Feature recognition and annotation method and system for 3D tunnel modeling

Abstract

An embodiment of the present invention discloses a feature recognition and annotation method and system for three-dimensional tunnel modeling. The method first pre-processes the acquired tunnel construction design drawings and point cloud scanning data of the corresponding construction stage to obtain a three-dimensional tunnel model data set with a spatial coordinate alignment relationship. The three-dimensional tunnel model data set includes structural point cloud units fused from multi-source data and contour line units vectorized from the design drawings; secondly, three-dimensional structural features are extracted from the data set to obtain geometric morphological features of the structural point cloud units and structural association features of the contour line units; then, these two features are jointly analyzed through a deep learning network to generate a feature analysis result including a feature response distribution; finally, based on this result, key structural areas in the three-dimensional tunnel model data set are identified and annotated to generate a feature recognition and annotation result including feature type labels and spatial position information, thereby providing a more accurate analysis and management basis for tunnel construction.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of computer data processing technology, and in particular to a feature recognition and annotation method and system for three-dimensional tunnel modeling. Background Art

[0002] During tunnel construction, accurately identifying and effectively labeling key structural areas is crucial for ensuring construction safety, improving construction quality, and optimizing construction progress. Currently, existing technologies offer a variety of approaches for processing tunnel construction data to identify and label structural areas.

[0003] A common existing technique relies solely on tunnel construction design drawings for analysis. However, construction design drawings only reflect theoretical design scenarios. During actual construction, the actual tunnel structure may deviate from the design drawings due to complex geological conditions, differences in construction techniques, and unforeseen factors. The drawings cannot reflect the dynamic changes in the structure during construction in real time, making it difficult to detect problems such as local deformation and structural damage, thus failing to provide timely and effective guidance for the construction process. Another existing technique utilizes only point cloud scanning data for tunnel structure analysis. However, point cloud data itself is discrete and massive, lacking clear semantic information and structural associations. Relying solely on point cloud data makes it difficult to determine the specific functions and relationships between various tunnel components, and it is difficult to accurately identify hidden structures or complex connections. Furthermore, the processing and analysis of point cloud data requires significant computing resources and time, resulting in low efficiency. Some existing techniques, while attempting to combine tunnel construction design drawings and point cloud scanning data for analysis, suffer from deficiencies in data fusion and feature extraction. The lack of systematic data processing prevents the full exploration of potential relationships between various data sources, resulting in incomplete and inaccurate extracted features.

[0004] Therefore, how to optimize the modeling feature recognition and annotation during the tunnel construction process to provide a more accurate analysis and management basis for tunnel construction is a technical challenge that needs to be overcome. Summary of the Invention

[0005] An embodiment of the present invention provides a feature recognition and annotation method and system for three-dimensional tunnel modeling.

[0006] In a first aspect, an embodiment of the present invention provides a feature recognition and annotation method for three-dimensional tunnel modeling, which is applied to a feature recognition and annotation system. The method includes:

[0007] Preprocessing the acquired tunnel construction design drawings and point cloud scan data from the corresponding construction phase to obtain a three-dimensional tunnel model dataset with spatial coordinate alignment, the three-dimensional tunnel model dataset comprising structural point cloud units fused from multi-source data and contour line units vectorized from the design drawings;

[0008] Performing three-dimensional structural feature extraction processing on the three-dimensional tunnel model data set to obtain geometric features of the structural point cloud unit and structural association features of the contour line unit;

[0009] Performing joint analysis and processing on the geometric features and the structural association features through a deep learning network to generate a feature analysis result including a feature response distribution;

[0010] Based on the feature analysis results, key structural areas in the three-dimensional tunnel model data set are identified and annotated to generate feature identification and annotation results including feature type labels and spatial position information.

[0011] In a second aspect, an embodiment of the present invention provides a feature recognition and annotation system, including:

[0012] processor;

[0013] a storage device having a computer program stored thereon,

[0014] When the computer program is executed by the processor, the processor implements any of the feature recognition and annotation methods applied to three-dimensional tunnel modeling.

[0015] An embodiment of the present invention provides a readable storage medium storing a program or instruction. When the program or instruction is executed by a processor, the steps of the feature recognition and annotation method applied to three-dimensional tunnel modeling are implemented.

[0016] The embodiment of the present invention processes tunnel construction design drawings and point cloud scanning data from corresponding construction phases to obtain a three-dimensional tunnel model dataset with spatial coordinate alignment, achieving effective fusion of multi-source data and ensuring that the structural point cloud units and the vectorized contour line units of the design drawings complement each other. Three-dimensional structural features are extracted from the three-dimensional tunnel model dataset to obtain the geometric features of the structural point cloud units and the structural association features of the contour line units, respectively. This provides in-depth analysis of the tunnel structure from different perspectives and provides rich feature information for accurately identifying key structural areas. A deep learning network is used to jointly analyze and process these two features, fully leveraging the powerful feature learning capabilities of deep learning. This network can uncover potential relationships between features and generate feature analysis results containing feature response distributions, thereby accurately determining the importance of different structural features. Based on the feature analysis results, key structural areas are identified and annotated, generating feature identification and annotation results containing feature type labels and spatial location information. This provides an intuitive and detailed basis for monitoring, management, and maintenance of tunnel construction, helping to improve construction efficiency and quality and ensure the safe and stable operation of tunnels. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a feature recognition and annotation method for three-dimensional tunnel modeling provided by an embodiment of the present invention.

[0018] Figure 2 A schematic diagram of the basic structure of a feature recognition and annotation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0020] See also Figure 1 As shown in FIG, this figure is a flow chart of a feature recognition and annotation method for three-dimensional tunnel modeling provided by an embodiment of the present invention. This method can be applied to a feature recognition and annotation system. Figure 1 As shown, the method may include steps 110 to 140.

[0021] Step 110: Preprocess the acquired tunnel construction design drawings and the point cloud scanning data of the corresponding construction stage to obtain a three-dimensional tunnel model dataset with a spatial coordinate alignment relationship. The three-dimensional tunnel model dataset includes structural point cloud units fused from multiple source data and contour line units vectorized from the design drawings.

[0022] In tunnel construction scenarios, the first step is to obtain the tunnel's construction design drawings and point cloud scan data from the corresponding construction phase. The construction design drawings detail the tunnel's design structure, dimensions, and other information, while the point cloud scan data captures the actual tunnel space using technologies such as 3D laser scanning.

[0023] Then, the two types of data are preprocessed. The preprocessing process includes vectorizing the construction design drawings, converting the graphics in the drawings into vector graphics that can be recognized and processed by the computer, and forming vectorized contour line units of the design drawings. For the point cloud scanning data, denoising, filtering and other operations are required to remove noise points and outliers and improve the quality of the data. Through the spatial coordinate alignment algorithm, the coordinate systems of the construction design drawings and the point cloud scanning data are unified so that the two have an aligned relationship in space, and finally a three-dimensional tunnel model data set consisting of structural point cloud units fused from multi-source data and vectorized contour line units of the design drawings is obtained. For example, by vectorizing the construction design drawings, the contour segments of each part of the tunnel can be obtained, and after processing, the point cloud scanning data can accurately reflect the actual spatial structure of the tunnel.

[0024] Step 120: performing three-dimensional structural feature extraction processing on the three-dimensional tunnel model data set to obtain geometric features of the structural point cloud unit and structural association features of the contour line unit.

[0025] After obtaining the 3D tunnel model dataset, it is necessary to extract its 3D structural features. For the structural point cloud units, their geometric features, namely the external shape of the actual tunnel structure, surface undulations, and other features, need to be analyzed. For the contour line units, the focus is on extracting their structural association features, namely the connection relationships and interactions between the various structural components of the tunnel. By extracting these two types of features, we can gain a deeper understanding of the structural characteristics and internal relationships of the tunnel. For example, the geometric features of the structural point cloud units can reflect whether the tunnel has deformations, bulges, etc., while the structural association features of the contour line units can determine whether the connections between the various parts of the tunnel are firm and reasonable.

[0026] In an optional embodiment, the performing of three-dimensional structural feature extraction processing on the three-dimensional tunnel model dataset to obtain the geometric features of the structural point cloud unit and the structural association features of the contour line unit includes:

[0027] Step 121: performing noise filtering processing on outlier points and redundant points on the structure point cloud unit to obtain a smoothed structure point cloud subset.

[0028] In the actual implementation process, there may be outliers and redundant points in the structural point cloud unit. These points will affect the accuracy of subsequent feature extraction, so the structural point cloud unit needs to be subjected to noise filtering. Outliers refer to points with obviously abnormal distances from surrounding points, and redundant points refer to points that are repeated or similar in space. Through noise filtering algorithms, such as density-based noise filtering algorithms, outliers and redundant points are removed to obtain a smoothed structural point cloud subset. For example, in the point cloud data, there may be some outliers caused by scanning errors, which are far away from the actual structure of the tunnel. Through noise filtering, these points can be removed, making the point cloud data smoother and accurately reflecting the actual structure of the tunnel.

[0029] Step 122: Perform neighborhood analysis on the smoothed structural point cloud subset, calculate the local curvature parameters and normal vector direction of each point cloud point, and generate geometric features that reflect the undulating state of the structural surface. The geometric features include the curvature distribution pattern and normal vector consistency index of the point cloud points.

[0030] After obtaining the smoothed structural point cloud subset, further neighborhood analysis processing is performed. For each point cloud point, the distribution of points in its neighborhood is analyzed, and the local curvature parameters and normal vector directions of the point are calculated. The local curvature parameters can reflect the degree of surface curvature at the location of the point, and the normal vector direction represents the normal direction of the surface where the point is located. By statistically analyzing the local curvature parameters and normal vector directions of all point cloud points, the curvature distribution pattern and normal vector consistency index of the point cloud points are generated. These indicators together constitute the geometric morphological features that reflect the undulating state of the structural surface. For example, in the vault part of the tunnel, the local curvature parameters and normal vector directions of the point cloud points will show corresponding distribution patterns. By analyzing these patterns, the surface undulations of the tunnel vault can be accurately understood.

[0031] Step 123: performing segment connectivity analysis on the contour line units, identifying contour line segments in the tunnel construction design drawings, and extracting length parameters and angle parameters of the contour line segments as basic structural features.

[0032] For contour line elements, segment connectivity analysis is required. First, the contour line segments in the tunnel construction design drawings are identified. These segments represent the boundaries of the tunnel's various structural components. Next, the length and angle parameters of these contour line segments are extracted. The length parameter reflects the length of the segment, while the angle parameter reflects the relative positional relationship between the segments. These length and angle parameters constitute the basic structural characteristics. For example, at the corners of a tunnel, the length and angle of the contour line segments will change significantly. By extracting these parameters, the structural characteristics of the tunnel corners can be accurately described.

[0033] Step 124: Structural correlation modeling is performed on the contour line units based on the basic structural features, and the intersection relationship and adjacency relationship between different contour line segments are analyzed to generate structural correlation features reflecting the connection mode of tunnel structural components. The structural correlation features include the distribution density of line segment intersections and the angular matching degree of adjacent line segments.

[0034] Based on the extracted basic structural features, the structural correlation modeling of the contour line units is performed. The intersection and adjacency relationships between different contour line segments are analyzed, that is, which line segments intersect with each other and which line segments are adjacent. By analyzing these relationships, the distribution density of line segment intersections and the angular matching degree of adjacent line segments are calculated. The distribution density of line segment intersections reflects the complexity of the connection between tunnel structural components, while the angular matching degree of adjacent line segments reflects the rationality and stability of the connection. These indicators together constitute the structural correlation characteristics that reflect the connection method of tunnel structural components. For example, at the support structure of the tunnel, different contour line segments will intersect and abut each other. By analyzing the distribution density of line segment intersections and the angular matching degree of adjacent line segments, it is possible to evaluate whether the connection of the support structure is firm and reasonable.

[0035] On the basis of steps 121 to 124, the method further includes:

[0036] Step 125: Perform unified dimension conversion processing on the geometric features and the structural association features to obtain a target feature set with the same feature dimension. The information entropy calculation result of the target feature set is used to adjust the feature fusion weight parameters of the deep learning network in the joint analysis process.

[0037] After obtaining the geometric features and structural association features, since their feature dimensions may differ, a unified dimensionality conversion process is required. Using a dimensionality conversion algorithm, the geometric features and structural association features are converted into a target feature set with the same feature dimensions. Information entropy is then calculated for the target feature set, which reflects the uncertainty and information content of the features. Based on the entropy calculation results, the feature fusion weight parameters of the deep learning network in the joint analysis process are adjusted. For example, if a feature has a high entropy, it indicates that it contains a large amount of information and can be given a higher weight during feature fusion to improve analysis accuracy.

[0038] Step 130: Perform joint analysis and processing on the geometric features and the structural association features through a deep learning network to generate a feature analysis result including a feature response distribution.

[0039] The obtained geometric features and structural association features are input into a deep learning network for joint analysis. Deep learning networks can exploit potential relationships between features based on their feature learning and analysis capabilities. In the deep learning network, geometric features and structural association features are processed and analyzed at multiple levels and scales. Through operations such as convolution, pooling, and activation, a feature analysis result containing a feature response distribution is ultimately generated. The feature response distribution reflects the importance of different structural features in the analysis, namely, the activation value and feature type identifier corresponding to each feature dimension. For example, through the joint analysis of geometric features and structural association features in the deep learning network, it can be found that certain key structural parts of the tunnel have strong feature responses, indicating that these parts play an important role in the tunnel structure.

[0040] As an optional embodiment, the joint analysis and processing of the geometric features and the structural association features by a deep learning network to generate a feature analysis result including a feature response distribution includes:

[0041] Step 131: Input the geometric features and the structural association features into the multi-scale feature fusion layer of the deep learning network, perform multi-resolution downsampling processing on the features in combination with a preset scale division rule, and generate a multi-scale feature stream.

[0042] In this embodiment of the present invention, geometric features and structural association features are input into the multi-scale feature fusion layer of a deep learning network. In this layer, features are subjected to multi-resolution downsampling according to preset scale division rules. Multi-resolution downsampling can analyze features at different scales, extract feature information at different scales, and generate a multi-scale feature stream. For example, by dividing features at different scales, global and local information of features can be obtained, and the multi-scale feature stream can contain feature information at different scales.

[0043] Step 132: For the local detail feature stream in the multi-scale feature stream, convolution kernels with different receptive fields are used to perform feature extraction processing to capture the local curvature change details of the point cloud in the geometric features and the line segment adjacency relationship details in the structural association features, and obtain local detail feature output.

[0044] In the multi-scale feature stream, convolution kernels with different receptive fields are used to extract features from the local detail feature stream. Convolution kernels with different receptive fields can capture feature information in different ranges. For geometric features, the focus is on capturing the details of the local curvature changes of the point cloud, that is, the local bending changes of the tunnel surface. For structural association features, attention is paid to the details of the line segment adjacency relationship, that is, the specific relationship between the adjacent line segments between the various structural components of the tunnel. Through the processing of the convolution kernel, a local detail feature output reflecting this detailed information is obtained. For example, using a convolution kernel with a small receptive field can capture subtle changes in the local curvature of the point cloud, while a convolution kernel with a large receptive field can obtain more extensive line segment adjacency information.

[0045] In a preferred embodiment, the local detail feature stream in the multi-scale feature stream is subjected to feature extraction processing using convolution kernels with different receptive fields to capture the local curvature change details of the point cloud in the geometric features and the line segment adjacency relationship details in the structural association features, thereby obtaining local detail feature output, including:

[0046] Step 1321: performing window division processing on the geometric features of the local detail feature stream to generate a plurality of local point cloud window units containing a preset number of point cloud points.

[0047] The geometric features in the local detail feature stream are windowed and divided into multiple local point cloud window units according to a preset window size and number. Each window unit contains a preset number of point cloud points. For example, the window size is set to a fixed spatial range, and a certain number of point cloud points are included within this range, thereby localizing the geometric features.

[0048] Step 1322: Perform statistical analysis on the point cloud point curvature distribution pattern and normal vector consistency index within each local point cloud window unit, and calculate the mean value and standard deviation of the curvature within the window as the statistical characteristics of the local curvature change.

[0049] For each local point cloud window unit, a statistical analysis is performed on the curvature distribution pattern and normal vector consistency index of the point cloud points within the window. The mean and standard deviation of the curvature of the point cloud points within the window are calculated. The mean reflects the overall level of curvature within the window, while the standard deviation reflects the degree of dispersion of the curvature. Together, these two indicators constitute the statistical characteristics of local curvature variation and can accurately describe the curvature changes of the local point cloud. For example, a large standard deviation of curvature within a window indicates that the surface fluctuations in that area are more severe.

[0050] Step 1323: performing line segmentation processing on the structural association features of the local detail feature stream to generate a plurality of local line segment group units containing continuous line segments.

[0051] Segment segmentation is performed on structurally related features in the local detail feature stream. Contour segments are segmented into multiple local segment group units according to specific rules, with each unit containing continuous segments. For example, adjacent and continuous segments are grouped into a local segment group unit based on their connectivity and spatial position. This allows for localized processing of structurally related features, facilitating analysis of segment adjacency relationships.

[0052] Step 1324: Perform sequence analysis on the segment length parameters and angle parameters in each local segment group unit, and extract the increasing and decreasing pattern of the segment length and the periodic change pattern of the angle as local segment adjacency relationship features.

[0053] A sequence analysis is performed on the segment length parameters and angle parameters within each local segment group unit. The changing trend of the segment length is analyzed to extract its increasing and decreasing pattern, that is, to determine whether the segment length is gradually increasing or decreasing. At the same time, the changes in the angle are analyzed to extract its periodic change pattern. These patterns constitute the local segment adjacency relationship characteristics and can reflect the characteristics of the adjacency relationship between local segments. For example, if the segment length within a local segment group unit shows an increasing pattern and the angle has a certain periodic change, it means that the structure of the area has a certain regularity.

[0054] Step 1325: Input the local curvature change statistical features and the local line segment adjacency relationship features into the local feature enhancement subnet to generate an enhanced local feature vector containing multi-source local detail information.

[0055] In this embodiment of the present invention, local curvature variation statistics and local line segment adjacency features are input into a local feature enhancement subnet. This subnet is a specially designed neural network module that further processes and enhances the input features, exploring potential relationships between them. Through this processing, an enhanced local feature vector containing multi-source local detail information is generated. This vector incorporates local detail information from both geometric features and structural association features.

[0056] Step 1326: Perform nonlinear activation processing on the enhanced local feature vector to obtain a local detail feature output reflecting the details of the local curvature change of the point cloud and the details of the line segment adjacency relationship.

[0057] Optionally, nonlinear activation processing is performed on the enhanced local feature vectors. Using an activation function, such as the ReLU function, the enhanced local feature vectors are nonlinearly transformed to highlight important features and suppress unimportant ones. After activation, local detail features are output that reflect details of local curvature variations in the point cloud and details of line segment adjacency. For example, the activation function can make features with large eigenvalues ​​more prominent, thereby improving analysis accuracy.

[0058] Step 133: For the global context feature stream in the multi-scale feature stream, the geometric features and the structural association features are subjected to global dependency modeling through a self-attention mechanism to generate a context feature vector reflecting the overall structural layout of the tunnel.

[0059] Optionally, a self-attention mechanism is used to model global dependencies within the global context feature stream within the multi-scale feature stream. The self-attention mechanism automatically calculates correlations between features and captures global dependencies. Using geometric features and structural association features as input, the self-attention mechanism calculates attention weights between different feature positions, thereby establishing a global dependency model. Ultimately, a context feature vector is generated that reflects the overall structural layout of the tunnel. This vector captures the interrelationships between various tunnel components and the overall layout. For example, certain features in different parts of the tunnel may have strong dependencies, and the self-attention mechanism can accurately capture these relationships.

[0060] In another preferred embodiment, the global context feature stream in the multi-scale feature stream is subjected to global dependency modeling processing on the geometric features and the structural association features through a self-attention mechanism to generate a context feature vector reflecting the overall structural layout of the tunnel, including:

[0061] Step 1331: Map the geometric features and the structural association features into a query vector, a key vector and a value vector respectively, wherein the query vector is used to represent the feature information of the target feature position, the key vector is used to represent the feature information of the non-target feature position, and the value vector is used to represent the feature value of the non-target feature position.

[0062] Optionally, the geometric features and structural association features are mapped separately to obtain a query vector, a key vector, and a value vector. The query vector represents the feature information of the target feature position, that is, the information of the feature position of current interest. The key vector represents the feature information of non-target feature positions and is used to match the query vector. The value vector represents the feature values ​​of the non-target feature positions and is used for subsequent weighted summation. For example, in the self-attention mechanism, by matching the query vector and the key vector, other feature positions related to the target feature position can be found, and then the value vector can be used to fuse this information.

[0063] Step 1332: Calculate the dot product similarity between the query vector and the key vector, and generate an attention weight matrix that reflects the degree of association between feature positions.

[0064] In this step, the dot product similarity between the query vector and the key vector is calculated. Dot product similarity can reflect the degree of similarity between the two vectors. By calculating the dot product similarity of all query vectors and key vectors, an attention weight matrix is ​​generated. This matrix reflects the degree of association between different feature positions. For example, if the dot product similarity between a query vector and a key vector is large, it indicates that the association between the two feature positions is strong, and it can be given a larger weight in subsequent processing.

[0065] Step 1333: Perform softmax normalization on the attention weight matrix to obtain the attention distribution probability of each feature position to the target feature position.

[0066] The softmax function converts the values ​​in the attention weight matrix into a probability distribution, so that the sum of the attention distribution probabilities of each feature position with respect to the target feature position is 1. Through normalization, the attention distribution probabilities of each feature position with respect to the target feature position are obtained. These probabilities can more accurately represent the degree of association between features. For example, a feature position with a larger attention distribution probability indicates that it has a greater influence on the target feature position.

[0067] Step 1334: Perform weighted summation processing on the value vector based on the attention distribution probability to generate a context-enhanced feature vector containing global feature position dependencies.

[0068] Furthermore, the value vectors are weighted and summed based on the attention distribution probability. Each value vector is multiplied by the corresponding attention distribution probability, and all the results are summed to generate a context-enhanced feature vector that contains global feature position dependencies. This vector incorporates information from different feature positions and reflects the global dependencies of features. For example, if the attention distribution probability of a feature position is large, then the corresponding value vector will account for a larger proportion in the weighted sum and contribute more to the context-enhanced feature vector.

[0069] Step 1335: Perform residual connection processing on the context-enhanced feature vector, the geometric morphological features, and the structural association features to obtain an initial context feature vector that integrates the global dependency relationship.

[0070] Optionally, a residual connection is performed on the context-enhanced feature vector with the geometric features and structural association features. This residual connection preserves the original feature information while incorporating global dependency information. By adding the context-enhanced feature vector to the original features, an initial context feature vector incorporating global dependencies is obtained. For example, residual connections can avoid issues such as vanishing gradients during information transfer, ensuring efficient information transfer.

[0071] Step 1336: Perform dimension compression processing on the initial context feature vector to generate a context feature vector that reflects the overall structural layout of the tunnel.

[0072] In embodiments of the present invention, dimensionality reduction can be implemented using dimensionality reduction algorithms, such as principal component analysis, to reduce the dimensionality of the feature vector while preserving its key information. This ultimately generates a contextual feature vector that reflects the overall structural layout of the tunnel. This vector is dimensionally more concise, facilitating subsequent processing and analysis. For example, dimensionality reduction can remove redundant information from the feature vector, improving computational efficiency and analytical accuracy.

[0073] Step 134: Input the local detail feature output and the context feature vector into a feature interaction network, perform information complementation through an element-by-element product operation, and generate an interactive feature map with spatiotemporal consistency constraints.

[0074] In this embodiment of the present invention, the local detail feature output and the context feature vector are input into the feature interaction network, where they are then complemented through an element-by-element multiplication operation. The local detail feature output contains detailed information about the local area of ​​the tunnel, such as local curvature changes in the point cloud and details about line segment adjacency, while the context feature vector reflects information about the overall structural layout of the tunnel. This element-by-element multiplication enables the fusion of local and global information, ensuring that the features at each location not only encompass local details but also consider their position and relationship within the overall structure.

[0075] During the element-by-element product operation, it is important to ensure that the dimensions of the local detail feature output and the context feature vector match. If the dimensions do not match, adjustments may be necessary, such as padding the lower-dimensional features or cropping the higher-dimensional features, to ensure smooth multiplication. The result of the element-by-element product operation is an interactive feature map with spatiotemporal consistency constraints. This interactive feature map integrates local and global information to more accurately reflect the characteristics of the tunnel structure.

[0076] For example, at a key structural location in a tunnel, the local detail feature output might reveal significant changes in the point cloud curvature, indicating significant surface undulation. Conversely, the context feature vector might indicate the importance of this location within the overall structure and its close connection to surrounding structures. Through element-by-element multiplication, the interactive feature map can simultaneously capture both aspects of information.

[0077] Step 135: performing channel normalization processing on the interaction feature map, and splicing the normalized interaction feature map along the channel dimension to generate a feature fusion vector containing multi-scale interaction information.

[0078] As you can understand, after obtaining the interaction feature map, it is subjected to channel normalization. The purpose of channel normalization is to standardize the data of each channel so that the data of different channels have a similar distribution range, which facilitates subsequent processing and comparison. During the channel normalization process, the mean and standard deviation of each channel data are calculated. The mean is then subtracted from the data of each channel and divided by the standard deviation to obtain the normalized channel data.

[0079] After channel normalization, the normalized interaction feature maps are concatenated along the channel dimension. This concatenation operation arranges the data from different channels in sequence along the channel dimension to form a new feature vector. This feature vector contains multi-scale interaction information and incorporates the information from each channel of the channel-normalized interaction feature map.

[0080] For example, an interaction feature map may contain multiple channels, each of which reflects different aspects of information. Through channel normalization and splicing, these different aspects of information can be integrated together to form a more comprehensive and richer feature fusion vector.

[0081] Step 136: Call the response distribution generation layer of the deep learning network to perform activation function mapping processing on the feature fusion vector to generate a feature response distribution that reflects the importance of different structural features. The feature response distribution includes the activation value of each feature dimension and the corresponding feature type identifier.

[0082] Optionally, the feature fusion vector is input into the response distribution generation layer of the deep learning network, where an activation function is used to map the feature fusion vector. The activation function performs a nonlinear transformation on each element in the feature fusion vector so that the output reflects the importance of different structural features.

[0083] After activation function mapping, a feature response distribution is generated. This distribution contains the activation value for each feature dimension and the corresponding feature type identifier. The activation value indicates the importance of that feature dimension in the analysis; a larger activation value indicates a more important structural feature. The feature type identifier is used to distinguish different types of structural features, such as the main tunnel structure and ancillary facility structures.

[0084] For example, if a feature dimension in the feature response distribution has a large activation value and the corresponding feature type is identified as the main tunnel structure, this indicates that the main tunnel structure feature represented by this feature dimension is highly important within the entire tunnel structure. Using the feature response distribution, we can quickly identify which structural features in the tunnel are critical.

[0085] Step 140: Identify and annotate key structural areas in the three-dimensional tunnel model dataset based on the feature analysis results, and generate feature identification and annotation results including feature type labels and spatial position information.

[0086] In this embodiment of the present invention, after obtaining feature analysis results, key structural regions within the 3D tunnel model dataset are identified and labeled based on these results. The feature response distribution in the feature analysis results can reflect the importance of different structural features. By analyzing the feature response distribution, it is possible to determine which structural regions corresponding to the feature dimensions are critical.

[0087] For identified critical structural areas, feature type labels and spatial location information are assigned. Feature type labels describe the structural and functional attributes of the critical structural areas, such as the tunnel main structure and ancillary facilities. Spatial location information specifies the specific location of the critical structural areas within the 3D tunnel model dataset, including spatial coordinate range and contour segment coverage.

[0088] For example, if the feature response distribution reveals a high activation value for a feature dimension, the corresponding structural region may be the tunnel support structure. This region is assigned the feature type label "tunnel support structure," and its spatial coordinate range in the point cloud scan data and the coverage of its contour segments in the tunnel construction design drawings are determined, thereby completing the identification and labeling of this key structural region.

[0089] As an implementation method, identifying and annotating key structural areas in the three-dimensional tunnel model dataset based on the feature analysis results to generate feature identification and annotation results including feature type labels and spatial location information includes:

[0090] Step 141: extracting feature dimensions whose activation values ​​exceed a preset threshold as key feature dimensions based on the feature response distribution in the feature analysis result, wherein the key feature dimensions correspond to key structural areas of the tunnel.

[0091] A preset threshold is set within the feature response distribution of the feature analysis results. Feature dimensions whose activation values ​​exceed this threshold are extracted and defined as key feature dimensions. The structural regions corresponding to these key feature dimensions are considered key structural regions of the tunnel, as the features in these regions are highly important in the analysis.

[0092] For example, a threshold is preset as a relatively large activation value standard. When the activation value of a feature dimension exceeds this threshold, it indicates that the structural area represented by this feature dimension plays a key role in the tunnel structure, such as the tunnel entrance and exit, important connection nodes, etc. By extracting key feature dimensions, key structural areas of the tunnel can be quickly located.

[0093] Step 142: assigning a corresponding feature type label to each key structural region according to the feature type identifier corresponding to the key feature dimension, wherein the feature type label includes a structural component type and a functional attribute category.

[0094] After determining the key feature dimensions, a feature type label is assigned to each key structural region based on its corresponding feature type identifier. Feature type identifiers are identifiers corresponding to each feature dimension in the feature response distribution and are used to distinguish different types of structural features. Feature type labels are further refined to include structural component types and functional attribute categories.

[0095] For example, if the feature type corresponding to the key feature dimension is identified as the tunnel main structure, then a more detailed feature type label can be assigned to the key structure area based on the specific feature information, such as "tunnel vault structure", where "tunnel vault" is the structural component type and "structure" reflects its functional attribute category.

[0096] As a preferred embodiment, the step of assigning a corresponding feature type label to each key structural region according to the feature type identifier corresponding to the key feature dimension includes:

[0097] Step 1421: Establish a mapping relationship between a feature type identifier and a preset label library, wherein the preset label library includes a tunnel main structure label, ancillary facility structure label, and a potential defect structure label.

[0098] First, establish a mapping relationship between feature type identifiers and a preset label library. This library is a predefined set of labels, including different types of labels, such as those for tunnel main structure, auxiliary facility structures, and potentially defective structures. By establishing this mapping relationship, feature type identifiers can be matched to specific labels.

[0099] For example, the feature type identifier is "main structure related", which can be mapped to "tunnel main structure label" in the preset label library. Therefore, the corresponding label category can be quickly found according to the feature type identifier.

[0100] Step 1422: Extract the activation value peak position in the feature response distribution of the key feature dimension, and determine the core feature type identifier corresponding to the key feature dimension; according to the core feature type identifier, search the corresponding basic label category in the preset label library, and the basic label category includes the first-level label of the structural component type and the second-level label of the functional attribute category.

[0101] For the feature response distribution of the key feature dimension, extract the peak location of the activation value. The feature type identifier corresponding to the activation value peak location is the core feature type identifier, representing the most important feature type for that key feature dimension. Based on the core feature type identifier, search the preset label library for the corresponding basic label category. The basic label category includes the primary label for the structural component type and the secondary label for the functional attribute category.

[0102] For example, the feature type corresponding to the peak position of the activation value in the feature response distribution of the key feature dimension is identified as "tunnel support". The corresponding basic label category can be found in the preset label library, such as the first-level label "support structure" and the second-level label "bearing function".

[0103] Step 1423: Based on the second highest peak position in the feature response distribution of the key feature dimension, extract the auxiliary feature type identifier reflecting the additional attributes of the feature; according to the auxiliary feature type identifier, search the corresponding supplementary label information in the preset label library, and the supplementary label information includes additional labels for structural material type and construction process category.

[0104] In addition to the activation value peak location, the second-highest peak location in the feature response distribution of the key feature dimension is also considered. The feature type identifier corresponding to the second-highest peak location is the auxiliary feature type identifier, which reflects the additional attributes of the feature. Based on the auxiliary feature type identifier, the corresponding supplementary label information is searched in the preset label library. The supplementary label information includes additional labels for structural material type and construction process category.

[0105] For example, the feature type corresponding to the second highest peak position is identified as "concrete material", and the corresponding supplementary label information can be found in the preset label library, such as "concrete structural material" and "cast-in-place construction technology" and other additional labels.

[0106] Step 1424: Combine and concatenate the basic label category and the supplementary label information to generate a feature type label including a multi-level label hierarchy, where the multi-level label hierarchy is used to describe the structural attributes and functional attributes of the key structural area.

[0107] The basic label category and supplementary label information are combined and spliced ​​to form a feature type label containing a multi-level label hierarchy. The multi-level label hierarchy can describe the structural and functional properties of key structural areas in more detail.

[0108] For example, the basic label category "supporting structure-bearing function" and the supplementary label information "concrete structure material-cast-in-place construction technology" are combined and spliced ​​to obtain the feature type label "supporting structure-bearing function-concrete structure material-cast-in-place construction technology", which comprehensively describes the properties and functions of the key structural area.

[0109] Step 143: extracting structural point cloud units and contour line units associated with the key feature dimensions in the three-dimensional tunnel model data set, and determining the spatial coordinate range of the key structural area in the point cloud scanning data and the contour line segment coverage in the tunnel construction design drawing.

[0110] After determining the key feature dimensions, structural point cloud units and contour line units associated with these key feature dimensions are extracted from the 3D tunnel model dataset. The structural point cloud units reflect the actual spatial structure of the tunnel, while the contour line units represent the tunnel's outline information as shown in the design drawings.

[0111] By analyzing the extracted structural point cloud units and contour line units, the spatial coordinate range of the key structural areas in the point cloud scan data and the contour line segment coverage in the tunnel construction design drawings are determined. The spatial coordinate range is determined by calculating the minimum bounding box of the point cloud unit, while the contour line segment coverage is determined by extracting the relevant contour line segments and determining their starting and ending point coordinates.

[0112] For example, in point cloud scan data, by analyzing point cloud units associated with key feature dimensions and calculating the vertex coordinates of their minimum bounding box, the spatial coordinate range of key structural areas can be determined. In tunnel construction design drawings, relevant contour segments are extracted, and their starting and ending point coordinates are recorded to determine the contour segment coverage.

[0113] Furthermore, extracting the structural point cloud units and contour line units associated with the key feature dimensions in the three-dimensional tunnel model data set, and determining the spatial coordinate range of the key structural area in the point cloud scanning data and the contour line segment coverage range in the tunnel construction design drawing, includes:

[0114] Step 1431: Perform feature dimension association analysis on the structural point cloud units in the three-dimensional tunnel model data set to establish a correspondence between the geometric features of each structural point cloud point and the key feature dimension; filter out the point cloud points to be processed whose geometric features contain the key feature dimension to form a point cloud subset of the key structure area, and calculate the vertex coordinates of the minimum bounding box of the point cloud subset as the spatial coordinate range.

[0115] A feature-dimensional correlation analysis is performed on the structural point cloud units in the 3D tunnel model dataset. A correspondence is established between the geometric features of each structural point cloud point and the key feature dimensions. This means determining whether the geometric features of each point cloud point contain the key feature dimensions. Point cloud points whose geometric features contain the key feature dimensions are selected for processing. These points constitute the point cloud subset for the key structural region.

[0116] For the point cloud subset, calculate the vertex coordinates of its minimum bounding box. The minimum bounding box is the smallest cube or cuboid that can completely enclose the point cloud subset. Its vertex coordinates represent the spatial coordinate range of the key structure area in the point cloud scan data.

[0117] For example, through feature dimension correlation analysis, we find that the geometric features of certain point cloud points match the key feature dimensions. These points are then selected to form a point cloud subset. The coordinates of the minimum bounding box vertices of this point cloud subset are then calculated. These coordinates can be used to determine the spatial extent of the key structural area in the point cloud scan data.

[0118] Step 1432: Perform feature dimension association analysis on the contour line units in the three-dimensional tunnel model data set to establish a correspondence between the structural association features of each contour line segment and the key feature dimension; filter out contour line segments whose structural association features contain the key feature dimension to form a line segment subset of the key structure area, and extract the starting point coordinates and end point coordinates of the line segment subset as the contour line segment coverage range.

[0119] A feature dimension association analysis is performed on the contour line elements in the 3D tunnel model dataset. A correspondence between the structural association features and the key feature dimensions of each contour line segment is established, and it is determined whether the structural association features of each contour line segment contain the key feature dimensions. Contour line segments whose structural association features contain the key feature dimensions are selected, and these line segments constitute a subset of line segments in the key structural area. The starting and ending point coordinates of the line segment subset are extracted; these coordinates determine the contour line segment coverage of the key structural area in the tunnel construction design drawings. For example, through feature dimension association analysis, contour line segments whose structural association features match the key feature dimensions are selected. The starting and ending point coordinates of these line segments are recorded to determine the contour coverage of the key structural area in the design drawings.

[0120] Based on this, the method further includes:

[0121] Step 1433: Perform spatial position alignment processing on the spatial coordinate range and the contour segment coverage range, so that the spatial coordinates in the point cloud scanning data and the contour segment coordinates in the tunnel construction design drawing have a corresponding relationship in the same coordinate system, and obtain a spatial position alignment result.

[0122] After determining the spatial coordinate range of the key structural area in the point cloud scan data and the coverage of the contour segments in the tunnel construction design drawings, the two are spatially aligned. Since the point cloud scan data and the tunnel construction design drawings may use different coordinate systems, their coordinate systems need to be aligned so that they have a corresponding relationship in the same coordinate system.

[0123] During the spatial alignment process, coordinate transformations, translations, and rotations may be required to ensure that the spatial coordinates in the point cloud scan data accurately correspond to the coordinates of the contour segments in the tunnel construction design drawings. This spatial alignment process yields a result. For example, the coordinate origin of the point cloud scan data may differ from that of the tunnel construction design drawings. Coordinate translations align the origins of the two, allowing the spatial coordinates and contour segment coordinates to be compared and analyzed in the same coordinate system.

[0124] Step 1434: Based on the spatial position alignment result, generate bimodal position description information including the point cloud coordinate range and the line segment coverage range, and the bimodal position description information is used to jointly locate the key structure area from the perspective of multi-source data.

[0125] Based on the spatial alignment results, a bimodal location description is generated. This bimodal location description includes both the point cloud coordinate range and the line segment coverage, allowing for the joint location of critical structural areas using multi-source data. By combining point cloud scan data with information from tunnel construction design drawings, the locations of critical structural areas can be more accurately determined.

[0126] For example, the bimodal location description information can be expressed as "point cloud coordinate range: (x1, y1, z1) - (x2, y2, z2); line segment coverage: starting point (x3, y3) - ending point (x4, y4)". Such description information can simultaneously reflect the location of key structural areas in point cloud scanning data and tunnel construction design drawings.

[0127] Step 144: Based on the spatial coordinate range and the contour line segment coverage range, the geometric center coordinates and bounding box parameters of the key structure area are calculated as the spatial position information of the key structure area.

[0128] The geometric center coordinates and bounding box parameters of the key structure area are calculated using the spatial coordinate range and contour segment coverage of the key structure area. The geometric center coordinates can represent the center position of the key structure area, while the bounding box parameters can describe the boundary range of the area.

[0129] When calculating the geometric center coordinates, it's necessary to consider both the spatial coordinate range of the point cloud scan data and the contour line segment coverage in the tunnel construction design drawings. For the bounding box parameters, in the point cloud scan data, the minimum bounding box size parameters for the spatial coordinate range are calculated; in the tunnel construction design drawings, the total length and maximum angle parameters for the contour line segment coverage are calculated.

[0130] For example, by calculating the coordinates of the vertices within the spatial coordinate range, the geometric center coordinates are obtained. At the same time, the length, width, and height of the minimum bounding box within the spatial coordinate range are calculated as the bounding box parameters in the point cloud scan data, and the total length and maximum angle of the contour line segments within the range covered are calculated as the bounding box parameters in the design drawing.

[0131] In the next step, the geometric center coordinates and bounding box parameters of the key structure area are calculated as the spatial position information of the key structure area based on the spatial coordinate range and the contour segment coverage range, including:

[0132] Step 1441: perform mean calculation processing on the coordinates of the minimum bounding box vertices in the spatial coordinate range to obtain the first geometric center coordinates of the key structure area in the point cloud scanning data; perform midpoint calculation processing on the starting point coordinates and the ending point coordinates of the contour segment coverage range to obtain the second geometric center coordinates of the key structure area in the tunnel construction design drawing; perform weighted averaging processing on the first geometric center coordinates in the point cloud scanning data and the second geometric center coordinates in the design drawing to generate the global geometric center coordinates of the fused multi-source data, and the weight coefficient of the weighted average is determined according to the credibility assessment result of the multi-source data.

[0133] The mean of the vertex coordinates of the minimum bounding box within the spatial coordinate range is calculated. The corresponding dimension values ​​of all vertex coordinates are added together and then divided by the number of vertices to obtain the first geometric center coordinates of the critical structural area in the point cloud scan data. For the starting and ending point coordinates of the contour segment coverage range, their midpoint coordinates are calculated. This is done by adding the corresponding dimension values ​​of the starting and ending point coordinates and dividing by 2 to obtain the second geometric center coordinates of the critical structural area in the tunnel construction design drawings.

[0134] Then, a weighted average is performed on the first and second geometric center coordinates. The weight coefficient for this weighted average is determined based on the credibility assessment results of the multi-source data. If the point cloud scan data has a high credibility, the first geometric center coordinate can be given a larger weight in the weighted average. Conversely, if the tunnel construction design drawings have a high credibility, the second geometric center coordinate can be given a larger weight. Through weighted averaging, the first and second geometric center coordinates are fused to generate a global geometric center coordinate that integrates the multi-source data. The global geometric center coordinate comprehensively considers information from the point cloud scan data and the tunnel construction design drawings, more accurately representing the center position of critical structural areas. For example, if the point cloud scan data has high acquisition accuracy and good data quality, and its credibility is determined to be high after credibility assessment, the first geometric center coordinate will be given a relatively large weight in the weighted average, making the global geometric center coordinate more inclined towards the center position reflected by the point cloud scan data.

[0135] Step 1442: Calculate the size parameters of the minimum bounding box in the spatial coordinate range as the bounding box parameters in the point cloud scanning data; calculate the total length and maximum angle parameters of the contour line segment coverage range as the bounding box parameters in the design drawing.

[0136] Optionally, for the spatial coordinate range in the point cloud scan data, calculate the size parameters of its minimum bounding box. The size parameters of the minimum bounding box include information such as length, width, and height. These parameters can describe the boundary range of the key structure area in the point cloud scan data. When calculating, first determine the maximum and minimum coordinate values ​​of the minimum bounding box in each coordinate axis direction, and then calculate the difference between them to obtain the length, width, and height. For example, in the X-axis direction, find the maximum and minimum values ​​of the X coordinates of all points in the spatial coordinate range, and the difference between the two is the length of the minimum bounding box in the X-axis direction.

[0137] For the contour line segment coverage area in the tunnel construction design drawings, the total segment length and maximum angle parameters are calculated. The total segment length is the sum of the lengths of all segments within the contour line segment coverage area, reflecting the overall length scale of the critical structural area in the design drawings. The maximum angle parameter is the maximum angle between segments within the contour line segment coverage area, reflecting the shape characteristics of the critical structural area. For example, a maximum angle close to 180 degrees indicates a long and narrow area; a smaller maximum angle indicates a compact area.

[0138] Step 1443: Combine the global geometric center coordinates with the bounding box parameters in the point cloud scanning data and the bounding box parameters in the tunnel construction design drawings to generate spatial position information including the center position and boundary range. The spatial position information is used to mark the spatial distribution of key structural areas.

[0139] The global geometric center coordinates obtained by fusing multi-source data are combined with the bounding box parameters from the point cloud scan data and the bounding box parameters from the tunnel construction design drawings. This combined processing integrates the center position and bounding range information of the critical structural areas, forming spatial location information containing both the center position and the bounding range. This spatial location information can comprehensively describe the distribution of the critical structural areas in three-dimensional space. For example, the spatial location information can be represented as "global geometric center coordinates: (x0, y0, z0); bounding box parameters from the point cloud scan data: length l1, width w1, height h1; bounding box parameters from the tunnel construction design drawings: total line segment length L, maximum angle α." This information can be used to accurately annotate the spatial distribution of the critical structural areas in the three-dimensional tunnel model.

[0140] Step 145: Associating and binding the feature type label with the spatial position information to generate a preliminary annotation result including a label-coordinate correspondence relationship.

[0141] The feature type labels assigned to key structural regions are associated and bound with the calculated spatial location information. Binding establishes a one-to-one correspondence between the feature type labels and spatial location information, ensuring that the feature type of each key structural region corresponds to its location in three-dimensional space. This binding process generates preliminary annotation results that include label-coordinate correspondences. For example, for a key structural region with the feature type label "tunnel vent structure" and spatial location information consisting of "global geometric center coordinates: (x0, y0, z0); point cloud scan data bounding box parameters: length l1, width w1, height h1; tunnel construction design drawing bounding box parameters: total line segment length L, maximum angle α," binding these two results creates a preliminary annotation result: "Label: Tunnel Ventilation Structure; Coordinates: (x0, y0, z0); Point cloud boundary (l1, w1, h1); Drawing boundary (L, α)." This preliminary annotation result provides the foundation for subsequent processing, clearly presenting the features and location information of each key structural region.

[0142] Step 146: performing duplicate region detection processing on the preliminary annotation results, merging the annotation information of key structural regions with overlapping spatial coordinate ranges, and generating a deduplicated feature recognition annotation result.

[0143] Optionally, there may be key structural areas with overlapping spatial coordinate ranges in the preliminary annotation results. These overlapping areas may be caused by data acquisition errors, inaccurate feature extraction, etc. In order to improve the accuracy and simplicity of the annotation results, it is necessary to perform duplicate area detection on the preliminary annotation results.

[0144] As an embodiment, performing duplicate region detection processing on the preliminary annotation results, merging annotation information of key structural regions with overlapping spatial coordinate ranges, and generating deduplicated feature recognition annotation results includes:

[0145] Step 1461: performing boundary expansion processing on the spatial coordinate range of each key structure area in the preliminary annotation result to generate an expanded coordinate range including a buffer area, wherein the width of the buffer area is adjusted according to the size of the key structure area.

[0146] The spatial coordinate range of each key structure area in the preliminary annotation results is expanded. The purpose of boundary expansion is to more accurately detect overlaps between areas and avoid missed detections due to data accuracy issues. During the expansion process, an expanded coordinate range containing a buffer area is generated. The width of the buffer area should be adjusted according to the size of the key structure area. Generally speaking, the larger the size of the key structure area, the larger the width of the buffer area. For example, for a larger key structure area, the width of its buffer area can be set relatively wide to ensure that potential overlaps with other areas can be detected; for smaller key structure areas, the width of the buffer area can be appropriately reduced. Through boundary expansion processing, the spatial relationship between key structure areas can be more comprehensively considered.

[0147] Step 1462: Calculate the ratio of the intersection area to the union area between the extended coordinate ranges of different key structure regions as the region overlap index; screen out key structure region pairs whose region overlap index exceeds a preset overlap threshold to form a set of repeated regions to be merged; perform consistency verification on the feature type labels of the key structure regions in the repeated region set, retain the repeated region pairs with completely consistent labels, and delete the repeated region pairs with inconsistent labels; perform union calculation on the spatial coordinate ranges of the retained repeated region pairs to generate a merged spatial coordinate range, wherein the merged spatial coordinate range contains all point cloud points and contour segments of the original two regions.

[0148] The ratio of the intersection area to the union area between the extended coordinate ranges of different key structure regions is calculated and used as the regional overlap index. The regional overlap index can quantify the degree of overlap between different key structure regions. By comparing the regional overlap index and the preset overlap threshold, the key structure region pairs whose regional overlap index exceeds the preset overlap threshold are screened out. These key structure region pairs constitute the set of repeated regions to be merged. The preset overlap threshold is a pre-set standard used to determine whether the overlap between regions has reached the degree that requires merging. For example, if the preset overlap threshold is 0.5, when the regional overlap index of two key structure regions is greater than 0.5, it is considered that their overlap degree is high and needs to be merged.

[0149] For the key structural regions in the set of duplicate regions to be merged, their feature type labels are checked for consistency. The duplicate region pair is retained only if the feature type labels of the two key structural regions are completely consistent; if the labels are inconsistent, the duplicate region pair is deleted. This is because inconsistent labels may indicate that the two regions actually represent different types of structures and should not be merged. For example, if the feature type label of one region is "tunnel main structure - vault" and the feature type label of another region is "tunnel ancillary facilities - lighting equipment", the labels of these two regions are inconsistent, and they should be deleted from the set of duplicate regions to be merged.

[0150] For the retained duplicate region pairs, a union calculation is performed on their spatial coordinate ranges. The union calculation combines all point cloud points and contour segments from the two original regions to generate a merged spatial coordinate range. The merged spatial coordinate range can contain all spatial information from the two original regions, avoiding information loss. For example, through the union calculation, the point cloud points and contour segments of the two overlapping regions are integrated into a new spatial coordinate range, which represents the spatial range of the merged key structure region.

[0151] Step 1463: De-duplication and merge the feature type labels of the retained duplicate region pairs to generate a unified feature type label; associate and bind the merged spatial coordinate range with the unified feature type label to generate a de-duplication feature recognition and labeling result.

[0152] The feature type labels of the retained duplicate region pairs are deduplicated and merged. Since the feature type labels of these duplicate region pairs have already been verified for consistency, they can be directly merged to generate a unified feature type label. For example, if the feature type label of two duplicate regions is "Tunnel Main Structure - Support Column", the unified feature type label after merging will still be "Tunnel Main Structure - Support Column".

[0153] The merged spatial coordinate range is associated and bound to a unified feature type label. This association and binding process aligns the merged spatial location information with the unified feature type label, generating deduplicated feature recognition and annotation results. Deduplicated feature recognition and annotation results are more accurate and concise, avoiding duplicate annotation issues and more clearly reflecting the characteristics and location information of key structural areas in the tunnel. For example, the deduplicated feature recognition and annotation results clearly display the unique feature type label and accurate spatial location information for each key structural area.

[0154] As a non-limiting embodiment, the method further includes: obtaining the correspondence between the spatial position information of each key structural area in the feature recognition and annotation results in the point cloud scanning data and the coverage range of the contour line segments in the tunnel construction design drawings, and constructing a multimodal annotation mapping table; performing cross-modal feature alignment processing on the multimodal annotation mapping table to extract the overlapping area features and the difference area features between the point cloud spatial coordinates and the drawing contour coordinates; based on the overlapping area features and the difference area features, respectively calculating the label consistency score and position deviation amount of the point cloud annotation and the drawing annotation; inputting the label consistency score and the position deviation amount into a preset credibility evaluation algorithm to generate a multimodal credibility score for each key structural area; screening and correcting the feature recognition and annotation results according to the credibility score, retaining the annotation areas with a credibility score higher than a preset threshold, adjusting the feature type label or spatial position information of the annotation areas with a credibility score lower than the threshold, and generating a feature recognition and annotation result after multimodal calibration.

[0155] First, the correspondence between the spatial location information of each key structural area in the point cloud scan data and the coverage of the contour segments in the tunnel construction design drawings is obtained from the feature recognition and annotation results. A multimodal annotation mapping table is then constructed. The multimodal annotation mapping table can associate the annotation information of the point cloud scan data and the tunnel construction design drawings, facilitating subsequent processing and analysis. For example, the multimodal annotation mapping table can record the correspondence between the spatial coordinate range of each key structural area in the point cloud scan data and the coverage of the contour segments in the tunnel construction design drawings, as well as the corresponding feature type label.

[0156] Secondly, the multimodal annotation mapping table is subjected to cross-modal feature alignment. Cross-modal feature alignment involves matching and aligning the point cloud spatial coordinates with the outline coordinates of the drawing to identify overlapping and discrepant areas. During this process, the overlapping and discrepant area features of the point cloud spatial coordinates and the outline coordinates of the drawing are extracted. The overlapping area features reflect the consistency of certain areas between the point cloud scan data and the tunnel construction design drawings, while the discrepant area features reflect the differences between the two. For example, in certain key structural areas, there may be certain deviations between the coordinates of the point cloud scan data and the tunnel construction design drawings. Cross-modal feature alignment can accurately identify these deviation areas.

[0157] Furthermore, based on the overlapping and differing region features, the label consistency score and position deviation of the point cloud annotation and the drawing annotation are calculated. The label consistency score measures the consistency of the feature type labels between the point cloud annotation and the drawing annotation, while the position deviation indicates the deviation in the spatial position between the point cloud annotation and the drawing annotation. For example, if the feature type labels of the point cloud annotation and the drawing annotation are exactly the same, the label consistency score will be high; if the spatial position deviation between the two is large, the position deviation will be large.

[0158] Furthermore, the label consistency score and positional deviation are input into a pre-set credibility assessment algorithm to generate a multimodal credibility score for each key structural region. The credibility assessment algorithm comprehensively considers the label consistency score and positional deviation to assess the credibility of the annotation information for each key structural region. The multimodal credibility score reflects the reliability of the annotation information for each key structural region. For example, a key structural region with a high label consistency score and a low positional deviation will have a high multimodal credibility score.

[0159] Finally, the feature recognition and annotation results are screened and corrected according to the credibility score. The annotation areas with a credibility score higher than the preset threshold are retained. The information of these annotation areas is relatively reliable and can be retained directly. For the annotation areas with a credibility score lower than the threshold, their feature type labels or spatial position information need to be adjusted. For example, if the credibility score of a key structural area is low, after analysis it is found that this is due to the large position deviation between the point cloud annotation and the drawing annotation, then its spatial position information can be corrected according to the actual situation. Through screening and correction processing, the feature recognition and annotation results after multimodal calibration are generated, and the feature recognition and annotation results after multimodal calibration are more accurate and reliable.

[0160] As another non-limiting embodiment, the method also includes: extracting feature type labels and spatial location information of all key structure areas from the feature recognition and annotation results, and constructing a structured annotation database containing area identifiers, label categories and spatial coordinates; performing regional association analysis on the structured annotation database to identify key structure area groups with functional collaborative relationships or spatial adjacency relationships; generating associated feature descriptions reflecting the interaction patterns between areas based on the semantic association of the feature type labels of each area in the key structure area group and the topological association of the spatial positions; fusing and expanding the associated feature descriptions with the original annotation information of each area in the key structure area group, and supplementing the key structure area group with additional labels containing interaction patterns, the additional labels are used to indicate the collaborative functions or spatial constraint relationships between areas, and obtain the fused and expanded annotation information; updating the structured annotation database based on the fused and expanded annotation information to generate enhanced feature recognition and annotation results containing area association information.

[0161] The feature type labels and spatial location information of all key structural regions are extracted from the feature recognition and annotation results to construct a structured annotation database. This database contains information such as region identifiers, label categories, and spatial coordinates. This allows for structured storage of annotation information for key structural regions, facilitating subsequent query and analysis. For example, the database can use the key structural region number as the region identifier, recording the feature type label and spatial location information for each region.

[0162] Regional association analysis is performed on the structured annotation database. By analyzing the characteristic type labels and spatial location information of key structural regions, groups of key structural regions with functional synergy or spatial adjacency are identified. Functional synergy refers to the functional coordination and collaboration between different key structural regions, while spatial adjacency refers to their spatial proximity or proximity. For example, the ventilation system and lighting system of a tunnel may have a functional synergy, while two adjacent support structures of the tunnel may have a spatial adjacency.

[0163] Based on the semantic relevance of the feature type labels and the topological relevance of the spatial locations of each region within a key structural region group, an associated feature description is generated to reflect the interaction patterns between regions. This associated feature description can provide a detailed description of the interaction methods and relationships between regions. For example, if a key structural region group includes ventilation and lighting systems, the associated feature description can explain how the operation of the ventilation system affects the heat dissipation of the lighting system, or how the layout of the lighting system cooperates with the air duct design of the ventilation system.

[0164] The associated feature descriptions are fused and expanded with the original annotation information for each key structural region within the group. Additional labels containing interaction patterns are added to the key structural region group. These additional labels indicate collaborative functions or spatial constraints between regions. For example, for a key structural region group containing ventilation and lighting systems, the additional label "Ventilation-Lighting Collaboration" can be added to indicate their collaborative function. This fusion and expansion process yields the fused and expanded annotation information.

[0165] The structured annotation database is updated based on the fused and expanded annotation information, generating enhanced feature recognition annotation results that include regional association information. These enhanced feature recognition annotation results not only include basic annotation information for key structural areas but also include inter-regional association information, providing a more comprehensive picture of the tunnel structure. For example, the enhanced feature recognition annotation results clearly show the relationship between each key structural area and other areas, providing richer information for tunnel maintenance, management, and optimization.

[0166] As another non-limiting embodiment, the method also includes: generating a historical annotation sequence arranged in chronological order based on the feature recognition and annotation results generated in each construction stage during the entire tunnel construction cycle; performing time series feature encoding processing on the historical annotation sequence to extract the feature type evolution trajectory and spatial position migration path of the key structural area in each time step; inputting the time series feature encoding results and the currently generated feature recognition and annotation results into the dynamic evolution prediction model, and combining them with time dependency modeling processing to predict the feature type evolution pattern and spatial position migration trend of each key structural area in the subsequent construction stage; using the feature type evolution pattern and spatial position migration trend as dynamic evolution information, associating them with the corresponding key structural area of ​​the current feature recognition and annotation result, and generating a time series extended feature recognition and annotation result including full-cycle evolution prediction.

[0167] Based on the feature recognition and annotation results generated at each construction stage throughout the tunnel construction cycle, a chronological historical annotation sequence is generated. This historical annotation sequence records the annotation information of key structural areas at different stages of the tunnel construction, including feature type labels and spatial location information. For example, the historical annotation sequence can record the annotation of key structural areas at each stage in the construction process, forming a time series data.

[0168] Temporal feature encoding is performed on the historical annotation sequence. Temporal feature encoding converts and encodes the information in the historical annotation sequence to facilitate subsequent analysis and processing. During this process, the feature type evolution trajectory and spatial position migration path of the key structural area at each time step are extracted. The feature type evolution trajectory reflects the changes in the feature type of the key structural area during different construction stages, while the spatial position migration path reflects the spatial position of the key structural area over time. For example, in the early stages of tunnel construction, a key structural area may be a temporary support structure. As construction progresses, it may evolve into a permanent structure. Temporal feature encoding can accurately record the evolution of this feature type.

[0169] The temporal feature encoding results and the currently generated feature recognition and annotation results are input into a dynamic evolution prediction model, which is then combined with temporal dependency modeling. The dynamic evolution prediction model, based on machine learning or deep learning, can learn temporal dependencies in historical data and, based on these relationships, predict the evolution of feature types and spatial migration trends in key structural areas during subsequent construction phases. Temporal dependency modeling analyzes how the feature types and spatial locations of key structural areas change over time. For example, the spatial location of some key structural areas may gradually shift toward the interior of the tunnel as construction progresses.

[0170] The feature type evolution pattern and spatial position migration trend are used as dynamic evolution information and associated with the corresponding key structure area of ​​the current feature recognition and annotation result. Through this association, the predicted dynamic evolution information is combined with the current annotation result to generate a time-series extended feature recognition and annotation result containing full-cycle evolution prediction. This result not only includes the annotation information of the current key structure area, but also adds the predicted information of its future evolution.

[0171] For example, in the current feature recognition and annotation results, a key structural area is labeled as "primary support structure." The dynamic evolution prediction model predicts that in the subsequent construction phase, this area will gradually evolve into a "secondary lining structure," and its spatial position will shift a certain distance into the tunnel as lining construction progresses. These feature type evolution patterns and spatial position migration trends are used as dynamic evolution information and associated with the "primary support structure" key structural area in the current annotation results. In this way, in the time-series extended feature recognition and annotation results, the annotation for this area not only includes the current "primary support structure" label and spatial position information, but also includes additional predictions of its future evolution into a "secondary lining structure" and position migration.

[0172] From the perspective of a dynamic evolution prediction model, its construction requires consideration of multiple aspects. First, the model's input data includes the time-series feature encoding results and the currently generated feature recognition and annotation results. The time-series feature encoding results are obtained by processing the historical annotation sequence, recording the feature evolution and positional migration of key structural areas during past construction phases. The currently generated feature recognition and annotation results reflect the latest status of the current construction phase.

[0173] In an embodiment of the present invention, the dynamic evolution prediction model includes multiple modules. The feature extraction module is responsible for extracting useful feature information from the input data, such as the feature type, spatial coordinates, and time step of key structural areas. These features are then passed to the time-dependent modeling module, which analyzes the time-dependent relationships in historical data to learn how the features and positions of key structural areas change over time. For example, by analyzing data from multiple construction phases, it was found that the evolution of the feature types of certain key structural areas exhibits a certain periodicity, or that there is a linear relationship between spatial position migration and construction progress.

[0174] Based on the learned temporal dependencies, the model's prediction module predicts the evolution of characteristic types and spatial migration trends for each key structural area during subsequent construction phases. Combining the current construction status with historical patterns, the prediction module outputs predictions for each key structural area at different future time steps.

[0175] Training a dynamic evolution prediction model requires a large amount of historically annotated data as training samples. This data can come from actual records of multiple tunnel construction projects or virtual data generated by simulating tunnel construction processes. During training, appropriate training parameters, such as the learning rate and number of iterations, must be set. The learning rate controls the step size for parameter updates at each iteration, while the number of iterations determines the number of rounds of model training. By continuously adjusting the training parameters, the model can accurately learn temporal dependencies, improving prediction accuracy.

[0176] The results of time-series extended feature recognition and annotation are also widely used. In tunnel construction planning, construction personnel can plan ahead for the procurement of construction materials and the deployment of equipment based on predicted feature type evolution patterns and spatial location migration trends. For example, if a critical structural area is predicted to require specialized materials in the future, the appropriate materials can be procured in advance to avoid construction delays.

[0177] In terms of construction quality monitoring, by comparing actual construction conditions with predicted results, deviations in the construction process can be promptly identified. If the actual feature type or spatial location of a critical structural area does not match the predicted results, the construction process needs to be inspected and adjusted to ensure that the construction quality meets the requirements.

[0178] During tunnel operation and maintenance, the results of time-series extended feature recognition and annotation also provide important insights. Based on the predicted evolution of feature types and location migration, a reasonable maintenance plan can be developed, allowing for proactive inspection and maintenance of potential problem areas, thereby extending the tunnel's service life.

[0179] Furthermore, the results of time-series extended feature recognition and annotation can support the digital management of tunnel construction. Integrating these annotation results into the tunnel construction management system enables real-time monitoring and dynamic management of the tunnel construction process. Through a visual interface, construction managers can intuitively view the current status and future forecast information of each key structural area, enabling timely decision-making.

[0180] Therefore, starting from obtaining tunnel construction design drawings and point cloud scanning data, after preprocessing, feature extraction, joint analysis, identification and annotation, the final result is a time-series extended feature recognition and annotation result that includes full-cycle evolution prediction. This result integrates information from multiple sources of data and can not only accurately reflect the characteristics and positions of key structural areas in the current tunnel construction stage, but also predict future evolution, thereby improving the efficiency, quality and safety of tunnel construction.

[0181] In an embodiment of the present invention, data preprocessing and feature extraction can be performed based on standard point cloud processing libraries (such as Point Cloud Library, PCL) and deep learning frameworks (such as PyTorch or TensorFlow) in the prior art. For example, in the noise filtering processing of the structural point cloud unit, the statistical outlier removal or radius filtering algorithm in PCL is referred to, and the neighborhood radius and standard deviation threshold are set according to the point cloud density to eliminate outliers and redundant points, and generate a smoothed subset; for the local curvature calculation of geometric features, the surface fitting method of PCL is used to directly output the curvature parameters and normal vector direction through the eigenvalue decomposition of the neighborhood point covariance matrix. In the structural association feature modeling of the contour line unit, the line segment detection and topology analysis tools of OpenCV are used to automatically extract the line segment length and angle, and the distribution density is calculated based on the line segment intersection spatial clustering algorithm. At the same time, the adjacency relationship is verified by combining angle histogram matching to ensure the repeatability of feature extraction.

[0182] The deep learning network is implemented based on a convolutional neural network architecture. The multi-scale feature fusion layer uses standard pyramid pooling or dilated convolution operations, with downsampling performed according to preset resolution rules. The windowing of the local detail feature stream directly uses sliding windows or region growing algorithms to segment point clouds and line segments. The local feature enhancement subnet integrates lightweight network modules for nonlinear activation processing, and explicitly selects public function types such as ReLU or Sigmoid when calling the activation function. The self-attention mechanism of the global context feature stream refers to the standard implementation of the Transformer model. After mapping features to query key-value vectors, similarity is calculated through dot product operations. Softmax normalization is used to generate the attention weight matrix, and residual connections are combined to optimize gradient propagation. Furthermore, principal component analysis or multi-layer perceptron dimensionality reduction is used for dimensionality reduction to ensure clear definition of the model structure and reproducibility.

[0183] The spatial position calculation of key structural areas can be integrated with GIS library functions. Point cloud and design drawing data are fused through weighted average coordinates. The weight coefficient is dynamically set based on the data acquisition accuracy report. The bounding box parameters are directly output by the bounding box generation algorithm. Spatial position alignment uses the ICP algorithm or coordinate transformation matrix to unify the coordinate system for dimensional consistency. During the duplicate area detection process, the buffer zone is expanded and adaptively adjusted in width based on the point cloud density. The area overlap calculation uses the IoU metric, and label consistency matching ensures logical rigor. Credibility assessment combines feature response distribution and position deviation for multimodal calibration.

[0184] The embodiment of the present invention processes tunnel construction design drawings and point cloud scanning data from corresponding construction phases to obtain a three-dimensional tunnel model dataset with spatial coordinate alignment, achieving effective fusion of multi-source data and ensuring that the structural point cloud units and the vectorized contour line units of the design drawings complement each other. Three-dimensional structural features are extracted from the three-dimensional tunnel model dataset to obtain the geometric features of the structural point cloud units and the structural association features of the contour line units, respectively. This provides in-depth analysis of the tunnel structure from different perspectives and provides rich feature information for accurately identifying key structural areas. A deep learning network is used to jointly analyze and process these two features, fully leveraging the powerful feature learning capabilities of deep learning. This network can uncover potential relationships between features and generate feature analysis results containing feature response distributions, thereby accurately determining the importance of different structural features. Based on the feature analysis results, key structural areas are identified and annotated, generating feature identification and annotation results containing feature type labels and spatial location information. This provides an intuitive and detailed basis for monitoring, management, and maintenance of tunnel construction, helping to improve construction efficiency and quality and ensure the safe and stable operation of tunnels.

[0185] See also Figure 2 As shown in FIG. 1 , this figure is a schematic diagram of the basic structure of a feature recognition and annotation system 200 provided by an embodiment of the present invention. The feature recognition and annotation system 200 includes:

[0186] Processor 201;

[0187] a storage device 202 having a computer program 2020 stored thereon;

[0188] When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the feature recognition and annotation methods applied to three-dimensional tunnel modeling.

[0189] Based on the above, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0190] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

Claims

1. A feature recognition and annotation method for three-dimensional tunnel modeling, characterized in that: The method comprises: Preprocessing the acquired tunnel construction design drawings and point cloud scan data from the corresponding construction phase to obtain a three-dimensional tunnel model dataset with spatial coordinate alignment, the three-dimensional tunnel model dataset comprising structural point cloud units fused from multi-source data and contour line units vectorized from the design drawings; Performing three-dimensional structural feature extraction processing on the three-dimensional tunnel model data set to obtain geometric features of the structural point cloud unit and structural association features of the contour line unit; Performing joint analysis and processing on the geometric features and the structural association features through a deep learning network to generate a feature analysis result including a feature response distribution; Identify and annotate key structural areas in the three-dimensional tunnel model dataset based on the feature analysis results, and generate feature identification and annotation results including feature type labels and spatial position information; The step of identifying and labeling key structural areas in the three-dimensional tunnel model dataset based on the feature analysis results to generate feature identification and labeling results including feature type labels and spatial location information includes: Extracting, based on the characteristic response distribution in the characteristic analysis result, characteristic dimensions whose activation values ​​exceed a preset threshold as key characteristic dimensions, wherein the key characteristic dimensions correspond to key structural areas of the tunnel; According to the feature type identifier corresponding to the key feature dimension, a corresponding feature type label is assigned to each key structural area, wherein the feature type label includes a structural component type and a functional attribute category; Extracting structural point cloud units and contour line units associated with the key feature dimensions from the three-dimensional tunnel model data set, and determining the spatial coordinate range of the key structural area in the point cloud scanning data and the contour line segment coverage range in the tunnel construction design drawing; Based on the spatial coordinate range and the contour line segment coverage range, calculating the geometric center coordinates and bounding box parameters of the key structure area as the spatial position information of the key structure area; Associating and binding the feature type label with the spatial position information to generate a preliminary annotation result including a label-coordinate correspondence relationship; The preliminary annotation results are subjected to duplicate region detection processing, and the annotation information of key structural regions with overlapping spatial coordinate ranges is merged to generate a deduplicated feature recognition annotation result.

2. The method according to claim 1, characterized in that The performing of three-dimensional structural feature extraction processing on the three-dimensional tunnel model data set to obtain the geometric features of the structural point cloud unit and the structural association features of the contour line unit includes: Performing noise filtering processing on outliers and redundant points on the structure point cloud unit to obtain a smoothed structure point cloud subset; Performing neighborhood analysis on the smoothed structural point cloud subset to calculate the local curvature parameter and normal vector direction of each point cloud point, and generating geometric features reflecting the undulation of the structural surface, wherein the geometric features include the curvature distribution pattern and normal vector consistency index of the point cloud points; Performing segment connectivity analysis on the contour line units, identifying contour line segments in the tunnel construction design drawings, and extracting length parameters and angle parameters of the contour line segments as basic structural features; Performing structural correlation modeling on the contour line units based on the basic structural features, analyzing the intersection relationship and adjacency relationship between different contour line segments, and generating structural correlation features reflecting the connection mode of tunnel structural components, wherein the structural correlation features include the distribution density of line segment intersection points and the angular matching degree of adjacent line segments; The method further comprises: The geometric features and the structural association features are uniformly dimensionalized to obtain a target feature set with the same feature dimension. The information entropy calculation result of the target feature set is used to adjust the feature fusion weight parameters of the deep learning network in the joint analysis process.

3. The method according to claim 1, characterized in that The joint analysis and processing of the geometric features and the structural association features by the deep learning network to generate a feature analysis result including a feature response distribution includes: Inputting the geometric features and the structural association features into the multi-scale feature fusion layer of the deep learning network, performing multi-resolution downsampling processing on the features in combination with a preset scale division rule, and generating a multi-scale feature stream; For the local detail feature stream in the multi-scale feature stream, convolution kernels with different receptive fields are used to perform feature extraction processing to capture the local curvature change details of the point cloud in the geometric morphological features and the line segment adjacency relationship details in the structural association features, thereby obtaining local detail feature output; For the global context feature stream in the multi-scale feature stream, the geometric features and the structural association features are modeled with a global dependency relationship through a self-attention mechanism to generate a context feature vector reflecting the overall structural layout of the tunnel; Inputting the local detail feature output and the context feature vector into a feature interaction network, performing information complementation through an element-by-element product operation, and generating an interactive feature map with spatiotemporal consistency constraints; Performing channel normalization on the interaction feature map, and concatenating the normalized interaction feature map along the channel dimension to generate a feature fusion vector containing multi-scale interaction information; The response distribution generation layer of the deep learning network is called to perform activation function mapping processing on the feature fusion vector to generate a feature response distribution reflecting the importance of different structural features, wherein the feature response distribution includes the activation value of each feature dimension and the corresponding feature type identifier.

4. The method according to claim 3, characterized in that The local detail feature stream in the multi-scale feature stream is subjected to feature extraction processing using convolution kernels with different receptive fields, capturing the local curvature change details of the point cloud in the geometric features and the line segment adjacency relationship details in the structural association features, and obtaining local detail feature output, including: Performing window division processing on the geometric features of the local detail feature stream to generate a plurality of local point cloud window units containing a preset number of point cloud points; Perform statistical analysis on the point cloud point curvature distribution pattern and normal vector consistency index within each local point cloud window unit, and calculate the mean and standard deviation of the curvature within the window as the statistical characteristics of local curvature change; Performing line segmentation processing on the structural correlation features of the local detail feature stream to generate a plurality of local line segment group units containing continuous line segments; Perform sequence analysis on the segment length parameters and angle parameters within each local segment group unit, extracting the increasing and decreasing pattern of segment length and the periodic change pattern of angle as the local segment adjacency relationship features; Inputting the local curvature change statistical features and the local line segment adjacency relationship features into a local feature enhancement subnet to generate an enhanced local feature vector containing multi-source local detail information; Performing nonlinear activation processing on the enhanced local feature vector to obtain a local detail feature output reflecting the details of the local curvature change of the point cloud and the details of the line segment adjacency relationship; The global context feature stream in the multi-scale feature stream is subjected to global dependency modeling processing on the geometric features and the structural association features through a self-attention mechanism to generate a context feature vector reflecting the overall structural layout of the tunnel, including: Mapping the geometric features and the structural association features into a query vector, a key vector, and a value vector, respectively, wherein the query vector is used to represent feature information of a target feature position, the key vector is used to represent feature information of a non-target feature position, and the value vector is used to represent feature values ​​of the non-target feature position; Calculating the dot product similarity between the query vector and the key vector to generate an attention weight matrix reflecting the degree of association between feature positions; Performing softmax normalization on the attention weight matrix to obtain the attention distribution probability of each feature position to the target feature position; Performing weighted summation processing on the value vector based on the attention distribution probability to generate a context-enhanced feature vector containing global feature position dependencies; Performing residual connection processing on the context enhanced feature vector, the geometric morphological feature and the structural association feature to obtain an initial context feature vector integrating the global dependency relationship; The initial context feature vector is subjected to dimension compression processing to generate a context feature vector reflecting the overall structural layout of the tunnel.

5. The method according to claim 1, characterized in that The step of assigning a corresponding feature type label to each key structural region according to the feature type identifier corresponding to the key feature dimension includes: Establishing a mapping relationship between a feature type identifier and a preset label library, wherein the preset label library includes a tunnel main structure label, ancillary facility structure label, and a potential disease structure label; Extracting the activation value peak position in the feature response distribution of the key feature dimension, and determining the core feature type identifier corresponding to the key feature dimension; Searching for a corresponding basic label category in the preset label library according to the core feature type identifier, wherein the basic label category includes a primary label of a structural component type and a secondary label of a functional attribute category; Extracting an auxiliary feature type identifier reflecting additional attributes of the feature based on the second highest peak position in the feature response distribution of the key feature dimension; Searching for corresponding supplementary label information in the preset label library according to the auxiliary feature type identifier, wherein the supplementary label information includes additional labels of structural material type and construction process category; The basic label category and the supplementary label information are combined and spliced ​​to generate a feature type label including a multi-level label hierarchy, where the multi-level label hierarchy is used to describe the structural attributes and functional attributes of the key structural area.

6. The method according to claim 1, characterized in that The step of extracting structural point cloud units and contour line units associated with the key feature dimensions from the three-dimensional tunnel model data set and determining the spatial coordinate range of the key structural area in the point cloud scanning data and the contour line segment coverage range in the tunnel construction design drawing includes: Performing feature-dimensional correlation analysis on the structural point cloud units in the three-dimensional tunnel model dataset to establish a correspondence between the geometric features of each structural point cloud point and the key feature dimension; screening out the point cloud points to be processed whose geometric features include the key feature dimension to form a point cloud subset of the key structural area; and calculating the vertex coordinates of the minimum bounding box of the point cloud subset as the spatial coordinate range; Performing feature dimension association analysis on the contour line units in the three-dimensional tunnel model dataset to establish a correspondence between the structural association feature of each contour line segment and the key feature dimension; screening contour line segments whose structural association features include the key feature dimension to form a line segment subset of the key structural area; extracting the starting point coordinates and the ending point coordinates of the line segment subset as the contour line segment coverage range; The method further comprises: Performing spatial position alignment processing on the spatial coordinate range and the contour segment coverage range so that the spatial coordinates in the point cloud scanning data and the contour segment coordinates in the tunnel construction design drawing have a corresponding relationship in the same coordinate system, thereby obtaining a spatial position alignment result; Based on the spatial position alignment result, bimodal position description information including the point cloud coordinate range and the line segment coverage range is generated, and the bimodal position description information is used to jointly locate the key structure area from the perspective of multi-source data.

7. The method according to claim 1, characterized in that The step of calculating the geometric center coordinates and bounding box parameters of the key structure area as the spatial position information of the key structure area based on the spatial coordinate range and the contour line segment coverage range includes: Performing mean calculation processing on the coordinates of the vertices of the minimum bounding box in the spatial coordinate range to obtain the first geometric center coordinates of the key structure area in the point cloud scanning data; Performing midpoint calculation processing on the coordinates of the starting point and the ending point of the range covered by the contour line segment to obtain the second geometric center coordinates of the key structural area in the tunnel construction design drawing; Performing weighted averaging processing on the first geometric center coordinates in the point cloud scan data and the second geometric center coordinates in the design drawing to generate a global geometric center coordinate of the fused multi-source data, wherein a weight coefficient of the weighted averaging is determined according to a credibility evaluation result of the multi-source data; Calculating size parameters of the minimum bounding box in the spatial coordinate range as parameters of the bounding box in the point cloud scanning data; Calculate the total length of the line segments and the maximum angle parameters of the contour line segments as the bounding box parameters in the design drawing; The global geometric center coordinates are combined with the bounding box parameters in the point cloud scanning data and the bounding box parameters in the tunnel construction design drawings to generate spatial position information including the center position and boundary range. The spatial position information is used to mark the spatial distribution of key structural areas.

8. The method according to claim 1, characterized in that The performing of duplicate region detection processing on the preliminary annotation results, merging the annotation information of key structural regions with overlapping spatial coordinate ranges, and generating a deduplicated feature recognition annotation result includes: Performing boundary expansion processing on the spatial coordinate range of each key structure area in the preliminary annotation result to generate an expanded coordinate range including a buffer area, wherein the width of the buffer area is adjusted according to the size of the key structure area; The ratio of the intersection area to the union area between the extended coordinate ranges of different key structural regions is calculated as the regional overlap index; Screen out key structural region pairs whose regional overlap index exceeds a preset overlap threshold to form a set of duplicate regions to be merged; Performing consistency check on the feature type labels of the key structural regions in the set of repeated regions, retaining repeated region pairs with completely consistent labels, and deleting repeated region pairs with inconsistent labels; Performing a union calculation on the spatial coordinate ranges of the retained repeated region pairs to generate a merged spatial coordinate range, wherein the merged spatial coordinate range includes all point cloud points and contour line segments of the original two regions; The feature type labels of the retained repeated region pairs are de-duplicated and merged to generate a unified feature type label; The merged spatial coordinate range is associated and bound with the unified feature type label to generate a deduplicated feature recognition and labeling result.

9. A feature recognition and annotation system, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the feature recognition and annotation method for three-dimensional tunnel modeling as described in any one of claims 1 to 8.

Images