficiency and complex illumination.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation technology for robots, and in particular to an autonomous navigation operation method and system for a micro drilling robot. Background Technology

[0002] When micro-drilling robots operate in complex underground tunnels, the light distribution on the surface of the borehole wall is highly uneven due to differences in lithology, fracture development, and water seepage. In some strata, the fine and uniform mineral particles lack texture features that can be used by the visual system. These two objective conditions together restrict the robot's stable perception ability in the tunnel, making image-based position estimation prone to degradation and failure in certain strata.

[0003] Regarding the accuracy of pose estimation, the propulsion depth estimation relying on a single visual odometry will accumulate systematic errors over the course of the journey. The drift in the depth direction and the radial direction are independent in origin but are treated together, lacking a differentiated correction mechanism for the drift characteristics in different directions. In terms of 3D modeling of the working environment, the elliptic deformation of the borehole wall geometry and the local distortion of the point cloud projection are difficult to eliminate together in existing methods. This results in significant deficiencies in the geometric consistency and coordinate system normalization of the final underground space model, which restricts the realization of reliable robot navigation in complex strata. Summary of the Invention

[0004] This invention discloses an autonomous navigation method and system for a micro-drilling robot. It involves performing illumination anomaly detection and sparse texture region localization on borehole image sequences, jointly annotating the borehole wall structure interface to generate an anchor point sequence, driving cross-frame pose transfer and borehole wall geometry measurement, and integrating elliptic deformation analysis and geometric constraints to construct a point cloud map. It introduces absolute anchoring of the drilling depth and frequency domain matching of the borehole wall structure to perform directional correction on the accumulated drift, generating directional correction commands. Based on the correction commands, it performs recalibration and consistency optimization on the pose sequence, ultimately outputting a navigation and positioning map with geographic coordinate normalization, achieving autonomous and reliable navigation of the micro-drilling robot in complex borehole environments.

[0005] The first aspect of this invention proposes an autonomous navigation operation method for a micro-drilling robot, comprising the following steps: A sequence of borehole inner wall images is obtained. Illumination anomaly response detection and feature sparse region localization are performed on the borehole inner wall image sequence to generate an illumination anomaly distribution map and a weak texture boundary map. Based on the illumination anomaly distribution map and the weak texture boundary map, joint annotation of wall structure interface transition is performed to generate a sequence of structural anchor points. Based on the structural anchor point sequence, cross-frame pose transfer is performed to generate an initial pose sequence. The borehole inner wall image sequence is subjected to borehole wall visual measurement to generate a borehole diameter deviation map. The initial pose sequence and the borehole diameter deviation map are geometrically constrained and fused to construct a borehole point cloud map. The cumulative drift in the initial pose sequence is decomposed directionally to identify axial drift and radial drift. The axial drift is corrected by sensor absolute depth anchoring to obtain axial correction amount. The radial drift and the weak texture boundary map are periodically matched with the wall structure to obtain radial constraint identifier. The axial correction amount and the radial constraint identifier are used to generate directional correction instructions. The initial pose sequence is recalibrated frame by frame according to the directional correction instruction to generate a calibrated pose sequence. Based on the calibrated pose sequence, the borehole point cloud map is optimized for global consistency to generate a navigation and positioning map.

[0006] A second aspect of this invention provides an autonomous navigation system for a micro-drilling robot, comprising: The structural annotation unit is used to acquire the borehole inner wall image sequence, perform illumination anomaly response detection and feature sparse region localization on the borehole inner wall image sequence to generate illumination anomaly distribution map and weak texture boundary map, and perform joint annotation of wall structure interface transition based on the illumination anomaly map and weak texture boundary map to generate structural anchor point sequence. The point cloud construction unit is used to generate an initial pose sequence by performing cross-frame pose transfer based on the structure anchor point sequence, perform hole wall visual measurement on the borehole inner wall image sequence to generate a hole diameter deviation map, and construct a borehole point cloud map by geometric constraint fusion positioning through the initial pose sequence and the hole diameter deviation map. The drift correction unit is used to perform directional decomposition on the cumulative drift in the initial pose sequence to identify axial drift and radial drift, perform sensor absolute depth anchoring correction on the axial drift to obtain axial correction amount, perform periodic matching of wall structure based on the radial drift and the weak texture boundary map to obtain radial constraint identifier, and generate directional correction command based on the axial correction amount and the radial constraint identifier. The positioning optimization unit is used to perform frame-by-frame pose recorrection on the initial pose sequence according to the directional correction instruction to generate a corrected pose sequence, and to perform global consistency optimization on the borehole point cloud map based on the corrected pose sequence to generate a navigation and positioning map.

[0007] The beneficial effects of this invention are reflected in the following points: First, the boundaries of the abnormal illumination response area and the sparse feature area are extracted from the borehole inner wall image sequence. A structural anchor point sequence is jointly generated through spatial overlay and cross-frame confidence assessment, solving the problem of visual perception degradation in sections with scarce borehole wall texture and illumination disturbances. The structural anchor point sequence drives cross-frame pose transfer, and the borehole diameter deformation distribution and stress direction are extracted by frame-by-frame borehole wall contour ellipse fitting. Borehole wall geometric constraints are integrated into pose estimation, achieving stable initial pose sequences and point cloud map construction in complex borehole environments. Second, the cumulative drift of the initial pose sequence is decomposed directionally, and axial drift and radial drift are processed independently. Axial drift is corrected by absolute depth anchoring after depth increment steady-state estimation and effective frame selection. Radial drift is frequency-domain correlated with the borehole wall structure periodic feature spectrum after transmission interference harmonic stripping. The differentiated correction strategy for the two types of drift solves the problem that a single correction method cannot simultaneously adapt to depth estimation errors and lateral offset errors. The generated directional correction instructions provide a physically clear correction basis for pose recalibration. Finally, based on the directional correction instructions, the initial pose sequence is subjected to frame-by-frame overlay and adjacent frame smoothing constraints. After global pose consistency verification, residual jumps are eliminated to generate a corrected pose sequence. The corrected pose sequence is used to drive the pose reprojection of the borehole point cloud map. Local distortion areas are located through overlap area consistency detection and iterative nearest point registration optimization is performed. Finally, the geographic coordinates are normalized by ground control point coordinate system transformation, and a navigation and positioning map with lithological properties and reliability labels is output. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating an autonomous navigation operation method for a micro drilling robot according to the present invention.

[0009] Figure 2 This is a structural block diagram of an autonomous navigation operation system for a miniature drilling robot according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0012] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides an autonomous navigation operation method for a micro drilling robot, including the following steps S110-S140: Step S110: Obtain the borehole inner wall image sequence, perform illumination anomaly response detection and feature sparse region localization on the borehole inner wall image sequence to generate illumination anomaly distribution map and weak texture boundary map, and perform joint annotation of wall structure interface transition based on illumination anomaly distribution map and weak texture boundary map to generate structural anchor point sequence.

[0015] Specifically, a sequence of borehole wall images is acquired. The miniature drilling robot, equipped with a ring-shaped LED array illumination module and a wide-angle fisheye camera, continuously acquires borehole wall images at a fixed frame rate during its advance. The frame rate is adjusted in conjunction with the robot's advance speed; when the advance speed is high, the frame rate automatically increases to ensure that the overlap between adjacent frames is not less than 60%. During pauses, the minimum frame rate is maintained to record static morphological changes in the borehole wall. Each frame of the borehole wall image sequence covers the entire 360° circumference of the borehole wall. The spatial resolution varies with the borehole diameter; the effective pixel density of a single frame is higher in the borehole narrowing section than in the borehole widening section. The difference in pixel density between the two borehole diameter sections is recorded in the image sequence's metadata as a borehole diameter estimation field. Frame number and encoder readings are recorded simultaneously during each frame acquisition. The encoder reading serves as the initial depth marker for each frame and is doubly bound to the frame number. If one marker fails, the depth assignment can still be recovered using the other marker. In the borehole inner wall image sequence, blurred frames caused by robot jitter or short pauses are judged by the average response of the Laplacian operator as the sharpness index. Frames below the threshold are marked as low-quality frames and do not participate in subsequent abnormal lighting response detection, but are retained in the sequence to maintain the continuity of frame sequence numbers. The marking information of low-quality frames and the frame sequence number are archived together for manual review.

[0016] Anomaly detection of illumination response was performed on the borehole inner wall image sequence to generate anomaly distribution maps. When the borehole advanced to the jointed and fractured section or the local seepage area of ​​the borehole wall, the illumination light from the ring LED array experienced enhanced diffuse reflection at the fracture opening and concentrated specular reflection in the seepage film-covered area. Both types of physical effects manifested as local brightness anomalies in the borehole inner wall image sequence, forming a clear contrast with the uniform illumination response of the normal rock wall area. The local brightness deviation of each pixel was calculated frame-by-frame for non-low-quality frames in the borehole inner wall image sequence. The local brightness deviation was defined as the difference between the pixel brightness value and the mean of the neighborhood centered on that pixel. Where I(x,y) represents the pixel brightness value, and Imean(x,y) represents the neighborhood mean. The neighborhood size matches the spatial resolution of the borehole inner wall image sequence. Pixels exceeding twice the standard deviation of the entire frame are considered as abnormal illumination pixels. Spatially clustered areas of abnormal illumination pixels are considered as abnormal illumination response zones. Isolated, scattered abnormal illumination pixels are considered sensor noise and are not included in the abnormal illumination response zones. Clustering of abnormal illumination response zones is performed using connected component analysis. Regions with connected component areas below a minimum area threshold are also considered noise and removed. Connected component areas exceeding the threshold have their spatial range recorded using the coordinates of their circumscribed rectangles. The abnormal illumination response zones of each frame are generated as a set of pixel coordinates arranged by frame number to form an abnormal illumination distribution map. The coordinate set of abnormal regions in each frame is bound and stored with the initial depth marker of the corresponding frame. The lighting anomaly distribution map performs temporal consistency verification between adjacent frames. Transient markers are added only to isolated anomaly areas that appear in a single frame to distinguish them from persistent anomaly areas. The decision window width of the transient markers is determined by the estimated number of frames of maximum robot jitter.

[0017] Feature sparse region localization is performed on the borehole inner wall image sequence to generate a weak texture boundary map. When the borehole traverses homogeneous mudstone or dense fine-grained sandstone, the mineral grains on the borehole wall surface are small and uniform, presenting large areas of smooth, low-contrast regions in the borehole inner wall image sequence, with extremely scarce visual features such as corner points. These regions are separated from the high-texture regions formed by adjacent coarse-grained sandstone or conglomerate sections by boundary lines in the weak texture boundary map. Harris corner response maps are extracted frame by frame from non-low-quality frames in the borehole inner wall image sequence. The local maximum density of the corner response map reflects the texture richness of the region. Continuous regions with density below a preset threshold are identified as feature sparse regions. The threshold is determined by the low percentile value of the global corner density distribution in the borehole inner wall image sequence to accommodate the texture benchmark differences of different strata. The boundary of the feature sparse region is defined as the gradient band where the corner response density transitions from the low-density region to the normal-density region. The midline of the gradient band is recorded as the boundary line in the weak texture boundary map, and the direction of the boundary line is determined by fitting the line connecting the density peak points of each column within the gradient band. The weak texture boundary map uses the coordinates of weak texture boundary lines in each frame. These boundary line coordinates exhibit depth continuity between frames, and the depth spacing between boundary line coordinates in adjacent frames inherits the resolution of the propulsion encoder. When the boundary line position jump exceeds a threshold between frames, boundary line continuity is restored. This restoration involves interpolating the median of the boundary line positions from the three frames before and after the jump to eliminate boundary jitter introduced by single-frame detection noise. In the borehole inner wall image sequence, the sparse feature areas and areas of abnormal illumination highly overlap within the fracture development zone. Rock wall spalling at the fracture opening leads to localized texture loss, while simultaneous illumination anomalies caused by the fractures occur.

[0018] Structural anchor point sequences are generated through joint annotation of wall structure interface transitions based on illumination anomaly distribution maps and weak texture boundary maps. When a borehole crosses a lithological interface, the differences in mineral composition and structure between the upper and lower rock strata on the borehole wall simultaneously cause changes in illumination response and abrupt changes in texture density. These two types of signals are highly coupled spatially. Joint annotation utilizes this coupling relationship to screen anchor points with reliable spatial positioning significance, while retaining candidate points triggered by a single signal with lower confidence. Illumination anomaly distribution maps and weak texture boundary maps are spatially superimposed in the same frame coordinate system. In the superimposed result, locations where both illumination anomaly areas and weak texture boundary lines appear simultaneously are identified as strong interface anchor point candidates; locations where only illumination anomaly areas appear but are not covered by weak texture boundary lines are identified as weak interface anchor point candidates; locations where only a persistent boundary line exists in the weak texture boundary map but is not covered by illumination anomaly areas are also included as weak interface anchor point candidates. The initial confidence value is lower than that of the case where both types of signals appear simultaneously. The two types of candidate points are stored in the structural anchor point sequence separately using a confidence field, with the initial confidence value of strong interface anchor point candidates being higher than that of weak interface anchor point candidates. In the illumination anomaly distribution map, single-frame anomaly regions carrying transient markers are removed from the candidate set if they do not reappear in adjacent frames during cross-frame verification. This is to eliminate the impact of transient illumination disturbances caused by robot propulsion jitter on the quality of the structural anchor point sequence. In the weak texture boundary map, boundary lines that maintain spatial stability across multiple consecutive frames are marked as persistent weak texture boundaries. When a persistent weak texture boundary overlaps with anomaly regions in the illumination anomaly distribution map, the confidence score of the corresponding anchor point is weighted with a bonus on top of its initial value. Weak interface anchor point candidates whose confidence scores exceed the strong interface judgment threshold are automatically upgraded to strong interface anchor point candidates. The strong interface judgment threshold is the mean of the initial confidence score distribution across the entire sequence plus one standard deviation to accommodate differences in the absolute confidence levels of anchor points in different strata. The structural anchor point sequence is managed using a dual index of frame number and initial depth marker. The confidence score field undergoes normalization across the entire sequence, unifying the confidence score range of each anchor point to 0 to 1 after normalization to support horizontal comparisons of anchor point quality at different depths.

[0019] Step S120: Based on the structural anchor point sequence, perform cross-frame pose transfer to generate an initial pose sequence, perform borehole wall visual measurement on the borehole inner wall image sequence to generate a borehole diameter deviation map, and construct a borehole point cloud map by geometric constraint fusion positioning through the initial pose sequence and the borehole diameter deviation map.

[0020] Specifically, an initial pose sequence is generated by cross-frame pose transfer based on the structural anchor point sequence. Each anchor point in the structural anchor point sequence carries a frame number, hole wall coordinates, and a confidence level field. Cross-frame pose transfer is based on the common-view anchor points in adjacent frames. The spatial coordinate difference of the common-view anchor points in the coordinate systems of the two frames is defined as the estimation basis for the inter-frame pose increment. Strong interface anchor points in the structural anchor point sequence have high confidence and participate in cross-frame matching first. Weak interface anchor points are included in the calculation as supplementary matching points when the number of strong interface anchor points is insufficient, to ensure that pose transfer is not interrupted in sparse sections with features such as homogeneous mudstone. When the number of available common-view anchor points between adjacent frames is lower than the minimum matching threshold, the pose increment is estimated by extrapolation of the motion model of the previous effective frame. Extrapolation frames are marked with extrapolation markers in the initial pose sequence. When the number of consecutive extrapolation frames exceeds the threshold, the robot is triggered to decelerate to wait for texture recovery. Cross-frame pose transfer uses the pose of the first frame as the origin of the world coordinate system. The initial pose sequence is generated by accumulating pose increments frame by frame. Each element of the initial pose sequence contains a translation vector and a rotation quaternion. The frame number strictly corresponds to the frame number of the borehole inner wall image sequence. During borehole advancement, the slight lateral displacement of the robot due to the lateral constraints of the borehole wall introduces a systematic deviation into the pose increment. After accumulating more than a set number of frames, cross-frame pose transfer performs a phased pose correction on the initial pose sequence using strong interface anchors from the structural anchor point sequence. The correction amount is calculated by minimizing the anchor point reprojection error.

[0021] In some embodiments, the step of performing borehole wall visual measurement on the borehole inner wall image sequence to generate a borehole diameter deviation map includes: performing frame-by-frame borehole wall contour detection on the borehole inner wall image sequence to generate a frame-by-frame borehole wall contour set; performing ellipse fitting calculation on the frame-by-frame borehole wall contour set to generate a frame-by-frame ellipse fitting parameter set; extracting the major and minor axis ratios of the frame-by-frame ellipse fitting parameter set to identify the soft and hard distribution areas of the wall surface; and performing stress direction calibration based on the soft and hard distribution areas of the wall surface and the frame-by-frame borehole wall contour set to generate a borehole diameter deviation map.

[0022] A frame-by-frame borehole wall contour detection is performed on the borehole inner wall image sequence to generate a frame-by-frame borehole wall contour set. Each frame of the borehole inner wall image sequence exhibits a significant brightness gradient in the annular region between the borehole wall and the drill pipe. This gradient band corresponds to the edge position of the borehole wall. The frame-by-frame borehole wall contour detection is based on this brightness gradient and uses an illumination anomaly distribution map as an auxiliary reference to extract the borehole wall edge contour of each frame. The Canny operator is applied to the borehole inner wall image sequence frame by frame to extract the edge response map. The dual threshold settings of the Canny operator are adaptively determined based on the high percentile value of the grayscale histogram of each frame to avoid missing edge detection when the brightness of the borehole wall changes with depth due to a fixed threshold. Edge chains extending continuously along the annular direction in the edge response map are identified as borehole wall contour candidates. When the arc length of a candidate chain is less than 60% of the frame perimeter, it is identified as an incomplete contour, triggering gap filling based on the illumination anomaly distribution map. The filling area uses the anomaly boundary at the corresponding position in the illumination anomaly distribution map as the surrogate contour of the gap segment. When drilling reaches the water-rich fracture zone, the mud film on the borehole wall obscures the rock texture boundaries. The obscuration range varies with mud concentration and fracture opening width. The outline of the obscured area is completed by depth interpolation of the outlines in adjacent frames. If continuous obscuration exceeds a threshold number of frames, the robot's progress is paused to allow the mud to settle before re-acquiring. The interpolated outline segments are marked with interpolation markers in the frame-by-frame borehole wall outline set. The distribution of these interpolation markers is recorded as a separate field for weight reduction processing during subsequent ellipse fitting. The frame-by-frame borehole wall outline set uses the frame number as the primary key. The borehole wall outline of each frame is stored in a pixel coordinate linked list. The starting point of the linked list is fixed at the top of the frame image to ensure consistency of azimuth angles between frames. The point spacing of the coordinate linked list is uniformly resampled to a fixed value based on the outline arc length to ensure consistent outline point density across frames.

[0023] Elliptic fitting calculations are performed on the frame-by-frame aperture wall contour set to generate a frame-by-frame elliptic fitting parameter set. The pixel coordinate list of the aperture wall contour in each frame of the aperture wall contour set is fitted using a least-squares elliptic fitting algorithm. The fitting objective is to solve for four parameters: the ellipse center coordinates (x0, y0), the major axis length a, the minor axis length b, and the major axis azimuth angle θ. The ellipse center coordinates (x0, y0) are used for aperture center offset quantization. These four parameters together constitute the elliptic fitting parameter record for that frame. The general algebraic form of least-squares elliptic fitting is Ax² + Bxy + Cy² + Dx + Ey + F = 0, where x and y are the contour pixel coordinates, and A, B, C, D, E, and F are the elliptic algebraic coefficients to be solved. The constraint condition is 4AC - B² = 1 to ensure the fitting result is an ellipse and to eliminate scale uncertainty. After substituting the pixel coordinates of each contour, an overdetermined system of equations is constructed, and the minimum residual solution is solved using singular value decomposition. The weight of the contour segments obtained by frame-by-frame borehole wall contour interpolation is reduced to 0.4 during ellipse fitting to minimize the skewness of the interpolation segments on the fitting results. The fitting residual is defined as the mean algebraic distance from each contour pixel to the fitted ellipse. Frames with residuals exceeding a threshold are identified as contour anomalous frames. Contour anomalous frames typically appear at borehole wall collapses or large joint openings. The ellipse fitting parameters of such frames are marked with anomalies in the frame-by-frame ellipse fitting parameter set. Anomaly-marked frames are not included in the statistical calculation of the λ sequence when identifying the soft and hard distribution areas of the borehole wall. The major-minor axis ratio λ = a / b serves as the core quantitative indicator of the ellipticity degree of each frame and is stored along with the fitting parameters. A larger λ value indicates more severe borehole wall ellipticity in that frame. The λ value in shallow soft rock sections is usually significantly higher than that in deep hard rock sections. This depth-dependent λ distribution pattern is directly reflected in the frame-by-frame ellipse fitting parameter set.

[0024] The major-minor axis ratio of the frame-by-frame elliptic fitting parameter set is extracted to identify the soft and hard distribution zones of the formation wall. The λ values ​​of each frame in the frame-by-frame elliptic fitting parameter set, arranged along the depth direction, show a trend corresponding to the soft and hard distribution of the formation. When the borehole traverses the mudstone and sandstone interlayers, the mudstone section has a consistently high λ value due to its strong plastic deformation capacity, while the sandstone section has a high λ value close to 1.0 due to its brittle rock mass and strong lateral pressure resistance. These two types of sections form alternating high-low ratio zones on the depth profile, and the alternation frequency is positively correlated with the formation cycle thickness. The λ sequence of the frame-by-frame elliptic fitting parameter set is smoothed using the mean of a sliding window with a width of W frames. Smoothing suppresses the interference of single-frame fitting noise on the zone boundary location. After smoothing, sections where λ continuously exceeds the judgment threshold λ_th are identified as soft wall distribution zones, and sections continuously below λ_th are identified as hard wall distribution zones. λ_th is determined by adding 0.5 times the standard deviation to the global mean of the λ sequence to accommodate the differences in the absolute ellipticization level of different borehole formations. The boundaries of the soft and hard distribution zones are located at the intersection of the smoothed λ sequence and λ_th. The depth coordinates of the intersection are accurate to the resolution of the propulsion encoder through linear interpolation. The accuracy of boundary positioning directly affects the accuracy of depth assignment in subsequent stress direction calibration. Isolated high λ segments within the soft wall distribution zone with a consecutive frame number below the minimum segment length threshold are considered noise and not separately classified. These are usually caused by boreholes penetrating thin soft interlayers or local rough borehole walls and are incorporated into the adjacent hard wall distribution zone. The depth intervals of the soft and hard distribution zones on the wall surface are associated with the frame number of the frame-by-frame elliptic fitting parameter group through initial depth marking.

[0025] A borehole diameter deviation map is generated by stress direction calibration based on the soft / hard distribution zone of the borehole wall and the frame-by-frame borehole wall profile set. The major axis azimuth θ of each frame in the soft wall distribution zone reflects the direction of the maximum horizontal principal stress in the formation. The angle between the borehole advance direction and the direction of the maximum principal stress determines the azimuth distribution of the borehole wall ellipticization. A systematic deflection of θ values ​​at different depths indicates a change in the formation stress direction with depth. Within the soft wall distribution zone, the θ values ​​of each frame generally remain stable in the depth direction. Frames with an θ difference exceeding a threshold are identified as locally disturbed frames. The azimuth angle of locally disturbed frames is substituted with the weighted average of adjacent stable frames, with the weights taken as the normalized value of the stable frame's λ value to ensure that frames with a higher degree of ellipticization have a higher weight. The aperture deviation map focuses on the radial deviation distribution of the aperture wall profile relative to the standard circle in each frame of the aperture wall profile set. The radial deviation is defined as the difference between the distance r(φ) from each profile point to the frame center and the standard aperture reference value r0: Δr(φ) = r(φ) - r0, where r0 is the average value of the stable aperture of the borehole inlet section, and φ is the azimuth angle of the profile point. Δr(φ) > 0 corresponds to the expansion direction, and Δr(φ) < 0 corresponds to the contraction direction. In the hard wall distribution area, the Δr(φ) distribution of each frame is close to zero and spatially uniform. In the soft wall distribution area, the Δr(φ) of each frame exhibits an alternating positive and negative elliptical distribution along the direction of the maximum principal stress. The expansion extreme value direction is aligned with the major axis azimuth angle θ of the stress direction. When the difference between the two azimuth angles exceeds 15°, the stress calibration result of that frame is re-verified. The aperture deviation map also includes the stress direction calibration results of each frame, stored as the major axis azimuth field. The depth coordinate system inherits the initial depth mark of the hole wall profile set of each frame. The radial deviation distribution corresponding to the interpolated frame in the aperture deviation map is supplemented with interpolation marks.

[0026] In some embodiments, the step of constructing a borehole point cloud map by geometric constraint fusion positioning using the initial pose sequence and the aperture deviation map includes: extracting local geometric feature points from the aperture deviation map to generate a set of geometric anchor points; performing spatiotemporal alignment of the set of geometric anchor points and the initial pose sequence to generate pose-geometric association pairs; performing constraint residual construction on the pose-geometric association pairs to generate a geometric constraint residual field; and performing borehole wall geometric constraint pose optimization based on the geometric constraint residual field to construct the borehole point cloud map.

[0027] Local geometric feature points are extracted from the aperture deviation map to generate a geometric anchor point set. The frame-by-frame mean sequence of Δr(φ) along the depth direction in the aperture deviation map exhibits local extrema at abrupt changes in borehole wall morphology. When the borehole crosses a lithological interface, the difference in compressive strength between the upper and lower rock strata causes abrupt changes in borehole diameter at the interface. The gradient jump between the diameter expansion extrema of the soft rock segment above the interface and the near-zero deviation of the hard rock segment below the interface forms stable and reproducible radial deviation extrema points in the aperture deviation map. The spatial location of these extrema points does not drift with repeated robot advancements, providing a reliable basis for use as geometric anchor points. The radial deviation extrema points of each frame's contour are extracted from the aperture deviation map frame by frame. The extraction method involves performing a first-order difference on the frame-by-frame mean sequence of Δr(φ) along the depth direction. Positions where the absolute value of the difference exceeds a multiple of the global standard deviation threshold are identified as candidate extrema points. Adjacent duplicate extrema are removed from the candidate points along the depth direction using a minimum spacing threshold, and the candidate points with the largest absolute difference values ​​in each local neighborhood are included in the geometric anchor point set. The azimuth coordinate φ and depth coordinate of an extreme point together define its three-dimensional position in the borehole wall coordinate system. The accuracy of the three-dimensional position is constrained by the spatial resolution of the borehole diameter deviation map. Extreme points in depth segments with stable stress direction calibration results and high λ values ​​in the borehole diameter deviation map have high reliability. Extreme points in depth segments with large azimuth fluctuations or carrying interpolation marks are marked with low reliability in the geometric anchor point set. The weight of anchor points with low reliability marks is reduced to 0.3 when constructing the constraint residuals in the subsequent process.

[0028] Spatiotemporal alignment of the geometric anchor point set with the initial pose sequence generates pose geometric association pairs. Each anchor point in the geometric anchor point set is associated with the pose record of the corresponding frame in the initial pose sequence using a dual index of frame number and depth coordinates. Spatiotemporal alignment uses the frame number of the anchor point as the primary key to search for the pose of the same frame in the initial pose sequence, binding the 3D position of the anchor point in the borehole wall coordinate system with the pose in the world coordinate system of the initial pose sequence to form a pose geometric association pair. The binding of the pose geometric association pairs is verified by coordinate system transformation. After the 3D coordinates of the anchor point are projected to the world coordinate system through the same frame pose transformation, the distance between the projected point and the projected point of the same anchor point in the adjacent frame should be less than a reasonable multiple of the pose increment magnitude. Binding pairs exceeding the reasonable range are judged as spatiotemporal alignment anomalies and are discarded. The pose geometric association pairs corresponding to frames carrying extrapolation markers in the initial pose sequence have reduced weights during spatiotemporal alignment. Within the long extrapolation interval where the borehole traverses homogeneous mudstone, the pose estimation error accumulates frame by frame. The pose geometric association pairs within this interval are decremented exponentially by the number of consecutive extrapolated frames k as a parameter. k=w0·e^(-αk) dynamically reduces the weight, where α is the attenuation coefficient, w0 is the baseline weight of the non-extrapolated frame, and α is calibrated based on the typical number of consecutive extrapolated frames in the homogeneous section of the borehole formation. The pose geometric association pairs generated after spatiotemporal alignment of the anchor points marked with low reliability in the geometric anchor point set synchronously inherit the low reliability markers. The weight fields of each record in the pose geometric association pair are uniformly normalized after spatiotemporal alignment. The sum of the normalized weights equals the total number of valid records to ensure the dimensional consistency of subsequent residual weighting.

[0029] The pose geometry association constructs a geometric constraint residual field by performing constraint residual construction. The three-dimensional position of each anchor point in the hole wall coordinate system of the pose geometry association is then determined by the corresponding frame pose T. k After transforming the projection to the world coordinate system, the projection point p is obtained. k w =T k ·p k c The projection point p of the same anchor point as the adjacent frame k+1 w The difference is defined as the constraint residual vector e k =p k w -p k+1 w If the anchor point is not visible in adjacent frames, the constraint residual component is set to zero and does not participate in the optimization. k The modulus reflects the degree of inconsistency between the initial pose sequence's pose estimation and the borehole geometry at that frame. When the drilling robot advances through homogeneous strata, the cumulative drift of the initial pose sequence is slow, and the constraint residuals of the pose-geometric correlation pairs remain low in homogeneous sections. In lithological interface sections with dense structural anchor point sequences, the cross-frame matching accuracy is high, and the constraint residuals are also relatively low. When the borehole traverses long-distance homogeneous mudstone sections and enters sandstone layers, the drift error accumulated by extrapolation is concentrated and released in the first frame where matchable features are re-acquired. The constraint residuals in this frame are usually significantly higher than those in adjacent frames, forming a peak response in the geometric constraint residual field. The geometric constraint residual field is represented by the constraint residual vector e of each frame. k Arranged according to frame number, the dimension of the residual vector is consistent with the number of effective anchor points in each frame of the pose geometric association pair. The residual components corresponding to the low reliability marker anchor points participate in subsequent optimization with a weight of 0.3. The spatial position and depth of the high residual peak are independently recorded in the peak coordinate field in the geometrically constrained residual field to support the delineation of key iteration intervals for pose optimization.

[0030] A borehole point cloud map is constructed using geometrically constrained pose optimization based on the geometrically constrained residual field. High residual frames in the geometrically constrained residual field correspond to locations with significant initial pose sequence estimation deviations. Borehole wall geometrically constrained pose optimization aims to minimize the global weighted mean of the geometrically constrained residual field. The pose increment of each frame in the initial pose sequence is used as the optimization variable. The pose of each frame is gradually adjusted through Levenberg-Marquardt iteration until the geometrically constrained residual field converges. The convergence criterion is that the change in the global weighted mean of the residual between two consecutive iterations is below a threshold. When the borehole traverses multiple lithological interfaces, the geometrically constrained residual field exhibits a multi-peak distribution. Pose optimization prioritizes frames with high residual peaks. The pose increment adjustment of frames surrounding each peak is determined by the proportional coefficient of the peak residual. The pose increment adjustment of low residual frames far from the residual peak is subject to a strict upper bound constraint to maintain pose stability in low-drift segments. The pose update of frames corresponding to strong interface anchor points in the structural anchor point sequence is also constrained by an upper bound during optimization iterations to prevent over-correction of high-confidence anchor point frames. After optimization, the pose of each frame is projected onto a unified world coordinate system by back-projecting the depth map point cloud generated from each frame of the borehole inner wall image sequence through the frame-by-frame borehole wall contour set and the borehole diameter deviation map. The geometric consistency of the point clouds in the overlapping area of ​​each frame is measured by the average point-to-point distance. Overlapping areas with an average value exceeding a threshold trigger local point cloud registration to eliminate residual deviations. After registration, the point clouds of each frame are merged to form a borehole point cloud map. The borehole point cloud map is desampled using voxel grids to remove redundant points. The voxel size matches the spatial resolution of the borehole inner wall image sequence. Within each voxel, the point closest to the voxel center is retained as a representative point. The borehole point cloud map is stored in the form of a three-dimensional point set in the world coordinate system. Each point is attached with three attributes: frame number, depth coordinates, and radial deviation value of the corresponding frame in the borehole diameter deviation map.

[0031] Step S130: The cumulative drift in the initial pose sequence is decomposed directionally to identify axial drift and radial drift. The axial drift is corrected by sensor absolute depth anchoring to obtain the axial correction amount. The radial drift and the weak texture boundary map are periodically matched to obtain the radial constraint identifier. The axial correction amount and the radial constraint identifier are used to generate the directional correction command.

[0032] Specifically, the cumulative drift in the initial pose sequence is directionally decomposed to identify axial and radial drift. The cumulative drift in the initial pose sequence is formed by the superposition of frame-by-frame errors from the pose transfer across frames. The depth estimation deviation in the drilling direction and the lateral displacement deviation perpendicular to the drilling direction are fundamentally different in their physical causes and correction methods. Axial drift corresponds to the estimation error in the depth direction, while radial drift corresponds to the lateral offset error perpendicular to the drilling axis. The two types of drift must be decomposed independently and then corrected accordingly. The translation vector of each frame in the initial pose sequence is decomposed into axial and radial components in the drilling coordinate system. The frame-by-frame cumulative value of the axial component constitutes the axial drift sequence, and the frame-by-frame cumulative value of the radial component constitutes the radial drift sequence. Axial drift increases rapidly with frame count in conglomerate sections where the robot advances quickly due to frequent failures in visual odometry feature matching, while its growth is slower in textured coarse sandstone sections where feature constraints inhibit its increase. Radial drift grows slowly in homogeneous mudstone with a simple borehole wall structure due to lateral constraints, but exhibits a step-like increase in fractured sections due to increased lateral robot displacement. The difference in the growth patterns of the two types of drift in the depth direction reflects the different effects of alternating borehole lithology on robot pose estimation. After directional decomposition, the axial drift sequence and the radial drift sequence are stored independently. Frames carrying extrapolation tags are simultaneously appended to both types of sequences for weight reduction during correction.

[0033] In some embodiments, the step of performing sensor absolute depth anchoring correction on the axial drift to obtain the axial correction amount includes: extracting frame-by-frame depth increments from the axial drift to generate a depth increment sequence; performing sliding window steady-state estimation on the depth increment sequence to generate a depth reference increment sequence; performing frame-by-frame deviation detection based on the depth increment sequence and the depth reference increment sequence to generate a valid frame marker; and performing absolute depth anchoring on the axial drift using the valid frame marker to obtain the axial correction amount.

[0034] A depth increment sequence is generated by extracting frame-by-frame depth increments from the axial drift. The difference between the axial components of adjacent frames in the axial drift sequence is defined as the frame-by-frame depth increment for that frame. The frame-by-frame depth increment reflects the estimated advance distance of the robot within the acquisition interval between adjacent frames. Under normal advance conditions, the frame-by-frame depth increment should be consistent with the difference in the advance encoder reading. When the drilling robot traverses a soft rock collapse section, the advance resistance drops sharply, and the robot's instantaneous acceleration causes a brief jump in the actual advance recorded by the encoder. The frame-by-frame depth increment of the axial drift shows an abnormal amplitude synchronously at the corresponding frame. When the drill bit wears down in a hard rock section, leading to a decrease in advance efficiency, the frame-by-frame depth increment shows a systematic downward trend. The two types of anomalies have different forms in the depth increment sequence: the former is a spike pulse, and the latter is a slowly varying bias. The extraction of the frame-by-frame depth increment is calculated using the first-order difference of the axial drift sequence. The difference operation makes the depth increment sequence one element less than the axial drift sequence. The missing part is filled into the first frame with the estimated advance speed of the robot entering the hole. After filling, the depth increment sequence is the same length as the axial drift sequence and is arranged by frame number to form the depth increment sequence. The increment values ​​of frames carrying extrapolation markers in the depth increment sequence are usually too large due to the lack of visual feature constraints. The extrapolation markers are retained in the depth increment sequence to reduce the weight during steady-state estimation. The large increment segments formed by consecutive extrapolated frames appear as a continuously high plateau region in the depth increment sequence, which can be distinguished from the single-frame spike pulses caused by velocity abrupt changes in morphology.

[0035] A sliding window steady-state estimation is performed on the depth increment sequence to generate a depth benchmark increment sequence. The depth increment sequence exhibits a low-variance, stable distribution in the stable section where the robot advances at a uniform speed, and a high-variance, fluctuating distribution in sections with abrupt changes in advance speed or extrapolation estimation. Intermittent acceleration in soft rock sections causes a brief plateau in the depth increment sequence, while decreased drilling efficiency in hard rock sections causes a persistently low depth increment sequence. The purpose of the sliding window steady-state estimation is to extract a representative advance speed benchmark from the distribution of each depth segment. The depth increment sequence is scanned frame by frame using a sliding window of fixed width. The depth increment values ​​within each window are averaged after removing the top and bottom 10% extreme values. The average value is used as the depth benchmark increment of the center frame of the window. Arranged by frame number, this constitutes the depth benchmark increment sequence. The removal of extreme values ​​ensures that abnormal advance speeds in a single frame do not affect the benchmark estimation. When drilling through fractured zones, the advance speed changes frequently, and the pruning ratio is appropriately increased in such sections to ensure the robustness of the benchmark estimation. The sliding window width is set to twice the number of frames required for the robot to complete a full borehole wall scan cycle at typical propulsion speeds. If the window is too narrow, short-cycle speed fluctuations are included in the baseline, causing instability. If the window is too wide, the actual speed trend is excessively smoothed. The width is calibrated based on typical drilling conditions. Frames carrying extrapolation markers in the depth increment sequence have their weight reduced to 0.3 when calculating the mean of their corresponding window. Without this weight reduction, the window mean is continuously inflated by the systematically large increment values ​​of the extrapolated segment, resulting in an artificially high overall depth baseline increment sequence in that segment. The depth coverage of the depth baseline increment sequence is consistent with that of the depth increment sequence.

[0036] For example, the step of generating valid frame markers by performing frame-by-frame deviation detection based on the depth increment sequence and the depth reference increment sequence includes: performing frame-by-frame synchronous alignment of the depth increment sequence and the depth reference increment sequence to generate a frame-level deviation sequence; extracting the steady-state deviation reference of the frame-level deviation sequence to generate a normal deviation range; performing deviation scanning on the frame-level deviation sequence and the normal deviation range to identify abnormal deviation frames; and generating valid frame markers by performing invalid frame marking based on the abnormal deviation frames.

[0037] A frame-by-frame synchronization alignment is performed between the depth increment sequence and the depth reference increment sequence to generate a frame-level deviation sequence. Both the depth increment sequence and the depth reference increment sequence are indexed by frame number, and the frame-by-frame synchronization alignment adheres to the principle of strict correspondence between frame numbers. The difference between the depth increment value of each frame and the corresponding depth reference increment value is defined as the frame-level deviation for that frame. These are arranged by frame number to form the frame-level deviation sequence. A positive value in the frame-level deviation sequence indicates that the actual depth advance in that frame exceeds the reference estimate, corresponding to a temporary increase in robot propulsion speed or over-matching of visual odometry features in textured coarse-grained sandstone sections, leading to an overestimation of depth. A negative value indicates that the actual depth advance is lower than the reference, corresponding to decreased propulsion efficiency when the drill bit encounters hard flint interlayers or a lack of features in homogeneous mudstone sections, leading to an underestimation of depth. The frame-level bias sequence simultaneously superimposes two types of components in the temporal domain: transient pulses caused by changes in propulsion conditions and slowly varying biases caused by systematic degradation of the visual odometry. The former appears when the propulsion speed changes abruptly at the sandstone-mudstone interface and usually recovers rapidly after only a few frames, while the latter accumulates slowly as the robot continuously traverses homogeneous strata. The temporal morphology of these two components in the frame-level bias sequence is significantly different. The bias values ​​of frames carrying extrapolation markers in the frame-level bias sequence are usually positive. Due to the lack of visual constraints in the extrapolation estimation, the depth increment is systematically overestimated. The distribution of extrapolation markers is retained as an independent field in the frame-level bias sequence, providing a distinguishing basis for calibrating the width of the normal bias interval.

[0038] The steady-state deviation benchmark of the frame-level deviation sequence is extracted to generate the normal deviation interval. The frame-level deviation sequence exhibits low-amplitude random fluctuations during the steady-state segment when the robot moves at a constant speed and the visual odometry is functioning normally. These fluctuations originate from camera exposure jitter, encoder quantization errors, and transient changes in propulsion resistance caused by minor protrusions in the borehole wall. The amplitude range of these normal fluctuations forms the basis for estimating the normal deviation interval. The steady-state deviation benchmark of the frame-level deviation sequence is calculated using the median of the entire sequence. The median is more robust to abnormal deviation frames than the mean, accurately reflecting the central trend of the sequence even when the borehole traverses multiple soft rock collapse zones, resulting in a high proportion of abnormal frames. In contrast, the mean is systematically skewed by high-amplitude abnormal frames in this situation. The normal deviation interval is defined as the steady-state deviation benchmark plus or minus a multiple threshold of the interquartile range of the frame-level deviation sequence. The half-width of the interval is denoted as σ. The interquartile range is insensitive to outliers and describes the normal fluctuation range of the sequence more robustly than the standard deviation. Its multiple coefficient is calibrated based on the speed fluctuation range under typical propulsion conditions of the drilling robot. When the multiplier coefficient is too small, normal velocity fluctuation frames are mistakenly labeled as invalid, resulting in insufficient effective frame density and excessive anchoring spacing for axial correction. When the multiplier coefficient is too large, abnormal real drift frames are missed, leading to unreliable frames being mixed into the anchoring reference. Both types of risks are addressed by jointly evaluating the effective frame rate of historical borehole data and anchoring error during multiplier coefficient calibration to determine a balance point. The upper and lower boundaries of the normal deviation interval are marked with horizontal reference lines in the frame-level deviation sequence. The interval width is dynamically adjusted according to the advancement conditions at different depths. The interval is appropriately widened when velocity fluctuations are large in soft rock sections and narrowed when advancing at a constant speed in hard rock sections. This dynamic adjustment is achieved by replacing the global interquartile range with the local interquartile range of the frame-level deviation sequence.

[0039] A deviation scan is performed on the frame-level deviation sequence and the normal deviation interval to identify abnormal deviation frames. The deviation scan uses the distance between the deviation value of each frame in the frame-level deviation sequence and the boundary of the normal deviation interval as the scan quantity. Frames whose deviation values ​​fall outside the normal deviation interval are included in the abnormal deviation frame candidate set. The normal deviation interval serves as a unified benchmark for frame-by-frame judgment in this scan. Isolated single-frame anomalies in the candidate set are usually caused by local rock cuttings in the borehole momentarily obstructing the camera's field of view. After the obstruction disappears, the next frame returns to normal. The identification of isolated single frames is based on the criterion that the two frames before and after the candidate frame are within the normal deviation interval. Systematic deviations that last for multiple frames correspond to continuous abnormalities in robot propulsion resistance or systematic degradation of visual odometry in a specific stratum. The difference in temporal continuity between the two types is used as a further distinction criterion. Isolated single-frame anomalies are downgraded in the final abnormal deviation frame judgment and do not trigger the entire segment marking. The deviation of the frame-level deviation sequence is defined as the ratio of the deviation value exceeding the normal deviation range by the magnitude Δd to the half-width σ of the range: ρ = Δd / σ. Frames with ρ > 1.0 are judged as abnormal deviation frames, and frames with ρ between 0.8 and 1.0 are marked as boundary frames. Boundary frames are retained with lower confidence during the generation of valid frame labels and are not directly judged as invalid. When the borehole traverses a water-bearing fracture zone, the robot's propulsion resistance changes abruptly and the borehole wall mud disturbance increases. In this section, the frame-level deviation sequence shows a concentrated anomaly with multiple consecutive frames where ρ exceeds the threshold. When the number of consecutive abnormal frames exceeds the threshold, it is judged as an entire abnormal interval. All frames within the entire abnormal interval are marked as abnormal deviation frames. The entire marking of abnormal deviation frames avoids the problem of discontinuous marking caused by the ambiguity of boundary frame classification when judging individual frames one by one.

[0040] Valid frame markers are generated based on invalid frame marking of abnormal deviation frames. Abnormal deviation frames correspond to frames with low depth estimation reliability in the initial pose sequence in terms of frame number. Frequent robot jolting during borehole advancement in loose gravel layers causes multiple consecutive frames to become abnormal deviation frames, while in dense limestone sections with uniform advancement, abnormal deviation frames are sporadically distributed. The density difference of abnormal frames in these two types of strata is directly reflected in the distribution of consecutive valid segment lengths in the valid frame markers. Abnormal deviation frames are marked as invalid frames one by one in the frame number list. Invalid frame markers are stored as binary fields: 1 indicates invalid, and 0 indicates valid. After all abnormal deviation frames have been traversed and marked, the invalid frame markers are inverted to obtain the valid frame markers. Boundary frames are recorded in the valid frame markers with a partial effective weight of 0.5. Their contribution weight is halved when participating in absolute depth anchoring, thus neither completely excluding the depth information of boundary frames nor diminishing their impact on the anchoring results. The axial correction amount of sections with concentrated boundary frame distribution is automatically downweighted during interpolation to reflect the low anchoring reliability of these sections. The start and end frame numbers of consecutive valid segments in the valid frame marker are indexed separately. Intervals with shorter consecutive valid segments correspond to depth segments with frequent formation disturbances. When the borehole crosses a fractured zone, the length of the consecutive valid segments may only be a few frames. When anchoring axial drift, such intervals are compensated by interpolation of the anchoring results of adjacent long consecutive valid segments. The axial correction amount of the interpolated interval is marked with a low confidence level to distinguish it from the directly anchored interval.

[0041] The axial correction is obtained by absolute depth anchoring of the axial drift using valid frame markers. The valid frames selected by the valid frame markers correspond to positions in the initial pose sequence where the depth estimation is relatively reliable. The absolute depth anchoring uses the encoder absolute depth reading at the valid frames selected by the valid frame markers as a reference value. The cumulative depth estimate of the axial drift sequence at the corresponding frame is aligned with the encoder reading. The alignment residual ε of the k-th frame is... k Defined as the difference between the visually estimated cumulative depth of that frame and the encoder's absolute depth reading. Alignment residual ε for each valid frame. kThe axial correction is extended to the entire sequence via linear interpolation along the frame number direction. The axial correction between adjacent valid frames is linearly transitioned by the residuals at the corresponding endpoints. The linear interpolation assumes that drift accumulates uniformly between adjacent valid frames. This assumption holds true during the robot's uniform speed advance. When the borehole traverses dense limestone and other formations with uniform resistance, the advance speed is stable, and the drift accumulation between adjacent valid frames is close to linear, resulting in a small interpolation error. However, during the soft rock collapse section, when the advance speed changes abruptly, the drift accumulation is nonlinear, and the interpolation error is larger. In this case, the interpolation error is partially suppressed by the weighting mechanism of the boundary frames. The interpolation span of the axial correction in sparsely distributed valid frames is large. This situation is common in depth sections where the borehole traverses long distances of homogeneous mudstone, resulting in a continuous lack of visual features and a dense appearance of consecutive extrapolated frames. The interval between adjacent valid frames can reach tens of frames. The larger the interpolation span, the lower the anchoring confidence. The axial correction in the corresponding interval is marked with a low confidence level. The axial component of the low confidence interval is weighted down when the directional correction command is generated. Each element of the axial correction sequence corresponds to the depth correction value of a frame in the axial drift sequence. The sign of the correction value indicates whether the depth estimate of that frame is too high or too low. The axial correction is managed with the frame number as the primary key.

[0042] In some embodiments, obtaining radial constraint identifiers by periodically matching the wall structure based on the radial drift and the weak texture boundary map includes: extracting the boundary spacing of the weak texture boundary map to generate a periodic feature spectrum of the wall structure; performing spectral analysis on the radial drift to generate a drift frequency sequence; performing frequency domain correlation matching between the drift frequency sequence and the periodic feature spectrum of the wall structure to identify a periodic correlation structure layer; and performing radial drift source tracing and localization based on the periodic correlation structure layer to generate radial constraint identifiers.

[0043] The boundary spacing of weakly textured boundary maps is periodically extracted to generate a periodic feature spectrum of wall structure. The spacing distribution of weakly textured boundary lines in the depth direction of each frame of the weakly textured boundary map reflects the periodicity of stratigraphic texture alternation. When homogeneous mudstone and coarse-grained sandstone are periodically interbedded, the spacing of the weakly textured boundary lines exhibits a stable repeating pattern. This pattern corresponds to the thickness of the sedimentary cycle and forms a significant periodic peak in the periodic feature spectrum of wall structure. The amplitude of the periodic peak is positively correlated with the lithological contrast of the cycle. The contrast of mudstone and conglomerate interbedded layers is higher than that of mudstone and siltstone interbedded layers, and the former forms a sharper main peak in the periodic feature spectrum of wall structure. In fault zones or mylonite sections with strong tectonic disturbance, the boundary line spacing is chaotic, and the amplitude of the periodic feature spectrum of wall structure in the corresponding frequency band is low and scattered, which does not affect the identification of the main period. The depth coordinate difference between adjacent boundary lines in a weakly textured boundary map constitutes the boundary spacing sequence. The depth coverage of the weakly textured boundary map determines the frequency resolution of the periodic characteristic spectrum of the wall structure. The boundary spacing sequence exhibits periodic fluctuations when arranged along the depth direction. The fluctuation amplitude reflects the consistency of the cycle thickness. Uniform thickness interlayers have low fluctuation amplitudes, while wedge-shaped sequences with gradually varying thickness show monotonically changing fluctuation amplitudes with depth. The boundary spacing sequence is used to calculate the spectrum using discrete Fourier transform. Frequency components with amplitudes exceeding a multiple of the global mean threshold are identified as significantly periodic structural frequencies. The structural frequencies and their amplitudes are arranged in ascending order to form the periodic characteristic spectrum of the wall structure. The physical depth period corresponding to each structural frequency is converted based on the resolution of the propulsion encoder.

[0044] For example, the step of performing spectral analysis on the radial drift to generate a drift frequency sequence includes: performing a discrete Fourier transform on the radial drift to generate a drift spectrum; extracting significant peaks from the drift spectrum to generate a peak frequency set; constructing an integer ratio relationship topology on the peak frequency set to identify harmonic association links and generate interference harmonic markers; and suppressing associated frequency bands on the drift spectrum based on the interference harmonic markers to generate a drift frequency sequence.

[0045] A discrete Fourier transform (DFT) was performed on the radial drift to generate a drift spectrum. The periodic components of the radial drift sequence in the depth direction originate from two mechanisms: one is the periodic disturbance exerted laterally on the borehole wall by the mechanical vibration of the robot's rotary transmission mechanism. This disturbance intensifies when the borehole traverses hard quartz sandstone, and appears in the drift spectrum as power frequencies and their harmonics, with the power frequency determined by the rotational speed of the transmission mechanism. The other is the regular fluctuation in the lateral estimation of the robot's visual odometry caused by the periodic changes in the stratigraphic texture. This component is significant when the borehole traverses thin interbedded sandstone and mudstone, with the frequency corresponding to the reciprocal of the interbedded cycle thickness, and the amplitude increasing with the increase of the interbedded contrast. Before applying the DFT, the radial drift sequence was weighted with a Hanning window. The cosine gradient characteristic of the Hanning window effectively suppresses spectral leakage caused by discontinuities at the beginning and end of the sequence. Compared with a rectangular window, it has a stronger ability to suppress sidelobes near strong peaks, ensuring that weak-amplitude stratigraphic structure components are not submerged by the sidelobes of strong-amplitude mechanical interference peaks. The mean of the weighted radial drift sequence is removed to eliminate the interference of the DC component on the low-frequency band. The mean corresponds to the average lateral offset trend of the robot during the entire propulsion process. This trend belongs to ultra-low frequency drift and is not a target component of periodic analysis. The amplitude spectrum of the discrete Fourier transform results is arranged in ascending order of frequency to form a drift spectrum. The frequency resolution is equal to the ratio of the radial drift sequence sampling rate to the sequence length. The amplitude of the drift spectrum is displayed on logarithmic coordinates to simultaneously present the details of strong and weak peaks. In the drift spectrum, the high-amplitude power frequency peaks caused by mechanical vibration and the low-amplitude low-frequency peaks caused by geological structures are clearly distinguishable on logarithmic amplitude coordinates.

[0046] Significant peaks are extracted from the drift spectrum to generate a peak frequency set. Frequency points in the drift spectrum with local maximum amplitude exceeding a multiple of the neighborhood mean are identified as significant peaks. Significant peak extraction is achieved through frequency-by-frequency scanning using a peak detection algorithm, with a scan step size of half the frequency resolution to ensure no narrowband peaks are missed. Adjacent peaks with a spacing less than the minimum resolution threshold are merged into a single peak to avoid peak splitting caused by spectral leakage. The frequency of the merged peak is the weighted average of the two peak amplitudes, with the larger of the two peak values ​​being the maximum amplitude. The power frequency vibration of the drilling robot's transmission mechanism typically manifests as a single, highest-amplitude peak in the drift spectrum. In hard, highly abrasive sandstone and conglomerate sections, the power frequency peak amplitude increases significantly as drill bit wear accelerates. The increasing trend of the power frequency peak amplitude during continuous drilling can serve as an indirect indicator of the deterioration of the transmission mechanism's health. Drift components caused by the periodicity of the geological structure have relatively low amplitudes and are concentrated in the low-frequency range. The low-frequency component amplitudes are most stable when the borehole traverses equal-thickness interbedded sections. When traversing gradually thick sequence bedding, the peak frequency is concentrated in the low-frequency peaks, which shift towards even lower frequencies with depth. The frequency values ​​and amplitudes of each significant peak are arranged in descending order of amplitude to form a peak frequency set. The number of records in the peak frequency set is determined by the total number of peaks whose amplitudes exceed the threshold of the global noise floor. The noise floor is estimated by the mean amplitude of the peakless intervals in the drift spectrum. Frequency bands in the drift spectrum with amplitudes lower than the global median are determined as peakless intervals.

[0047] The peak frequency set is topologically constructed using integer ratio relationships to identify harmonic association links and generate interference harmonic markers. Peak values ​​caused by mechanical vibration within the peak frequency set appear as the fundamental frequency and its integer multiples of harmonics. The fundamental frequency is determined by the rotational speed of the transmission mechanism. The amplitude of higher-order harmonics attenuates with increasing order, and the amplitude attenuation law is typically approximated as a negative power function of the order. When drilling through soft rock, the transmission mechanism rotates at higher speeds, resulting in a higher fundamental frequency. In hard conglomerate sections, as the rotational speed decreases, the fundamental frequency shifts to lower frequencies, and the frequency distribution of harmonic links shifts accordingly. The fundamental frequency shift of harmonic links at different depths reflects the actual operating rotational speed changes of the transmission mechanism in different rock types. The frequency ratios between peak pairs within the peak frequency set are approximated using rational numbers. A frequency ratio deviation below a tolerance threshold is considered a harmonic association pair. The tolerance threshold is set as a multiple of the frequency resolution to accommodate spectral estimation errors. A directed graph is constructed between harmonic correlation pairs according to the fundamental frequency-harmonic direction. The set of peaks connected in the directed graph constitutes a harmonic correlation link. The link length is measured by the number of nodes contained. Harmonic correlation links with a length exceeding 2 are identified as sources of mechanical interference. All peaks within the link are marked as interfering harmonics in the peak frequency set. The interfering harmonic markers are generated by arranging the frequency values ​​of each interfering peak and the fundamental frequency of its respective link in ascending order. Isolated peaks in the peak frequency set that are not included in any harmonic correlation link are retained as non-harmonic peaks. Non-harmonic peaks correspond to non-harmonic disturbances caused by the periodicity of the formation structure or the contact between the robot and the borehole wall. In the frequency domain correlation matching, they are further distinguished by the periodic characteristic spectrum of the wall structure. Successfully matched non-harmonic peaks are classified as formation structure components, while unmatched isolated peaks are classified as contact disturbances and recorded separately.

[0048] Based on interference harmonic markers, the drift spectrum is correlated with frequency bands to generate a drift frequency sequence. The frequencies of each interference harmonic peak in the interference harmonic markers and their adjacent frequency bands are subjected to amplitude suppression in the drift spectrum. The suppression range extends half a width to both sides of each interference peak frequency, with the half-width being half the distance between adjacent non-interference peaks to avoid the suppression range encroaching on adjacent effective peak frequency bands. When two interference peaks are close together, the suppression intervals are merged into a continuous segment for unified processing. The amplitude of the interference frequency band marked by the interference harmonic markers in the drift spectrum is reduced to the noise floor level by multiplying it by a suppression coefficient. The suppression coefficient is determined by the reciprocal of the ratio of the interference peak amplitude to the noise floor, ensuring that the suppressed interference frequency band amplitude is close to the background noise without introducing negative values. The suppression coefficient takes its minimum value at the center of each interference peak and gradually changes to 1.0 towards the frequency band edge to avoid amplitude steps at the suppression zone boundary. After suppression by the associated frequency band, the drift spectrum is subjected to inverse discrete Fourier transform on the residual significant peaks. During reconstruction, only the peak frequency components that pass the significance threshold are retained. Frequency components below the threshold are considered background noise and do not participate in the reconstruction to prevent noise components from superimposing and introducing random disturbances after the inverse transform. The amplitude values ​​of each frame of the reconstructed signal are arranged according to the frame number to form a drift frequency sequence. The drift frequency sequence is a depth-domain radial drift signal after interference harmonic suppression, stored with the frame number as the primary key. When the borehole traverses thick homogeneous mudstone, the amplitude of the drift frequency sequence is close to zero. When traversing thin interbedded sandstone and mudstone, the amplitude shows a regular oscillation corresponding to the interbedded period. In the fracture zone, the amplitude of the drift frequency sequence is high and there is no obvious periodicity. The differences in the drift frequency sequence morphology of the three types of stratigraphic scenarios provide an intuitive signal basis for frequency domain correlation matching.

[0049] Frequency domain correlation matching was used to identify periodically associated structural layers by matching the drift frequency sequence with the periodic characteristic spectrum of the wall structure. When the period of the drift frequency sequence amplitude in the depth direction corresponds to the structural frequency in the periodic characteristic spectrum of the wall structure, it indicates that the periodicity of the radial drift originates from the periodic constraint of the borehole wall strata structure. This correspondence is most significant when the borehole traverses alternating layers of sandstone and mudstone of equal thickness. The main peak frequency of the drift frequency sequence in the alternating layer segment highly matches the structural frequency of the corresponding alternating layer thickness in the periodic characteristic spectrum of the wall structure, and the frequency deviation between the two is usually less than half a time of the frequency resolution. The power spectrum of the drift frequency sequence and the periodic characteristic spectrum of the wall structure are used to calculate the normalized cross-spectral amplitude at each structural frequency using a local frequency band window. The width of the local frequency band window is twice the frequency resolution to cover the inherent broadening of the spectrum estimation. Frequency points where the normalized cross-spectral amplitude exceeds a threshold are identified as frequency domain correlation matching points. The threshold is determined by adding twice the standard deviation to the mean of the normalized cross-spectral amplitude of the non-periodic frequency band in the periodic characteristic spectrum of the wall structure. The physical depth period corresponding to the matching point, after being converted by the encoder resolution, is compared with the boundary spacing sequence of the weak texture boundary map. Continuous intervals in the boundary spacing sequence that match the matching period are determined as periodic associated structure layers. The matching criterion is that the relative deviation between the boundary spacing and the matching period is less than a set proportion, and the number of consecutive matching boundary line pairs is not less than a minimum logarithmic threshold. The periodic associated structure layer is stored using the top and bottom boundary depth coordinates as the recording unit. The depth range of the periodic associated structure layer is determined by the intervals where the corresponding periodic boundary lines are concentrated in the weak texture boundary map. The interval boundaries are located by the local peak of the boundary line density. The depth coordinates where the density drops to 50% of the peak on both sides of the density peak are used as the top and bottom boundaries of the periodic associated structure layer. The positioning accuracy of the top and bottom boundaries inherits the depth resolution of the weak texture boundary map.

[0050] Radial drift tracing and localization based on periodically correlated structural layers generates radial constraint markers. Periodically correlated structural layers define the spatial range of radial drift constrained by the periodic stratigraphic structure in the depth direction. When drilling through interbedded sandstone and mudstone, these layers typically cover the entire interbedded section. However, in thick massive strata, they degenerate into scattered, thin-layered constraint intervals, resulting in significant differences in depth anchoring density for radial drift tracing in these two scenarios. Radial drift tracing and localization uses the top and bottom boundary depths of the periodically correlated structural layers as anchoring references. Drift frequency sequences are truncated within the depth range of each periodically correlated structural layer, and the mean amplitude of each subsequence reflects the systematic offset level of radial drift within that structural layer. The difference between the mean amplitude and the theoretical value of zero drift is defined as the radial drift offset of that structural layer. The sign of the offset indicates the dominant lateral displacement orientation of the robot. Periodically correlated structural layers with larger offsets correspond to depth segments with significant cumulative lateral displacement of the robot. The intervals between adjacent periodically associated structural layers lack structural constraints. The radial drift offset of these intervals is estimated by linear interpolation of the offsets of the adjacent periodically associated structural layers on both sides. The interpolation results are marked with an interpolation flag in the radial constraint identifier to distinguish directly constrained sections. The radial constraint identifier uses the depth interval of the periodically associated structural layer as the recording unit. The constraint weight is calculated by normalizing the normalized cross-spectral amplitude of the frequency domain correlation matching point. The higher the normalized cross-spectral amplitude, the more certain the correspondence between the radial drift of that layer and the stratigraphic structural period is. Recording units with higher constraint weights in the radial constraint identifier contribute more to the radial correction component when generating directional correction commands.

[0051] Axial correction commands are generated based on axial correction values ​​and radial constraint identifiers. The axial correction value sequence provides the correction values ​​in the depth direction for each frame, while the radial constraint identifier provides the lateral drift offset and constraint weights for each periodically associated structural layer. These two are merged into a frame-by-frame axial correction vector based on the frame number and depth coordinate mapping. The frame-by-frame depth correction values ​​of the axial correction values ​​are directly converted into the axial components of the axial correction commands. The axial components of low-confidence marked frames are multiplied by a weighting factor to reduce the impact of unreliable correction values. The weighting factor is determined by normalizing the inverse of the axial correction value interpolation span; a larger interpolation span corresponds to a lower anchoring density and lower correction confidence. The radial drift offset of each periodically associated structural layer in the radial constraint identifier is uniformly distributed across frames within the layer's depth range and converted into a frame-by-frame radial correction value. For structural layers with high constraint weights in the radial constraint identifier, the radial correction value for the corresponding frame is directly taken as the offset, while the radial correction value for structural layers with low constraint weights is multiplied by the constraint weight coefficient and reduced. When drilling through complex fracture zones, the constraint weights of periodically associated structural layers are generally low. The radial correction component in this depth range is reduced overall to avoid introducing new deviations with low-confidence correction values. The directional correction command consists of two components: a frame-by-frame axial correction component and a radial correction component. The frame number strictly corresponds to the initial pose sequence. The two components are stored independently in the directional correction command to support the directional application of pose recorrection.

[0052] Step S140: Perform frame-by-frame pose recalibration on the initial pose sequence according to the directional correction instruction to generate a calibrated pose sequence, and perform global consistency optimization on the borehole point cloud map based on the calibrated pose sequence to generate a navigation and positioning map.

[0053] In some embodiments, the step of performing frame-by-frame pose recalibration on the initial pose sequence according to the directional correction instruction to generate a calibrated pose sequence includes: extracting axial correction components and radial correction components based on the directional correction instruction to generate a directional correction vector; superimposing the poses of each frame in the initial pose sequence according to the directional correction vector to generate a frame-by-frame recalibrated pose; performing inter-frame smoothing constraint elimination correction on the frame-by-frame recalibrated pose to generate a smooth pose sequence; and performing global pose consistency verification based on the smooth pose sequence to generate a calibrated pose sequence.

[0054] Axial and radial correction components are extracted based on the directional correction command to generate a directional correction vector. In the directional correction command, the axial correction component of each frame is stored as a scalar value along the depth coordinate axis, while the radial correction component is stored as a two-dimensional planar vector perpendicular to the propulsion axis. Both components are independently indexed in the directional correction command using the frame number as the primary key. The axial correction component is extracted by directly reading the axial correction scalar value corresponding to each frame in the directional correction command. For frames marked with low confidence, the axial correction scalar value is multiplied by a weighting factor during extraction. The weighting factor is determined by normalizing the inverse of the interpolation span of the axial correction value; a larger interpolation span indicates a lower anchoring density for that frame, and a smaller corresponding weighting factor. The radial correction component is extracted based on the radial drift offset of the periodically associated structural layer to which each frame belongs. The offset is evenly distributed across frames within the depth range of that layer. The radial correction component of frames corresponding to structural layers with low constraint weights is reduced overall. When the borehole traverses a fault zone with complex structures, the constraint weights of the periodically associated structural layers are generally low, resulting in the largest reduction in radial correction components in these sections. The axial and radial correction components of each frame are concatenated to form a three-dimensional correction vector, which constitutes the directional correction vector of that frame. The directional correction vectors are arranged by frame number to form a set of directional correction vectors for the entire sequence. The dimensional structure of the vector set is consistent with the coordinate system of the translation vector of the initial pose sequence, and the coordinate axes of the two are aligned to ensure the physical correctness of the superposition operation.

[0055] The poses of each frame in the initial pose sequence are superimposed based on the directional correction vector to generate a frame-by-frame recorrected pose. The translation vector T of each frame in the initial pose sequence... k The directional correction vector c of the corresponding frame k Perform vector addition: T k ' = T k + c k The superimposed translation vector T k 'and the original rotation quaternion q k The combination is the frame-by-frame recalibrated pose for this frame, where q kThe rotation quaternion is the corresponding frame of the initial pose sequence. The rotation component remains unchanged during the superposition process. During drilling, the change in robot attitude angle is much smaller than the change in displacement. The impact of attitude angle estimation error on the point cloud projection accuracy is negligible at the current aperture scale. The superposition operation is performed in ascending order of frame number. Each frame is superimposed independently and without dependence on others, avoiding the error propagation introduced by frame-by-frame cumulative superposition. The translation vector of the frame carrying the extrapolation mark in the initial pose sequence has been reduced in weight by the directional correction vector before superposition to weaken the impact of extrapolation error. After superposition, the pose correction amplitude of the extrapolated frame in the frame-by-frame recorrected pose is lower than that of the effective frame. The extrapolated frame mark is retained in the frame-by-frame recorrected pose to identify the extrapolated segment boundary during subsequent smooth constraint. The radial component of the corresponding frame of the depth segment with added interpolation mark in the radial constraint mark has been reduced in weight during superposition. The interpolation mark information is synchronously inherited to the frame-by-frame recorrected pose. During the drilling process, the robot's lateral displacement increases sharply in the soft rock collapse section due to the sudden weakening of the borehole wall constraint. The radial component of the corresponding frame's directional correction vector is usually large. After superposition, the displacement correction amplitude is most significant in the frame-by-frame re-correction of pose in this type of frame. The abrupt change in the correction amplitude between adjacent frames is eliminated in the subsequent smoothing constraint step.

[0056] The frame-by-frame recalibrated pose is smoothed by inter-frame smoothing constraints to eliminate correction jumps and generate a smooth pose sequence. The independent superposition of the directional correction vectors in each frame may introduce amplitude jumps between adjacent frames, especially at the boundary between valid and low-confidence frames where the correction amplitude difference is significant. Directly superimposed frame-by-frame recalibrated poses exhibit abrupt pose changes at these boundaries, causing geometric misalignment at the point cloud stitching during borehole point cloud map projection. The smoothing constraint uses a Gaussian kernel to perform a weighted moving average on the translation vector sequence of the frame-by-frame recalibrated pose. The half-width of the Gaussian kernel is determined by the number of frames required for the robot to complete one pose correction cycle at a typical propulsion speed. If the kernel width is too small, correction jump suppression is insufficient; if the kernel width is too large, the true pose change is over-smoothed, leading to distortion of borehole orientation information. The smoothing operation applies positional constraints to the frames corresponding to the strong interface anchors in the structural anchor sequence. The translation vectors of the strong interface anchor frames are fixed in the smoothing calculation as the superposition of frame-by-frame recorrected poses and do not participate in the neighborhood weighted average, ensuring that the pose accuracy of high-confidence anchor frames is not skewed by neighborhood smoothing. After smoothing, the translation vectors of each frame are recombined with the original rotation quaternions to form the frame records of the smoothed pose sequence. The difference between the translation vectors before and after smoothing is stored as a smoothing correction amount in a scalar format in an additional field of the smoothed pose sequence, used to quantify the magnitude of the correction jump in each frame.

[0057] A corrected pose sequence is generated based on a global pose consistency check using a smoothed pose sequence. The smoothed pose sequence may contain residual inconsistencies that were not eliminated by local smoothing across the entire depth range. This manifests as discrepancies in the projection positions of the same borehole wall structural features between two frames that are far apart. When the borehole re-enters a conglomerate section after traversing a long distance of homogeneous mudstone, the drift error accumulated during extrapolation is released in a concentrated manner, making this type of cross-segment inconsistency most prominent. The global pose consistency check uses the reprojection positions of strong interface anchors in the structural anchor sequence under each frame of the smoothed pose sequence as a benchmark. The distribution range of the reprojection coordinates of the same strong interface anchor under different frames is measured by the area of ​​the covariance ellipse. Strong interface anchor segments with areas exceeding a threshold are identified as globally inconsistent intervals. Secondary smoothing correction is performed on the globally inconsistent intervals. Secondary smoothing uses the strong interface anchor frames at the beginning and end of the inconsistent interval as fixed endpoints. The translation vectors of each frame within the interval are linearly interpolated according to the frame number to the line connecting the two endpoints. The interpolation result replaces the translation vector of the corresponding frame in the original smoothed pose sequence. The global consistency check and secondary smoothing correction are iterated until the area of ​​the reprojection covariance ellipse of all strong interface anchor points is lower than the threshold. In complex multi-lithological interaction sections, the number of iterations is usually more than in homogeneous strata. The smooth pose sequence after iteration convergence is output as the corrected pose sequence. Each frame of the corrected pose sequence is managed with the frame number as the primary key and strictly corresponds to the frame number of the initial pose sequence.

[0058] In some embodiments, the step of performing global consistency optimization on the borehole point cloud map based on the corrected pose sequence to generate a navigation and positioning map includes: performing pose reprojection on each frame of the point cloud in the borehole point cloud map using the corrected pose sequence to generate a reprojected point cloud set; performing overlap region consistency detection on the reprojected point cloud set to identify local distortion regions; performing point cloud registration optimization on the local distortion regions to eliminate geometric inconsistencies and generate an optimized point cloud map; and performing global coordinate system unification on the optimized point cloud map to generate a navigation and positioning map.

[0059] A reprojected point cloud set is generated by performing pose reprojection on each frame of the borehole point cloud map using a corrected pose sequence. Each frame of the borehole point cloud map is projected onto the world coordinate system using an initial pose sequence during acquisition. The corrected pose sequence corrects the translation vector of the initial pose sequence frame by frame. Each frame of the borehole point cloud map must be reprojected based on the corrected pose to eliminate the spatial misalignment introduced by the initial drift. The 3D point coordinates p in each frame of the point cloud are... i c Transformation matrix T of the frame corresponding to the corrected pose sequence k Reprojection to world coordinate system: p i w =T k '·p i cAfter reprojection, the spatial positions of each frame's point cloud in the world coordinate system shift compared to the original projection result. The shift is equal to the projection of the difference between the translation vectors of the corrected pose sequence and the initial pose sequence in the world coordinate system. Point clouds corresponding to frames carrying low-confidence markers in the corrected pose sequence are appended with low-reliability markers after reprojection. Low-reliability point clouds are distinguished by an independent marker field in the reprojected point cloud set. When the borehole traverses a fractured zone, these low-reliability point clouds typically exhibit large spatial dispersion, and their positional accuracy is lower than that of the reprojected point clouds in stable rock strata. The reprojected point clouds are arranged by frame number and then merged to form a reprojected point cloud set. The reprojected point cloud set retains three attributes for each point: frame number, depth coordinates, and radial deviation value of the borehole diameter deviation map. The reprojected coordinates replace the original projection coordinates and are written into the corresponding attribute fields.

[0060] Overlapping region consistency detection is performed on the reprojected point cloud to identify local distortion regions. Adjacent frame point clouds spatially overlap in the drilling direction. Points from different frames within the overlapping region should geometrically describe the 3D structure of the same borehole wall segment. If there is a systematic deviation in the positions of points in the overlapping region of two frame point clouds, it indicates residual pose inconsistencies between the corresponding frames. This deviation is particularly significant when the borehole traverses formations with drastic changes in drilling speed. Overlapping region identification is calculated by the intersection of the coverage intervals of adjacent frame point clouds in the depth direction within the reprojected point cloud. Frame pairs with non-empty intersections constitute overlapping frame pairs. Within the overlapping region of each overlapping frame pair, nearest neighbor search is used to find corresponding points for each point. The average distance between corresponding points is defined as the overlap consistency error of the frame pair. Frame pairs with overlap consistency errors exceeding a threshold are marked as locally inconsistent frame pairs. When locally inconsistent frame pairs are continuously distributed in the depth direction, they are merged into local distortion regions. The depth range of the local distortion region is calculated by the union of the depth intervals of the inconsistent frame pairs. When drilling progresses through soft rock collapse sections, the borehole wall morphology changes rapidly, and the geometric structure of the overlapping area between adjacent frames differs significantly. The overlap consistency error in this type of section is inherently high. The threshold for determining local distortion areas in this type of section is dynamically widened using the radial deviation amplitude of the corresponding frame in the borehole diameter deviation diagram as a parameter to avoid misjudging normal borehole wall deformation as pose inconsistency.

[0061] To address localized distortion regions, point cloud registration optimization is performed to eliminate geometric inconsistencies and generate optimized point cloud maps. The pose inconsistencies between point clouds in different frames within these regions stem from insufficient local correction accuracy in the directional correction commands. Point cloud registration optimization aims to maximize the geometric fit between adjacent point clouds within the localized distortion region. This is achieved through an iterative nearest-neighbor algorithm to fine-tune the poses of each frame within the region. The iterative nearest-neighbor algorithm uses the first frame of the point cloud within the localized distortion region as a fixed reference and performs a rigid body transformation search frame-by-frame for each subsequent frame. The rigid body transformation parameters are represented by the rotation matrix R and the translation vector t. The optimization objective is to minimize the average distance between the nearest neighbor pairs within the overlapping area of ​​each transformed point cloud and the reference frame. Point cloud registration optimization within localized distortion regions is constrained by the pose of the boundary frames. The boundary frame pose remains fixed throughout the optimization iterations to prevent geometric discontinuities between the local optimization results and the stable point clouds outside the region. After optimization and iteration convergence, the point clouds of each frame are reprojected based on the updated pose. The overlap consistency error after reprojection is recalculated. Frame pairs that still exceed the threshold trigger secondary registration. Secondary registration uses a denser initial transform candidate set to escape local extrema. After all local distortion regions are processed, the point clouds of each frame are merged into an optimized point cloud map. The optimized point cloud map is downsampled twice using the same voxel grid parameters as the borehole point cloud map to unify the point density. Within each voxel, the single point closest to the voxel center is retained as a representative.

[0062] A global coordinate system unification is performed on the optimized point cloud map to generate a navigation and positioning map. The world coordinate system of the optimized point cloud map is defined with the pose of the first frame as the origin. There is an origin offset and axial rotation deviation between it and the geographic coordinate system used for borehole engineering surveying. In inclined or directional borehole operations, there is a known angle between the advance axis and the vertical direction. The depth axis of the world coordinate system must be rotated according to the corresponding angle to align with the elevation axis of the geographic coordinate system. Direct translation without rotation correction will cause the deep point cloud coordinates to accumulate lateral deviation with depth. The ground measurement control points are provided by the total station measurement results at the borehole entrance. The three-dimensional coordinates of the entrance are accurately known in the geographic coordinate system and correspond to the origin of the first frame of the optimized point cloud map. The rigid body transformation matrix between the two is solved by the coordinate relationship of the control points. All three-dimensional points of the optimized point cloud map are uniformly mapped to the geographic coordinate system after transformation matrix and stored in geodetic coordinate format. The deviation of the advance trajectory caused by fault displacement or changes in the dip angle of the rock strata is represented as a three-dimensional curve in the geographic coordinate system. Depth segments with large deviations are marked as trajectory curvature anomalies in the navigation and positioning map. The navigation and positioning map is supplemented with the stress direction calibration results of the aperture deviation map and the lithological classification information of the soft and hard distribution areas of the wall. Soft rock sections are marked as high deformation risk areas, and hard rock sections are marked as stable advancement areas. Each point records five attributes: geographic coordinates, frame number, radial deviation value, lithological classification label, and point cloud reliability label.

[0063] To implement the autonomous navigation operation method of a micro-drilling robot corresponding to the above method embodiments, so as to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 This paper shows a structural block diagram of an autonomous navigation system 200 for a micro-drilling robot according to an embodiment of this application, including: The structural annotation unit 201 is used to acquire the borehole inner wall image sequence, perform illumination anomaly response detection and feature sparse region localization on the borehole inner wall image sequence to generate illumination anomaly distribution map and weak texture boundary map, and perform joint annotation of wall structure interface transition based on the illumination anomaly distribution map and weak texture boundary map to generate structural anchor point sequence. The point cloud construction unit 202 is used to generate an initial pose sequence by performing cross-frame pose transfer based on the structural anchor point sequence, perform hole wall visual measurement on the borehole inner wall image sequence to generate a hole diameter deviation map, and construct a borehole point cloud map by geometric constraint fusion positioning through the initial pose sequence and the hole diameter deviation map. The drift correction unit 203 is used to perform directional decomposition on the cumulative drift in the initial pose sequence to identify axial drift and radial drift, perform sensor absolute depth anchoring correction on the axial drift to obtain axial correction amount, perform periodic matching of wall structure based on the radial drift and the weak texture boundary map to obtain radial constraint identifier, and generate directional correction command based on the axial correction amount and the radial constraint identifier. The positioning optimization unit 204 is used to perform frame-by-frame pose recorrection on the initial pose sequence according to the directional correction instruction to generate a corrected pose sequence, and to perform global consistency optimization on the borehole point cloud map based on the corrected pose sequence to generate a navigation and positioning map.

[0064] The aforementioned autonomous navigation system 200 for a micro-drilling robot can implement the autonomous navigation method for a micro-drilling robot described in the above-described method embodiments. The options described in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0065] The above description is only a part or preferred embodiment of this application. Neither the text nor the drawings should limit the scope of protection of this application. All equivalent structural transformations made using the content of this application's specification and drawings under the overall concept of this application, or direct / indirect applications in other related technical fields, are included within the scope of protection of this application.

Claims

1. An autonomous navigation operation method for a micro-drilling robot, characterized in that, include: A sequence of borehole inner wall images is obtained. Illumination anomaly response detection and feature sparse region localization are performed on the borehole inner wall image sequence to generate an illumination anomaly distribution map and a weak texture boundary map. Based on the illumination anomaly distribution map and the weak texture boundary map, joint annotation of wall structure interface transition is performed to generate a sequence of structural anchor points. Based on the structural anchor point sequence, cross-frame pose transfer is performed to generate an initial pose sequence. The borehole inner wall image sequence is subjected to borehole wall visual measurement to generate a borehole diameter deviation map. The initial pose sequence and the borehole diameter deviation map are geometrically constrained and fused to construct a borehole point cloud map. The cumulative drift in the initial pose sequence is decomposed directionally to identify axial drift and radial drift. The axial drift is corrected by sensor absolute depth anchoring to obtain axial correction amount. The radial drift and the weak texture boundary map are periodically matched with the wall structure to obtain radial constraint identifier. The axial correction amount and the radial constraint identifier are used to generate directional correction instructions. The initial pose sequence is recalibrated frame by frame according to the directional correction instruction to generate a calibrated pose sequence. Based on the calibrated pose sequence, the borehole point cloud map is optimized for global consistency to generate a navigation and positioning map.

2. The method according to claim 1, characterized in that, The step of performing borehole wall visual measurement on the borehole inner wall image sequence to generate a borehole diameter deviation map includes: Perform frame-by-frame borehole wall contour detection on the borehole inner wall image sequence to generate a frame-by-frame borehole wall contour set; Ellipse fitting calculations are performed on the frame-by-frame hole wall contour set to generate a frame-by-frame ellipse fitting parameter set; The major-minor axis ratio is extracted and the wall surface hardness distribution area is identified based on the frame-by-frame ellipse fitting parameter set. Based on the wall hardness distribution area and the frame-by-frame hole wall contour set, stress direction calibration is performed to generate a hole diameter deviation map.

3. The method according to claim 1, characterized in that, The step of constructing a borehole point cloud map by geometrically constraining fusion positioning using the initial pose sequence and the aperture deviation map includes: Local geometric feature points are extracted from the aperture deviation map to generate a set of geometric anchor points; Spatiotemporal alignment of the geometric anchor point set with the initial pose sequence generates pose-geometric association pairs. The pose geometric association is used to construct a geometric constraint residual field from the execution constraint residual; Based on the geometric constraint residual field, borehole point cloud map is constructed by optimizing the borehole wall geometric constraint pose.

4. The method according to claim 1, characterized in that, The process of obtaining the axial correction amount by performing absolute depth anchoring correction on the axial drift actuator includes: A depth increment sequence is generated by extracting frame-by-frame depth increments from the axial drift. A sliding window steady-state estimation is performed on the depth increment sequence to generate a depth reference increment sequence; Valid frame markers are generated by performing frame-by-frame deviation detection based on the depth increment sequence and the depth reference increment sequence. The axial correction amount is obtained by anchoring the axial drift to absolute depth using the effective frame marker.

5. The method according to claim 1, characterized in that, The step of obtaining radial constraint identifiers by periodically matching the wall structure based on the radial drift and the weak texture boundary map includes: The boundary spacing of the weak texture boundary map is extracted periodically to generate the periodic feature spectrum of the wall structure; The radial drift is subjected to spectral analysis to generate a drift frequency sequence; Frequency domain correlation matching is performed between the drift frequency sequence and the periodic feature spectrum of the wall structure to identify the periodic associated structural layer; Based on the aforementioned periodic correlation structure layer, radial drift tracing and localization are performed to generate radial constraint identifiers.

6. The method according to claim 1, characterized in that, The step of performing frame-by-frame pose recorrection on the initial pose sequence according to the directional correction instruction to generate a corrected pose sequence includes: Based on the aforementioned directional correction command, axial correction components and radial correction components are extracted to generate a directional correction vector; Based on the directional correction vector, the poses of each frame in the initial pose sequence are superimposed to generate a frame-by-frame recorrected pose; The frame-by-frame recorrected pose is subjected to inter-frame smoothing constraint elimination correction jump to generate a smooth pose sequence. A corrected pose sequence is generated by performing a global pose consistency check based on the smooth pose sequence.

7. The method according to claim 1, characterized in that, The step of performing global consistency optimization on the borehole point cloud map based on the corrected pose sequence to generate a navigation and positioning map includes: The corrected pose sequence is used to perform pose reprojection on each frame of the borehole point cloud map to generate a reprojected point cloud set. Perform overlap region consistency detection on the reprojection point set to identify local distortion regions; For the local distortion region, point cloud registration optimization is performed to eliminate geometric inconsistencies and generate an optimized point cloud map; The optimized point cloud map is then unified with a global coordinate system to generate a navigation and positioning map.

8. The method according to claim 4, characterized in that, The step of generating valid frame markers by performing frame-by-frame deviation detection based on the depth increment sequence and the depth reference increment sequence includes: The depth increment sequence and the depth reference increment sequence are synchronized and aligned frame by frame to generate a frame-level deviation sequence; Extract the steady-state deviation benchmark from the frame-level deviation sequence to generate the normal deviation range; Perform a deviation scan between the frame-level deviation sequence and the normal deviation interval to identify abnormal deviation frames; Based on the abnormal deviation frame, invalid frame marking is performed to generate valid frame marking.

9. The method according to claim 5, characterized in that, The step of generating a drift frequency sequence by performing spectral analysis on the radial drift includes: The radial drift is subjected to a discrete Fourier transform to generate a drift spectrum. Extract significant peaks from the drift spectrum to generate a set of peak frequencies; The peak frequency set is used to construct an integer ratio relationship topology to identify harmonic-related links and generate interference harmonic markers; Based on the interference harmonic markers, the drift spectrum is correlated with frequency bands to suppress and generate a drift frequency sequence.

10. An autonomous navigation system for a miniature drilling robot, characterized in that, include: The structural annotation unit is used to acquire the borehole inner wall image sequence, perform illumination anomaly response detection and feature sparse region localization on the borehole inner wall image sequence to generate illumination anomaly distribution map and weak texture boundary map, and perform joint annotation of wall structure interface transition based on the illumination anomaly map and weak texture boundary map to generate structural anchor point sequence. The point cloud construction unit is used to generate an initial pose sequence by performing cross-frame pose transfer based on the structure anchor point sequence, perform hole wall visual measurement on the borehole inner wall image sequence to generate a hole diameter deviation map, and construct a borehole point cloud map by geometric constraint fusion positioning through the initial pose sequence and the hole diameter deviation map. The drift correction unit is used to perform directional decomposition on the cumulative drift in the initial pose sequence to identify axial drift and radial drift, perform sensor absolute depth anchoring correction on the axial drift to obtain axial correction amount, perform periodic matching of wall structure based on the radial drift and the weak texture boundary map to obtain radial constraint identifier, and generate directional correction command based on the axial correction amount and the radial constraint identifier. The positioning optimization unit is used to perform frame-by-frame pose recorrection on the initial pose sequence according to the directional correction instruction to generate a corrected pose sequence, and to perform global consistency optimization on the borehole point cloud map based on the corrected pose sequence to generate a navigation and positioning map.

Images