Mine geological radar image adaptive denoising method based on generative adversarial network

Abstract

The present application relates to a mine geological radar image adaptive denoising method based on a generative adversarial network, belonging to the technical field of image enhancement. In view of the problem that the anchor rod diffraction signal is similar to the geological abnormal signal, which is easy to cause false enhancement or weakening of abnormal signal, the mine geological radar B-scan image, support design parameters, antenna moving track and surrounding rock reference wave velocity are obtained, the anchor rod vertex coordinates and curvature are calculated, and the spatial prior mask is generated combined with the historical position deviation; the hyperbolic curve area is extracted and matched, the anchor rod diffraction area and suspected geological abnormal area are identified, the anchor rod diffraction contribution is separated through diffraction template matching tracking, and the residual image is obtained; the suspected geological abnormal area is subjected to depth difference and axial periodicity discrimination, the confirmed area is set as a protection area and F-K domain image enhancement is implemented, and the geological signal is separated in the overlapping area by using conditional generative adversarial network; the method is used for image enhancement of mine surrounding rock crack, separation and water-bearing abnormality.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to an adaptive denoising method for mine ground radar images based on generative adversarial networks. Background Technology

[0002] Ground-penetrating radar (GPR) in mines transmits electromagnetic waves into the surrounding rock and receives reflected signals, generating B-scan images that reflect anomalies such as fractures, delamination, and water-bearing structures within the rock. Due to the inhomogeneity of the surrounding rock medium, equipment interference, and the influence of metal support components such as anchor bolts, the acquired B-scan images typically contain noise, background clutter, and strong diffraction signals, necessitating image enhancement processing to improve the discernibility of geological anomaly signals. Image denoising is one specific technique or processing step in achieving image enhancement.

[0003] Existing technologies typically employ filtering, wavelet transform, background subtraction, or deep learning methods to denoise and enhance mine ground-penetrating radar images. However, anchor bolts in mine roadways generate strong hyperbolic diffraction signals in B-scan images, and these diffraction signals may share similar image characteristics with geological anomaly signals. Existing methods often fail to adequately incorporate anchor bolt support design parameters and the deviation of the actual anchor bolt position from its theoretical position, making it difficult to accurately identify and separate the anchor bolt diffraction signals.

[0004] The inventors of this application have discovered that the above-mentioned technology has difficulty in accurately determining the anchor bolt diffraction area based on the deviation of the actual installation position of the anchor bolt, and effectively separating the anchor bolt diffraction contribution component from the B-scan image. This may result in the anchor bolt diffraction signal being mistakenly enhanced into a geological anomaly signal, or the real geological anomaly signal being mistakenly weakened during the denoising process, thereby affecting the accuracy of mine geological radar image enhancement and geological anomaly identification. Summary of the Invention

[0005] This invention provides an adaptive denoising method for mine ground-penetrating radar images based on generative adversarial networks. The technical problem to be solved is: how to accurately identify and separate the anchor diffraction contribution component in the mine ground-penetrating radar B-scan image when there is a deviation in the actual installation position of the anchor bolt, so as to avoid the real geological anomaly signal being erroneously weakened or enhanced during the denoising and image enhancement process.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: An adaptive denoising method for mine geological radar images based on generative adversarial networks includes: Acquire mine geological radar B-scan images of the current detection section, support design parameters, antenna movement trajectory, surrounding rock reference wave velocity, and historically confirmed anchor bolt sets of previously detected sections; Based on the three-dimensional spatial coordinates of each anchor in the current detection segment, the trajectory of the antenna movement, and the reference wave velocity of the surrounding rock, calculate the theoretical vertex coordinates and theoretical curvature of each anchor in the mine geological radar B-scan image; Based on the anchor bolt number, the anchor bolt column number, and the roadway mileage, establish the spatial correspondence between each anchor bolt in the current detection segment and the historically confirmed anchor bolts. Based on the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of the corresponding historically confirmed anchor bolts, correct the theoretical vertex coordinates of the corresponding anchor bolts in the current detection segment. Determine the search window with the corrected theoretical vertex coordinates as the center, and generate a pixel-level spatial prior mask based on each search window. Extract hyperbolic regions from the mine geological radar B-scan image, determine the actual vertex coordinates and actual curvature of each hyperbolic region, mark hyperbolic regions whose actual vertex coordinates fall into the corresponding search window and whose actual curvature is within the preset curvature tolerance range as anchor diffraction zones, and add the corresponding anchors to the confirmed anchor set, and mark hyperbolic regions that do not match any anchors as suspected geological anomaly zones; Based on the three-dimensional spatial coordinates of each anchor in the confirmed anchor set, the antenna movement trajectory, and the surrounding rock reference wave velocity, a corresponding theoretical diffraction template is generated. The diffraction contribution component of each confirmed anchor is separated from the mine ground radar B-scan image by matching and tracking to obtain the residual B-scan image. For suspected geological anomaly areas in the residual B-scan image, a secondary discrimination is performed based on the depth difference between the suspected geological anomaly area and the confirmed anchor bolt and the periodic verification along the tunnel axis. The area confirmed by the secondary discrimination to contain geological anomaly signals is marked as a protection zone, and image enhancement is performed on the protection zone to obtain an enhanced image.

[0007] Furthermore, the pixel-level spatial prior mask is iteratively updated, including: The matching confidence is determined based on the vertex position deviation, curvature deviation, and template relevance of the candidate hyperbolic regions within the search window corresponding to each unconfirmed anchor. The template relevance is the normalized cross-correlation between the candidate hyperbolic region and the theoretical diffraction template of the corresponding anchor. The matching confidence decreases as the vertex position deviation and curvature deviation increase, and increases as the template relevance increases. According to the order of matching confidence from high to low, a preset number of unconfirmed anchor bolts are selected for spatial matching; Accumulate the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of newly confirmed anchor bolts in the current round and previous rounds. When the number of accumulated positional deviation samples reaches the preset number of samples, interpolate the accumulated positional deviation samples along the channel number direction and time depth direction of the B-scan image to obtain the deviation field. The search window positions of the remaining unconfirmed anchors and the pixel-level spatial prior mask are corrected based on the deviation field, and the spatial matching is re-executed until no new confirmed anchors are added in this round or the number of iterations reaches the preset number.

[0008] The beneficial effects of this invention are as follows: 1. This invention utilizes the positional deviation between the actual and theoretical vertex coordinates of historically confirmed anchor bolts to correct the theoretical vertex coordinates and search window of the anchor bolts in the current detection segment. It can adapt to the deviation of the actual installation position of the anchor bolt from the design position, improve the positioning accuracy of the anchor bolt diffraction area, and provide an accurate spatial basis for subsequent anchor bolt interference separation and targeted image enhancement of geological anomaly areas.

[0009] 2. This invention generates a theoretical diffraction template based on the confirmed spatial location of anchor bolts and separates the anchor bolt diffraction contribution components from mine ground-penetrating radar B-scan images through matching and tracking. Compared with uniform filtering or overall denoising, this invention can specifically suppress deterministic diffraction interference generated by anchor bolts, reduce the possibility of anchor bolt diffraction signals being mistakenly enhanced as geological anomaly signals during image enhancement, and improve the accuracy of image enhancement results.

[0010] 3. This invention performs secondary discrimination based on the depth difference between suspected geological anomaly areas and confirmed anchor bolts, as well as the periodicity along the tunnel axis. This can distinguish between the diffraction response formed by the regular arrangement of anchor bolts and the irregular response formed by geological anomalies, reducing the probability of false anomaly areas entering the image enhancement processing range and improving the reliability of image enhancement object selection.

[0011] 4. This invention marks only the areas containing geological anomaly signals as protected areas after secondary discrimination and performs image enhancement on the protected areas. This avoids indiscriminate enhancement of the entire B-scan image, reduces the situation where real geological anomaly signals are mistakenly weakened during the denoising process and non-target interference is synchronously amplified during the image enhancement process, thereby improving the targeting of image enhancement and the recognizability of geological anomaly signals. Attached Figure Description

[0012] Figure 1 A flowchart of the core processing flow of the adaptive denoising method for mine geological radar images; Figure 2 Flowchart for iterative update of pixel-level spatial prior masks; Figure 3A flowchart illustrating the secondary identification process for suspected geological anomaly areas; Figure 4 Flowchart of the process for matching, tracking and separating components contributing to anchor bolt diffraction. Detailed Implementation

[0013] The technical solution of this application will be further described below with reference to the embodiments. It should be understood that the following embodiments are only used to illustrate this application and are not intended to limit the scope of protection of this application.

[0014] The hardware environment system described in this application includes a mine geological radar host, a transmitting antenna, a receiving antenna, a data acquisition and processing unit, and a support design parameter database. The transmitting and receiving antennas move along the tunnel axis, exciting and receiving electromagnetic wave signals sequentially at equal intervals or time intervals to form a mine geological radar B-scan image. The mine geological radar B-scan image uses the horizontal direction as the tunnel number direction, corresponding to the tunnel axis position; and the vertical direction as the time-depth direction, corresponding to the two-way propagation time of the electromagnetic waves. The data acquisition and processing unit is communicatively connected to the mine geological radar host and is used to acquire, store, and process the B-scan image data in real time. The support design parameter database stores the support design parameters for the current detection section and historically detected sections, including the three-dimensional spatial coordinates of each anchor bolt, anchor bolt number, anchor bolt row number, anchor bolt spacing, and tunnel mileage. The three-dimensional spatial coordinates include axial coordinates along the tunnel axis, depth coordinates along the tunnel radial direction, and height coordinates along the tunnel height direction. The center frequency of the mine geological radar host is usually tens to hundreds of megahertz to meet the detection depth requirements of several meters to tens of meters in the surrounding rock of mine roadways.

[0015] This application provides an adaptive denoising method for mine geological radar images based on generative adversarial networks, including the following steps: The system acquires the mine geological radar B-scan image, support design parameters, antenna movement trajectory, surrounding rock reference wave velocity, and historically confirmed anchor bolt set for the current detection section. The support design parameters include the three-dimensional spatial coordinates, anchor bolt number, anchor bolt row number, anchor bolt spacing, and roadway mileage of each anchor bolt in the current detection section. The historically confirmed anchor bolt set includes the actual vertex coordinates and theoretical vertex coordinates of each historically confirmed anchor bolt.

[0016] Specifically, this step is the data acquisition and preparation step. The mine geological radar B-scan image of the current detection section is acquired channel by channel by the data acquisition and processing unit as the transmitting antenna and the receiving antenna move along the tunnel axis. Each channel of the B-scan image corresponds to a single channel A-scan waveform received at an antenna position. The channel spacing between adjacent channels constitutes the spatial sampling step size of the antenna movement trajectory in the channel number direction of the B-scan image. The support design parameters are derived from the support design parameter database. The three-dimensional spatial coordinates of each anchor bolt in the current detection section are the anchor bolt installation positions pre-planned in the tunnel construction design. The anchor bolt row number represents the arrangement sequence number of the anchor bolt along the tunnel axis, the anchor bolt column number represents the arrangement sequence number of the anchor bolt along the circumferential direction of the tunnel cross section, the anchor bolt spacing includes the row spacing along the tunnel axis and the column spacing along the circumferential direction of the tunnel cross section, and the tunnel mileage represents the excavation mileage value corresponding to the tunnel cross section where the anchor bolt is located. The reference wave velocity of the surrounding rock is the nominal electromagnetic wave propagation velocity of the surrounding rock medium in the tunnel, which is usually determined empirically based on the lithology of the surrounding rock or calculated by laboratory measurement of the dielectric constant of the borehole core. The historically detected section refers to the tunnel section that has undergone ground-penetrating radar detection and tunnel excavation before the current detection section. The anchor bolts included in the historically detected section have had their actual spatial positions confirmed in the field after tunnel excavation. Each record in the historically confirmed anchor bolt set contains the actual three-dimensional coordinates, actual vertex coordinates, and theoretical vertex coordinates of a confirmed anchor bolt. The actual three-dimensional coordinates of the anchor bolt are the true three-dimensional position of the anchor bolt in the tunnel coordinate system obtained by field measurement after construction completion. The actual vertex coordinates are the tunnel number and time position corresponding to the actual hyperbola vertex of the anchor bolt in the historical mine ground-penetrating radar B-scan image. The theoretical vertex coordinates are the tunnel number and time position corresponding to the theoretical hyperbola vertex of the anchor bolt in the historical mine ground-penetrating radar B-scan image.

[0017] Based on the three-dimensional spatial coordinates of each anchor in the current detection segment, the trajectory of the antenna movement, and the reference wave velocity of the surrounding rock, the theoretical vertex coordinates and theoretical curvature of each anchor in the mine ground radar B-scan image are calculated.

[0018] Specifically, this step maps the three-dimensional physical position of the anchor bolt in the tunnel space to theoretical hyperbolic feature parameters in the B-scan image domain. For any anchor bolt in the current detection section, let its three-dimensional spatial coordinates be the axial coordinates. Depth coordinates and height coordinates The antenna movement trajectory determines the antenna axial coordinates at each antenna position. Antenna depth coordinates Antenna height coordinates The anchor bolt to the first One-way propagation distance of electromagnetic waves at the location of the antenna Calculated by the following formula: Corresponding two-way propagation time for: In the formula, The reference wave velocity for the surrounding rock. When the antenna axial coordinate... With respect to the axial coordinate of the anchor rod When they overlap, the one-way propagation distance reaches its minimum value. : The theoretical vertex coordinates include the theoretical vertex Dao number in the Dao number dimension. The theoretical peak time in the time depth dimension ,in , According to the axial coordinate of the anchor rod The theoretical hyperbola is calculated by converting the channel spacing of the antenna's movement trajectory. The theoretical hyperbola unfolds along the channel number direction in the B-scan image, and its trajectory satisfies the two-way propagation time. Follow The relationship between the changes in the theoretical curvature. The degree of curvature of the theoretical hyperbola at its vertex is characterized by the following formula: The theoretical curvature The physical meaning is as follows: the closer the anchor is to the antenna survey line, the lower the reference wave velocity of the surrounding rock, the narrower the opening of the theoretical hyperbola, and the greater the curvature. The theoretical vertex coordinates and the theoretical curvature together constitute two key parameters describing the theoretical diffraction characteristics of the anchor in the B-scan image domain, providing a theoretical reference benchmark for subsequent anchor diffraction area matching and identification.

[0019] Based on the anchor bolt number, the anchor bolt column number, and the tunnel mileage, a spatial correspondence is established between each anchor bolt in the current detection segment and the historically confirmed anchor bolts. Based on the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of the corresponding historically confirmed anchor bolts, the theoretical vertex coordinates of the corresponding anchor bolts in the current detection segment are corrected. The search window is determined with the corrected theoretical vertex coordinates as the center, and a pixel-level spatial prior mask is generated based on each search window.

[0020] Specifically, this step utilizes the positioning deviation information of confirmed anchor bolts in the historically detected segments to perform cross-segment correction on the theoretical vertex coordinates of unconfirmed anchor bolts in the B-scan image in the current detection segment, and generates a spatial prior mask to guide subsequent anchor bolt matching and identification.

[0021] First, establish the spatial correspondence. Due to the regular arrangement pattern of the roadway support design along the roadway axis, anchors with the same row number or column number have the same design cross-sectional position and similar surrounding rock conditions at different roadway mileages. Based on the anchor row number, anchor column number, and roadway mileage, establish a one-to-one correspondence between each anchor in the current detection segment and the confirmed anchors in the historically confirmed anchor set that have the same row number or similar column number and the closest roadway mileage. For an anchor with a specific row and column number in the current detection segment, its corresponding historically confirmed anchor is the one with the same or adjacent row number, the same column number, and the smallest difference in roadway mileage.

[0022] For each anchor pair after establishing the correspondence, calculate the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of the historically confirmed anchor. The positional deviation includes the offset in the track direction. Offset in the time depth direction The positional deviation originates from the difference between the actual and designed positions of the anchor bolts due to surrounding rock deformation during tunnel excavation, as well as the wave velocity deviation caused by the non-uniformity of the surrounding rock medium. Applying this positional deviation to the theoretical vertex coordinates of the corresponding anchor bolt in the current detection section yields the corrected theoretical vertex coordinates: Corrected theoretical vertex tunnel number. Corrected theoretical peak time .

[0023] Centered on the corrected theoretical vertex coordinates, a neighborhood window with a preset range is selected in both the trace number dimension and the time depth dimension of the B-scan image to form a search window targeting the anchor bolt. The half-width of the trace number dimension and the half-width of the time depth dimension of the search window are set according to the upper limit of the allowable anchor bolt position deviation and the possible range of change of the surrounding rock wave velocity, respectively. Pixel-level mapping is performed on the search windows of all anchor bolts in the current detection segment, and a mask label value is assigned to each pixel in the B-scan image: pixels falling into any search window are labeled with a first label value, indicating that the pixel is in the expected occurrence area of ​​the anchor bolt diffraction signal; pixels not falling into any search window are labeled with a second label value, indicating that the pixel is in the low probability area of ​​the anchor bolt diffraction signal. The mask label values ​​of all pixels constitute the pixel-level spatial prior mask.

[0024] Extract hyperbolic regions from the mine geological radar B-scan image, determine the actual vertex coordinates and actual curvature of each hyperbolic region, mark hyperbolic regions whose actual vertex coordinates fall into the corresponding search window and whose deviation between actual curvature and theoretical curvature is within the preset curvature tolerance range as anchor diffraction zones, and add the corresponding anchors to the confirmed anchor set, and mark hyperbolic regions that do not match any anchors as suspected geological anomaly zones.

[0025] Specifically, this step extracts all hyperbolic reflection or diffraction regions from the B-scan image and classifies them into anchor diffraction zones or suspected geological anomaly zones by comparing their positions and curvatures with those of the anchor search window in the pixel-level spatial prior mask. The extraction of hyperbolic regions can be achieved using a hyperbolic detection algorithm based on Hough transform, a hyperbolic enhancement detection algorithm based on phase symmetry, or a semantic segmentation network based on deep learning. Each extracted hyperbolic region is parameterized by its vertex position and curvature. The actual vertex coordinates are the trace number and temporal depth value corresponding to the vertex pixel of the hyperbolic region in the B-scan image. The actual curvature is the curvature parameter extracted after hyperbolic fitting of the trajectory of the hyperbolic region; its extraction method is consistent with the definition of theoretical curvature, i.e., it also uses curvature values. In the form of.

[0026] For each hyperbolic region, determine whether its actual vertex coordinates fall within the search window corresponding to a certain anchor. If they do, and the difference between the actual curvature of the hyperbolic region and the theoretical curvature of the corresponding anchor is considered... If the curvature does not exceed the preset curvature tolerance range, the hyperbolic region is confirmed to be generated by radar diffraction of the anchor bolt. This hyperbolic region is marked as an anchor bolt diffraction zone, and the anchor bolt is moved from its unconfirmed state to the confirmed anchor bolt set. The preset curvature tolerance range is set comprehensively based on the uncertainty of the surrounding rock reference wave velocity and the anchor bolt installation position tolerance. For hyperbolic regions where the actual vertex coordinates do not fall within any search window, or where the actual vertex coordinates fall within a search window but the curvature deviation exceeds the preset curvature tolerance range, it is determined that the region is not generated by a known anchor bolt and is marked as a suspected geological anomaly zone. The suspected geological anomaly zone may correspond to geological anomalies such as fissures, cavities, aquifers, or faults in the surrounding rock, or it may correspond to sporadic metallic reflectors or false anomalies formed by random noise outside the support design parameters.

[0027] Based on the three-dimensional spatial coordinates of each anchor in the confirmed anchor set, the antenna movement trajectory, and the reference wave velocity of the surrounding rock, a corresponding theoretical diffraction template is generated. By matching and tracking, the diffraction contribution components of each confirmed anchor are separated from the mine ground radar B-scan image to obtain the residual B-scan image.

[0028] Specifically, this step utilizes the precise spatial location information of the confirmed anchor bolts to generate their theoretical diffraction response templates, and employs a matching pursuit algorithm to sequentially separate the diffraction signals of the confirmed anchor bolts from the original B-scan image. For each confirmed anchor bolt in the set, based on its three-dimensional spatial coordinates and the antenna movement trajectory, combined with the surrounding rock reference wave velocity, a theoretical diffraction template for that anchor bolt in the B-scan image domain is generated through electromagnetic wave propagation geometry. The theoretical diffraction template is a two-dimensional image block centered on the theoretical vertex coordinates of the anchor bolt, where the amplitude of each pixel is weighted according to the degree to which the pixel deviates from the theoretical hyperbolic trajectory, and the theoretical diffraction template embeds a waveform matching the wavelet emitted by the mine ground-penetrating radar in the time-depth direction.

[0029] The matching and tracking separation process is as follows: using the mine ground-penetrating radar B-scan image as the original B-scan image, and using the original B-scan image as the initial residual image; forming a theoretical diffraction template dictionary from the theoretical diffraction templates of each confirmed anchor; in each iteration, calculating the normalized cross-correlation between each template in the theoretical diffraction template dictionary and the current residual image within the corresponding search window, selecting the theoretical diffraction template with the highest cross-correlation value and its corresponding optimal amplitude scaling factor and optimal phase offset, and reconstructing the diffraction contribution component of the anchor in the current residual image; subtracting the diffraction contribution component from the current residual image to obtain the updated residual image and entering the next iteration; terminating the iteration when the residual energy decrease is less than a preset energy threshold or the number of iterations reaches a preset upper limit, and determining the final residual image as the residual B-scan image.

[0030] For suspected geological anomaly areas in the residual B-scan image, a secondary discrimination is performed based on the depth difference between the suspected geological anomaly area and the confirmed anchor bolt and the periodic verification along the tunnel axis. The area confirmed by the secondary discrimination to contain geological anomaly signals is marked as a protection zone, and image enhancement is performed on the protection zone to obtain an enhanced image.

[0031] Specifically, this step performs secondary screening on the suspected geological anomaly areas initially marked in the residual B-scan image to eliminate false anomalies caused by non-anchor metal interference bodies or random noise, and provides protective enhancement to areas confirmed as genuine geological anomalies.

[0032] The secondary discrimination is comprehensively evaluated from two dimensions. The first dimension is depth discrimination: calculating the time depth value corresponding to the actual vertex coordinates of the suspected geological anomaly area and the time depth value corresponding to the theoretical vertex coordinates of the confirmed anchor bolt that is spatially closest to the suspected geological anomaly area. If the depth difference is greater than a preset depth threshold, it indicates that the anomaly source is located in the deeper surrounding rock behind the anchor bolt, which is consistent with the typical occurrence location characteristics of a geological anomaly. The second dimension is periodic discrimination: establishing the mapping relationship between the tunnel number and the axial mileage of the B-scan image based on the antenna movement trajectory, obtaining the anchor bolt spacing in the support design parameters, and searching for other suspected geological anomaly areas within a preset neighborhood of the suspected geological anomaly area along the tunnel axis to obtain a sequence of vertex positions of multiple suspected geological anomaly areas. Analyzing whether the sequence of vertex positions of multiple suspected geological anomaly areas exhibits a regular periodic distribution characteristic at integer multiples of the anchor bolt spacing. If the multiple suspected geological anomaly areas repeat strictly at intervals of the anchor bolt spacing or its integer multiples, these anomaly sources are more likely to be regular metallic reflectors related to the anchor bolt support structure rather than real geological anomalies. If the depth difference is greater than the preset depth threshold and the periodic distribution feature is not significant, i.e., the periodic matching degree is lower than the preset periodic matching threshold, then the suspected geological anomaly area is confirmed to contain a real geological anomaly signal, and it is marked as the protected area. The protected area is given priority protection in subsequent image enhancement processes to prevent the enhancement operation from suppressing or distorting the weak geological anomaly signal.

[0033] The image enhancement operation performed on the protected area is as follows: Extract the image data sub-matrix corresponding to the protected area; perform a two-dimensional Fourier transform on the sub-matrix to obtain FK domain data; in the FK domain data, determine the target frequency-wavenumber band corresponding to the diffraction energy of the geological anomaly based on the surrounding rock reference wave velocity, the center frequency of the mine ground-penetrating radar, and the antenna movement trajectory; apply gain to the frequency-wavenumber components within the target frequency-wavenumber band and attenuate the frequency-wavenumber components outside the target frequency-wavenumber band to obtain enhanced FK domain data; perform a two-dimensional inverse Fourier transform on the enhanced FK domain data to obtain enhanced protected area image data; merge the enhanced protected area image data with the unmarked protected area image data in the residual B-scan image to form the enhanced image.

[0034] In summary, the technical solution described in this application transforms the three-dimensional spatial information of anchor bolts in the roadway support design parameters into a spatial prior mask in the B-scan image domain. It then uses the actual-theoretical position deviation of confirmed anchor bolts in historically detected sections to perform cross-segment correction on the theoretical position of anchor bolts in the current detection section, achieving precise separation of anchor bolt diffraction signals from geological anomaly signals. The spatial prior mask provides strong prior constraints for anchor bolt matching and identification, significantly reducing the probability of false matches between the anchor bolt diffraction hyperbola and other hyperbolic reflection events. The matching and tracking separation mechanism uses confirmed anchor bolts as reliable anchor points, successively stripping away the anchor bolt diffraction contribution components, effectively exposing geological anomaly signals in the residual B-scan image. The secondary discrimination mechanism based on depth discrimination and axial periodicity discrimination effectively suppresses false anomalies formed by non-anchor bolt metallic interference and random noise. The FK domain selective enhancement operation performed on the protected area improves the identifiability of geological anomaly signals while avoiding damage to weak signals, resulting in an enhanced image that combines the dual effects of anchor bolt interference suppression and geological anomaly fidelity enhancement.

[0035] In this embodiment, iteratively updating the pixel-level spatial prior mask includes the following steps: The matching confidence is determined based on the vertex position deviation, curvature deviation, and template relevance of the candidate hyperbolic regions within the search window corresponding to each unconfirmed anchor. The template relevance is the normalized cross-correlation between the candidate hyperbolic region and the theoretical diffraction template of the corresponding anchor. The matching confidence decreases as the vertex position deviation and curvature deviation increase, and increases as the template relevance increases.

[0036] Specifically, for each unconfirmed anchor, within the search window corresponding to that anchor in the pixel-level spatial prior mask, there may be one or more candidate hyperbolic regions obtained through hyperbolic extraction. For each candidate hyperbolic region, a three-dimensional matching index is calculated: First, the Euclidean distance between the actual vertex position of the candidate hyperbolic region and the center of the anchor's search window (i.e., the corrected theoretical vertex coordinates). The vertex position deviation is obtained by dividing the Euclidean distance by the normalized half-width of the search window. The value ranges from zero to one. The smaller the value, the more accurate the position matching; secondly, the absolute value of the difference between the actual curvature of the candidate hyperbolic region and the theoretical curvature of the anchor rod, after normalization by the preset curvature tolerance range, yields the curvature deviation. The value ranges from zero to one. The smaller the value, the better the curvature shape matches; third, the B-scan image sub-block corresponding to the candidate hyperbolic region is subjected to normalized cross-correlation operation with the theoretical diffraction template of the anchor rod. The value range of the normalized cross-correlation is from negative one to positive one, and the larger the value, the higher the waveform similarity.

[0037] The matching confidence This is a comprehensive function of the vertex position deviation, the curvature deviation, and the template relevance. In this embodiment, the matching confidence... Determined by the following formula: In the formula, The deviation of the vertex position. The curvature deviation is mentioned. The template correlation is the normalized cross-correlation value between the candidate hyperbolic region and the theoretical diffraction template. and These are the position deviation attenuation coefficient and the curvature deviation attenuation coefficient, respectively, used to control the attenuation rate of confidence due to position deviation and curvature deviation; , and The preset weighting coefficients satisfy... The specific values ​​of each weighting coefficient can be adjusted according to the relative importance of position accuracy, curvature accuracy and waveform similarity in the actual detection scenario.

[0038] According to the order of matching confidence from high to low, a preset number of unconfirmed anchor bolts are selected for spatial matching.

[0039] Specifically, the matching confidence scores of all unconfirmed anchor bolts are sorted in descending order, and the bolts with the highest matching confidence scores are selected. The unconfirmed anchors undergo spatial matching in this round. The spatial matching operation is consistent with the matching process between the hyperbolic region and the anchor in claim 1, that is, the candidate hyperbolic region with the highest matching confidence within the selected search window of the unconfirmed anchor is marked as the anchor diffraction zone of that anchor, and the anchor is moved into the set of confirmed anchors. The preset number... The value is set based on the total number of anchor bolts covered in a single tunnel detection and the desired convergence speed. The typical value range is 5% to 20% of the total number of anchor bolts. Selecting only a small number of anchor bolts with high confidence for initial confirmation can avoid the accumulation of bias caused by the incorrect confirmation of anchor bolts with low matching confidence.

[0040] The positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of newly confirmed anchor bolts in the current round and previous rounds is accumulated. When the number of accumulated positional deviation samples reaches the preset number of samples, the accumulated positional deviation samples are interpolated along the channel number direction and time depth direction of the B-scan image to obtain the deviation field.

[0041] Specifically, after each round of spatial matching is completed, the newly confirmed anchor bolt position deviation records for that round are added to the cumulative position deviation sample set. The position deviation includes the offset in the track number direction. Offset in the time depth direction When the number of cumulative position deviation samples reaches a preset number of samples. At the same time, two-dimensional spatial interpolation is performed on the accumulated positional deviation samples across the entire area of ​​the B-scan image to construct a continuous deviation field covering the entire B-scan image region of the current detection segment. The deviation field includes the track number orientation deviation field. and time-depth direction deviation field ,in For the coordinates of the channel number, The time-depth coordinates are used. The two-dimensional spatial interpolation can be achieved using thin-plate spline interpolation or radial basis function interpolation to obtain a smooth and continuous deviation distribution surface. The deviation field reflects the systematic law of the anchor bolt B-scan image position offset caused by the spatial non-uniformity of the surrounding rock medium, as well as the gradual trend of roadway surrounding rock deformation along the axial direction.

[0042] The search window positions of the remaining unconfirmed anchors and the pixel-level spatial prior mask are corrected based on the deviation field, and the spatial matching is re-executed until no new confirmed anchors are added in this round or the number of iterations reaches the preset number.

[0043] Specifically, using the constructed deviation field, the center position of the search window for the remaining unconfirmed anchors is corrected: let the center position of the search window for an unconfirmed anchor in the previous iteration be the track number coordinate. and time depth coordinates The channel number offset of the deviation field at that location and time depth offset The center of the search window is shifted, and the updated center of the search window is... and Simultaneously, the pixel-level spatial prior mask is regenerated based on all updated search windows. Based on the updated pixel-level spatial prior mask and the search windows, spatial matching for this round is re-executed. The iterative cycle of accumulating deviation samples, constructing the deviation field, correcting the search window, and spatial matching is repeated until no unconfirmed anchor bolts are successfully confirmed as newly confirmed anchor bolts in a certain iteration, or the number of iterations reaches the preset maximum number of iterations. Through the iterative construction of the deviation field and mask updates, as the number of confirmed anchor bolts increases round by round, the accuracy of the deviation field in characterizing the non-uniformity of the surrounding rock and the deformation of the tunnel continuously improves, enabling the positioning accuracy of the search window to gradually improve with each iteration, forming a self-feedback calibration-enhanced anchor bolt recognition closed loop.

[0044] In summary, the aforementioned iterative update mechanism transforms the positional deviations of confirmed anchor bolts into a global deviation field, and uses this deviation field to dynamically correct the search window for unconfirmed anchor bolts. This significantly improves the accuracy and robustness of anchor bolt matching and identification in roadways with complex surrounding rock conditions and limited support construction precision. The confidence-first matching strategy avoids deviation contamination introduced by low-confidence anchor bolts, while the deviation field-driven mask iterative update continuously refines the prior space constraints as information accumulates.

[0045] In this embodiment, the secondary discrimination of the suspected geological anomaly area includes the following steps: Based on the antenna movement trajectory, a mapping relationship is established between the tunnel number and the axial mileage of the mine geological radar B-scan image.

[0046] Specifically, the antenna movement trajectory records the axial mileage value of the tunnel corresponding to each antenna transmit-receive position. Based on the axial mileage corresponding to each tunnel in the antenna movement trajectory, a tunnel number sequence is established. relative to the axial mileage of the tunnel One-to-one mapping function between For the case of equal channel spacing acquisition, the mapping function is a linear relationship: ,in This refers to the axial mileage of the starting roadway. The channel spacing is used. For non-equal spacing acquisition, the mapping function is obtained through piecewise linear interpolation or spline interpolation.

[0047] Obtain the anchor spacing in the support design parameters, and obtain multiple suspected geological anomaly zones within a preset neighborhood along the tunnel axis of the target suspected geological anomaly zone.

[0048] Specifically, the anchor bolt spacing is extracted from the support design parameters. The anchor bolt spacing can be the row spacing along the tunnel axis, the column spacing along the tunnel cross-section circumferentially, or a combination of both. A currently identified suspected geological anomaly area is designated as the target suspected geological anomaly area. A preset neighborhood range is extended along both the positive and negative directions of the tunnel axis, centered on the tunnel axis position of the target suspected geological anomaly area. All suspected geological anomalies within this neighborhood range are obtained, forming a set of neighborhood suspected geological anomaly areas. The preset neighborhood range is several times the anchor bolt spacing to ensure that a sufficient number of suspected geological anomalies are captured for periodic analysis.

[0049] Calculate the depth difference between the actual vertex coordinates of the target suspected geological anomaly zone and the theoretical vertex coordinates of the most recently confirmed anchor bolt, and determine the periodic matching degree between the multiple suspected geological anomaly zones and the anchor bolt spacing based on the vertex position sequence of the multiple suspected geological anomaly zones.

[0050] Specifically, the depth difference is calculated as follows: The confirmed anchor bolt that is closest to the target suspected geological anomaly area along the roadway axis is identified, and the time depth value corresponding to the theoretical vertex coordinates of that confirmed anchor bolt is obtained. Obtain the time depth value corresponding to the actual vertex coordinates of the suspected geological anomaly area. The depth difference The depth difference If the value is positive and greater than the preset depth threshold, it indicates that the source of the anomaly in the suspected geological anomaly zone is located in the surrounding rock area at a depth of the anchor bolt.

[0051] The degree of periodic matching is determined as follows: The axial mileage of each suspected geological anomaly zone in the neighboring suspected geological anomaly zone set is extracted and arranged in ascending order to form a mileage sequence. For the mileage sequence, multiple periodic detection grids are constructed with the anchor bolt spacing as the interval step size. For each periodic detection grid, the root mean square deviation (RMSD) of the mileage value of each suspected geological anomaly zone and the distance between the nearest grid point is calculated. The smaller the RMSD, the closer the distribution of the suspected geological anomaly zones is to a regular arrangement with the anchor bolt spacing as the period, i.e., the higher the degree of periodic matching. The reciprocal of the minimum RMSD under all periodic detection grids is used as the quantitative index of the degree of periodic matching. If the degree of periodic matching is high, it indicates that the suspected geological anomaly zone is likely a regular metallic reflector related to the anchor bolt arrangement, rather than a real geological anomaly.

[0052] When the depth difference is greater than a preset depth threshold and the period matching degree is lower than a preset period matching threshold, it is confirmed that the target suspected geological anomaly area contains geological anomaly signals, and the target suspected geological anomaly area is marked as a protected area.

[0053] Specifically, the logic of the secondary discrimination is as follows: if the depth difference is greater than the preset depth threshold, the possibility of anomalies caused by the anchor bolt itself or shallow metal bodies adjacent to the anchor bolt is ruled out; if the period matching degree is lower than the preset period matching threshold, the possibility of anomalies caused by regularly arranged support metal components is ruled out. When both conditions are met simultaneously, the target suspected geological anomaly area is marked as the protected area with a high confidence level of the diffraction response of a real geological anomaly.

[0054] In summary, the aforementioned secondary discrimination mechanism uses the known depth of the anchor bolts as a depth reference benchmark and the spatial periodicity of the anchor bolt arrangement as a periodic reference benchmark. It performs dual verification on the initially screened suspected anomalies from both the spatial location dimension and the spatial distribution pattern dimension, effectively distinguishing between regular disturbances caused by the anchor bolt support structure and irregular anomaly responses caused by geological anomalies.

[0055] In this embodiment, separating the diffraction contribution components of each confirmed anchor bolt from the mine ground-penetrating radar B-scan image by matching and tracking includes the following steps: The theoretical diffraction templates of each confirmed anchor bolt are used to form a theoretical diffraction template dictionary.

[0056] Specifically, for each confirmed anchor in the confirmed anchor set, a corresponding theoretical diffraction template is constructed according to its three-dimensional spatial coordinates, the antenna movement trajectory, and the surrounding rock reference wave velocity, using the theoretical diffraction template generation method described above. Let the confirmed anchor set include... For each confirmed anchor bolt, the theoretical diffraction template dictionary is... Depend on A theoretical diffraction template is formed: Each template It is a two-dimensional image block matrix centered on the theoretical vertex coordinates of the corresponding anchor rod, with the same size as the search window. The pixel values ​​in the template are assigned according to the theoretical hyperbolic trajectory and the emitted wavelet waveform.

[0057] In each round of matching and tracking, the target theoretical diffraction template with the highest correlation to the current residual image is selected from the theoretical diffraction template dictionary. The magnitude scaling factor and phase offset are used as optimization variables to construct the residual minimization objective function between the target theoretical diffraction template and the current residual image within the corresponding search window.

[0058] Specifically, matching pursuit is an iterative decomposition algorithm. Let the... The current residual image during round iteration is Initial residual image This refers to the original B-scan image. In the... In each iteration, for each template in the theoretical diffraction template dictionary In the search window area corresponding to the template Within this context, calculate the normalized cross-correlation between the template and the current residual image: In the formula This indicates that the current residual image is in the search window. The pixel value submatrix within, This represents the matrix inner product operation. Let Frobenius norm be denoted. The template with the highest normalized cross-correlation value is selected as the target theoretical diffraction template for this round, denoted as . The corresponding search window is denoted as .

[0059] With amplitude scaling factor and phase offset To optimize the variables, the following residual minimization objective function is constructed: In the formula, This indicates that the target theoretical diffraction template has undergone phase shift. The template after, The Frobenius norm is represented. The phase offset is... The amplitude scaling factor reflects the potential phase difference between the actual electromagnetic reflection response and the theoretical diffraction response of the anchor bolt. This reflects the proportional relationship between the actual reflection intensity of the anchor bolt and the preset amplitude of the template.

[0060] Solve the objective function of minimizing the residual to obtain the optimal amplitude scaling factor and optimal phase offset corresponding to the target theoretical diffraction template. Determine the diffraction contribution component of the corresponding anchor rod based on the target theoretical diffraction template, the optimal amplitude scaling factor, and the optimal phase offset.

[0061] Specifically, the objective function for minimizing the residual is about and This is a nonlinear least squares problem. For a given phase offset... The optimal amplitude scaling factor has a closed-form solution: Will Substituting back into the objective function, it is transformed into a function about The one-dimensional search problem is solved by traversing a pre-defined set of phase offset candidates. Determine the optimal phase offset that minimizes the objective function. and the corresponding optimal amplitude scaling factor The diffraction contribution component of the corresponding anchor bolt. In the search window Refactor internally as follows: In the search window In areas other than those mentioned above, the diffraction contribution component is zero.

[0062] The diffraction contribution component is subtracted from the current residual image to update the current residual image until the residual energy decrease of the current round is less than a preset energy threshold or the number of matching and tracking reaches a preset number. The final current residual image is then determined as the residual B-scan image.

[0063] Specifically, the current residual image is updated in the following way: Calculate the decrease in residual energy before and after this iteration. .when When the energy level is less than a preset energy threshold, it indicates that there are no significant anchor diffraction contribution components in the remaining residual that can be separated, and the iteration stops. Iteration also stops if the number of matching and tracking attempts reaches a preset number. The current residual image at the end of the iteration is the residual B-scan image.

[0064] In summary, the aforementioned matching pursuit separation scheme transforms anchor bolt diffraction signal separation into a successive single-template matching decomposition problem by constructing a theoretical diffraction template dictionary, thus avoiding signal crosstalk that may occur when separating multiple anchor bolt diffraction signals simultaneously. The residual minimization objective function simultaneously optimizes both amplitude and phase degrees of freedom, enabling the reconstructed diffraction contribution components to more accurately reproduce the intensity and phase characteristics of the actual anchor bolt diffraction response.

[0065] In this embodiment, performing image enhancement on the protected area includes the following steps: A two-dimensional Fourier transform is performed on the image data corresponding to the protected area to obtain FK domain data.

[0066] Specifically, the protected area is the two-dimensional pixel region in the B-scan image marked as requiring protective enhancement. The pixel value matrix corresponding to the protected area is extracted, and a two-dimensional discrete Fourier transform is performed on the pixel value matrix to transform the image data from the time-depth-channel spatial domain to the frequency-wavenumber domain. The FK domain data is then processed in terms of frequency... The first dimension of the coordinate system is the wave number. The second dimension is the coordinate system, where each coordinate position stores a complex number of frequency-wavenumber component values, including the amplitude spectrum and the phase spectrum.

[0067] Based on the surrounding rock reference wave velocity, the center frequency of the mine ground-penetrating radar, and the antenna movement trajectory, the target frequency-wavenumber band corresponding to the diffraction energy of the geological anomaly is determined in the FK domain data.

[0068] Specifically, the hyperbolic diffraction trajectory of geological anomalies in B-scan images exhibits a specific frequency-wavenumber correspondence in its energy distribution within the FK domain. Based on the propagation laws of electromagnetic waves in surrounding rock media, the wavenumber... With frequency The reference wave velocity of the surrounding rock The angle between the direction of diffraction wave propagation and the direction of the antenna survey line The relationship between them is satisfied For diffracted signals generated by geological anomalies, the angle between their propagation direction and the direction of the antenna survey line is... Typically distributed over a large area, corresponding to a wide range of wavenumbers. The target frequency-wavenumber band is defined by the following inequality: In the formula, and These are the low-frequency cutoff frequency and the high-frequency cutoff frequency, respectively, based on the center frequency of the mine ground-penetrating radar. Take 40% to 60% of the center frequency. Take 140% to 160% of the center frequency; and These are the minimum and maximum incident angles of the diffraction signal from the geological anomaly, respectively, set according to the geometric parameters of the tunnel cross-section and the height of the antenna's moving trajectory.

[0069] Gain is applied to the frequency-wavenumber components within the target frequency-wavenumber band, and attenuation is applied to the frequency-wavenumber components outside the target frequency-wavenumber band to obtain enhanced FK domain data.

[0070] Specifically, in the FK domain data, for each frequency-wavenumber component located within the target frequency-wavenumber band, its amplitude value is multiplied by a preset gain factor. , The value of is greater than one, with a typical range of 1.5 to 3.0. For each frequency-wavenumber component located outside the target frequency-wavenumber band, its amplitude value is multiplied by a preset attenuation factor. , The value of is less than one, with a typical range of 0.1 to 0.5. A smooth transition band can be set for the transition region between the gain factor and the attenuation factor to avoid ringing effects caused by frequency domain truncation. The processed FK domain data retains the original phase information of the frequency-wavenumber components, with only selective adjustments made to the amplitude spectrum.

[0071] The enhanced FK domain data is subjected to a two-dimensional inverse Fourier transform and then merged with the non-protected area image data in the residual B-scan image to form the enhanced image.

[0072] Specifically, a two-dimensional inverse discrete Fourier transform is performed on the enhanced FK domain data to obtain enhanced protected area image data. In the residual B-scan image, the original pixel values ​​corresponding to the protected area are replaced with the enhanced protected area image data, while the non-protected area image data remains unchanged. The merged complete image is the enhanced image.

[0073] In summary, the aforementioned FK domain enhancement scheme utilizes the frequency-wavenumber domain distribution differences between the diffraction signals of geological anomalies and those of anchor bolts to selectively enhance the characteristic frequency-wavenumber components of geological anomalies while suppressing the frequency-wavenumber components of anchor bolt residues and other interferences, achieving precise signal-level denoising and enhancement. The defined protected area ensures that the enhancement operation is applied only to areas confirmed as geological anomalies, avoiding signal distortion that might be introduced by applying enhancement operations to the entire map.

[0074] In this embodiment, the method further includes the following steps: Detect the blurred overlapping area where the pixel coverage of the anchor bolt diffraction zone overlaps with the pixel coverage of the protected area.

[0075] Specifically, in the B-scan image, when the extended tail of the anchor diffraction hyperbola intersects spatially with the diffraction hyperbola of the geological anomaly, the pixel coverage of the anchor diffraction area and the pixel coverage of the protected area may overlap. Within this overlapping area, the anchor diffraction signal and the geological anomaly signal are linearly superimposed at the pixel level, making it impossible to completely separate the signals from the two sources using the aforementioned matching tracing method. The method for detecting the blurred overlapping area is as follows: for each pixel in the B-scan image, check whether it is simultaneously marked as the anchor diffraction area or the extended influence area of ​​the anchor diffraction template, and simultaneously marked as the protected area. Pixels that simultaneously satisfy both conditions are merged to obtain the blurred overlapping area.

[0076] Extract residual image blocks containing the blurred overlapping regions from the residual B-scan image, and extract spatial prior mask blocks corresponding to the positions of the residual image blocks from the pixel-level spatial prior mask.

[0077] Specifically, taking the geometrically bounding rectangle of the blurred overlapping region as the center, a preset margin is extended outward to form the residual image block. The residual image block contains the aliasing signal that remains after the matching and tracking processing. At the same coordinate position of the pixel-level spatial prior mask, a spatial prior mask block of the same size is cropped. The spatial prior mask block retains the prior distribution information of the anchor's spatial position.

[0078] The residual image block and the spatial prior mask block are input together into a pre-trained conditional generative adversarial network to obtain the geological signal component image corresponding to the blurred overlapping region.

[0079] Specifically, the conditional generative adversarial network (GAN) uses multi-channel data stitched together along the channel dimension from the residual image patch and the spatial prior mask block as conditional input, and outputs a geological signal component image of the same size as the input. The generator network structure of the GAN adopts a U-Net architecture, including an encoder path composed of alternating stacked convolutional and pooling layers and a decoder path composed of alternating stacked convolutional and upsampling layers. Spatial detail information is passed between the encoder and decoder paths via skip connections. The spatial prior mask block enables the generator network to use prior spatial distribution information of anchor positions to determine which pixel positions are more likely to originate from anchor diffraction and which pixel positions are more likely to originate from geological anomaly diffraction when processing blurred and overlapping regions, thereby performing targeted signal allocation during signal separation.

[0080] The training method of the pre-trained conditional generative adversarial network is described in detail later. The conditional generative adversarial network is only triggered to execute in the blurred overlapping region; image data in other regions are not processed by the conditional generative adversarial network to control computational overhead.

[0081] The geological signal component image and the corresponding region image data in the enhanced image are weighted and fused to update the blurred overlapping region in the enhanced image.

[0082] Specifically, a fusion weighting coefficient is set, and a weighted average is applied pixel-by-pixel to the image data of corresponding regions in the geological signal component image and the enhanced image. The fused pixel values ​​replace the original pixel values ​​of the blurred overlapping regions in the enhanced image, resulting in an updated enhanced image.

[0083] In summary, the above scheme addresses the extreme condition where the anchor bolt diffraction zone and the geological anomaly protection zone overlap at the pixel level by introducing a conditional generative adversarial network (GAN) to achieve sub-pixel-level signal separation. The GAN uses residual image blocks and spatial prior mask blocks as conditional inputs, enabling the signal separation process to utilize not only the statistical features of the residual image but also the spatial location prior of the anchor bolt for guidance. This results in superior separation performance compared to linear decomposition methods in the overlapping areas of the anchor bolt and geological anomaly.

[0084] In this embodiment, the method further includes the following steps: Obtain the calibrated three-dimensional coordinates of each anchor in the confirmed anchor set, and recalculate the theoretical two-way propagation time of each anchor's corresponding theoretical vertex coordinates based on the calibrated three-dimensional coordinates, the antenna movement trajectory, and the surrounding rock reference wave velocity.

[0085] Specifically, the calibrated three-dimensional coordinates are the actual spatial coordinates of the anchor bolts obtained on-site using measurement methods such as a total station or a 3D laser scanner after the tunnel excavation and support construction are completed. Their accuracy is higher than the design three-dimensional spatial coordinates in the support design parameters. For each anchor bolt in the confirmed anchor bolt set, its calibrated three-dimensional coordinates replace the original three-dimensional spatial coordinates in the support design parameters. Following the same geometric propagation distance calculation method as in claim 1, the theoretical two-way propagation time corresponding to the theoretical vertex coordinates in the B-scan image for that anchor bolt is recalculated. .

[0086] Determine whether the calibrated three-dimensional coordinates and the radar echo of the confirmed anchor bolt meet the preset wave velocity calibration conditions.

[0087] Specifically, the preset wave velocity calibration conditions include a joint evaluation of the completeness of the actual spatial position data of the anchor bolt and the quality of the radar echo signal, and the specific criteria are detailed later.

[0088] When the preset wave velocity calibration conditions are met, the reference wave velocity of the surrounding rock is multiplied by the ratio of the theoretical two-way propagation time and the actual two-way propagation time corresponding to each confirmed anchor bolt to obtain the local actual wave velocity of the surrounding rock at each confirmed anchor bolt location. A sparse wave velocity calibration point set is then constructed based on the calibration three-dimensional coordinates of each confirmed anchor bolt and the local actual wave velocity of the surrounding rock.

[0089] Specifically, the actual two-way propagation time is the two-way propagation time corresponding to the vertex of the diffraction hyperbola actually observed by the anchor in the B-scan image. This value can be directly read from the time-depth value of the actual vertex coordinates of the confirmed anchor bolt diffraction zone. The actual local wave velocity of the surrounding rock. Determined by the following formula: In the formula, The reference wave velocity for the surrounding rock is given. The physical meaning of this formula is that the difference between the theoretical two-way propagation time and the actual two-way propagation time stems from the deviation between the reference wave velocity for the surrounding rock and the true average wave velocity of the surrounding rock along the anchor-antenna path. By scaling the reference wave velocity for the surrounding rock proportionally, the theoretical propagation time can be aligned to the actual observed value, thereby deriving the actual local wave velocity of the surrounding rock at the anchor location. Each record in the sparse wave velocity calibration point set includes the calibration three-dimensional coordinates of a confirmed anchor and the corresponding actual local wave velocity of the surrounding rock. These records are sparsely distributed in the three-dimensional space of the tunnel.

[0090] Using the sparse wave velocity calibration point set as spatial interpolation constraint points, anisotropic kriging interpolation is used to reconstruct the three-dimensional wave velocity field of the surrounding rock of the excavated section, and the depth difference described in claim 3 is recalculated using the three-dimensional wave velocity field.

[0091] Specifically, the anisotropic kriging interpolation allows for different range parameters in three spatial directions: a larger range along the tunnel axial direction, reflecting the slow axial change of the surrounding rock wave velocity; a moderate range along the tunnel radial direction (depth direction); and a range along the tunnel height direction, which can be set according to the tunnel cross-sectional dimensions and the surrounding rock stratification. The semi-variogram model for the anisotropic kriging interpolation can be an exponential model, a spherical model, or a Gaussian model. By performing kriging interpolation on the sparse wave velocity calibration point set, the interpolated wave velocity at any spatial location within the excavated section of the surrounding rock area is obtained, constituting the three-dimensional wave velocity field. Using this three-dimensional wave velocity field, when calculating the depth difference in the secondary discrimination, the interpolated wave velocity at the target location is used instead of the surrounding rock reference wave velocity for depth conversion, resulting in a depth difference value that more closely approximates the true wave velocity distribution of the surrounding rock.

[0092] When the preset wave velocity calibration conditions are not met, the depth difference as described in claim 3 is calculated using the surrounding rock reference wave velocity.

[0093] Specifically, when the calibrated three-dimensional coordinate data of the anchor bolt is incomplete or the radar echo signal-to-noise ratio does not meet the wave velocity calibration requirements, the reference wave velocity of the surrounding rock is used for depth conversion to avoid introducing a larger depth estimation error due to insufficient calibration data.

[0094] In summary, the aforementioned wave velocity calibration scheme uses the confirmed anchor bolts as calibration points for natural wave velocities in the surrounding rock of the roadway. It inverts the local actual wave velocity using the known spatial location of the anchor bolts and the time-distance relationship of radar observation responses, and then extends this to the entire space through anisotropic kriging interpolation, thus upgrading the surrounding rock wave velocity from a nominal reference value to a spatial distribution field. The depth difference calculation based on the three-dimensional wave velocity field is more accurate than the calculation based on a single reference wave velocity, thereby improving the reliability of the depth criterion in the secondary discrimination.

[0095] In this embodiment, the calibrated three-dimensional coordinates are the actual three-dimensional coordinates of the anchor bolts obtained after completion. The preset wave velocity calibration conditions include: the data integrity of the actual three-dimensional coordinates of the anchor bolts reaches a preset integrity threshold; the proportion of anchor bolts in the confirmed anchor bolt set whose signal-to-noise ratio exceeds a preset signal-to-noise ratio threshold reaches a preset proportion threshold; and the number of anchor bolts that meet the signal-to-noise ratio requirement reaches a preset number of calibration points.

[0096] In specific implementation, the data integrity is defined as the percentage of anchors with acquired calibration 3D coordinates out of the total number of anchors in the confirmed anchor set, and the preset integrity threshold can be set to 40% to 60%. The signal-to-noise ratio is defined as the ratio of the maximum amplitude of the anchor diffraction signal within a local window centered on the theoretical vertex coordinates of the anchor in the B-scan image to the root mean square amplitude of the background noise within that window. The preset signal-to-noise ratio threshold can be ten to twenty times, the preset ratio threshold can be set to 60% to 80%, and the preset number of calibration points can be five to ten. Before constructing the sparse wave velocity calibration point set, abnormal anchors with a deviation between the theoretical two-way propagation time and the actual two-way propagation time exceeding a preset time deviation threshold are removed. A typical value for the preset time deviation threshold is 1.5 to 3 times the main period of the wavelet transmitted by the mine ground-penetrating radar.

[0097] In this embodiment, the training method of the conditional generative adversarial network includes the following steps: Based on the support design parameters, the surrounding rock reference wave velocity, and the preset geological anomaly parameters, a time-domain finite-difference forward modeling simulation is performed to generate a pure anchor bolt B-scan image containing only anchor bolt diffraction signals, a pure geological anomaly B-scan image containing only geological anomaly diffraction signals, and a hybrid B-scan image containing both anchor bolt diffraction signals and geological anomaly diffraction signals.

[0098] Specifically, the finite-difference time-domain forward modeling is a numerical method for solving Maxwell's equations. A two-dimensional or three-dimensional forward modeling mesh is established, comprising a tunnel cross-section geometric model, a surrounding rock medium model, a rock bolt model, and a geological anomaly model. The surrounding rock medium model uses the relative permittivity corresponding to the reference wave velocity of the surrounding rock as the background medium parameter, and can be superimposed with preset random medium perturbations to simulate the non-homogeneity of the surrounding rock. The rock bolt model adopts a geometric model of a metal cylinder or metal rod, arranged according to the three-dimensional spatial coordinates and spacing of the rock bolts in the support design parameters. The geological anomaly model adopts a geometry defined by preset geological anomaly parameters, including the anomaly type (cavity, water-bearing fissure, or fault fracture zone), the anomaly's geometric dimensions, spatial location, and relative permittivity. Excitation sources and receiving points for transmitting and receiving antennas are set in the forward modeling mesh, and the antennas are moved sequentially according to their movement trajectories, recording the received waveform of each trajectories to form a frame of B-scan image.

[0099] Three B-scan images were generated through three independent finite-difference temporal modeling simulations: First, a pure anchor B-scan image was generated by retaining only the anchor model and removing the geological anomaly model in the forward model; second, a pure geological anomaly B-scan image was generated by retaining only the geological anomaly model and removing the anchor model in the forward model; third, a hybrid B-scan image was generated by retaining both the anchor model and the geological anomaly model in the forward model. These three B-scan images were perfectly aligned in terms of trace spacing, temporal sampling interval, and antenna movement trajectory, forming a pixel-level registered training sample triplet.

[0100] A corresponding training spatial prior mask is generated based on the spatial location of the anchor diffraction signal in the pure anchor B-scan image.

[0101] Specifically, in the pure anchor B-scan image, the position of the anchor diffraction signal is entirely determined by the anchor spatial coordinates in the forward model, eliminating the measurement uncertainty present in the actual data. Therefore, based directly on the anchor's three-dimensional spatial coordinates in the support design parameters, the antenna's movement trajectory, and the surrounding rock reference wave velocity, a search window for each anchor is generated according to the calculation method for the theoretical vertex coordinates in claim 1, thereby generating the training spatial prior mask. The generation accuracy of the training spatial prior mask is higher than that of the pixel-level spatial prior mask in the actual data, providing accurate spatial guidance information for the conditional generative adversarial network.

[0102] Using the hybrid B-scan image and the corresponding training space prior mask as conditional inputs, and the corresponding pure geological anomaly B-scan image as the supervision target, the conditional generative adversarial network is trained to separate geological signal components from the hybrid B-scan image.

[0103] Specifically, the conditional generative adversarial network includes a generator network and a discriminator network. The generator network takes the multi-channel data (consisting of the hybrid B-scan image and the training space prior mask along the channel dimension) as conditional input and outputs a predicted geological signal component image. The discriminator network takes the predicted geological signal component image output by the generator, or the actual pure geological anomaly B-scan image, and concatenates them with the conditional input as the judgment object, outputting a probability value for determining whether the image is true or false. During training, the loss functions of the generator network and the discriminator network are alternately optimized. The loss function of the generator network consists of a weighted combination of adversarial loss and pixel-level reconstruction loss: the adversarial loss is the probability that the discriminator network judges the generator output as false; the pixel-level reconstruction loss is the pixel-wise mean square error between the predicted geological signal component image output by the generator and the pure geological anomaly B-scan image used as the supervision target. The loss function of the discriminator network is a standard binary classification cross-entropy loss. After being fully trained with a large number of randomly generated training sample triples, the conditional generative adversarial network has the ability to recover geological signal components from mixed B-scan images, and the training spatial prior mask enables the generator network to learn to use anchor spatial location information to assist in signal separation.

[0104] In this embodiment, the method further includes the following steps: The system gain calibration parameters, antenna coupling correction parameters, and geometric reflection parameters of each confirmed anchor bolt are obtained from the mine ground-penetrating radar. Based on the matching tracking described in claim 4, the amplitude scaling factor of each confirmed anchor bolt in multiple frequency bands is determined.

[0105] Specifically, the system gain calibration parameter is a quantitative conversion relationship between the transmit power and the digital value of the received signal of the mine ground-penetrating radar host under different gain settings, which can be obtained through factory calibration or on-site metal plate reflection calibration. The antenna coupling correction parameter is used to eliminate the system background signal caused by direct coupling and surface coupling between the transmitting and receiving antennas, and can be obtained through antenna-to-air calibration measurement in an open area. The geometric reflection parameters include the cross-sectional dimensions of the anchor rod, the material conductivity, and the azimuth angle relative to the antenna polarization direction. As a long and slender metal cylinder, the geometric reflection parameters of the anchor rod determine the scattering directionality of radar electromagnetic waves on the anchor rod surface. The multiple frequency bands are obtained by dividing the effective frequency band range of the mine ground-penetrating radar into equal intervals or octaves, and the number of frequency bands can be three to five. For each frequency band, after bandpass filtering of the B-scan image data, the amplitude scaling factor of each confirmed anchor rod in that frequency band is recalculated according to the matching tracking method described in claim 4, to obtain the amplitude scaling factor sequence of the anchor rod in multiple frequency bands.

[0106] Based on the propagation distance of each confirmed anchor bolt, the actual local wave velocity of the surrounding rock, the system gain calibration parameters, the antenna coupling correction parameters, the geometric reflection parameters, and the amplitude scaling factors in the multiple frequency bands, the local relative permittivity and electromagnetic attenuation coefficient of the surrounding rock at each confirmed anchor bolt location are jointly inverted.

[0107] Specifically, for a given anchor bolt, the amplitude scaling factor within a specific frequency band. There is a quantitative relationship between the amplitude scaling factor and the propagation and reflection process of electromagnetic waves in the surrounding rock medium. According to the radar equation, the amplitude scaling factor can be expressed as: In the formula, The center frequency of the frequency band; The equivalent gain factor for the system gain calibration parameters; This is the equivalent correction factor for the antenna coupling correction parameters; The radar reflection coefficient of the anchor bolt is the relative permittivity of the surrounding rock. Functions related to the geometric reflection parameters; The electromagnetic attenuation coefficient is the frequency. The function; The one-way propagation distance from the anchor to the antenna position is calculated based on the calibrated three-dimensional coordinates of the anchor and the antenna's movement trajectory. For multiple frequency bands, this is configured as follows: A system of simultaneous equations is formed, with the frequency dependence parameter of the electromagnetic attenuation coefficient being unknown. The local relative permittivity of the surrounding rock and the electromagnetic attenuation coefficient at the anchor bolt location are obtained by solving the system of simultaneous equations using least-squares fitting or nonlinear optimization. The electromagnetic attenuation coefficient can be determined using an exponential frequency dependence model. Perform parameterization, where This is a reference value for the attenuation coefficient. The frequency dependence index is used as both inversion variables and solved together.

[0108] Using the local relative permittivity and electromagnetic attenuation coefficient of the surrounding rock at each confirmed anchor location as interpolation constraint points, the distribution fields of the relative permittivity and electromagnetic attenuation coefficient of the surrounding rock are reconstructed respectively. Based on the joint anomaly characteristics of the distribution fields of the relative permittivity and electromagnetic attenuation coefficient of the surrounding rock, candidate regions for water-bearing anomalies and candidate regions for fracture development are identified.

[0109] Specifically, using the same interpolation strategy as the anisotropic kriging interpolation described in claim 7, spatial interpolation is performed on the local relative permittivity of the surrounding rock and the electromagnetic attenuation coefficient to obtain the distribution fields of the relative permittivity and electromagnetic attenuation coefficient of the surrounding rock. Joint anomaly analysis is performed on the two distribution fields: high-value anomaly regions in the relative permittivity distribution field indicate high water content or abrupt changes in lithological composition in the surrounding rock; high-value anomaly regions in the electromagnetic attenuation coefficient distribution field indicate enhanced absorption of electromagnetic waves by the surrounding rock, typically associated with water-bearing fractures or clay mineral enrichment areas. Spatial locations showing positive anomalies in both distribution fields are marked as candidate areas for water-bearing anomalies; spatial locations showing positive anomalies in the relative permittivity distribution field and local strip-shaped anomalies in the electromagnetic attenuation coefficient distribution field are marked as candidate areas for fracture development. The candidate areas for water-bearing anomalies and the candidate areas for fracture development serve as target areas for subsequent detailed exploration or drilling verification, providing auxiliary decision-making basis for tunnel excavation safety assessment.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive denoising method for mine ground-penetrating radar images based on generative adversarial networks, characterized in that, include: Acquire mine geological radar B-scan images, support design parameters, antenna movement trajectory, surrounding rock reference wave velocity, and historically confirmed anchor bolt sets for the current detection section; Based on the three-dimensional spatial coordinates of each anchor in the current detection segment, the trajectory of the antenna movement, and the reference wave velocity of the surrounding rock, calculate the theoretical vertex coordinates and theoretical curvature of each anchor in the mine geological radar B-scan image; Based on the anchor bolt number, the anchor bolt column number, and the roadway mileage, establish the spatial correspondence between each anchor bolt in the current detection segment and the historically confirmed anchor bolts. Based on the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of the corresponding historically confirmed anchor bolts, correct the theoretical vertex coordinates of the corresponding anchor bolts in the current detection segment. Determine the search window with the corrected theoretical vertex coordinates as the center, and generate a pixel-level spatial prior mask based on each search window. Extract hyperbolic regions from the mine geological radar B-scan image, determine the actual vertex coordinates and actual curvature of each hyperbolic region, mark hyperbolic regions whose actual vertex coordinates fall into the corresponding search window and whose actual curvature is within the preset curvature tolerance range as anchor diffraction zones, and add the corresponding anchors to the confirmed anchor set, and mark hyperbolic regions that do not match any anchors as suspected geological anomaly zones; Based on the three-dimensional spatial coordinates of each anchor in the confirmed anchor set, the antenna movement trajectory, and the surrounding rock reference wave velocity, a corresponding theoretical diffraction template is generated. The diffraction contribution component of each confirmed anchor is separated from the mine ground radar B-scan image by matching and tracking to obtain the residual B-scan image. For suspected geological anomaly areas in the residual B-scan image, a secondary discrimination is performed based on the depth difference between the suspected geological anomaly area and the confirmed anchor bolt and the periodic verification along the tunnel axis. The area confirmed by the secondary discrimination to contain geological anomaly signals is marked as a protection zone, and image enhancement is performed on the protection zone to obtain an enhanced image.

2. The method according to claim 1, characterized in that, Iteratively updating the pixel-level spatial prior mask includes: The matching confidence is determined based on the vertex position deviation, curvature deviation, and template relevance of the candidate hyperbolic regions within the search window corresponding to each unconfirmed anchor. The template relevance is the normalized cross-correlation between the candidate hyperbolic region and the theoretical diffraction template of the corresponding anchor. The matching confidence decreases as the vertex position deviation and curvature deviation increase, and increases as the template relevance increases. According to the order of matching confidence from high to low, a preset number of unconfirmed anchor bolts are selected for spatial matching; Accumulate the positional deviation between the actual vertex coordinates and the theoretical vertex coordinates of newly confirmed anchor bolts in the current round and previous rounds. When the number of accumulated positional deviation samples reaches the preset number of samples, interpolate the accumulated positional deviation samples along the channel number direction and time depth direction of the B-scan image to obtain the deviation field. The search window positions of the remaining unconfirmed anchors and the pixel-level spatial prior mask are corrected based on the deviation field, and the spatial matching is re-executed until no new confirmed anchors are added in this round or the number of iterations reaches the preset number.

3. The method according to claim 1, characterized in that, The secondary discrimination of the suspected geological anomaly area includes: Based on the antenna movement trajectory, establish the mapping relationship between the tunnel number and the axial mileage of the mine geological radar B-scan image; Obtain the anchor bolt spacing in the support design parameters, and obtain multiple suspected geological anomaly areas within a preset neighborhood along the tunnel axis of the target suspected geological anomaly area; Calculate the depth difference between the actual vertex coordinates of the target suspected geological anomaly zone and the theoretical vertex coordinates of the most recently confirmed anchor bolt, and determine the periodic matching degree between the multiple suspected geological anomaly zones and the anchor bolt spacing based on the vertex position sequence of the multiple suspected geological anomaly zones; When the depth difference is greater than a preset depth threshold and the period matching degree is lower than a preset period matching threshold, it is confirmed that the target suspected geological anomaly area contains geological anomaly signals, and the target suspected geological anomaly area is marked as a protected area.

4. The method according to claim 1, characterized in that, The step of separating the diffraction contribution components of each confirmed anchor bolt from the mine ground-penetrating radar B-scan image by matching and tracking includes: The theoretical diffraction templates of each confirmed anchor bolt are used to form a theoretical diffraction template dictionary; In each round of matching and tracking, the target theoretical diffraction template with the highest correlation to the current residual image is selected from the theoretical diffraction template dictionary. The magnitude scaling factor and phase offset are used as optimization variables to construct the residual minimization objective function between the target theoretical diffraction template and the current residual image within the corresponding search window. Solve the objective function of minimizing the residual to obtain the optimal amplitude scaling factor and optimal phase offset corresponding to the target theoretical diffraction template. Determine the diffraction contribution component of the corresponding anchor based on the target theoretical diffraction template, the optimal amplitude scaling factor, and the optimal phase offset. The diffraction contribution component is subtracted from the current residual image to update the current residual image until the residual energy decrease of the current round is less than a preset energy threshold or the number of matching and tracking reaches a preset number. The final current residual image is then determined as the residual B-scan image.

5. The method according to claim 1, characterized in that, The image enhancement performed on the protected area includes: Perform a two-dimensional Fourier transform on the image data corresponding to the protected area to obtain FK domain data; Based on the surrounding rock reference wave velocity, the center frequency of the mine ground radar and the antenna movement trajectory, the target frequency-wavenumber band corresponding to the diffraction energy of the geological anomaly is determined in the FK domain data. Gain is applied to the frequency-wavenumber components within the target frequency-wavenumber band, and attenuation is applied to the frequency-wavenumber components outside the target frequency-wavenumber band to obtain enhanced FK domain data. The enhanced FK domain data is subjected to a two-dimensional inverse Fourier transform and then merged with the non-protected area image data in the residual B-scan image to form the enhanced image.

6. The method according to claim 1, characterized in that, Also includes: Detect the blurred overlapping area where the pixel coverage of the anchor bolt diffraction area overlaps with the pixel coverage of the protected area; Extract residual image blocks containing the blurred overlapping regions from the residual B-scan image, and extract spatial prior mask blocks corresponding to the positions of the residual image blocks from the pixel-level spatial prior mask; The residual image block and the spatial prior mask block are input together into a pre-trained conditional generative adversarial network to obtain the geological signal component image corresponding to the blurred overlapping region. The geological signal component image and the corresponding region image data in the enhanced image are weighted and fused to update the blurred overlapping region in the enhanced image.

7. The method according to claim 4, characterized in that, Also includes: Obtain the calibrated three-dimensional coordinates of each anchor in the confirmed anchor set, and recalculate the theoretical two-way propagation time of each anchor corresponding to the theoretical vertex coordinates based on the calibrated three-dimensional coordinates, the antenna movement trajectory, and the surrounding rock reference wave velocity; Determine whether the calibrated three-dimensional coordinates and the radar echo of the confirmed anchor bolt meet the preset wave velocity calibration conditions; When the preset wave velocity calibration conditions are met, the reference wave velocity of the surrounding rock is multiplied by the ratio of the theoretical two-way propagation time and the actual two-way propagation time corresponding to each confirmed anchor bolt to obtain the local actual wave velocity of the surrounding rock at each confirmed anchor bolt position, and a sparse wave velocity calibration point set is constructed based on the calibration three-dimensional coordinates of each confirmed anchor bolt and the local actual wave velocity of the surrounding rock. Using the sparse wave velocity calibration point set as spatial interpolation constraint points, anisotropic kriging interpolation is used to reconstruct the three-dimensional wave velocity field of the surrounding rock of the excavated section, and the depth difference described in claim 3 is recalculated using the three-dimensional wave velocity field. When the preset wave velocity calibration conditions are not met, the depth difference as described in claim 3 is calculated using the surrounding rock reference wave velocity.

8. The method according to claim 7, characterized in that: The calibrated three-dimensional coordinates are the actual three-dimensional coordinates of the anchor bolts obtained after completion of the project. The preset wave velocity calibration conditions include: the data integrity of the actual three-dimensional coordinates of the anchor bolt reaches a preset integrity threshold, the proportion of anchor bolts in the confirmed anchor bolt set whose signal-to-noise ratio exceeds a preset signal-to-noise ratio threshold reaches a preset proportion threshold, and the number of anchor bolts that meet the signal-to-noise ratio requirement reaches a preset calibration point number. Before constructing the sparse wave velocity calibration point set, abnormal anchors with deviations exceeding a preset time deviation threshold between the theoretical two-way propagation time and the actual two-way propagation time are removed.

9. The method according to claim 6, characterized in that, The training methods for the conditional generative adversarial network include: Based on the support design parameters, the surrounding rock reference wave velocity, and the preset geological anomaly parameters, a time-domain finite difference forward modeling simulation is performed to generate a pure anchor bolt B-scan image containing only anchor bolt diffraction signals, a pure geological anomaly B-scan image containing only geological anomaly diffraction signals, and a hybrid B-scan image containing both anchor bolt diffraction signals and geological anomaly diffraction signals. A corresponding training spatial prior mask is generated based on the spatial location of the anchor diffraction signal in the pure anchor B-scan image. Using the hybrid B-scan image and the corresponding training space prior mask as conditional inputs, and the corresponding pure geological anomaly B-scan image as the supervision target, the conditional generative adversarial network is trained to separate geological signal components from the hybrid B-scan image.

10. The method according to claim 7, characterized in that, Also includes: The system gain calibration parameters, antenna coupling correction parameters, and geometric reflection parameters of each confirmed anchor bolt are obtained from the mine ground radar. According to the matching tracking described in claim 4, the amplitude scaling factor of each confirmed anchor bolt in multiple frequency bands is determined respectively. Based on the propagation distance of each confirmed anchor bolt, the actual local wave velocity of the surrounding rock, the system gain calibration parameters, the antenna coupling correction parameters, the geometric reflection parameters, and the amplitude scaling factors in the multiple frequency bands, the local relative permittivity and electromagnetic attenuation coefficient of the surrounding rock at each confirmed anchor bolt location are jointly inverted. Using the local relative permittivity and electromagnetic attenuation coefficient of the surrounding rock at each confirmed anchor location as interpolation constraint points, the distribution fields of the relative permittivity and electromagnetic attenuation coefficient of the surrounding rock are reconstructed respectively. Based on the joint anomaly characteristics of the distribution fields of the relative permittivity and electromagnetic attenuation coefficient of the surrounding rock, candidate regions for water-bearing anomalies and candidate regions for fracture development are identified.

Images