A tunnel surrounding rock intelligent evaluation method based on fusion of TGS multi-channel images and geological sketches

Abstract

The application provides a tunnel surrounding rock intelligent evaluation method based on TGS multi-channel image and geological sketch fusion, and belongs to the technical field of tunnel advanced geological prediction and surrounding rock stability evaluation. In view of the problems that the existing TGS slice manual interpretation difference is large, the space does not match the geological sketch, the structure surface input is unstable, and the surrounding rock evaluation and block risk cannot be linked, the method collects TGS three types of slice data, the structure surface parameters of the working face and the construction conditions, after unified coordinate system registration preprocessing, extracts the spatial features of the TGS image, standardizes and sorts the structure surface parameters and encodes, obtains two types of fusion features through the cross-modal fusion constrained by geometric correlation, and finally outputs the surrounding rock integrity grade, the top arch block condition risk and the review prompt in sections. The application improves the automation level and the result repeatability of the surrounding rock evaluation, and can provide technical support for the safety warning of the tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel advanced geological prediction and surrounding rock stability evaluation technology, and in particular to an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches. Background Technology

[0002] In the field of advanced geological prediction for tunnels and underground engineering, the integrity of the surrounding rock is the core basis for evaluating the stability of the tunnel face, selecting initial support parameters, and controlling construction risks. The TGS advanced geological prediction system can output three types of parameter slices based on seismic wave detection: stress gradient, water-bearing probability, and P-wave velocity. These slices can be used to assist in analyzing the abnormal response, water-bearing probability, and elastic wave propagation characteristics of the rock mass ahead. These three types of results can provide information from different aspects. The geological sketch of the tunnel face can record geometric and occurrence parameters such as the attitude, spacing, and roughness of structural surfaces, which is a direct basis for judging the stability of local blocks.

[0003] In current engineering applications, surrounding rock evaluation mainly relies on geological engineers interpreting various TGS slices separately and combining them with geological sketches to draw conclusions manually. This approach has several technical shortcomings: First, the three types of TGS slices are often interpreted independently, making it difficult to quantitatively identify complex anomaly patterns coupled with multiple parameters. Manual interpretation is also greatly influenced by experience. Second, geological sketches of the tunnel face are mostly saved in text and image format, without establishing a unified engineering coordinate system and quantitative geometric correlation with the TGS prediction window ahead. This makes it impossible to accurately locate the spatial position of segmented anomalies ahead and their corresponding control structural surfaces. Third, the number of multiple sets of structural surfaces is variable, and the input order is not fixed, making it difficult to form a stable and repeatable evaluation input. At the same time, existing evaluations mostly output conclusions on the overall integrity of long sections, without combining excavation methods, support conditions, and other construction conditions to output risk indicators for roof collapse and data quality verification prompts, which is insufficient to meet the needs of refined construction safety management.

[0004] Therefore, it is necessary to propose an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches to solve or at least alleviate the above-mentioned defects. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches, in order to solve the problems of existing technologies such as lack of automated spatial joint interpretation, inconsistency with the geological sketch coordinates of the tunnel face, unstable input order of structural surfaces leading to poor evaluation repeatability, and inability to quantitatively output the integrity of the surrounding rock and the risk of roof collapse conditions.

[0006] To achieve the above objectives, this invention provides an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches, comprising the following steps: S1. Collect TGS slice data and obtain the structural surface parameters and construction conditions of the geological sketch of the working face. Perform registration and preprocessing of the TGS slice data and the structural surface parameters under a unified engineering coordinate system, and encode the construction conditions. The TGS slice data includes stress gradient slice map, water content probability slice map and P-wave velocity slice map. S2, extract features from the processed TGS slice data to obtain spatial feature maps and feature grid quality masks; S3. The parameters of each group of structural surfaces are concatenated to obtain structural surface feature vectors. The structural surface feature vectors of each group are then standardized and sorted. The encoded features of each group of structural surfaces are extracted based on the standardized and sorted structural surface feature vectors. S4. Based on the geometric relationship between the structural surface and the TGS grid under the unified engineering coordinate system, the association constraints for the evaluation of the surrounding rock integrity and the evaluation of the risk of the top arch falling off are constructed respectively. Under the association constraints, the spatial feature map is fused with the coding features of each group of structural surfaces to obtain the integrity fusion feature and the falling off risk fusion feature. S5, the integrity fusion feature and the rockfall risk fusion feature are weighted and converged using the feature mesh quality mask to form integrity segmentation feature and rockfall risk segmentation feature, and the tunnel surrounding rock evaluation result is output according to the integrity segmentation feature and the rockfall risk segmentation feature.

[0007] Preferably, step S1 specifically includes the following steps: S11, the TGS advanced geological prediction system is used to conduct detection at the current tunnel face station number to obtain stress gradient slice map, water content probability slice map and P-wave velocity slice map within the prediction window ahead. S12 records the structural parameters and construction conditions of multiple structural surfaces at the current tunnel face; wherein, the structural surface parameters include the dip direction, dip angle, average spacing, roughness, opening, extension length, infill type, groundwater status, and visible traces of the structural surface; the construction conditions include the excavation method, blasting parameters, initial support status, and unsupported exposure time; at the same time, the tunnel axis orientation, tunnel face station number, excavation outline, and arch crown range are recorded as engineering geometric information; S13, a unified engineering coordinate system is established with the tunnel axis mileage as the X-axis, the cross-section transverse direction as the Y-axis, and the vertical direction as the Z-axis. The TGS slice data is mapped to the engineering coordinates within the prediction window, and the structural surface attitude and visible traces are converted into structural surface normal vectors and spatial positions. S14, normalize the mapped TGS slice data and generate a data quality mask and channel validity markers.

[0008] Preferably, the step between step S13 and step S14 further includes: S131, regarding the first The first TGS grid and the first The basic correlation coefficients for structural surfaces are used to represent the reliable extrapolation of structural surfaces, potential intersections with the cavity wall, and the relationship between the mesh positions. S132, for candidate structural surface combinations consisting of two or three sets of structural surfaces and the free face of the cavern. Calculate the potential wedge geometric compatibility coefficient and combine it with the basic correlation coefficient to form the block drop risk correlation coefficient, as follows: in, The correlation coefficient for the risk of block dropping. Based on the correlation coefficient, is the geometric compatibility coefficient of the potential wedge.

[0009] Preferably, step S2 specifically includes the following steps: S21, stack the normalized stress gradient slice, water content probability slice, and P-wave velocity slice into a three-channel image according to the channel dimension; S22, perform masked global average pooling on each channel, calculate the original channel weights through the channel attention module, and re-normalize the channel weights by combining the effective channel labels to obtain the normalized channel weights, and weight each channel image based on the normalized channel weights. S23 employs three parallel convolutional branches: 3×3, 5×5, and 7×7. The features from each branch are concatenated and then processed by adaptive pooling and 1×1 convolution to obtain a spatial feature map. S24, Project the basic correlation coefficient, the block drop risk correlation coefficient, and the data quality mask onto the feature grid where the spatial feature map is located to obtain the feature grid quality mask.

[0010] Preferably, step S3 specifically includes the following steps: S31 converts the dip and dip angle in each group of structural surface parameters into sine and cosine representations, normalizes the spacing, roughness, opening and extension length, and concatenates them with the filling material, groundwater state, visible trace spatial location encoding, observation confidence and corresponding missing marker vector to form a structural surface feature vector. S32, determine the geometric sorting key in sequence according to the projection position of the visible trace of the structural surface on the outline of the working face, the distance from the potential intersection line of the top arch, the dip angle, the inclination angle and the spacing; S33, standardize and sort the feature vectors of each group of structural surfaces based on the geometric sorting key; S34. Input the standardized and sorted feature vectors of each group of structural surfaces into a bidirectional LSTM to obtain the encoded features of each group of structural surfaces.

[0011] Preferably, step S4 specifically includes the following steps: S41, the spatial feature map is unfolded according to a grid to obtain image features; S42, Construct task correlation coefficients; wherein, for the surrounding rock integrity evaluation task, the task correlation coefficient adopts the basic correlation coefficient; for the roof collapse risk evaluation task, the task correlation coefficient adopts the collapse risk correlation coefficient. S43, for each TGS grid When the feature mesh quality mask is effective and the task correlation coefficient is greater than 0, the geometric constraint attention weight is calculated, and the encoded features are weighted and converged in combination with the geometric constraint attention weight to obtain the structural surface fusion feature of the task corresponding to the mesh. S44, the image features, the structural surface fusion features of the corresponding task, and the feature mesh quality mask are stitched together and then subjected to light convolution processing to obtain the integrity fusion features and the block drop risk fusion features, respectively.

[0012] Preferably, step S5 specifically includes the following steps: S51, along the TGS prediction window, divide into several evaluation segments according to the preset mileage length, and divide the cross-sections according to the top arch and sidewalls. S52, the integrity fusion feature and the block drop risk fusion feature in each segment are weighted and converged using the feature mesh quality mask to obtain the integrity segment feature and the block drop risk segment feature of each segment; S53, for the first with an effective grid Each evaluation segment outputs the tunnel surrounding rock evaluation results; wherein, the tunnel surrounding rock evaluation results include the distribution of surrounding rock integrity levels, the risk of roof collapse under construction conditions, confidence index, abnormal grid locations and information on major control structural surfaces.

[0013] Preferably, the method further includes the following steps before step S1: A model was constructed, and a weighted sum of the surrounding rock integrity level classification loss and the rockfall condition risk loss was used to jointly train the model, as follows: in, The weighted cross-entropy loss is the result of the surrounding rock integrity level. The weighted binary cross-entropy loss is used to calculate the risk of block drop conditions. λ Risk task weighting.

[0014] Preferably, step S131 is specifically calculated using the following formula: in, Based on the correlation coefficient, For the first The reliability of observations of the group structure surface. For indicator functions, The structural surface is defined by the visible traces or theoretical intersection lines from the working face towards the mesh. extrapolated distance, To correlate the scale parameter that decays with extrapolated distance, To achieve the maximum reliable extrapolation distance, It is the coefficient by which the extended plane of the structural surface intersects or approaches the open surface of the cavern within a preset spatial tolerance.

[0015] Preferably, step S53 is followed by the following step: S54 calculates the integrity level confidence index, the reliability index of block drop risk output, the proportion of valid TGS data, and the structural surface information quality index when the model outputs, and outputs the corresponding reason codes so that engineers can classify and handle the situation in conjunction with the on-site log.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches. It achieves registration and preprocessing of TGS three-type slices, tunnel face structural parameters, and construction conditions through a unified engineering coordinate system, establishing verifiable spatial correspondences between multi-source data, which helps reduce spatial misalignment during manual processing. Through multi-scale feature extraction processing of TGS multi-channel image data, it preserves the engineering mileage, cross-sectional lateral position, and slice layer information of anomaly areas, providing input for joint analysis of multi-parameter composite anomalies. By standardizing and sorting the structural surface feature vectors before extracting coded features, it improves the evaluation of input parameters and sorting... Under the same order rules, the influence of manual numbering and the order of entry on the input sequence is eliminated, ensuring the repeatability and stability of the variable number of structural surfaces input. Through a cross-modal fusion mechanism constrained by geometric association, integrity fusion features and rockfall risk fusion features are generated respectively, so that the two types of evaluation adopt the corresponding structural surface association constraints and retain spatial position. Finally, based on the feature grid quality mask weighted convergence, segmented features are formed, and the segmented surrounding rock integrity level, crown rockfall condition risk and quality review prompts are output simultaneously. This helps to improve the automation level and traceability of the evaluation process, and provides segmented information for on-site review. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the evaluation method in one embodiment of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0023] Please see the appendix Figure 1 An embodiment of the present invention provides an intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches, comprising the following steps: S1. Acquire TGS slice data and obtain structural surface parameters and construction conditions from the geological sketch of the tunnel face. Register and preprocess the TGS slice data and structural surface parameters under a unified engineering coordinate system, and encode the construction conditions. The TGS slice data includes stress gradient slices, water-bearing probability slices, and P-wave velocity slices. The preprocessing process first establishes a unified engineering coordinate system to complete the spatial alignment of multi-source data, then calculates the geometric correlation coefficient between the structural surface and the TGS grid, and finally completes numerical normalization and quality mask generation. This step incorporates geophysical results, geological logging information, and construction status information into the same standardized processing flow, unifying the input format, spatial reference, and normalization caliber. This provides input data with accurate spatial coordinates, quality indicators, and geometric correlations for subsequent cross-modal feature fusion, reducing differences from manual processing and retaining prompts for insufficient data quality.

[0024] S2 involves feature extraction from the processed TGS slice data to obtain spatial feature maps and feature grid quality masks. During implementation, multi-scale spatial feature extraction is performed on the processed TGS slice data. First, channel weights are adaptively adjusted through channel attention. Then, multi-scale parallel convolution branches extract spatial features at different scales, ultimately yielding spatial feature maps carrying corresponding spatial location information and feature grid quality masks for the corresponding feature grids. The feature extraction process preserves the two-dimensional spatial dimension throughout, avoiding direct compression into a single global feature vector. This ensures that each extracted feature grid corresponds one-to-one with the actual engineering mileage and cross-sectional lateral position. This step completes deep feature extraction from multi-channel images while preserving spatial location. It enhances the weight of effective information through channel attention and covers anomalous responses of different sizes through multi-scale convolution, providing a precise spatial feature foundation for subsequent grid-by-grid cross-modal fusion.

[0025] S3 involves concatenating the structural surface parameters of each group to obtain structural surface feature vectors. These feature vectors are then standardized and sorted. Based on the standardized and sorted feature vectors, the encoded features of each structural surface are extracted. This process, involving standardized sorting and sequence coding, handles a variable number of structural surface parameters. First, the discrete structural surface parameters are converted into standardized feature vectors. Then, a geometric sorting rule, independent of manual numbering, is used to standardize and sort all structural surface feature vectors. Finally, a variable-length sequence coding network is employed to extract the encoded features and combined features of each structural surface group. Standardized sorting ensures a consistent input sequence for the same working face when parameters and sorting rules are consistent. Variable-length coding adapts to field scenarios with a variable number of structural surfaces. This step eliminates the influence of manual numbering and input order on the structural surface coding results, ensuring the repeatability and stability of the variable number of structural surface inputs, while effectively extracting combined correlation features between multiple groups of structural surfaces.

[0026] S4. Based on the geometric association between the structural surfaces and the TGS mesh under the unified engineering coordinate system, association constraints are constructed for the evaluation of surrounding rock integrity and the risk assessment of roof collapse, respectively. Under these association constraints, the spatial feature map is fused with the coded features of each group of structural surfaces to obtain integrity fusion features and collapse risk fusion features. Task-oriented fusion based on geometric constraints achieves feature decoupling between the two evaluation tasks. During implementation, the spatial feature map is unfolded grid by grid. For the two tasks of surrounding rock integrity and roof collapse risk, the corresponding geometric association coefficients are used as constraints to calculate the structural surface fusion features corresponding to each feature grid. Then, the image features and structural surface fusion features are concatenated and convolved to obtain the integrity fusion features for the corresponding surrounding rock integrity evaluation and the collapse risk fusion features for the corresponding roof collapse risk evaluation. The fusion process is only performed between grids and structural surfaces with effective geometric association; invalid grids and unassociated structural surfaces do not participate in the normalization calculation. This step uses corresponding geometric correlation coefficients to constrain different evaluation tasks, ensuring that structural surface features only participate in fusion within grids with effective correlations. This guarantees the spatial rationality of the fusion process and allows the two types of evaluations to use their respective structural surface correlation constraints and output branches.

[0027] S5, the integrity fusion feature and the rockfall risk fusion feature are weighted and converged using the feature mesh quality mask to form integrity segmentation features and rockfall risk segmentation features. The tunnel surrounding rock evaluation results are then output based on these features. Multi-dimensional surrounding rock evaluation results are output through segmented weighted convergence. During implementation, the evaluation is divided into segments along the prediction window according to a preset mileage. The integrity fusion feature and rockfall risk fusion feature within each segment are weighted and converged using the feature mesh quality mask to form the corresponding segment's integrity segmentation features and rockfall risk segmentation features. Finally, an independent output head outputs tunnel surrounding rock evaluation results containing integrity level, rockfall risk, confidence level, and auxiliary information based on the two types of segmentation features. This step outputs evaluation results segmented according to commonly used engineering mileage, aligning with the segmented control requirements of on-site construction management. Simultaneously, the weighted quality mask ensures the effectiveness of the segmented evaluation results.

[0028] In a preferred embodiment of the present invention, step S1 specifically includes the following steps: S11, the TGS advanced geological prediction system is used to conduct detection at the current tunnel face station to obtain stress gradient slices, water-bearing probability slices, and P-wave velocity slices within the prediction window; during implementation, the TGS advanced geological prediction system is used at the current tunnel face station. Location can be determined by deploying seismic sources and detectors to generate seismic waves and receive reflected and transmitted wave signals. The stress gradient slice map within the forward prediction window can then be obtained through data inversion processing. Water content probability slice diagram P-wave velocity slice diagram The existing color scale ranges, units, coordinate labels, and valid plotting areas from the results report are retained, and the data processing version is recorded when available. The stress gradient slice, water probability slice, and P-wave velocity slice adopt the names from the TGS results report, providing corresponding geophysical interpretation information as input for subsequent feature extraction and fusion evaluation. This method does not directly determine the surrounding rock integrity level or rockfall risk based on a single slice parameter. If the original numerical matrix can be directly exported during acquisition, numerical data is read first; if only report images are available, complete color scale bars are retained for subsequent parameter value estimation based on color scales.

[0029] It is worth noting that when TGS results are provided in the form of forecast report PDFs or images, the frame, mileage axis along the tunnel axis, transverse axis of the cross section, slice height or layer identifier, color scale, and effective drawing area are extracted for the three types of slice maps, and the conversion relationship between pixel coordinates and engineering coordinates is established. For fixed-height slices, the vertical axis represents the transverse position of the cross section rather than the elevation; the height or layer of the slice is saved as independent metadata. The pixel color is converted into the corresponding parameter range using the color scale marked in the report, and the representative values ​​of the corresponding range are used to form a single-channel numerical map. If the original exported data is available, the numerical matrix is ​​read first; color screenshots from different reports that have not been color-scaled must not be directly mixed for training.

[0030] S12 records the structural parameters and construction conditions of multiple structural surfaces at the current working face; wherein, the structural surface parameters include the dip direction, dip angle, average spacing, roughness, opening, extension length, infill type, groundwater status, and visible traces of the structural surface; the construction conditions include the excavation method, blasting parameters, initial support status, and unsupported exposure time; at the same time, the tunnel axis orientation, working face station number, excavation outline, and arch crown range are recorded as engineering geometric information; wherein, the infill type and groundwater status can be encoded using unique thermal coding, and continuous parameters are normalized using training set statistics.

[0031] S13. A unified engineering coordinate system is established with the tunnel axis mileage as the X-axis, the transverse direction of the cross-section as the Y-axis, and the vertical direction as the Z-axis. The TGS slice data is mapped to the engineering coordinates within the prediction window. The orientation of the structural surface and the visible traces are converted into structural surface normal vectors and spatial positions. A verifiable coordinate mapping relationship is established between the image pixel coordinate system and the on-site engineering coordinate system, so that the geophysical slices and geological sketches have a unified spatial benchmark. This provides a spatial basis for the subsequent quantitative geometric correlation between the structural surface and the TGS grid, and reduces the spatial misalignment of different data sources.

[0032] Furthermore, the step between step S13 and step S14 includes: S131, regarding the first The first TGS grid and the first The basic correlation coefficient for structural surfaces is used to represent the reliable extrapolation of structural surfaces, their potential intersection with the tunnel wall, and their positional relationship with the grid. The basic correlation coefficient comprehensively considers three dimensions: the reliability of structural surface observations, the attenuation law of extrapolation distance, and the degree of intersection with the tunnel wall. When the maximum reliable extrapolation distance is exceeded, the basic correlation coefficient is directly set to zero. This coefficient characterizes the geometric reliability of the structural surface observation and the grid, does not involve wedge-shaped block drop determination, and can cover single or multiple sets of parallel or intersecting structural surfaces within all effective extrapolation ranges.

[0033] S132, for candidate structural surface combinations consisting of two or three sets of structural surfaces and the free face of the cavern. Calculate the potential wedge geometric compatibility coefficient and combine it with the basic correlation coefficient to form the block drop risk correlation coefficient, as follows: in, The correlation coefficient for the risk of block dropping. Based on the correlation coefficient, is the geometric compatibility coefficient of the potential wedge.

[0034] Specifically, the basic correlation coefficient is calculated using the following formula: in, Indicates an indicator function, This indicates the structural surface extending from the visible trace or theoretical intersection line at the working face towards the mesh. extrapolated distance, Indicates the maximum reliable extrapolation distance. Indicates the first The observation reliability of the group structure surface can be determined by the weighted result of the visible trace length, measurement integrity, exposure degree of the working face, and parameter integrity. Each item is first normalized to 0 to 1, and the weights are non-negative, with the sum of the weights being 1. This represents the scale parameter that indicates the decay of the correlation with extrapolated distance. It represents the coefficient that indicates the intersection or proximity of the extended plane of the structural surface and the open surface of the cavern within a preset spatial tolerance.

[0035] Furthermore, the dip and tilt angle are converted into unit normal vectors. For combinations of two or three candidate structural surfaces Based on the direction of intersection and the exposure relationship between the half-space of each structural surface and the open surface of the cavern, it is determined whether a candidate block can form a finitely closed structure that can be exposed into the cavern. If the conditions for finite closure or exposure are not met, then... When satisfied, We can set it to 1, and let the potential wedge geometric compatibility coefficient be 1. Basic correlation coefficient Used for integrity branching and shared structure responses, and the correlation coefficient for block drop risk. Used only for branches with risk of block drops. Wedge geometry compatibility coefficient. It only characterizes the geometric compatibility of potential blocks and does not replace mechanical stability calculations involving cohesion, friction angle, rock bridge, and anchoring effects. , and The values ​​of are all between 0 and 1; when Or when there is no geometric relationship between the structural surface and the mesh, let and The value is 0. For segments exceeding the reliable extrapolation range, no structural surface modes are introduced, and integrity evaluation can be completed solely by TGS image modes; for the risk of block drop, an insufficient structural information prompt is output.

[0036] S14: Normalize the mapped TGS slice data and generate a data quality mask and channel validity markers; normalize the three types of TGS results according to the training set statistics, using the same version of the training set statistics in the training, validation, testing, and field inference stages; truncate the normalization results to [0,1]. When the training set range of a channel is less than the preset tolerance or the current results cannot reliably reconstruct the channel, mark the channel as invalid and mask it using the channel validity marker and quality mask. For reports with different prediction interval lengths, resample according to the actual mileage or use adaptive pooling, while retaining the original mileage, cross-sectional lateral position, and slice height or layer mapping: In the formula, c represents the stress gradient, water content probability, or P-wave velocity channel. To prevent normalization of positive numbers with a denominator of zero, clip (·,0,1) means truncate the result to [0,1]. Indicates the first Channel in coordinates The dimensionless value after normalization and truncation of the training set statistics. Indicates the first Channel in coordinates The parameter values ​​are estimated based on the color scale or directly derived values. Indicates that only the first [number] in the training set The minimum value obtained from the statistics of the effective area of ​​the channel. Indicates that only the first [number] in the training set The maximum value obtained from the effective area statistics of each channel; normalization statistics only use the effective area of ​​the training set. The three types of color results are converted into single-channel numerical images according to their respective color scales, and then stacked by channel; areas that cannot be reliably converted are treated as invalid data. A data quality mask is generated. Mark areas outside the valid drawing area, areas with missing data, and areas that cannot be reliably converted using color codes; also use these markings when quality anomalies are already reported. =1 indicates a valid mesh. =0 indicates an invalid mesh. Generate a channel validity flag. When the entire channel is invalid, command =0. Subsequent channel statistics, attention calculation, convolutional feature aggregation, and segmentation evaluation all adopt... and Explicit weighting or masking ensures that invalid regions do not participate in the aggregation of valid features, rather than simply weighting or masking. This is used as a normal input channel. When all channels are invalid or the valid grid set is empty, the normalized automatic evaluation is not performed, and a TGS data shortage reason code is output.

[0037] In a preferred embodiment of the present invention, step S2 specifically includes the following steps: S21, the normalized stress gradient slice, water content probability slice, and longitudinal wave velocity slice are stacked into a three-channel image according to the channel dimension; the three independent TGS slices are integrated into a unified multi-channel image format to adapt to the input format requirements of the subsequent convolutional neural network, while strictly maintaining the spatial alignment relationship between each channel to avoid feature matching errors caused by channel misalignment.

[0038] S22, perform masked global average pooling on each channel, calculate the original channel weights through the channel attention module, and re-normalize the channel weights by combining the effective channel labels to obtain the normalized channel weights, and weight each channel image based on the normalized channel weights. Specifically, global average pooling of the mask is expressed as follows: in, Indicates the first The mask-weighted average response of the channel within the effective region. Representing coordinates The data quality mask value at this location is set to 1 for valid and 0 for invalid. Used to avoid the denominator being zero during space pooling; The formula for calculating the original channel weight is: in, Indicates the original weight of the channel. Element-wise logic The activation function maps the input to the range of 0 to 1. This represents the trainable weight matrix of the first layer of the channel attention module. This represents the trainable weight matrix of the second layer of the channel attention module, with an output dimension of 3 (the number of channels). This represents the element-wise linear rectified activation function, which outputs the larger of the input value and 0. This represents a column vector composed of the average mask responses of the three channels in channel order; the dimensions of the two-layer weight matrix match the input and output dimensions of the layers before and after it.

[0039] The channel normalized weights are then calculated using the original channel weights: in, Indicates the first The valid flag for a channel: 1 indicates that the channel can participate in the calculation, and 0 indicates that the channel is invalid. This is the index for summing the three channels. This is used to avoid the denominator being zero when normalizing channel weights; Finally passed Complete adaptive weighting of channels.

[0040] S23 employs three parallel convolutional branches: 3×3, 5×5, and 7×7. The features from each branch are concatenated and then processed by adaptive pooling and 1×1 convolution to obtain the spatial feature map. The three parallel convolutional branches extract local anomaly boundaries, medium-scale physical properties or water-bearing anomaly areas, and the variation trend along the mileage direction, respectively. Each branch can use 64 and 128 output channels respectively, and the resulting spatial feature map is obtained after concatenation and adaptive pooling and 1×1 convolution. This step preserves the grid spatial location and does not directly compress all features into a single global vector here. Each convolutional branch uses the aforementioned channel-weighted image; invalid pixels are masked before convolution input and continue to be applied with a quality mask of the appropriate scale after convolution and during feature convergence.

[0041] S24, Project the basic correlation coefficient, the block drop risk correlation coefficient, and the data quality mask onto the feature grid where the spatial feature map is located to obtain the feature grid quality mask; Based on the convolution downsampling ratio and the original mileage-cross-section-slice layer mapping, the basic correlation coefficient is... Correlation coefficient of block drop risk and quality mask Projecting onto the nearest valid mesh using area-weighted pooling or the nearest valid mesh mapping method The feature mesh containing the feature mesh is used to obtain the feature mesh quality mask. Area-weighted pooling calculates the area-weighted average of the correlation coefficients and quality mask in the original pixel grids corresponding to each feature grid, using this average as the value for the feature grid. Nearest valid grid mapping takes the value of the nearest valid original grid at the center of the feature grid as the mapping result. After projection, the feature grid quality mask is obtained, and the starting mileage, ending mileage, starting lateral position, ending lateral position, and slice height layer information corresponding to each feature grid are recorded. The area-weighted pooled quality mask represents the effective coverage ratio, with a value ranging from 0 to 1, and subsequent weighting is performed according to this ratio. The nearest valid grid mapping is limited to the original coverage area corresponding to the feature grid; if there are no valid grids within this range, the value is set to 0. The correlation coefficients and quality mask are projected separately.

[0042] In a preferred embodiment of the present invention, step S3 specifically includes the following steps: S31 converts the dip and inclination angles in each group of structural surface parameters into sine and cosine representations, normalizes the spacing, roughness, opening, and extension length, and concatenates them with the infill material, groundwater status, spatial location encoding of visible traces, observation confidence, and corresponding missing marker vectors to form a structural surface feature vector. During implementation, the angle parameters are first processed to be continuous, and the dip and inclination angles in each group of structural surface parameters are converted into sine and cosine representations to avoid feature discontinuities caused by numerical jumps between 0° and 360°. Spacing, roughness, opening, and extension length are normalized using training set statistics and mapped to the 0-1 interval. Infill type and groundwater status are converted into numerical vectors using one-hot encoding. The spatial location encoding of visible traces includes the normalized coordinates of the two endpoints of the trace on the tunnel face profile, the engineering coordinates of the trace midpoint, and the horizontal and vertical distances from the trace midpoint to the centerline of the crown arch. Observation confidence and corresponding missing marker vectors are added simultaneously. The missing marker vector marks each parameter as either a measured or filled value. Missing continuous parameters are filled using training set statistics, but are clearly distinguished by the missing markers. All the above parameters are concatenated along the feature dimension to obtain the structural surface feature vector corresponding to each set of structural surfaces. Trigonometric functions are calculated in radians; degrees should be converted to radians first. Specifically, this is expressed as follows: Among them, for the first Group structure surface, Represents the eigenvectors of the structural surface. This indicates a tendency to transform into sine and cosine. This represents the angle of inclination converted to sine and cosine. This represents the interval after normalization based on the training set statistics. This represents the roughness after normalization based on the statistics of the training set. This represents the opening size after normalization based on training set statistics. This represents the extension length after normalization based on training set statistics. Indicates the state of the filling material. Indicates the state of groundwater. This represents the spatial location encoding of the visible trace. Indicates the reliability of the observation. This represents the corresponding missing marker vector. The visible trace spatial location encoding is also included. It should include at least the normalized projection position of the trace on the tunnel face profile, the position of the midpoint of the trace in the unified engineering coordinate system, and the distance from the center line of the crown arch. Each parameter is labeled with either a measured value or an imputed value. Missing continuous parameters are imputed using statistics from the training set. Distinguish between measured values ​​and filled values.

[0043] S32, determine the geometric sorting key in sequence according to the projection position of the visible trace of the structural surface on the outline of the working face, the distance from the potential intersection line of the top arch, the dip angle, the inclination angle and the spacing; this ensures that the input order is consistent when entering data repeatedly.

[0044] S33, based on the geometric sorting key, the feature vectors of each group of structural surfaces are standardized and sorted; when the geometric sorting keys are parallel within the preset measurement tolerance, the stable order is determined according to the fixed lexicographical order of the normalized parameter vectors, and the structural surface number is only used as an identifier and does not participate in the sorting; when visible traces are missing, the theoretical intersection projection position of the structural surface and the tunnel face contour is used as a substitute sorting quantity; after sorting, the differences between manual numbering and the order of input are reduced, and the feature sequence of the same group of structural surfaces will not change due to the change of input order under the same conditions of parameters, tolerance processing method and sorting rules. When the information required to determine the position of a structural surface is missing, the structural surface does not participate in the position-related geometric fusion, and a prompt of insufficient structural information is output.

[0045] S34. The standardized and sorted feature vectors of each group of structural surfaces are input into a bidirectional LSTM to obtain the encoded features of each group of structural surfaces. The combined features of the structural surfaces are then extracted through a bidirectional long short-term memory network. In practice, the standardized and sorted feature vectors of each group of structural surfaces are input into the bidirectional long short-term memory network, which uses a two-layer bidirectional LSTM structure. The network outputs the hidden state vector corresponding to each structural surface, i.e., the encoded features of each group of structural surfaces. It can also output the comprehensive features of the structural surfaces. When the number of structural faces is insufficient, an effective length mask is used to shield the filling positions, and the filling positions are not included in feature calculation; when geological sketching confirms that there are no obvious structural faces, a preset "no structural face" marker is input; when no sketching has been conducted or the record is missing, a missing marker is used. This step extracts the bidirectional correlation features between multiple sets of structural faces through a sequence coding network, and is adapted to field scenarios with a variable number of structural faces, effectively expressing the combined effect of different numbers of structural faces on the stability of the surrounding rock.

[0046] Furthermore, step S4 specifically includes the following steps: S41, the spatial feature map is unfolded according to a grid to obtain image features; the grid-by-grid image feature vector is obtained through spatial unfolding. In practice, the two-dimensional spatial feature map is unfolded according to the grid dimension to form multiple TGS grids. Each image feature vector corresponds to a TGS grid. Synchronously retain each TGS grid The corresponding feature mesh quality mask values ​​and engineering coordinate information are used to convert the two-dimensional spatial feature map into a mesh-by-mesh feature vector form, which is adapted to the subsequent mesh-by-mesh weighted convergence operation of structural surface features, enabling the fusion process to achieve accurate spatial correspondence at the mesh scale.

[0047] S42, Construct task correlation coefficients; wherein, for the surrounding rock integrity evaluation task, the task correlation coefficient adopts the basic correlation coefficient; for the roof collapse risk evaluation task, the task correlation coefficient adopts the collapse risk correlation coefficient; construct differentiated correlation constraints through task matching, and match the corresponding task correlation coefficients for different evaluation tasks during implementation. For the task of evaluating the integrity of surrounding rock, the task correlation coefficient is the basic correlation coefficient projected onto the feature grid, ensuring that all structural surfaces within the effective extrapolation range participate in the integrity feature fusion, i.e., making = For the task of assessing the risk of roof collapse, the task correlation coefficient is the collapse risk correlation coefficient projected onto the feature mesh. Only structural surfaces with identified potential wedge geometric correlations and within a reliable extrapolation range participate in the structural feature fusion of this branch, i.e., let = This step involves matching the corresponding geometric constraints for the two types of evaluation tasks, ensuring that the integrity evaluation incorporates the influence of all valid structural surfaces, while the block drop risk evaluation only incorporates the influence of structural surfaces with wedge-shaped body conditions, thus differentiating the constraints for the two types of tasks.

[0048] S43, for each TGS grid When the feature mesh quality mask is effective (i.e. When the task correlation coefficient is greater than 0, the geometric constraint attention weight is calculated, and the encoded features are weighted and converged in combination with the geometric constraint attention weight to obtain the structural surface fusion features of the task corresponding to the grid. Specifically, first calculate the normalization factor of the effective structure surface corresponding to mesh u under task t: in, Indicates task Lower grid The normalization factor corresponding to the effective structural surface, Represents the j-th group of structural surfaces and meshes The corresponding task correlation coefficient, Indicates projection onto the grid Feature mesh quality mask, For indicator functions, Represents a grid Transpose of the image feature column vector. This represents the trainable feature projection matrix. This represents the non-negative geometric correlation weight coefficient. Represents a positive number to prevent logarithmic invalidation.

[0049] Then geometrically constrained attention weights The calculation is as follows: in, ; and through The encoded features are weighted and converged to obtain the structural surface fusion features of the task corresponding to the grid. .

[0050] It should be noted that when When, define and order The vector is either a zero vector or a pre-defined unstructured surface embedding vector, thus avoiding division by zero. Each mesh can correspond to multiple structured surfaces, and each structured surface can also be associated with multiple TGS meshes, forming a clear many-to-many spatial relationship.

[0051] S44, the image features, the structural surface fusion features corresponding to the task, and the feature mesh quality mask are stitched together and then subjected to lightweight convolution processing to obtain the integrity fusion feature and the block drop risk fusion feature, respectively; , and After concatenation, lightweight convolution is applied to obtain complete fusion features. Features of combining block drop risk Invalid grids are re-evaluated after the convolution output. Masking. During implementation, for each feature grid, the image features, the corresponding task's structural surface fusion features, and the feature grid quality mask scalar are concatenated along the feature dimension to obtain a concatenated feature vector. The concatenated features of all grids are then restored to a two-dimensional feature map form, and input into two independent lightweight convolutional networks for deep fusion of image modal and structural surface modal features, ultimately yielding complete fusion features. Features integrated with block drop risk .

[0052] Furthermore, step S5 specifically includes the following steps: S51 divides the TGS prediction window into several evaluation segments according to a preset mileage length, and then partitions the cross-sections according to the crown arch and sidewalls. Each evaluation segment corresponds to a set of feature grids, that is, all feature grids within the mileage range that fall within that segment. This divides the continuous prediction window into segmented units that match construction control, ensuring that the evaluation results correspond to clear mileage intervals, facilitating on-site segmented design of support parameters and construction risk management. Cross-section partitioning is determined based on the slice layer and engineering coordinates. If the crown arch is not effectively covered, a verification prompt is output as shown in S54.

[0053] S52, the integrity fusion feature and the block drop risk fusion feature within each segment are weighted and converged using the feature mesh quality mask to obtain the integrity segment feature and block drop risk segment feature of each segment; let the first... The feature grid set corresponding to each evaluation segment is: Employing a feature mesh quality mask To each and Weighted aggregation forms complete segmentation features. Block drop risk segmentation features Automatic evaluation is not performed when there is no valid grid within a segment; specifically: .

[0054] S53, for the first with an effective grid Each evaluation segment outputs the tunnel surrounding rock evaluation results; wherein, the tunnel surrounding rock evaluation results include the distribution of surrounding rock integrity levels. Construction conditions Risk of the top arch falling off Confidence indices, abnormal grid locations, and key control surface information; a shared image coding foundation, task-corresponding structure fusion, and independent output headers can be used. Only after the data has been independently verified to complete probability calibration and meet the preset calibration error requirements is it recorded as conditional probability. Before calibration, it is only used as a 0-1 conditional risk score. The integrity of the surrounding rock can be divided into five levels: intact, relatively intact, moderately intact, relatively broken, and broken. The label is determined based on the development of the structural surface after excavation, the rock mass integrity index, and the unified engineering logging rules, and is not directly replaced by the engineering surrounding rock category.

[0055] Among them, the distribution of surrounding rock integrity levels is as follows: Softmax The classifier outputs a probability distribution of five levels: complete, fairly complete, moderately complete, fairly fragmented, and fragmented. The level corresponding to the highest probability is the predicted completeness level, expressed as: in, The trainable mapping function represents the integrity classification output head, which maps integrity segmentation features to unnormalized classification scores for five integrity levels.

[0056] The risk of the top arch collapsing is caused by... Sigmoid The classifier output, taking the segmented features of the block drop risk and the encoded construction condition vector as input, outputs a conditional risk score in the range of 0 to 1, represented as: in, The trainable mapping function representing the block drop risk output header maps the block drop risk segmentation features and encoded construction conditions together into a scalar score, which is then processed... Sigmoid The function yields a conditional risk score.

[0057] It should be noted that the risk event of rockfall is defined as: within a specified mileage section, the crown arch range, and a preset observation window, under given construction conditions. This refers to rockfall events that occur when the excavation method, blasting parameters, initial support status, and unsupported exposure time exceed the preset single-block volume threshold or cumulative volume threshold. The observation window and volume threshold are preset based on the engineering cross-section, construction management system, and risk control requirements. The surrounding rock integrity level label and rockfall event label are established based on the results of post-excavation geological logging, construction logs, hazard mitigation records, supervision records, or on-site image confirmation, and do not directly replace the rockfall label with the surrounding rock integrity level.

[0058] Furthermore, the procedure prior to step S1 includes: A model was constructed, and a weighted sum of the surrounding rock integrity level classification loss and the rockfall condition risk loss was used to jointly train the model, as follows: in, The weighted cross-entropy loss is the result of the surrounding rock integrity level. The weighted binary cross-entropy loss is used to calculate the risk of block drop conditions. λ Risk task weighting.

[0059] Before implementing the on-site evaluation step, the model was built and parameters were fixed through multi-task joint training. During implementation, an end-to-end evaluation model was constructed, consisting of a TGS multi-channel feature extraction network, a structural surface sequence encoding network, a dual-task cross-modal fusion network, and independent output heads. Joint training was completed using a weighted sum of the surrounding rock integrity level classification loss and the rockfall condition risk loss. The total loss value was the product of the weighted cross-entropy loss of the surrounding rock integrity level plus the weighted binary cross-entropy loss of the risk task weights and the rockfall condition risk. Before training, the training set, validation set, and test set were divided according to the overall project source. Adjacent working faces and overlapping prediction windows from the same project were grouped into the same data subset. To address the imbalance of integrity level samples, classification weights were set inversely proportional to the number of samples in each category. To address the imbalance of positive and negative samples in rockfall events, binary classification weights were set according to the ratio of positive to negative samples. Risk task weights... Based on the validation set, performance parameters for both tasks were tuned to balance integrity classification accuracy and rockfall risk ranking capability. The training process employed the AdamW optimizer, cosine annealing learning rate scheduling, early stopping strategy, and gradient pruning constraints, along with data augmentation techniques such as parameter perturbations consistent with measurement errors and local channel failures to improve model generalization. After independent validation, the model parameters were fixed, allowing direct use of the fixed model during the field evaluation phase without repeated training. This training method enables the model to simultaneously perform surrounding rock integrity classification and rockfall risk assessment. The two tasks share an image encoding foundation while retaining independent constraints and output branches. Training requires consistent surrounding rock integrity labels and rockfall event labels that match the evaluation segments and observation windows, and records corresponding construction conditions. Category weighting and grouping are used to mitigate the impact of sample imbalance and reduce the risk of leakage of homogeneous information, respectively.

[0060] Furthermore, step S53 is followed by the following step: S54, when the model outputs, it calculates the integrity level confidence index, the reliability index of the block drop risk output, the proportion of valid TGS data, and the structural surface information quality index, and outputs the corresponding reason codes so that engineers can perform graded handling in conjunction with on-site logging; among them, the integrity level confidence index Distribution of acceptable surrounding rock integrity levels The maximum value, the reliability index of block drop risk output. The structural surface information quality index can be determined based on the calibration error, prediction entropy, or model ensemble dispersion on the independent validation set. The factors include the number of effective structural surfaces, observation reliability, crown coverage, and foundation correlation coefficient. Correlation coefficient with risk The effective association statistics were determined. During the independent validation and probability calibration phases, the integrity level confidence index was determined based on the validation results. Reliability indicators for block drop risk output Risk output of conditions for arch collapse Interpretive definitions and application thresholds. Proportion of valid TGS data. Insufficient, the corresponding confidence index does not meet its criteria, structural surface is missing, or the maximum credible extrapolation distance is exceeded. Or the crown is not within the effective slice level of TGS or the structural surface information quality index When data is insufficient, the integrity assessment or the risk of missing blocks is marked as "requires manual review" or "insufficient structural information," respectively. Insufficient structural information does not affect the TGS image branch from continuing to provide integrity assessments. The system also outputs reason codes such as "insufficient TGS data," "missing structural surface information," "outside the extrapolation range of structural surfaces," "crown arch not effectively covered," "probability calibration status," or "low model confidence," so that engineers can perform graded handling in conjunction with on-site logging. Calibration error is used for overall evaluation during the verification phase; the upper limit criterion is used for piecewise prediction entropy or integrated dispersion, and the lower limit criterion is used for the maximum category probability.

[0061] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent evaluation of tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches, characterized in that, Includes the following steps: S1. Collect TGS slice data and obtain the structural surface parameters and construction conditions of the geological sketch of the working face. Perform registration and preprocessing of the TGS slice data and the structural surface parameters under a unified engineering coordinate system, and encode the construction conditions. The TGS slice data includes stress gradient slice map, water content probability slice map and P-wave velocity slice map. S2, extract features from the processed TGS slice data to obtain spatial feature maps and feature grid quality masks; S3. The parameters of each group of structural surfaces are concatenated to obtain structural surface feature vectors. The structural surface feature vectors of each group are then standardized and sorted. The encoded features of each group of structural surfaces are extracted based on the standardized and sorted structural surface feature vectors. S4. Based on the geometric relationship between the structural surface and the TGS grid under the unified engineering coordinate system, the association constraints for the evaluation of the surrounding rock integrity and the evaluation of the risk of the top arch falling off are constructed respectively. Under the association constraints, the spatial feature map is fused with the coding features of each group of structural surfaces to obtain the integrity fusion feature and the falling off risk fusion feature. S5, the integrity fusion feature and the rockfall risk fusion feature are weighted and converged using the feature mesh quality mask to form integrity segmentation feature and rockfall risk segmentation feature, and the tunnel surrounding rock evaluation result is output according to the integrity segmentation feature and the rockfall risk segmentation feature.

2. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S11, the TGS advanced geological prediction system is used to conduct detection at the current tunnel face station number to obtain stress gradient slice map, water content probability slice map and P-wave velocity slice map within the prediction window ahead. S12 records the structural parameters and construction conditions of multiple structural surfaces at the current tunnel face; wherein, the structural surface parameters include the dip direction, dip angle, average spacing, roughness, opening, extension length, infill type, groundwater status, and visible traces of the structural surface; the construction conditions include the excavation method, blasting parameters, initial support status, and unsupported exposure time; at the same time, the tunnel axis orientation, tunnel face station number, excavation outline, and arch crown range are recorded as engineering geometric information; S13, a unified engineering coordinate system is established with the tunnel axis mileage as the X-axis, the cross-section transverse direction as the Y-axis, and the vertical direction as the Z-axis. The TGS slice data is mapped to the engineering coordinates within the prediction window, and the structural surface attitude and visible traces are converted into structural surface normal vectors and spatial positions. S14, normalize the mapped TGS slice data and generate a data quality mask and channel validity markers.

3. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches as described in claim 2, characterized in that, Between step S13 and step S14, the following is also included: S131, regarding the first The first TGS grid and the first The basic correlation coefficients for structural surfaces are used to represent the reliable extrapolation of structural surfaces, potential intersections with the cavity wall, and the relationship between the mesh positions. S132, for candidate structural surface combinations consisting of two or three sets of structural surfaces and the free face of the cavern. Calculate the potential wedge geometric compatibility coefficient and combine it with the basic correlation coefficient to form the block drop risk correlation coefficient, as follows: in, The correlation coefficient for the risk of block dropping. Based on the correlation coefficient, is the geometric compatibility coefficient of the potential wedge.

4. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 3, characterized in that, Step S2 specifically includes the following steps: S21, stack the normalized stress gradient slice, water content probability slice, and P-wave velocity slice into a three-channel image according to the channel dimension; S22, perform masked global average pooling on each channel, calculate the original channel weights through the channel attention module, and re-normalize the channel weights by combining the effective channel labels to obtain the normalized channel weights, and weight each channel image based on the normalized channel weights. S23 employs three parallel convolutional branches: 3×3, 5×5, and 7×7. The features from each branch are concatenated and then processed by adaptive pooling and 1×1 convolution to obtain a spatial feature map. S24, Project the basic correlation coefficient, the block drop risk correlation coefficient, and the data quality mask onto the feature grid where the spatial feature map is located to obtain the feature grid quality mask.

5. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31 converts the dip and dip angle in each group of structural surface parameters into sine and cosine representations, normalizes the spacing, roughness, opening and extension length, and concatenates them with the filling material, groundwater state, visible trace spatial location encoding, observation confidence and corresponding missing marker vector to form a structural surface feature vector. S32, determine the geometric sorting key in sequence according to the projection position of the visible trace of the structural surface on the outline of the working face, the distance from the potential intersection line of the top arch, the dip angle, the inclination angle and the spacing; S33, standardize and sort the feature vectors of each group of structural surfaces based on the geometric sorting key; S34. Input the standardized and sorted feature vectors of each group of structural surfaces into a bidirectional LSTM to obtain the encoded features of each group of structural surfaces.

6. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 3, characterized in that, Step S4 specifically includes the following steps: S41, the spatial feature map is unfolded according to a grid to obtain image features; S42, Construct task correlation coefficients; wherein, for the surrounding rock integrity evaluation task, the task correlation coefficient adopts the basic correlation coefficient; for the roof collapse risk evaluation task, the task correlation coefficient adopts the collapse risk correlation coefficient. S43, for each TGS grid When the feature mesh quality mask is effective and the task correlation coefficient is greater than 0, the geometric constraint attention weight is calculated, and the encoded features are weighted and converged in combination with the geometric constraint attention weight to obtain the structural surface fusion feature of the task corresponding to the mesh. S44, the image features, the structural surface fusion features of the corresponding task, and the feature mesh quality mask are stitched together and then subjected to light convolution processing to obtain the integrity fusion features and the block drop risk fusion features, respectively.

7. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 1, characterized in that, Step S5 specifically includes the following steps: S51, along the TGS prediction window, divide into several evaluation segments according to the preset mileage length, and divide the cross-sections according to the top arch and sidewalls. S52, the integrity fusion feature and the block drop risk fusion feature in each segment are weighted and converged using the feature mesh quality mask to obtain the integrity segment feature and the block drop risk segment feature of each segment; S53, for the first with an effective grid Each evaluation segment outputs the tunnel surrounding rock evaluation results; wherein, the tunnel surrounding rock evaluation results include the distribution of surrounding rock integrity levels, the risk of roof collapse under construction conditions, confidence index, abnormal grid locations and information on major control structural surfaces.

8. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 1, characterized in that, The procedure preceding step S1 also includes: A model was constructed, and a weighted sum of the surrounding rock integrity level classification loss and the rockfall condition risk loss was used to jointly train the model, as follows: in, The weighted cross-entropy loss is the result of the surrounding rock integrity level. The weighted binary cross-entropy loss is used to calculate the risk of block drop conditions. λ Risk task weighting.

9. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 3, characterized in that, The specific calculation method used in step S131 is as follows: in, Based on the correlation coefficient, For the first The reliability of observations of the group structure surface. For indicator functions, The structural surface is defined by the visible traces or theoretical intersection lines from the working face towards the mesh. extrapolated distance, To correlate the scale parameter that decays with extrapolated distance, To achieve the maximum reliable extrapolation distance, It is the coefficient by which the extended plane of the structural surface intersects or approaches the open surface of the cavern within a preset spatial tolerance.

10. The intelligent evaluation method for tunnel surrounding rock based on the fusion of TGS multi-channel images and geological sketches according to claim 7, characterized in that, The step S53 is followed by the following step: S54 calculates the integrity level confidence index, the reliability index of block drop risk output, the proportion of valid TGS data, and the structural surface information quality index when the model outputs, and outputs the corresponding reason codes so that engineers can classify and handle the situation in conjunction with the on-site log.

Images