Quantitative monitoring method for water leakage of excavation face of mountain tunnel based on deep learning

Abstract

The invention provides a method for quantitatively monitoring water leakage of a mountain tunnel excavation face based on deep learning. The method comprises the following steps: 1, synchronously acquiring multi-source data; step 2, identifying a water leakage area based on image deep learning; 3, mapping and fusing the image and the point cloud data; 4, performing quantitative calculation on the total water leakage area based on the fusion point cloud; and 5, generating a monitoring report and early warning. According to the method, the technical bottlenecks that a traditional monitoring method depends on manual operation, the quantification difficulty is large, and the real-time performance is poor are effectively solved, full-process automatic and accurate monitoring of leakage water from recognition, positioning and quantification to early warning is achieved, the risk prevention and control capacity of leakage water of the excavation face of the mountain tunnel is remarkably improved, and the method has important engineering significance for guaranteeing tunnel construction safety.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring technology, and in particular to a method for quantitative monitoring of water seepage at the excavation face of a mountain tunnel based on deep learning. Background Technology

[0002] In mountain tunnel construction, water seepage at the excavation face is a core hidden danger that can induce face instability, support structure deformation, and even collapse. Traditional monitoring methods have significant drawbacks: manual inspections are highly subjective and lack real-time accuracy, making it difficult to detect hidden seepage points; single-point sensors have limited monitoring dimensions and cannot reflect the spatial distribution of seepage; single visual monitoring can only perform qualitative identification, and its misjudgment rate exceeds 15% due to interference from lighting and rock wall texture, and it lacks spatial positioning capabilities; while single laser point clouds provide high-precision spatial information, they lack temporal optical characteristics, making it difficult to distinguish between seepage and naturally moist rock walls, and unable to track dynamic development trends. Summary of the Invention

[0003] This invention proposes a method for quantitative monitoring of water seepage at the excavation face of mountain tunnels based on deep learning, which solves the problems of low accuracy, poor real-time performance, and single data dimension in the monitoring of water seepage in mountain tunnels.

[0004] The specific technical solution of the present invention is as follows: A method for quantitative monitoring of water seepage at the excavation face of a mountain tunnel based on deep learning, comprising the following steps: Step 1: Synchronous Acquisition of Multi-Source Data Image data and point cloud data of the mountain tunnel excavation face are acquired simultaneously using a video imaging device and a 3D laser scanner; Step 2: Leakage Area Identification Based on Image Deep Learning (21): Verification of the commonality of the triangle in the leakage area The images acquired in step 1 are denoised, cropped, and scaled. The leakage areas are then segmented using manual annotation or a preliminary semantic segmentation model. The outlines of the segmented leakage areas are fitted with polygons, and geometric features are extracted, including outline shape, vertices and angles, symmetry and directionality. Finally, the geometric features are statistically analyzed to clarify the common patterns of the "triangle" shape of the leakage areas, providing a theoretical basis and feature support for the design of the leakage area identification model. (22): Design of a triangular receptive field convolution module A triangular parametric receptive field convolutional module was designed for detecting water seepage areas at the excavation face of mountain tunnels. By setting morphological and scale parameters, a "square convolutional kernel + triangular mask" technical solution was adopted to extract triangular features. Finally, a modular structure containing triangular convolutional layers, feature enhancement layers and adaptive pooling layers was constructed to improve the model's feature extraction capability for water seepage areas. (23): Construction and training of deep learning networks Using the U-Net network with embedded triangular convolution modules as the core architecture, the model training and parameter optimization are completed through data augmentation, DiceLoss loss function optimization and multi-index performance verification. Finally, a high-precision binary segmentation map of the seepage area is output, providing an accurate target area for subsequent spatial mapping. Step 3: Image and point cloud data mapping and fusion A combination of statistical filtering and radius filtering was used to denoise the original point cloud data, and the origin of the point cloud coordinates was shifted to the center of the excavation face. SIFT 3D feature points were extracted from the denoised point cloud, and SIFT 2D feature points were extracted from the image segmentation results after Gaussian blurring. Then, the similarity between the SIFT 3D and 2D feature points was calculated using the BRIEF descriptor. Mismatched pairs were removed using the Ratio test, and outliers were removed using the least squares method. The initial registration parameters were solved using the PnP algorithm, and finally, a high-precision registration matrix was obtained through ICP optimization. Based on the registration matrix, the image RGB texture, semantic labels, and point cloud 3D coordinates and reflection intensity were associated to generate and store a multi-dimensional fused point cloud. Metadata such as timestamps and registration errors were also recorded to provide a basis for subsequent quantitative calculations and accuracy verification. Step 4: Quantitative calculation of total seepage area based on fused point cloud (41): Construction of Triangular Network The fused point cloud is projected onto the XY plane and sorted by coordinates. A two-dimensional triangular mesh of the excavation face is generated by constructing super triangles and inserting them point by point. The two-dimensional triangular mesh is mapped back to three-dimensional coordinates to generate a three-dimensional triangular mesh. Invalid triangles are then removed and the structure of the three-dimensional triangular mesh is optimized. At the same time, it is verified whether the coverage error of the three-dimensional triangular mesh meets the engineering accuracy requirements.

[0005] (42): Calculation of the area of ​​the seepage triangle First, the semantic labels "seepage / non-seepage" are mapped to a 3D triangular mesh to mark the seepage triangles. Second, the area of ​​a single seepage triangle is calculated using the vector cross product method. Finally, the areas of all seepage triangles are summed to obtain the total seepage area. The area change rate is obtained by calculating the difference between the total seepage areas at adjacent times. A threshold for the change rate is set to determine whether the seepage is in a rapid development state.

[0006] (43): Verification of computational accuracy The total leakage area calculated in step (42) is compared with the total leakage area calculated by the manual "grid paper method" to ensure that the estimation error of the total leakage area is ≤5%; at the same time, data is collected three times on the same excavation face, and the standard deviation of the total leakage area is calculated three times. The standard deviation is required to be <0.05m² to verify the stability and reliability of the monitoring results. Step 5: Generate monitoring reports and early warnings Generate a monitoring report to monitor the total leakage area and the rate of change of the area. When the set threshold is exceeded, an alarm is issued.

[0007] Furthermore, in step 1, the video imaging device adopts an industrial-grade high-definition video imaging device, which uses an AI video monitoring and measuring instrument. The AI ​​video monitoring and measuring instrument is installed on a steel support 5-10m behind the excavation face to ensure that the field of view covers ≥80% of the excavation face area and the overlap area with the three-dimensional laser point cloud scanning area is ≥20m². Image data is continuously acquired and timestamps are recorded synchronously. The acquired images are preprocessed: Wiener filtering is used to eliminate image blur, grayscale world algorithm is used to complete white balance correction, histogram equalization is used to achieve brightness normalization, and finally Zhang's calibration method is used to complete the external parameter calibration of the video imaging device to ensure that the image clarity and parameter accuracy meet the requirements of subsequent analysis. The 3D laser scanner is a phase-type 3D laser scanner, specifically the Southern Surveying and Mapping MicroIII. The 3D laser scanner and video imaging device are coaxially mounted to ensure overlapping fields of view. The scan line spacing is set to 5mm to ensure a point cloud density of ≥100 points / m². Point cloud data is continuously acquired at a frequency of 1.2MHz, with approximately 5 million points per frame. The data is stored in LAS format, and the timestamp error between the point cloud data and image data is strictly controlled to ≤10ms to ensure the spatiotemporal synchronization of multi-source data.

[0008] Furthermore, the verification of the commonalities of the leakage area triangle in step (21) specifically includes: The Douglas-Peucker algorithm was used to fit polygons to the segmented leakage area contours, and the following geometric features were extracted: Outline shape: Statistically analyze the number of sides and interior angle distribution of the fitted polygons, with a focus on the proportion of triangles or approximate triangles; Vertices and Angles: Calculate the coordinates of the polygon vertices to determine the leakage source point and the endpoints of the two diffusion boundaries, and analyze the concentrated range of the included angle formed by the above three points; Symmetry and Directionality: Verify the "vertex-base" triangular structure trend of the leakage area along the direction of gravity or the direction of the crack, and clarify the correlation between its morphological characteristics and physical causes; Statistical verification of the commonalities of triangles: Feature statistical analysis: Statistical analysis of the geometric features of a large number of samples. The core verification indicators include: ① More than 80% of the leakage areas have simplified polygon fitting results that are triangles or approximate triangles; ② The included angle formed by the three key vertices of the leakage area outline is mainly acute, which is consistent with the diffusion law of leakage water under the action of gravity. Explanation of common patterns: Combining the physical causes of tunnel water leakage, the rationality of the "triangle" feature is verified from the mechanism level, eliminating the influence of accidental statistical errors on common patterns.

[0009] Furthermore, the design of the triangular receptive field convolution module in step (22) specifically includes: (1) Parametric design of receptive field Based on the common characteristics of the triangular leakage area, the parameters of the triangular receptive field are set: Morphological parameters: apex angle 30°-60°, waist length ratio 1:1.5; Scale parameters: shallow networks use small triangular receptive fields, while deep networks use large triangular receptive fields; (2) Triangular receptive field implementation method The technical solution of "square convolution kernel + triangle mask" is adopted to break through the limitations of the traditional square receptive field: in the weight matrix of 3×3 and 5×5 square convolution kernel, only the weight of the triangular region is retained, and the weight of the non-triangular region is set to 0, so that the convolution kernel only calculates the triangular feature region in the image, which is equivalent to realizing the function of the triangular receptive field. (3) Module structure construction The core of the triangular receptive field convolution module consists of three parts: Triangular convolutional layer: Embedding the above-mentioned parameterized triangular receptive field, replacing the traditional convolutional layer, to achieve accurate extraction of triangular features of the seepage area; Feature enhancement layer: The number of channels in the feature map is adjusted by 1×1 convolution, and the non-linear mapping capability of key features such as triangle edges and vertices is enhanced by combining BatchNormalization and ReLU6 activation function. Adaptive pooling layer: It adopts the "vertex-preserving pooling" strategy to record the coordinates of triangle vertices. During the pooling process, it assigns higher weights to the vertex position features to avoid the loss of key structural information.

[0010] Furthermore, the deep learning network construction and training in step (23) specifically includes: (1) Infrastructure Setup Based on the U-Net encoder-decoder symmetrical structure, the following optimization design is performed: Encoder: Set 4-5 encoding blocks, each encoding block contains "triangle convolutional layer + ReLU6 activation function + vertex-preserving pooling layer", and extract multi-scale seepage triangle features by progressively compressing spatial dimensions; Decoder: 4-5 decoding blocks are set up corresponding to the encoder. Each decoding block contains "transposed convolution + feature concatenation + 1×1 convolution", which restores the feature map resolution and fuses the multi-scale features output by the encoder. Skip connections: Skip connections are established between corresponding layers of the encoder and decoder to pass the triangular detail features extracted by the shallow network and make up for the feature loss problem during the upsampling process; (2) Model training optimization Dataset partitioning: The labeled water leakage images were divided into training, validation and test sets in a ratio of 7:2:1; the dataset size was expanded to 3 times its original size by data augmentation techniques such as rotation, translation and adding Gaussian noise to avoid model overfitting. Parameter settings: DiceLoss is selected as the loss function, Adam is used as the optimizer, and the total number of training epochs is set to 50. Performance Validation: IoU, Dice coefficient, and false negative rate were used as performance evaluation metrics for the model. If small-scale false negatives occurred, the small-scale receptive field parameters of the shallow network were optimized. If background false positives occurred, the feature fusion weights of the deep network were adjusted until the model segmentation accuracy was ≥92%. (3) Output of leakage area The network output layer uses a 1×1 convolutional kernel to compress the number of feature map channels to 2; the pixel values ​​are mapped to the 0-1 range through the Sigmoid activation function, and a threshold of 0.5 is set to generate a binary segmentation map, which clearly marks the two-dimensional pixel positions of the leakage area and provides an accurate target area for subsequent spatial mapping.

[0011] Furthermore, step 3, the mapping and fusion of image and point cloud data, specifically includes: (a) Data preprocessing (1) Point cloud preprocessing A combination of statistical filtering and radius filtering was used to denoise the original point cloud: the statistical filtering parameters were set to K nearest neighbors of 50 and standard deviation threshold of 1.0; the radius filtering parameters were set to radius of 0.01m and minimum number of points of 5 to ensure that more than 90% of the effective point cloud was retained; the origin of the point cloud coordinates was translated to the center of the excavation face to simplify the subsequent registration and calculation process. (2) Feature point extraction SIFT 3D feature points are extracted from the preprocessed point cloud, prioritizing high-frequency texture areas such as steel support nodes and anchor holes, ensuring that the number of feature points extracted from each region of the arch, sidewall, and invert is ≥5; SIFT 2D feature points are extracted from the image segmentation results, and Gaussian blur preprocessing is performed before extraction to reduce the interference of tunnel reflections on the feature point extraction accuracy. (II) Feature point matching and registration (1) Preliminary matching The similarity between SIFT 3D feature points and 2D feature points is calculated using BRIEF descriptors, and the Ratio test is used to remove false matching pairs to ensure that the number of valid matching points retained is ≥20 pairs. (2) Registration optimization Outliers were eliminated using the least squares method, and the initial registration parameters were solved using the PnP algorithm. Based on this, the registration parameters were further optimized using the ICP algorithm, and finally a high-precision registration matrix was obtained.

[0012] (III) Generation of Fusion Point Clouds Based on the final registration matrix, the RGB texture features and "seepage / non-seepage" semantic labels of the seepage area in the image are associated with the three-dimensional coordinates and reflection intensity attributes of the point cloud to generate a fused point cloud containing multi-dimensional information of "spatial-intensity-optical-semantic", which is stored in PLY format. At the same time, metadata such as timestamps and registration errors are recorded to provide a basis for subsequent quantitative calculations and accuracy verification.

[0013] Furthermore, the construction of the triangulation network in step (41) specifically includes: (1) Point cloud projection and sorting Projecting the fused point cloud onto the XY plane yields a two-dimensional point set. Sort by x-coordinate in ascending order; if x-coordinates are the same, sort by y-coordinate in ascending order to reduce the computational load of subsequent triangulation. (2) Generation of triangular network Construct a super triangle: Generate a super-large triangle that contains all projection points, ensuring coverage of the entire excavation area; Point-by-point insertion: Insert the sorted 2D points into the super triangle one by one. Use a search algorithm to find the "influence triangle" containing the point, delete the edges in the triangle and connect the inserted point to the boundary vertex of the triangle to form a new triangle. Repeat the above operation until all points are inserted. (3) Triangle network optimization verification Triangles containing the vertices of the super triangle are removed to obtain a two-dimensional triangular mesh of the excavation face; the two-dimensional triangular mesh is mapped back to three-dimensional coordinates to generate a three-dimensional triangular mesh, the aspect ratio of each triangular face in the three-dimensional triangular mesh is calculated, and narrow triangular faces with aspect ratio > 3 are removed; the coverage of the triangular mesh is verified to ensure that its error with the designed area of ​​the excavation face is ≤ 2%, which meets the engineering accuracy requirements.

[0014] Furthermore, in step (42), the calculation of the area of ​​the seepage triangle specifically includes: (1) Semantic tag association Using the registration relationship obtained in step 3, the semantic labels of "water seepage / non-water seepage" identified by deep learning in the image are mapped to a 3D triangular mesh, completing the labeling of all water seepage triangles.

[0015] (2) Calculation of the area of ​​a single triangular face For the triangular facet marked "seepage" The vertex coordinates are respectively , , The area is calculated using the cross product method. The specific steps are as follows: Calculate the vector: , ; Calculate the cross product modulus: , where c = a × b, and c is the cross product of vectors a and b; Calculate the area of ​​a single triangular face: ; (3) Calculation of total seepage area and rate of change Total seepage area: Traverse all triangles marked as "seepage", sum up the areas of each triangle, and get the total seepage area S_total; Area change rate: Based on time-series fused point cloud data, with the data acquisition time interval set to Δt, the difference in the total leakage area at adjacent time points is calculated to obtain the area change rate. A positive value indicates that the leakage area is expanding, while a negative value indicates that the leakage area is shrinking. A threshold of V_S≥0.1m² / h is set to determine whether the leakage is in a rapid development state.

[0016] Furthermore, the accuracy verification in step (43) specifically includes: (1) Comparative verification The manual "grid paper method" is used as a reference standard: a 10cm×10cm grid is drawn on a high-resolution photograph of the excavation face, the number of grids covering the seepage area is counted, and the seepage area is estimated; the total seepage area calculated by the method of this invention is compared with the manual estimation result to ensure that the error is ≤5%; (2) Repeated verification Data from multiple sources was collected three times from the same excavation face, with an interval of ≥30 minutes between each collection to ensure stable monitoring conditions. The total leakage area corresponding to the three data points was calculated. The standard deviation of the three areas was calculated. If the standard deviation was <0.05m², it proved that the monitoring results of the method of the present invention were stable and reliable.

[0017] Furthermore, step 5 generates a monitoring report, which specifically includes: Generate a "Quantitative Monitoring Report on Water Leakage at the Excavation Face of a Mountain Tunnel". The report should include: Basic information: monitoring time, excavation face mileage, data acquisition equipment parameters, point cloud and image registration error; Key results: total leakage area, area and location of the largest leakage area, and rate of change of leakage area; Accuracy metrics: manual verification error rate, standard deviation of repeated monitoring; It also includes visual displays: The 3D triangular mesh model was rendered using CloudCompare software, with the seepage triangles rendered in red and the non-seepage triangles rendered in gray. The location and area of ​​each seepage area were marked in the model to generate a 3D visualization model. At the same time, the triangular mesh model was cut along the Y-axis of the center of the excavation face to generate a 2D cross-sectional view, which intuitively shows the distribution characteristics of seepage water on the excavation face. The warning includes: Based on the total leakage area S_total and the area change rate V_S, thresholds are set to classify four warning levels, as follows: Normal level: S_total < 0.5m² and V_S < 0.05m² / h, only routine monitoring and recording are required, and no special response measures are needed; Attention level: If 0.5m²≤S_total<1m² or 0.05m² / h≤V_S<0.1m² / h, a blue warning light will be activated, and a monitoring report will be sent to the construction technicians simultaneously, increasing the frequency of patrols of the excavation face; Warning level: 1m²≤S_total<2m² or 0.1m² / h≤V_S<0.2m² / h, activate the yellow warning light and low-frequency buzzer, send a warning text message to project management personnel, organize technical personnel to conduct on-site verification, and formulate targeted prevention and control measures; Emergency Level: If S_total≥2m² or V_S≥0.2m² / h, activate the red warning light and high-frequency buzzer, simultaneously generate an emergency response plan, and push it to the construction team's decision-making level. If necessary, suspend excavation operations to ensure construction safety.

[0018] The beneficial effects of this invention are as follows: This invention effectively solves the technical bottlenecks of traditional monitoring methods, such as reliance on manual operation, difficulty in quantification, and poor real-time performance. It realizes automated and accurate monitoring of water leakage from identification, location, quantification to early warning, significantly improving the risk prevention and control capabilities of water leakage at the excavation face of mountain tunnels, and has important engineering significance for ensuring the safety of tunnel construction. Attached Figure Description

[0019] 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 these drawings without creative effort.

[0020] Fig. 1 This is a flowchart of the method of the present invention; Fig. 2 This is a comparison image of the identification of the water leakage area before and after in Embodiment 2 of the present invention. Detailed Implementation

[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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0022] A method for quantitative monitoring of water seepage at the excavation face of a mountain tunnel based on deep learning, comprising the following steps: Step 1: Synchronous Acquisition of Multi-Source Data Image data and point cloud data of the mountain tunnel excavation face are acquired simultaneously using a video imaging device and a 3D laser scanner.

[0023] Step 2: Leakage Area Identification Based on Image Deep Learning (21): Verification of the commonality of the triangle in the leakage area The images acquired in step 1 are denoised, cropped, and scaled. The leakage areas are then segmented using manual annotation or a preliminary semantic segmentation model. The outlines of the segmented leakage areas are fitted with polygons, and geometric features are extracted, including outline shape, vertices and angles, symmetry and directionality. Finally, the geometric features are statistically analyzed to clarify the common patterns of the "triangle" shape of the leakage areas, providing a theoretical basis and feature support for the design of the leakage area identification model. (22): Design of a triangular receptive field convolution module A triangular parametric receptive field convolutional module was designed for detecting water seepage areas at the excavation face of mountain tunnels. By setting morphological and scale parameters, a "square convolutional kernel + triangular mask" technical solution was adopted to extract triangular features. Finally, a modular structure containing triangular convolutional layers, feature enhancement layers and adaptive pooling layers was constructed to improve the model's feature extraction capability for water seepage areas. (23): Construction and training of deep learning networks Using the U-Net network with embedded triangular convolution modules as the core architecture, the model training and parameter optimization are completed through data augmentation, DiceLoss loss function optimization and multi-index performance verification. Finally, a high-precision binary segmentation map of the seepage area is output, providing an accurate target area for subsequent spatial mapping. Step 3: Image and point cloud data mapping and fusion A combination of statistical filtering and radius filtering was used to denoise the original point cloud data, and the origin of the point cloud coordinates was shifted to the center of the excavation face. SIFT 3D feature points were extracted from the denoised point cloud, and SIFT 2D feature points were extracted from the image segmentation results after Gaussian blurring. Then, the similarity between the SIFT 3D and 2D feature points was calculated using the BRIEF descriptor. Mismatched pairs were removed using the Ratio test, and outliers were removed using the least squares method. The initial registration parameters were solved using the PnP algorithm, and finally, a high-precision registration matrix was obtained through ICP optimization. Based on the registration matrix, the image RGB texture, semantic labels, and point cloud 3D coordinates and reflection intensity were associated to generate and store a multi-dimensional fused point cloud. Metadata such as timestamps and registration errors were also recorded to provide a basis for subsequent quantitative calculations and accuracy verification. Step 4: Quantitative calculation of total seepage area based on fused point cloud (41): Construction of Triangular Network The fused point cloud is projected onto the XY plane and sorted by coordinates. A two-dimensional triangular mesh of the excavation face is generated by constructing super triangles and inserting them point by point. The two-dimensional triangular mesh is mapped back to three-dimensional coordinates to generate a three-dimensional triangular mesh. Invalid triangles are then removed and the structure of the three-dimensional triangular mesh is optimized. At the same time, it is verified whether the coverage error of the three-dimensional triangular mesh meets the engineering accuracy requirements.

[0024] (42): Calculation of the area of ​​the seepage triangle First, the semantic labels "seepage / non-seepage" are mapped to a 3D triangular mesh to mark the seepage triangles. Second, the area of ​​a single seepage triangle is calculated using the vector cross product method. Finally, the areas of all seepage triangles are summed to obtain the total seepage area. The area change rate is obtained by calculating the difference between the total seepage areas at adjacent times. A threshold for the change rate is set to determine whether the seepage is in a rapid development state.

[0025] (43): Verification of computational accuracy The total leakage area calculated in step (42) is compared with the total leakage area calculated by the manual "grid paper method" to ensure that the estimation error of the total leakage area is ≤5%; at the same time, data is collected three times on the same excavation face, and the standard deviation of the total leakage area is calculated three times. The standard deviation is required to be <0.05m² to verify the stability and reliability of the monitoring results. Step 5: Generate monitoring reports and early warnings Generate a monitoring report to monitor the total leakage area and the rate of change of the area. When the set threshold is exceeded, an alarm is issued. Example 2

[0026] Reference Figs. 1-2 A method for quantitative monitoring of seepage water at the excavation face of a mountain tunnel based on deep learning, comprising the following steps: Step 1: Multi-source data synchronous acquisition High-definition image data acquisition Industrial-grade high-definition video imaging equipment (such as an AI video monitoring and measuring instrument) should be used, with the following technical parameters: resolution no less than 1920×1080, frame rate 25fps, and horizontal field of view ≥160°. The device should be installed on a steel support 5-10m behind the excavation face, ensuring that the field of view covers ≥80% of the excavation face area, and the overlap area with the 3D laser point cloud scanning area is ≥20m² (to avoid blind spots). Image data should be continuously acquired, and timestamps (accurate to milliseconds) should be recorded synchronously. The acquired images should be preprocessed: Wiener filtering should be used to eliminate image blur (de-blurring), a grayscale world algorithm should be used for white balance correction, histogram equalization should be used for brightness normalization, and finally, Zhang's calibration method should be used to calibrate the camera's extrinsic parameters to ensure that the image clarity and parameter accuracy meet the requirements of subsequent analysis.

[0027] 3D laser point cloud acquisition A phase-detection 3D laser scanner (such as the Southern Surveying and Mapping MicroIII) should be selected, with the following technical parameters: scanning accuracy ±2mm and scanning frequency 1.2MHz. The scanner and the high-definition image acquisition device should be coaxially mounted (ensuring overlapping fields of view), with a scan line spacing of 5mm, ensuring a point cloud density ≥100 points / m². Point cloud data (including 3D coordinates, reflection intensity, etc.) should be continuously acquired at a frequency of 1.2MHz, with approximately 5 million points per frame, stored in LAS format. The timestamp error between the point cloud data and image data should be strictly controlled to ≤10ms to ensure spatiotemporal synchronization of multi-source data.

[0028] Step 2: Leakage Area Identification Based on Image Deep Learning Based on the high-definition image data acquired in step 1, a deep learning model is used to extract the common triangular features of the leakage area, achieving high-precision semantic segmentation of the leakage area and providing accurate target area information for subsequent spatial mapping. Specifically, this includes the following steps: Step 21: Verification of the commonalities of the leakage area triangle Data preprocessing and region segmentation The image dataset collected in step 1 (which needs to cover different geological conditions, leakage levels, and environmental interference scenarios) is denoised, cropped, and scaled. The leakage area is segmented by manual annotation or a preliminary semantic segmentation model (such as traditional thresholding or edge detection algorithms) to generate a contour mask and eliminate interference from background elements such as tunnel lining and pipelines.

[0029] Geometric feature extraction The Douglas-Peucker algorithm was used to fit polygons to the segmented leakage area contours, and the following key geometric features were extracted: Outline morphology: Statistically analyze the number of sides and interior angle distribution of fitted polygons, focusing on the proportion of triangles (3 sides) or approximate triangles (4 sides, but 3 of which form a significant triangular structure); Vertex and Angle: Calculate the coordinates of the polygon vertices to determine the source of leakage and the endpoints of the diffusion boundaries on both sides. Analyze the concentrated range of the angle formed by the above three points (usually 30°-60°, which conforms to the diffusion law of leakage water under the action of gravity). Symmetry and directionality: Verify the "vertex-base" triangular structure trend of the leakage area along the direction of gravity or the direction of the crack, and clarify the correlation between its morphological characteristics and physical causes.

[0030] Statistical verification of commonalities in triangles Feature statistical analysis: Statistical analysis of the geometric features of a large number of samples (at least hundreds to thousands of images) was performed. The core verification indicators include: ① More than 80% of the leakage areas have simplified polygon fitting results that are triangles (3 sides) or approximate triangles (e.g., 4 sides, but 3 of the sides form a significant triangular structure); ② The included angle formed by the three key vertices of the leakage area contour (the leakage source point and the endpoints of the diffusion boundaries on both sides) is mainly acute (concentrated between 30° and 60°), which is consistent with the diffusion law of leakage water under the action of gravity.

[0031] Explanation of common patterns: Combining the physical causes of tunnel water leakage (such as water seeping along cracks and spreading to both sides under the action of gravity, forming a triangular area with the crack outlet as the vertex and spreading downward), the rationality of the "triangle" feature is verified from the mechanism level, eliminating the influence of accidental statistical errors on common patterns.

[0032] Step 22: Design of the triangular receptive field convolution module (1) Parametric design of receptive field Based on the common characteristics of the triangular leakage area, multi-scale triangular receptive field parameters are set: Morphological parameters: Apex angle 30°-60° (matching the leakage diffusion angle), waist length ratio 1:1.5 (adapting to different leakage areas); Scale parameters: The shallow network uses a small triangular receptive field (16×16 pixels, used to detect small dripping areas), while the deep network uses a large triangular receptive field (64×64 pixels, used to detect large seepage areas).

[0033] (2) Receptive field implementation method The technical solution of "square convolution kernel + triangle mask" is adopted to overcome the limitations of the traditional square receptive field: in the weight matrix of 3×3 and 5×5 square convolution kernel, only the weight of the triangular region is retained (the weight of the non-triangular region is set to 0), so that the convolution kernel only calculates the triangular feature region in the image, which is equivalent to realizing the function of the triangular receptive field.

[0034] (3) Module structure construction The core of the triangular receptive field convolution module consists of three parts: Triangular convolutional layer: Embedding the above-mentioned parameterized triangular receptive field, replacing the traditional convolutional layer, to achieve accurate extraction of triangular features of the seepage area; Feature enhancement layer: Adjusts the number of channels in the feature map through 1×1 convolution, and combines BatchNormalization and ReLU6 activation function to enhance the non-linear mapping capability of key features such as triangle edges and vertices; Matching pooling layer: The "vertex-preserving pooling" strategy is adopted to record the coordinates of the triangle vertices. During the pooling process, the vertex position features are given higher weights to avoid the loss of key structural information.

[0035] Step 23: Deep Learning Network Construction and Training (1) Infrastructure Setup Based on the U-Net encoder-decoder symmetrical structure, the following optimization design is performed: Encoder (downsampling): Set 4-5 encoding blocks, each encoding block contains "triangle convolutional layer + ReLU6 activation function + vertex-preserving pooling layer", and extract multi-scale seepage triangle features by gradually compressing the spatial dimension (from 512×512 to 16×16); Decoder (upsampling): 4-5 decoding blocks are set up corresponding to the encoder. Each decoding block contains "transposed convolution + feature concatenation + 1×1 convolution", which restores the feature map resolution while fusing the multi-scale features output by the encoder. Skip connections: Skip connections are established between corresponding layers of the encoder and decoder to pass the detailed features of triangles (such as edges and vertices) extracted by the shallow network, thus compensating for the feature loss problem during the upsampling process.

[0036] (2) Model training optimization Dataset partitioning: The labeled water leakage images (including binary masks) were divided into training, validation and test sets in a ratio of 7:2:1; the dataset size was expanded to 3 times its original size by data augmentation techniques such as rotation (±5°), translation (±0.1m) and adding Gaussian noise to avoid model overfitting; Parameter settings: DiceLoss is selected as the loss function (to solve the imbalance problem between leakage water samples and background samples), Adam is used as the optimizer (initial learning rate 1e-4, decaying by 0.8 times every 10 rounds), and the total number of training rounds is set to 50 rounds; Performance validation: IoU (Intersection over Union), Dice coefficient, and false negative rate were used as performance evaluation metrics for the model. If small-scale false negatives occurred, the small-scale receptive field parameters of the shallow network were optimized. If background false positives occurred, the feature fusion weights of the deep network were adjusted until the model segmentation accuracy was ≥92%.

[0037] (3) Output of leakage area The network output layer uses a 1×1 convolutional kernel to compress the number of feature map channels to 2 (corresponding to the background and water leakage respectively); the pixel values ​​are mapped to the 0-1 range through the Sigmoid activation function, and a threshold of 0.5 is set to generate a binary segmentation map, which clearly marks the two-dimensional pixel positions of the water leakage area, providing an accurate target area for subsequent spatial mapping.

[0038] Step 3: Image and point cloud data mapping and fusion 1. Data Preprocessing (1) Point cloud preprocessing A combined approach of statistical filtering and radius filtering was used to denoise the original point cloud: the statistical filtering parameters were set to K-nearest neighbors of 50 and a standard deviation threshold of 1.0; the radius filtering parameters were set to a radius of 0.01m and a minimum number of points of 5, ensuring that more than 90% of the valid point cloud was retained. The origin of the point cloud coordinates was shifted to the center of the excavation face (X-axis set horizontal, Y-axis vertical, and Z-axis along the tunnel axis) to simplify the subsequent registration and calculation process.

[0039] (2) Feature point extraction SIFT (Scale-Invariant Feature Transform) 3D feature points are extracted from the preprocessed point cloud, prioritizing high-frequency texture areas such as steel support nodes and anchor holes, ensuring that the number of feature points extracted from each region of the arch, sidewall, and invert is ≥5; SIFT 2D feature points are extracted from the image segmentation results, and Gaussian blur preprocessing is performed before extraction to reduce the interference of tunnel reflections on the accuracy of feature point extraction.

[0040] 2. Feature point matching and registration (1) Preliminary matching The similarity between SIFT 3D feature points and SIFT 2D feature points is calculated using BRIEF (Binary Robust Independent Elementary Features) descriptors (5×5 neighborhood, 256-bit binary). The Ratio test (with a threshold of 0.8) is used to remove mismatched pairs, ensuring that the number of valid matching points retained is ≥20 pairs.

[0041] (2) Registration optimization The least squares method (1000 iterations, interior point threshold of 1.5 pixels) is used to remove outliers, and the initial registration parameters (rotation matrix) are solved using the PnP (Perspective-n-Point) algorithm. With translation vector Based on this, the ICP algorithm (50-100 iterations from point to surface, Euclidean distance error ≤0.5mm) is used to further optimize the registration parameters, and finally a high-precision registration matrix (R_final, T_final) is obtained.

[0042] 3. Point cloud fusion generation Based on the final high-precision registration matrix, the RGB texture features and "seepage / non-seepage" semantic labels of the seepage area in the image are associated with the three-dimensional coordinates (x, y, z) and reflection intensity attributes of the point cloud to generate a fused point cloud containing multi-dimensional information of "spatial-intensity-optical-semantic", which is stored in PLY format. At the same time, metadata such as timestamps and registration errors are recorded to provide a basis for subsequent quantitative calculations and accuracy verification.

[0043] Step 4: Quantitative calculation of total seepage area based on fused point cloud By constructing a triangular network topology of the excavation face and combining it with semantic tags from image recognition, the system can accurately calculate the leakage area and perform dynamic trend analysis.

[0044] Step 41: Constructing the Triangular Network (1) Point cloud projection and sorting Projecting the fused point cloud onto the XY plane (the cross-section of the excavation face perpendicular to the tunnel axis) yields a two-dimensional point set. Sort by x-coordinate in ascending order (if x-coordinates are the same, sort by y-coordinate in ascending order) to reduce the computational load of subsequent triangulation.

[0045] (2) Generation of triangular network Construct a super triangle: Generate a super triangle containing all projection points (vertices coordinates are set to a range of 10 times outside the boundary of the point set) to ensure coverage of the entire excavation area; Point-by-point insertion: Insert the sorted 2D points into the super triangle in sequence. Use a search algorithm to find the "influence triangle" containing the point, delete the edges within the triangle and connect the inserted point to the boundary vertex of the triangle to form a new triangle. Repeat the above operation until all points are inserted.

[0046] (3) Triangle network optimization verification Triangles containing vertices of the super triangle are removed to obtain a 2D triangulation network (Delaunay triangulation network) of the excavation face. This 2D triangulation network is then mapped back to 3D coordinates to generate a 3D triangulation network. The aspect ratio (length of longest side / length of shortest side) of each triangle in the 3D triangulation network is calculated, and narrow triangles with an aspect ratio greater than 3 are removed (ensuring a removal rate ≤ 5%). The coverage rate of the triangulation network is verified, ensuring that its error compared to the designed area of ​​the excavation face is ≤ 2%, meeting the engineering accuracy requirements.

[0047] Step 42: Calculation of the area of ​​the seepage triangle (1) Semantic tag association Using the registration relationship obtained in step 3, the semantic labels of "water seepage / non-water seepage" identified by deep learning in the image are mapped to a 3D triangular mesh, completing the labeling of all water seepage triangles.

[0048] (2) Calculation of the area of ​​a single triangular face For the triangular facet marked "seepage" (The vertex coordinates are respectively) , , The area is calculated using the vector cross product method, and the specific steps are as follows: Calculate the vector: , ; Calculate the cross product modulus: , where c = a × b (c is the cross product of vectors a and b); Calculate the area of ​​a single triangular face: (Unit: m²)

[0049] (3) Calculation of total seepage area and rate of change Total seepage area: Traverse all triangles marked "seepage", sum up the areas of each triangle, and get the total seepage area S_total (keep 2 decimal places, such as 0.92m²). Area change rate: Based on time-series fused point cloud data (collection time interval set to Δt, unit: h), the difference in total leakage area at adjacent time points is calculated to obtain the area change rate. (A positive value indicates that the leakage area is expanding, and a negative value indicates that the leakage area is shrinking); set V_S≥0.1m² / h as the threshold to determine whether the leakage is in a rapid development state.

[0050] Step 43: Verification of Calculation Accuracy (1) Comparative verification The manual "grid paper method" is used as a reference standard: a 10cm×10cm grid is drawn on a high-resolution photo of the excavation face, the number of grids covering the seepage area is counted, and the total seepage area is estimated; the total seepage area calculated by the method of this invention is compared with the result of the manual estimation of the total seepage area to ensure that the error is ≤5%.

[0051] (2) Repeated verification Data from multiple sources was collected three times on the same excavation face (with an interval of ≥30 minutes between each collection to ensure stable monitoring conditions). The total leakage area corresponding to the three data points was calculated. The standard deviation of the three areas was calculated. If the standard deviation was <0.05m², it proved that the monitoring results of the method of this invention were stable and reliable.

[0052] Step 5: Generate monitoring reports and early warnings 1. Quantification result output (1) Monitoring report generation Generate a "Quantitative Monitoring Report on Water Leakage at the Excavation Face of a Mountain Tunnel". The report should include: Basic information: monitoring time, excavation face mileage, data acquisition equipment parameters, point cloud and image registration error; Key results: total leakage area, area and location of the largest leakage area, and rate of change of leakage area; Accuracy indicators: manual verification error rate, standard deviation of repeated monitoring.

[0053] (2) Visualization The 3D triangular mesh model was rendered using CloudCompare software, with the seepage triangles rendered in red and the non-seepage triangles rendered in gray. The location and area of ​​each seepage area were marked in the model to generate a 3D visualization model. At the same time, the triangular mesh model was cut along the Y-axis of the center of the excavation face to generate a 2D cross-sectional view, which intuitively shows the distribution characteristics of seepage water on the excavation face.

[0054] 2. Four-level early warning mechanism Based on the "Technical Specification for Highway Tunnel Construction" (JTG / T3660-2020) and engineering practice experience, and combining the total leakage area (S_total) and the area change rate (V_S), thresholds are set to classify four warning levels, as follows: Normal level: S_total < 0.5m² and V_S < 0.05m² / h, only routine monitoring and recording are required, and no special response measures are needed; Attention level: If 0.5m²≤S_total<1m² or 0.05m² / h≤V_S<0.1m² / h, a blue warning light will be activated, and a monitoring report will be sent to the construction technicians simultaneously, increasing the frequency of patrols of the excavation face; Warning level: 1m²≤S_total<2m² or 0.1m² / h≤V_S<0.2m² / h, activate the yellow warning light and low-frequency buzzer, send a warning text message to project management personnel, organize technical personnel to conduct on-site verification, and formulate targeted prevention and control measures; Emergency Level: If S_total≥2m² or V_S≥0.2m² / h, activate the red warning light and high-frequency buzzer, simultaneously generate an emergency response plan (such as local grouting to plug water, strengthening the support structure, etc.), and push it to the construction party's decision-making level. If necessary, suspend excavation operations to ensure construction safety.

[0055] This invention designs a deep learning model for the common triangular features of seepage areas, overcoming the limitations of traditional square receptive fields. It achieves high segmentation accuracy and strong resistance to interference from uneven lighting and rock texture. Through high-precision point cloud registration (error ≤ 0.5mm), the image recognition results are mapped to three-dimensional space, and the area is calculated using the vector cross product method to control errors, solving the problem of "qualitative analysis being difficult to quantify" in traditional methods. Simultaneous acquisition of multi-source data and fully automated processing (from recognition to calculation time ≤ 10 minutes) allow for tracking the dynamic development trend of seepage, avoiding missed or misjudged cases during manual inspections. A four-level early warning mechanism combined with emergency response plans provides accurate decision-making support for construction teams.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based mountain tunnel excavation face seepage water quantitative monitoring method, characterized in that: Comprising the following steps: Step 1: Multi-source data synchronous acquisition Synchronously acquire image data and point cloud data of the mountainous tunnel excavation face through a video imaging device and a three-dimensional laser scanner; Step 2: Identification of seepage water area based on image deep learning (21): Triangle commonality verification of seepage water area After denoising, cropping and scale normalization of the image data collected in Step 1, the seepage water area is segmented by artificial labeling or a preliminary semantic segmentation model. The outline of the segmented seepage water area is then polygonally fitted, and geometric features including contour shape, vertex and angle, symmetry and directionality are extracted. Finally, statistical analysis of the geometric features is performed to determine the commonality of the "triangle" shape of the seepage water area, providing a theoretical basis and feature support for the design of the seepage water area identification model; (22): Design of triangle receptive field convolution module A triangle parameterized receptive field convolution module is designed for detection of the seepage water area of the mountainous tunnel excavation face. By setting the shape and scale parameters, the "square convolution kernel + triangle mask" technical solution is used to extract triangle features, and finally a modular structure including triangle convolution layer, feature enhancement layer and adaptive pooling layer is constructed to improve the feature extraction capability of the model for the seepage water area; (23): Deep learning network construction and training With the U-Net network embedded with the triangle convolution module as the core architecture, the model training and parameter optimization are completed through data augmentation, DiceLoss loss function optimization and multi-index performance verification, and finally the output of high-precision binary segmentation map of the seepage water area is realized, providing accurate target area for subsequent spatial mapping; Step 3: Mapping and fusion of image and point cloud data The original point cloud data is denoised using a combination of statistical filtering and radius filtering, and the point cloud coordinate origin is translated to the center of the excavation face. SIFT three-dimensional feature points are extracted from the denoised point cloud, and SIFT two-dimensional feature points are extracted from the image segmentation result after Gaussian blur processing. Then the similarity between SIFT three-dimensional feature points and SIFT two-dimensional feature points is calculated using BRIEF descriptor, and the false matching pairs are removed through Ratio test. The outliers are then removed using least squares method, and the initial registration parameters are solved using PnP algorithm. Finally, the high-precision registration matrix is obtained through ICP algorithm optimization; Based on the registration matrix, the image RGB texture and semantic label are associated with the point cloud three-dimensional coordinates and reflectance intensity to generate multi-dimensional fused point cloud and store it, while recording the metadata of timestamp and registration error, providing a basis for subsequent quantitative calculation and accuracy verification; Step 4: Quantitative calculation of total seepage water area based on fused point cloud (41): Construction of triangular mesh The fused point cloud is projected onto the XY plane and sorted by coordinates. A two-dimensional triangular mesh of the excavation face is generated by constructing super triangles and inserting points one by one. The two-dimensional triangular mesh is mapped back to three-dimensional coordinates to generate a three-dimensional triangular mesh. Invalid triangles are then removed and the three-dimensional triangular mesh structure is optimized. At the same time, it is verified whether the coverage error of the three-dimensional triangular mesh meets the engineering accuracy requirements; (42): Calculation of seepage triangular face area Firstly, the "water permeable / non-permeable" semantic label is mapped to a three-dimensional triangular mesh, and the marking of water permeable triangles is completed; secondly, the vector cross product method is used to calculate the area of a single water permeable triangle; finally, the total water leakage area is obtained by accumulating the areas of all water permeable triangles, and the area change rate is obtained by calculating the difference between the total water leakage areas at adjacent times, and the change rate threshold is set to determine whether the water leakage is in a rapid development state; (43): calculation accuracy verification The total water leakage area calculated in step (42) is compared with the total water leakage area calculated by the "grid paper method" to ensure that the error of the total water leakage area estimation is ≤5%; at the same time, the same excavation surface is collected 3 times, and the standard deviation of the total water leakage area of 3 times is calculated, and the standard deviation is required to be <0.05 m², to verify the stability and reliability of the monitoring results; Step 5: generate monitoring report and warning Generate a monitoring report, monitor the total water leakage area and area change rate, and issue an alarm when the set threshold is exceeded.

2. The deep learning-based mountain tunnel excavation face seepage water quantification monitoring method of claim 1, characterized in that: In step 1, the video imaging device uses an industrial-grade high-definition video imaging device, which uses an AI video monitoring measurement instrument installed on a steel support 5-10 m behind the excavation surface, ensuring that the field of view covers ≥80% of the excavation surface area and the overlapping area with the three-dimensional laser point cloud scanning area is ≥20 m²; Continuous image data collection, synchronized timestamp recording; image preprocessing: image blurring is removed by Wiener filtering, white balance correction is completed by using the gray world algorithm, brightness normalization is achieved by histogram equalization, and finally the external parameter calibration of the video imaging device is completed by Zhang's calibration method, ensuring that the image clarity and parameter accuracy meet the subsequent analysis requirements; The three-dimensional laser scanner uses a phase three-dimensional laser scanner, which uses a South Surveying and Mapping MircoIII, and installs the three-dimensional laser scanner and the video imaging device coaxially to ensure that their fields of view overlap, sets the scanning line spacing to 5 mm to ensure that the point cloud density is ≥100 points / m², continuously collects point cloud data at a frequency of 1.2 MHz, with about 5 million points per frame, and stores it in LAS format. Strictly control the timestamp error of point cloud data and image data ≤10 ms to ensure the spatio-temporal synchronization of multi-source data. 3.The deep learning-based mountain tunnel face seepage water quantitative monitoring method of claim 1, wherein: Step (21) water leakage area triangle commonality verification, specifically including: Douglas-Peucker algorithm is used to fit the polygon of the segmented water leakage area contour, and the following geometric features are extracted: Contour morphology: count the number of edges and internal angle distribution of the fitted polygon, and focus on analyzing the proportion of triangles or approximately triangles; Vertex and angle: calculate the polygon vertex coordinates, determine the leakage source point, both sides of the diffusion boundary endpoint, and analyze the concentration range of the included angle formed by the above three points; Symmetry and directionality: verify the "vertex-bottom edge" triangular structure trend of the leakage area along the gravity direction or the fracture strike, and clarify the relevance of its morphological characteristics and physical causes; Triangle commonality statistical verification: Feature statistical analysis: statistical analysis of the geometric features of a large number of samples, the core verification indicators include: ① more than 80% of the leakage area, the simplified polygon fitting result is a triangle or an approximate triangle; ② the included angle formed by the three key vertices of the leakage area contour is mainly acute, which conforms to the diffusion law of leakage water under the action of gravity; Common law explanation: combined with the physical causes of tunnel leakage water, the rationality of the "triangle" feature is verified from the mechanism level, and the influence of accidental statistical errors on the common law is excluded. 4.The deep learning-based mountain tunnel face seepage water quantitative monitoring method of claim 1, wherein: Step (22) of the triangular receptive field convolution module design, specifically includes: (1) receptive field parameterization design Combined with the triangular common feature of the leakage water area, set the triangular receptive field parameters: Morphology parameters: vertex angle 30°-60°, waist length ratio 1:1.5; Scale parameters: shallow network adopts small triangular receptive field, deep network adopts large triangular receptive field; (2) Triangular receptive field implementation Using the technical scheme of "square convolution kernel + triangular mask", breaking through the limitations of traditional square receptive field: in the weight matrix of 3x3, 5x5 square convolution kernel, only the weights of the triangular region are retained, and the weights of the non-triangular region are set to 0, so that the convolution kernel only calculates the triangular feature area in the image, which is equivalent to realizing the function of triangular receptive field; (3) Module structure building The triangular receptive field convolution module core includes three parts: Triangular convolution layer: embed the above parameterized triangular receptive field, replace the traditional convolution layer, and realize the accurate extraction of the triangular feature of the leakage water area; Feature enhancement layer: adjust the feature map channel number through 1x1 convolution, combined with BatchNormalization and ReLU6 activation function, strengthen the nonlinear mapping ability of key features such as triangular edges and vertices; Adaptive pooling layer: adopt "vertex reservation pooling" strategy, record the triangular vertex coordinates, give higher weight to the vertex position features in the pooling process, and avoid the loss of key structural information. 5.The deep learning based mountain tunnel face seepage water quantitative monitoring method according to claim 1, characterized in that: Step (23) of deep learning network construction and training, specifically includes: (1) Basic framework building Based on the U-Net encoder-decoder symmetric structure as the basic framework, the following optimization design is made: Encoder: set 4-5 encoding blocks, each encoding block contains "triangular convolution layer + ReLU6 activation function + vertex reservation pooling layer", extract multi-scale triangular features of leakage water by gradually compressing the spatial dimension; Decoder: corresponding to the encoder, set 4-5 decoding blocks, each decoding block contains "transpose convolution + feature splicing + 1x1 convolution", recover the resolution of the feature map while fusing the multi-scale features output by the encoder; Skip connection: establish a skip connection between the corresponding levels of the encoder and the decoder to pass the triangular detail features extracted by the shallow network, and make up for the feature loss problem in the upsampling process; (2) Model training optimization Dataset division: divide the labeled leakage water images into training set, validation set and test set according to the ratio of 7:2:1; through data enhancement methods such as rotation, translation and adding Gaussian noise, the data set size is expanded to 3 times the original size, avoiding model overfitting; Parameter setting: DiceLoss is selected as the loss function, Adam is used as the optimizer, and the total number of model training rounds is set to 50 rounds; Performance verification: IoU, Dice coefficient, and missed detection rate are used as model performance evaluation indicators. If there is a small range of leakage and missed detection problem, optimize the small scale receptive field parameters of the shallow network; if there is a background false detection problem, adjust the feature fusion weight of the deep network until the model segmentation accuracy is ≥92%; (3) Leaking water area output The output layer of the network uses a 1x1 convolution kernel to compress the number of feature map channels to 2; through the Sigmoid activation function, the pixel value is mapped to the 0-1 interval, and a threshold of 0.5 is set to generate a binary segmentation map to clearly mark the two-dimensional pixel position of the leaking water area, providing accurate target area for subsequent spatial mapping. 6.The deep learning based mountain tunnel face seepage water quantitative monitoring method according to claim 1, characterized in that: In step 3, image and point cloud data mapping fusion, specifically including: (1) Point cloud preprocessing The combination scheme of "statistical filtering + radius filtering" is adopted to denoise the original point cloud: the statistical filtering parameters are set as K nearest neighbor number 50 and standard deviation threshold 1.0; the radius filtering parameters are set as radius 0.01 m and minimum point number 5, ensuring that more than 90% of the effective point cloud is retained; the point cloud coordinate origin is translated to the center of the excavation surface to simplify the subsequent registration and calculation process; (2) Feature point extraction SIFT three-dimensional feature points are extracted from the preprocessed point cloud, and high-frequency texture areas such as steel support nodes and anchor rod holes are preferentially selected to ensure that the number of feature points extracted from each area of the vault, side wall, and inverted arch is ≥5; SIFT two-dimensional feature points are extracted from the image segmentation results, and Gaussian blur preprocessing is performed before extraction to reduce the interference of reflections inside the tunnel on feature point extraction accuracy; (2) Feature point matching and registration (1) Preliminary matching The similarity of SIFT three-dimensional feature points and two-dimensional feature points is calculated through BRIEF descriptor, and Ratio test is used to remove false matching pairs to ensure that the number of effective matching points retained is ≥20 pairs; (2) Registration optimization Least squares method is used to remove outliers, and PnP algorithm is used to solve the initial registration parameters; on this basis, ICP algorithm is used to further optimize the registration parameters, and finally a high-precision registration matrix is obtained; (Three) Fusion point cloud generation Based on the final registration matrix, the RGB texture features and "water leakage / non-water leakage" semantic labels of the leaking water area in the image are associated with the three-dimensional coordinates and reflectivity attributes of the point cloud to generate a fusion point cloud containing "space-intensity-optical-semantic" multi-dimensional information, stored in PLY format; at the same time, time stamp, registration error and other metadata are recorded to provide basis for subsequent quantitative calculation and accuracy verification. In step (41), triangular mesh construction, specifically including:

7. The deep learning-based mountain tunnel face seepage water quantitative monitoring method of claim 1, wherein: (1) Point cloud projection and sorting (2) Triangular mesh generation Projecting the fused point cloud to the XY plane to obtain a two-dimensional point set ; sorting the points in ascending order of x coordinate and, if the x coordinates are the same, in ascending order of y coordinate, to reduce the amount of calculation in subsequent triangulation Construct a super triangle: generate a super large triangle containing all the projected points to ensure coverage of the entire excavation surface area; Point-by-point insertion: the sorted two-dimensional points are inserted into the super triangle one by one, and the "influence triangle domain" containing the point is found through the search algorithm, the domain edges are deleted, and the inserted point and the domain boundary vertex are connected to form a new triangle; repeat the above operation until all points are inserted; ​ (3) Triangulation optimization verification Remove the triangles containing the vertices of the super triangle to obtain the two-dimensional triangular mesh of the excavation surface; map the two-dimensional triangular mesh back to three-dimensional coordinates to generate a three-dimensional triangular mesh, and calculate the "aspect ratio" of each triangular face in the three-dimensional triangular mesh. Remove the long and narrow triangular faces with an aspect ratio > 3; verify the coverage rate of the triangular mesh to ensure that the error with the design area of the excavation surface is ≤2%, meeting the engineering precision requirements. 8.The deep learning based mountain tunnel face seepage water quantitative monitoring method according to claim 1, characterized in that: In step (42), the area calculation of the water seepage triangular face includes: (1) Semantic label association Through the registration relationship obtained in step 3, the "water seepage / non-water seepage" semantic label identified by image deep learning is mapped to the three-dimensional triangular mesh, and the labeling of all water seepage triangular faces is completed. (2) Single triangular face area calculation triangles marked as "water-infiltrating" , whose vertex coordinates are , , , the area is calculated by vector cross product method, the specific steps are as follows: Computing the vector: , ; Computing the cross product module length: where c = a x b, c is the cross product of vectors a and b; Calculating the area of a single triangle face: ; (3) Total water seepage area and change rate calculation Total water seepage area: traverse all triangular faces labeled as "water seepage", and accumulate the area of each triangular face to obtain the total water seepage area S_total. Area change rate: based on the fusion of point cloud data in time series, the time interval is set as Δt, the difference of total leakage water area between adjacent time is calculated, and the area change rate is obtained , positive value indicates that the leakage water area expands, negative value indicates that the leakage water area shrinks; set V_S≥0.1 m² / h as the threshold value to determine whether the leakage water is in a rapid development state. 9.The deep learning based mountain tunnel face seepage water quantitative monitoring method according to claim 1, wherein: In step (43), the calculation accuracy verification includes: (1) Comparative verification Use the manual "grid paper method" as the reference standard: draw a 10cm×10cm grid on the high-definition photo of the excavation surface, count the number of grids covered by the water seepage area, and estimate the water seepage area. Compare the total water seepage area calculated by the method of the present application with the manual estimation result to ensure that the error is ≤5%; (2) Repeated verification Collect multi-source data for the same excavation surface for 3 times with an interval of ≥30 minutes to ensure stable monitoring conditions, calculate the total water seepage area corresponding to the 3 times of data respectively; calculate the standard deviation of the 3 times of area, if the standard deviation <0.05m², it proves that the monitoring result of the method of the present application is stable and reliable.

10. The deep learning-based mountain tunnel face seepage water quantitative monitoring method of claim 1, wherein: In step 5, the monitoring report is generated, which includes: Generate a "mountain tunnel excavation surface water seepage quantitative monitoring report", the report content needs to include: Basic information: monitoring time, excavation surface mileage, data acquisition equipment parameters, point cloud and image registration error; Core results: total water seepage area, maximum water seepage area and position, water seepage area change rate; Accuracy index: manual verification error rate, repeated monitoring standard deviation; Also includes visualization: Use CloudCompare software to render the three-dimensional triangular mesh model, with water seepage triangular faces rendered in red and non-water seepage triangular faces rendered in gray; label the position and area of each water seepage area in the model to generate a three-dimensional visualization model; at the same time, cut the triangular mesh model along the center Y axis of the excavation surface to generate a two-dimensional profile, which intuitively shows the distribution characteristics of water seepage in the excavation surface; Early warning includes: Set thresholds for total water seepage area S_total and area change rate V_S to divide four levels of early warning, as follows: Normal level: S_total <0.5m² and V_S <0.05m² / h, only routine monitoring is required, no special response measures are needed; Attention level: 0.5m²≤S_total<1m² or 0.05m² / h≤V_S<0.1m² / h, turn on the blue early warning light, push the monitoring report to the construction technical personnel at the same time, and strengthen the patrol frequency of the excavation surface; Early warning level: 1 m²≤S_total<2 m² or 0.1 m² / h≤V_S<0.2 m² / h, activate yellow warning light and low-frequency buzzer, send early warning message to project managers, organize technical personnel to check on site, and develop targeted prevention and control measures; Emergency level: S_total≥2 m² or V_S≥0.2 m² / h, start red warning light and high-frequency buzzer, generate emergency disposal scheme synchronously, push to the decision-making layer of the construction party, and suspend excavation work if necessary to ensure construction safety.

Images