A multi-modal three-dimensional analysis method and system for construction geology environment

Abstract

This invention discloses a three-dimensional analysis method and system for multimodal construction geological environments, relating to the field of three-dimensional analysis technology for construction geology. The method includes: collecting multi-dimensional data from the construction area and preprocessing the data to form a multimodal dataset; performing small-sample augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors; mapping the multimodal feature vectors to a unified three-dimensional mesh space, constructing a graph structure, and performing cross-modal feature fusion to generate a three-dimensional geological model; performing adaptive analysis on the three-dimensional geological model, generating voxel-level dynamic adjustment vectors based on loose body conditions, adaptively adjusting graph structure edge weights, neighborhood ranges, and node features, and identifying local anomalies and small-scale loose body structures; dynamically updating the three-dimensional geological model by combining real-time data and historical geological information, adjusting node features and global structure through incremental learning; and visualizing the analysis results of the three-dimensional geological model to generate construction auxiliary decision-making.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional analysis technology of construction geology, specifically to a three-dimensional analysis method and system for multimodal construction geological environment. Background Technology

[0002] In the field of construction geological analysis, with the rapid development of sensing technology and data acquisition methods, multimodal data acquisition has become a research focus. In recent years, advancements in various sensor technologies, including image processing, radar, acoustics, and vibration sensors, have enabled the acquisition of rich spatial, physical, and mechanical information about construction areas. Simultaneously, the integration of computer vision, signal processing, and machine learning methods has driven the development of 3D geological modeling and analysis techniques. Through the integrated processing of multimodal data, it is possible to achieve a certain degree of 3D visualization and analysis of complex geological environments, improving the accuracy of construction planning and risk assessment.

[0003] Current 3D geological analysis techniques for construction have several shortcomings. For multimodal data fusion, existing methods often process image, radar, and acoustic information separately, lacking a unified high-dimensional feature representation and modal weight adaptive mechanism. This results in the underutilization of information on the heterogeneity and physical properties of loose bodies. The ability to identify local anomalies and small-scale loose body structures is limited, failing to achieve dynamic adjustment based on density, porosity, and historical anomaly information, easily overlooking potential risk areas. Existing methods lack real-time data updates and incremental learning mechanisms, failing to integrate new observation data in a timely manner during construction, leading to discrepancies between the 3D geological model and the actual situation. Existing visualization methods are mostly static displays, lacking the ability to comprehensively transform anomaly indicators, density gradients, and porosity into potential sliding risks, and also failing to provide construction support decisions by combining logical judgments and threshold conditions. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing three-dimensional geological analysis methods for construction have insufficient multimodal information fusion, limited ability to identify local anomalies and small-scale loose structures, and are unable to dynamically fuse real-time data for incremental updates. The invention also addresses how to achieve adaptive feature adjustment, dynamic updating of the three-dimensional geological model, and provide construction auxiliary decision-making support.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a three-dimensional analysis method for multimodal construction geological environments, comprising: collecting multidimensional data from the construction area and preprocessing the multidimensional data to form a multimodal dataset; performing small-sample augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors; mapping the multimodal feature vectors to a unified three-dimensional mesh space, constructing a graph structure, and performing cross-modal feature fusion to generate a three-dimensional geological model; performing adaptive analysis on the three-dimensional geological model, generating voxel-level dynamic adjustment vectors based on loose body conditions, adaptively adjusting graph structure edge weights, neighborhood ranges, and node features, and identifying local anomalies and small-scale loose body structures; dynamically updating the three-dimensional geological model by combining real-time data and historical geological information, adjusting node features and global structure through incremental learning; and visualizing the analysis results of the three-dimensional geological model to generate construction auxiliary decision-making.

[0007] As a preferred embodiment of the three-dimensional analysis method for the multimodal construction geological environment described in this invention, the formation of the multimodal dataset includes: deploying multiple sensor acquisition nodes in the construction area to collect image data, radar data, acoustic data, vibration data, borehole and soil survey information; standardizing the spatial coordinates of the borehole and soil survey information as a global coordinate reference; aligning the image, radar, and acoustic data to the global coordinate reference using a coordinate transformation matrix and an iterative nearest-point algorithm; and performing denoising and format unification processing on various types of data to output a multimodal dataset that has undergone spatiotemporal alignment and preliminary registration.

[0008] As a preferred embodiment of the three-dimensional analysis method for the multimodal construction geological environment described in this invention, the generation of multimodal feature vectors includes: analyzing and filtering multimodal data of the construction area using historical survey records in the multimodal dataset to identify and obtain sparse anomalous areas and small-scale loose body structures within the construction area; generating small-sample enhanced data through rotation, mirroring, noise injection, and spectral transformation; inputting the small-sample enhanced data into a feature encoding network for mapping, encoding and fusing modal information in a unified high-dimensional feature space to generate multimodal feature vectors that characterize the heterogeneity and physical properties of loose bodies; quantizing the porosity, particle distribution, and compressibility modulus of loose bodies into high-dimensional tensors and concatenating them with the multimodal feature vectors in the channel dimension, and dynamically adjusting the contribution of each modal feature in the feature vector through a modal weight adaptive mechanism.

[0009] As a preferred embodiment of the three-dimensional analysis method for multimodal construction geological environment described in this invention, the generation of the three-dimensional geological model includes: projecting multimodal feature vectors onto three-dimensional mesh nodes to construct a graph structure in which node attributes include physical parameters, amplitude, and texture features; capturing spatial dependencies between nodes through a multi-scale graph convolutional network, while introducing additional edge connections based on physical spatial distance and topological neighborhood into the graph structure, weighting the extended connections in attention calculation, and handling cross-modal and long-distance dependencies; based on the connectivity between nodes and attention weights, accumulating and summing the feature values ​​of each node in the spatial neighborhood to obtain the aggregated feature vector of each node in the current layer; mapping the aggregated feature vector to the node features according to the three-dimensional mesh voxel resolution to generate the three-dimensional geological model.

[0010] As a preferred embodiment of the three-dimensional analysis method for the multimodal construction geological environment described in this invention, the identification of local anomalies and small-scale loose structures includes: for any voxel in the three-dimensional geological model, generating a voxel-level dynamic adjustment vector based on the voxel's loose density, porosity, density spatial gradient, and historical anomaly continuity index; using the dynamic adjustment vector to weight and adjust the node features and channel features of the current layer, while adaptively adjusting the edge weights and neighborhood range of the graph structure; quantizing and integrating the multidimensional features to construct a composite anomaly index; the multidimensional features include the adjusted node features, density gradient, porosity, and historical anomaly continuity; evaluating the composite anomaly index of each voxel in the three-dimensional geological model, calculating the adaptive learning rate of the local subgraph within the anomaly propagation domain of the voxel based on the index magnitude; dynamically determining the background field smoothing coefficient based on the magnitude of the voxel composite anomaly index relative to the global maximum value; and in the anomaly propagation domain corresponding to... Within a local subgraph, weight optimization is performed, or a smoothing constraint is applied to the background field. When the composite anomaly index exceeds a preset composite anomaly threshold, weight optimization is performed within the local subgraph corresponding to the anomaly propagation domain. Weight optimization includes determining the weight update step size based on the correlation between voxel features and anomaly propagation increments within the local subgraph, dynamically adjusting the weight amplitude according to the anomaly degree of each voxel, and strengthening the expression of relevant voxel features along the anomaly propagation direction. When the composite anomaly index does not exceed the preset composite anomaly threshold, a smoothing constraint is applied to the background field. Smoothing constraint includes weighting each voxel feature in the spatial neighborhood based on a voxel-level dynamic adjustment vector, using stable voxel features in the neighborhood as the constraint benchmark, applying stronger constraints to voxels with low anomaly degrees, and combining density gradient, porosity, and historical anomaly continuity information to weight and adjust the global fusion features to suppress the influence of local noise. Through dynamic adjustment and local weight optimization or smoothing constraints, an updated 3D geological model is output.

[0011] As a preferred embodiment of the three-dimensional analysis method for multimodal construction geological environment described in this invention, the step of adjusting node features and global structure through incremental learning includes: inputting newly acquired multimodal data in real time into the updated three-dimensional geological model to obtain new feature vectors for each voxel; calculating real-time composite anomaly indices based on the new feature vectors and comparing them with historical composite anomaly indices to obtain the change in composite anomaly indices; calculating dynamic weight adjustment coefficients based on the stability of real-time composite anomaly indices and historical features, and using the dynamic weight adjustment coefficients to fuse the new feature vectors with historical node features to obtain fused voxel node features; determining whether to perform incremental fine-tuning on the local subgraph corresponding to the anomaly propagation domain based on the change in composite anomaly indices; if incremental fine-tuning is not triggered, then performing smooth updates on the global fused features; and outputting the dynamically updated three-dimensional geological model based on the incremental fine-tuning results or the global smooth update results.

[0012] As a preferred embodiment of the three-dimensional analysis method for the multimodal construction geological environment described in this invention, the generation of construction auxiliary decision includes: visualizing the output of the three-dimensional geological model as anomaly heat map, loose body density distribution profile and potential sliding area map; and generating construction early warning by combining logical judgment and threshold triggering conditions.

[0013] As a preferred embodiment of the multimodal 3D analysis system for construction geological environment described in this invention, the system includes: a multidimensional data acquisition module, a small-sample enhancement module, a 3D mapping modeling module, an adaptive analysis module, a dynamic update module, and a visualization decision-making module. The multidimensional data acquisition module collects multidimensional data from the construction area and preprocesses the data to form a multimodal dataset. The small-sample enhancement module performs small-sample enhancement and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors. The 3D mapping modeling module maps the multimodal feature vectors to a unified 3D mesh space, constructs a graph structure, and performs cross-modal feature fusion to generate a 3D geological model. The adaptive analysis module performs adaptive analysis on the 3D geological model, generates voxel-level dynamic adjustment vectors based on loose body conditions, adaptively adjusts graph structure edge weights, neighborhood ranges, and node features, and identifies local anomalies and small-scale loose body structures. The dynamic update module dynamically updates the 3D geological model by combining real-time data and historical geological information, adjusting node features and global structure through incremental learning. The visualization decision-making module visualizes the analysis results of the 3D geological model and generates construction auxiliary decisions.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a three-dimensional analysis method for a multimodal construction geological environment.

[0015] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a three-dimensional analysis method for a multimodal construction geological environment.

[0016] The beneficial effects of this invention are as follows: The multimodal three-dimensional analysis method for construction geological environment provided by this invention collects image, radar, acoustic, vibration, and borehole soil information, and performs standardization and preprocessing to solve the problems of inconsistent data scales and spatial reference systems among different modalities, forming a multimodal dataset that can be directly fused; through small sample enhancement and physical perception feature learning, the features of sparse anomaly regions and small-scale loose body structures are effectively represented in high-dimensional space, improving the sensitivity to local heterogeneity and small anomalies; multimodal features are mapped to a unified three-dimensional mesh, and cross-modal fusion is performed using multi-scale graph convolution and Transformer to construct a dense three-dimensional geological model, achieving a unified expression of local structure and overall spatial dependence; through adaptive adjustment and incremental learning, the three-dimensional geological model dynamically responds to loose body conditions and real-time data changes, accurately identifies local anomalies, and updates three-dimensional geological information; anomaly indicators and sliding risk quantification are used for visualization and construction auxiliary decision-making, providing an operational basis for construction sequence, support design, and risk prevention and control. Attached Figure Description

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

[0018] Figure 1 The overall flowchart of a three-dimensional analysis method for multimodal construction geological environment provided by the present invention is shown.

[0019] Figure 2 Loose identification map for a three-dimensional analysis method of multimodal construction geological environment provided by the present invention.

[0020] Figure 3 A schematic diagram of a computer device for a three-dimensional analysis method of a multimodal construction geological environment provided by the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Reference Figures 1-2 As an embodiment of the present invention, a three-dimensional analysis method for multimodal construction geological environment is provided, comprising: S1: Collect multidimensional data of the construction area and preprocess the multidimensional data to form a multimodal dataset.

[0023] Furthermore, the formation of a multimodal dataset includes: deploying multiple sensor acquisition nodes in the construction area to collect image data, radar data, acoustic data, vibration data, and borehole and soil survey information; standardizing the spatial coordinates of the borehole and soil survey information as a global coordinate reference; aligning the image, radar, and acoustic data to the global coordinate reference using a coordinate transformation matrix and an iterative nearest-point algorithm; and performing denoising and format unification processing on various types of data to output a multimodal dataset that has undergone spatiotemporal alignment and preliminary registration.

[0024] It should be noted that one specific scheme for forming a multimodal dataset includes deploying multiple sensor acquisition nodes in the construction area to collect image data, radar data, acoustic data, vibration data, and borehole and soil survey information. The collected data... The data from each sensor is , Indicates the sensor index. , This indicates the total number of sensors; the data type can be a two-dimensional image matrix, radar amplitude matrix, acoustic waveform sequence, or vibration signal sequence. Borehole and soil survey information includes spatial coordinates. and corresponding geological attributes , Indicates the borehole point index. , Indicates the number of boreholes.

[0025] To ensure the comparability and fusion of multimodal data within the same spatial reference frame, the spatial coordinates of borehole and soil survey information are standardized to form a global coordinate benchmark. , is represented as: in, They represent the first Standardized drilling points coordinate, Representing the coordinates of all boreholes respectively The mean, Representing coordinates The standard deviation is used to eliminate differences between different measurement scales and to standardize the coordinate range. Standardized coordinates This will serve as a benchmark for subsequent multimodal data spatial alignment. Image data... Radar data Harmony and acoustic data Mapping to global coordinate reference In the mapping process, a coordinate transformation matrix is ​​constructed. An affine transformation is performed on the data from each type of sensor, as follows: in, Indicates the first Homogeneous coordinate representation of sensor data points Indicates the first The coordinates of each sensor data point mapped to the global coordinate system Indicates the first The affine or rigid coordinate transformation matrix of each sensor data point from the original local coordinate system to the global coordinate system. , Represents the real number field. It belongs to a 4×4 real matrix space. To improve registration accuracy, the Iterative Closest Point (ICP) algorithm is used to align each sensor point cloud with the global coordinate reference. By minimizing the spatial Euclidean distance, it can be expressed as: in, Indicates the first Alignment error function for individual sensor data points This indicates the number of reference points participating in the ICP iteration. Indicates the index of the iteration point. Indicates the first Spatial coordinate vectors of reference points, Indicates the first After the sensor data points are aligned to the global coordinate system by coordinate transformation, the first... The spatial coordinate vectors of each reference point are obtained through iterative optimization. The optimal solution is found, enabling preliminary registration of multimodal data in the global coordinate system. Noise reduction and format unification are then performed on various data types to obtain a standardized multimodal dataset. , is represented as: in, Indicates the first The result after preprocessing of sensor data points Indicates the first The raw data of each sensor data point after global coordinate alignment This represents the data preprocessing function, including denoising, sampling rate unification, and data type standardization operations, to ensure that data from different modalities can be directly fused in both spatial and numerical scales.

[0026] It should also be noted that by deploying multiple sensor acquisition nodes in the construction area to obtain image, radar, acoustic, vibration, and borehole soil survey information, and standardizing the borehole and soil coordinates to construct a global coordinate benchmark, the data from various sensors are aligned using a coordinate transformation matrix and an iterative nearest-point algorithm, achieving spatial alignment and denoising of multimodal data. This solves the problems of inconsistent scales, inconsistent spatial reference systems, and significant signal-noise interference in traditional construction geological monitoring, enabling direct fusion of data from different modalities under a unified spatial and numerical scale. A high-precision, quantifiable, and fusionable multimodal dataset is generated, providing a reliable data foundation for feature extraction and 3D modeling, and improving the accuracy and repeatability of construction geological environment analysis.

[0027] S2: Perform few-sample augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors.

[0028] Furthermore, generating multimodal feature vectors involves analyzing and filtering multimodal data of the construction area using historical survey records in the multimodal dataset to identify and obtain sparse anomalous regions and small-scale loose body structures within the construction area; generating small-sample augmented data through rotation, mirroring, noise injection, and spectral transformation; inputting the small-sample augmented data into a feature encoding network for mapping, encoding and fusing modal information in a unified high-dimensional feature space to generate multimodal feature vectors that characterize the heterogeneity and physical properties of loose bodies; quantizing the porosity, particle distribution, and compressibility modulus of loose bodies into high-dimensional tensors and concatenating them with the multimodal feature vectors in the channel dimension, and dynamically adjusting the contribution of each modal feature in the feature vector through a modal weight adaptive mechanism.

[0029] It should be noted that one specific approach to generating multimodal feature vectors includes using standardized multimodal datasets. The system identifies and learns features from sparse and anomalous regions and small-scale loose structures within the construction area; it also analyzes multimodal datasets based on historical survey records. Analysis and screening were conducted to identify sparse and anomalous areas and small-scale loose structures within the construction area. The screening results were denoted as follows: , Indicates the first An anomalous region or a small-scale loose structure. This indicates the identified structure index. This represents the total number of structures identified. (For the set) Each area in Data augmentation was performed using a small-sample augmentation method. Specifically, for the first... Multimodal data of anomaly regions Perform rotation, mirroring, noise injection, and spectral transformation operations to generate an enhanced sample set, represented as follows: in, Indicates the number of samples after small sample augmentation. An enhanced multimodal data sample set for each anomalous region. Indicates an enhanced sample index. Indicates the first The number of augmented samples in each anomalous region Indicates the first The first abnormal region Augmentation samples. Augmentation methods apply spatial or frequency domain perturbations to the original data to ensure that the model can generalize to sparse anomalies under different poses and noise conditions during feature encoding. The total set of augmented samples is then calculated. Input Feature Encoding Network In a unified high-dimensional feature space Internal encoding and modal fusion are performed, represented as: in, This represents the multimodal feature vector of a single augmented sample. Represents the global index of the augmented sample. The parameters represent the feature coding network. express The real space of dimension 1 This represents the dimension of the feature vector. To further enhance the physical perception features, the porosity of the loose material is used. Particle distribution and compressive modulus Quantized into high-dimensional tensors , express The real space, Represent the dimension of the physical property tensor and associate it with the multimodal feature vector. When concatenated along the channel dimension, it is represented as: in, Indicates the first The final feature vector set after concatenating the anomaly regions along the channel dimension contains multimodal features and physical attribute features. Indicates the first The physical property tensor vector corresponding to each anomalous region. express The real space of dimension .

[0030] For each mode The final feature vector Calculate the first Weight coefficients for each mode The weighting coefficients reflect the reliability of modal data under current environmental and construction conditions, based on the signal-to-noise ratio (SNR) and modal correlation index. Normalized calculation, expressed as: in, Indicates the first The weighting coefficients of each modality Indicates the current mode or channel number being calculated. Indicates the total number of modes or channels. Indicates the first Signal-to-noise ratio of each mode, Indicates the first Correlation of modal features Indicates another mode or channel number currently being calculated. Indicates the first Signal-to-noise ratio of each mode, Indicates the first The correlation of each modal feature is determined. Normalization is applied to ensure that the sum of all modal weights is 1, thus maintaining consistency in magnitude during feature fusion. According to the... Weight coefficients for each mode The feature vectors of each modality are weighted and summed to obtain the final feature vector after adaptive adjustment. , is represented as: in, Indicates the first The final adaptive feature vector after adaptive adjustment of each abnormal region Indicates the first The first abnormal region The final eigenvectors of each modality.

[0031] It should also be noted that, based on a standardized multimodal dataset, by identifying sparse anomalous regions and small-scale loose structures within the construction area, small-sample data augmentation is performed using generative adversarial networks or self-supervised contrastive learning. The augmented data is then input into a feature encoding network to generate multimodal feature vectors in a unified high-dimensional feature space. Physical properties such as porosity, particle distribution, and compressibility modulus are quantized into tensors and concatenated with the multimodal feature vectors. A modal weight adaptive mechanism is then used to fuse the features of each modality. This addresses the problem of traditional methods struggling to accurately characterize microscopic loose structures when anomalous bodies are sparse and data is insufficient. For local anomalies or small-scale loose bodies, the network can still obtain stable and high-fidelity feature representations, enhancing the feature vectors' ability to perceive physical properties while ensuring the model's generalization ability, thus achieving accurate quantification of microstructures in complex construction geological environments.

[0032] S3: Map multimodal feature vectors to a unified 3D mesh space, construct a graph structure, and perform cross-modal feature fusion to generate a 3D geological model.

[0033] Furthermore, generating a 3D geological model involves projecting multimodal feature vectors onto 3D mesh nodes to construct a graph structure whose node attributes include physical parameters, amplitude, and texture features; capturing spatial dependencies between nodes through a multi-scale graph convolutional network, while introducing additional edge connections based on physical spatial distance and topological neighborhood into the graph structure, weighting the extended connections in attention calculation to handle cross-modal and long-distance dependencies; based on node connectivity and attention weights, summing the feature values ​​of each node within its spatial neighborhood to obtain the aggregated feature vector of each node in the current layer; and mapping the aggregated feature vector to the node features onto the voxel mesh according to the voxel resolution of the 3D mesh to generate the 3D geological model.

[0034] It should be noted that one specific approach to generating a 3D geological model involves mapping multimodal feature vectors to a unified 3D mesh space to construct a graph structure for 3D geological modeling. Let the 3D mesh space consist of a set of nodes. constitute, Indicates the first One grid node, Indicates the grid node index, This represents the total number of grid nodes. The mapping relationship is achieved through spatial coordinate functions. Completed, indicated as: in, Indicates the first Normalized 3D coordinates of the anomaly region Represents a mapping function. Represents the 3-dimensional real number space. The... Feature vectors of grid nodes Includes physical parameter tensors Signal amplitude and texture features , is represented as: in, Indicates the first The physical parameter tensor of an anomalous region Indicates the first The signal amplitude in each abnormal region Indicates the first Texture features of anomaly regions express The real space of dimension 1 This represents the dimension of the node feature vectors. A graph structure is constructed to capture the spatial dependencies between nodes. edge set This indicates the connection relationships between nodes. This represents the identifier of an edge element in the edge set of a graph structure. Represents nodes The index of another node used for edge weight calculation. Edge weight Combining physical distance and topological neighborhood, it can be represented as: in, Represents a node With nodes The edge weights between them Represents a node Three-dimensional spatial coordinates, Represents a node Three-dimensional spatial coordinates, This represents the distance scaling parameter. This represents the neighborhood radius threshold. In a graph structure... Above, a multi-scale graph convolutional network (GCN) is used to process node features, as shown below: in, Indicates the first The grid node at the ... Layer node feature vectors Indicates the first The set of neighboring nodes of a grid node. Indicates the first Layer learnable weight matrix, Represents a non-linear activation function. Indicates the first The grid node at the ... Layer node feature vectors This represents the layer index of a graph convolutional network. This represents the number of network layers. By introducing an attention mechanism to weight neighborhood features, long-distance cross-modal dependencies are handled, as shown below: in, Indicates the first The attention-weighted node feature vector of each grid node. Represents a node With nodes Attention weights between them Indicates the first The node at the th Layer node feature vectors This represents the learnable attention parameter matrix. Indicates the first The node at the th Layer node feature vectors This indicates a transpose operation. This attention weighting allows cross-modal feature information from distant nodes to influence the representation of the current node, thereby improving the ability to perceive local anomalies. Through neighborhood accumulation and voxelization operations, node features are mapped onto a dense 3D voxel mesh to generate a 3D geological model. , is represented as: in, Indicates mapping to voxels The set of nodes, Indicates voxels in Directional index, Voxel representation The final feature vector, These represent the nodes in the 3D mesh. , , Axis position numbering, This represents a three-dimensional geological model on a dense three-dimensional voxel mesh. express A three-dimensional geological model is a real number space. A voxel grid, each voxel containing 3D eigenvectors These represent the resolution of the 3D mesh along the three directions.

[0035] It should also be noted that by projecting multimodal feature vectors onto 3D mesh nodes, a graph structure containing physical parameters, signal amplitude, and texture features is constructed. Furthermore, a multi-scale graph convolutional network is used in conjunction with a Transformer for cross-modal fusion, achieving comprehensive capture of spatial dependencies between nodes and long-distance cross-modal features. This addresses the difficulty in simultaneously handling local structural information and global spatial dependencies in traditional 3D modeling methods, enabling the effective embedding of microscopic features into the 3D spatial model. This generates a dense 3D voxel mesh model, preserving details of local anomalies and reflecting the spatial continuity of the overall geological environment, thus achieving comprehensive perception of both minute anomalies and macroscopic structures in complex geological environments.

[0036] S4: Adaptive analysis of the three-dimensional geological model, generating voxel-level dynamic adjustment vectors based on the loose body conditions, adaptively adjusting graph structure edge weights, neighborhood ranges and node features, and identifying local anomalies and small-scale loose body structures.

[0037] Furthermore, referring to Figure 2 The identification of local anomalies and small-scale loose body structures includes: for any voxel in the 3D geological model, generating a voxel-level dynamic adjustment vector based on the voxel's loose body density, porosity, density spatial gradient, and historical anomaly continuity index; using the dynamic adjustment vector to weight and adjust the node features and channel features of the current layer, while adaptively adjusting the edge weights and neighborhood range of the graph structure; quantizing and integrating multidimensional features to construct a composite anomaly index; the multidimensional features include adjusted node features, density gradient, porosity, and historical anomaly continuity; evaluating the composite anomaly index of each voxel in the 3D geological model, calculating the adaptive learning rate of the local subgraph within the anomaly propagation domain of the voxel based on the index magnitude; dynamically determining the background field smoothing coefficient based on the magnitude of the voxel composite anomaly index relative to the global maximum value; performing weight optimization within the local subgraph corresponding to the anomaly propagation domain or performing background field smoothing. Smoothing constraints are applied to the background field. When the composite anomaly index exceeds the preset composite anomaly threshold, weight optimization is performed in the local subgraph corresponding to the anomaly propagation domain. Weight optimization includes determining the weight update step size based on the correlation between the voxel features and the anomaly propagation increment in the local subgraph, dynamically adjusting the weight amplitude according to the anomaly degree of each voxel, and strengthening the expression of relevant voxel features along the anomaly propagation direction. When the composite anomaly index does not exceed the preset composite anomaly threshold, smoothing constraints are applied to the background field. Smoothing constraints include weighting each voxel feature in the spatial neighborhood based on the voxel-level dynamic adjustment vector, using the stable voxel features in the neighborhood as the constraint benchmark, applying stronger constraints to voxels with low anomaly degree, and combining density gradient, porosity, and historical anomaly continuity information to weight and adjust the global fusion features to suppress the influence of local noise. Through dynamic adjustment and local weight optimization or smoothing constraints, the updated three-dimensional geological model is output.

[0038] It should be noted that one method for identifying local anomalies and small-scale loose body structures specifically includes, in order to enable the three-dimensional geological model to adaptively change the information propagation method of the map structure according to the loose body conditions, for any voxel in the three-dimensional geological model... Based on the loose bulk density of the voxel Porosity Density spatial gradient and historical anomaly continuity indicators Generate a voxel-level dynamic adjustment vector, represented as: in, Represents voxels in a sparse 3D voxel mesh The dynamic adjustment vector, Voxel representation Loose mass density, Voxel representation porosity, Voxel representation The density spatial gradient, Voxel representation The degree of persistence of anomalies in consecutive historical moments. This represents the fusion function. By dynamically adjusting the vector, the local porosity, porosity development state, density abrupt change, and historical anomaly persistence of the loose body are uniformly mapped into voxel-level feature adjustment quantities that can participate in network regulation.

[0039] Furthermore, the dynamic adjustment vector is used not only to adjust the node features and channel features in the current layer, but also to synchronously adjust the edge weights and neighborhood ranges in the graph structure. For voxels... and adjacent voxels Based on the spatial distance between the two, the difference in loose body density, the difference in porosity, and the difference in the continuity of historical anomalies, the updated edge weights are calculated and expressed as follows: in, Voxel representation With voxels The edge weights between them are adaptively adjusted according to the loose body working condition. Voxel representation With voxels The base edge weights before adjustment Indicates voxel Another voxel index for determining neighborhood connectivity. They represent voxels in , , Directional index number, This represents the edge weight adjustment function, used to adjust the basic edge weights based on the differences in physical properties and anomalous continuity between two voxels. Voxel representation Loose mass density, Voxel representation porosity, Voxel representation The degree of anomaly persistence in continuous historical moments. By updating edge weights, voxels with similar physical properties and anomalous trends are given stronger information transmission weights, while voxels with large differences in physical states have reduced connection weights. This allows the information propagation path of the graph structure to adapt to the heterogeneous distribution and local flow trends of loose bodies. Furthermore, the adaptive neighborhood range is determined based on the voxel-level dynamic adjustment vector, expressed as: in, Voxel representation The adaptive neighborhood radius, This represents the neighborhood expansion coefficient. When the density of loose material in a certain region changes significantly, the porosity increases, or historical anomalies continue to intensify, its neighborhood range is appropriately expanded to capture relevant voxels in the direction of anomalous diffusion; when the region is in a stable state, its neighborhood range remains small to avoid interference from irrelevant regional features on the current voxels.

[0040] After adjusting the edge weights and neighborhood ranges, a dynamic adjustment vector is used to weight and adjust the features of the current layer nodes and channels, effectively mapping local physical and historical information to the network feature space to achieve the adaptive response of the 3D geological model to local anomalies and loose body structures, expressed as: in, Represents the weighted vector after dynamic adjustment. Laminolith The node feature vectors, Indicates the first layer The original node feature vector, Represents the weighted vector after dynamic adjustment. layer The channel feature vector, Indicates the first layer The original channel feature vector.

[0041] To avoid instability in anomaly detection due to relying solely on the deviation of a single feature, a composite anomaly index is constructed. This index jointly quantifies the deviation of the adjusted node features relative to the background field, density gradient, porosity, and historical anomaly continuity, and is expressed as: in, Voxel representation Composite abnormal indicators, Indicates the first Laminolith The node feature vectors, Voxel representation The mean of the background field features, These represent the weighting coefficients for characteristic deviation, density gradient, porosity, and historical anomaly continuity, respectively. This composite anomaly index simultaneously reflects the characteristic deviation of local anomalies, abrupt changes in loose body density, porosity development, and anomaly persistence, providing a more accurate description of loose body structural changes than a single threshold determination.

[0042] Furthermore, to ensure that composite anomaly indicators not only serve as a basis for anomaly judgment but also participate in the subsequent model parameter update process, this embodiment dynamically determines the preset composite anomaly threshold, the local weight optimization learning rate, and the background field smoothing coefficient based on the distribution of composite anomaly indicators of each voxel in the 3D geological model. Specifically, the preset composite anomaly threshold is not a fixed value but is adaptively determined based on the mean and dispersion of the composite anomaly indicators of all voxels, expressed as: in, This indicates the preset composite anomaly threshold. This represents the mean value of the composite anomaly index of all voxels in the three-dimensional geological model. This represents the standard deviation of the composite abnormality index of all voxels. This represents the threshold adjustment coefficient. When composite abnormal indicators... Exceeding the preset composite anomaly threshold When triggering local subgraph optimization in the anomaly propagation domain, the learning rate for local subgraph weight optimization is dynamically determined based on the relative relationship between the composite anomaly index of the corresponding voxel and the mean of the global composite anomaly index, expressed as: in, Voxel representation The adaptive learning rate corresponds to the local subgraph of the anomaly propagation domain. Indicates the basic learning rate. This represents a smoothing constant to prevent the denominator from reaching zero. By dynamically determining the learning rate, voxels with higher composite anomaly levels receive larger optimization step sizes for their corresponding local subgraphs, allowing the model to prioritize strengthening feature representations in regions with higher anomaly levels; regions with lower composite anomaly levels have smaller update steps to avoid over-optimization. When the composite anomaly index... The pre-set composite anomaly threshold was not exceeded. Based on the magnitude of the voxel composite anomaly index relative to the global maximum composite anomaly index, the background field smoothing coefficient is dynamically determined, expressed as: in, Voxel representation The adaptive smoothing coefficient, Represents the basic smoothing coefficient. This represents the maximum value among all voxel composite abnormality indices. This represents the smoothing constant. By dynamically determining the background field smoothing coefficient, stable regions with lower composite anomaly levels receive stronger background field smoothing constraints to suppress noise disturbances and maintain spatial continuity; regions with higher composite anomaly levels but not yet exceeding the threshold have reduced smoothing intensity to avoid over-smoothing of potential anomalies.

[0043] When composite abnormal indicators Exceeding the preset composite anomaly threshold At that time, an abnormal propagation domain is constructed centered on this voxel. The anomaly propagation domain consists of neighborhood voxels that satisfy edge weight correlation and physical attribute similarity, and is represented as: in, Indicated by voxels The central anomalous propagation domain, This represents the threshold for edge weight correlation. Voxel representation spatial coordinates, Voxel representation Spatial coordinates. Preset composite anomaly threshold. Experimental optimization was conducted based on the specific characteristics of the 3D geological model data, voxel density, anomaly index distribution, and geological conditions of the construction area. By constructing an anomaly propagation domain, local optimization is no longer limited to a single voxel, but rather weight optimization is performed within a local subgraph that has a physical association and spatial propagation relationship with the anomaly, expressed as: in, Indicates the first In-layer anomaly propagation domain The weight matrix corresponding to the local subgraph, Indicates the abnormal propagation domain The corresponding abnormal propagation increment, This represents the partial derivative. Local subgraph optimization enables the model to enhance the feature representation of relevant voxels along the direction of loose body anomaly propagation, rather than adjusting only isolated voxels. When composite anomaly indicators... The pre-set composite anomaly threshold was not exceeded. At this time, local weight optimization of the anomaly propagation domain is not triggered; instead, a smoothing constraint is applied to the background field, as follows: in, Indicates the first Laminolith The node feature vectors are used. Through voxel-level dynamic vector adjustment, graph structure edge weight adjustment, adaptive expansion of neighborhood range, composite anomaly index judgment, and local subgraph optimization of the anomaly propagation domain, the 3D geological model can adaptively change the information propagation path and feature update range according to the loose body conditions. It strengthens feature expression in anomaly areas and maintains background continuity in stable areas, thereby realizing the identification of local anomalies and small-scale loose body structures, and outputting an updated 3D geological model with annotated anomaly areas. , is represented as: in, Indicates the three-dimensional mesh in The total number of elements in the direction, Indicates the three-dimensional mesh in The total number of elements in the direction, Indicates the three-dimensional mesh in The total number of elements in the direction.

[0044] It should also be noted that traditional graph convolution or Transformers typically rely on pre-defined adjacency relationships or uniform attention weights, and the information propagation path between nodes is basically fixed after modeling. However, the risks of loose bodies often manifest as local density reduction, increased porosity, abrupt changes in density gradient, and continuous development of anomalies between adjacent voxels. Simply judging by node feature deviation or fixed thresholds can easily misidentify noise as anomalies and may also miss small loose structures propagating along space. This invention uses voxel-level dynamic adjustment vectors to unify loose body density, porosity, density spatial gradient, and historical anomaly continuity into network-executable adjustment quantities, and further simultaneously adjusts graph edge weights, neighborhood radii, node features, and channel features, so that the model's information propagation path is no longer fixed but changes with local geological conditions. By constructing composite anomaly indices and anomaly propagation domains, local optimization is upgraded from "single-point voxel parameter tuning" to "anomaly-related local subgraph update," which can strengthen the expression of relevant voxels along the direction of loose body anomaly expansion, while stable regions maintain continuity through background field smoothing. This enables the three-dimensional geological model to enhance the sensitivity of identifying local anomalies and small-scale loose structures, while avoiding interference from irrelevant areas and instability in the background field, thus forming an adaptive analysis model that better conforms to the geological evolution of construction.

[0045] S5: Combines real-time data and historical geological information to dynamically update the 3D geological model, and adjusts node features and global structure through incremental learning.

[0046] Furthermore, the incremental learning adjustment of node features and global structure includes: inputting newly acquired multimodal data in real time into the updated 3D geological model to obtain new feature vectors for each voxel; calculating real-time composite anomaly indices based on the new feature vectors and comparing them with historical composite anomaly indices to obtain the change in composite anomaly indices; calculating dynamic weight adjustment coefficients based on the stability of real-time composite anomaly indices and historical features, and using the dynamic weight adjustment coefficients to fuse the new feature vectors with historical node features to obtain fused voxel node features; determining whether to perform incremental fine-tuning on the local subgraph corresponding to the anomaly propagation domain based on the change in composite anomaly indices; if incremental fine-tuning is not triggered, then smoothing the global fused features; and outputting the dynamically updated 3D geological model based on the incremental fine-tuning results or the global smoothing update results.

[0047] It should be noted that one approach to adjusting node features and global structure through incremental learning specifically includes, based on an adaptive 3D geological model. Combined with new multimodal data acquired in real time The 3D geological model is dynamically updated. To achieve incremental learning, [the following steps are taken]. Each voxel Node features Composite abnormal indicators Abnormal propagation domain and adaptive graph edge weights Save it as historical state information for subsequent fusion of old and new features and local subgraph updates.

[0048] New multimodal data acquired in real time After preprocessing and feature encoding, voxels are obtained. New feature vector The real-time composite anomaly index of the voxel is recalculated based on the new feature vector. The change in composite anomaly indicators of real-time data relative to historical states is calculated and expressed as: in, Voxel representation The amount of composite anomaly change relative to historical states after real-time data input. Voxel representation A composite anomaly index calculated based on real-time new data. This is used to determine whether real-time data causes anomaly expansion, enhanced loose body flow, or further development of pore structure. To avoid over-covering historical stable geological information while incorporating real-time data, a dynamic weighting adjustment coefficient is calculated based on composite anomaly indices and historical stability, expressed as: in, Voxel representation The dynamic weight adjustment coefficient, and , Voxel representation Historical stability measure This represents a smoothing constant to prevent the denominator from reaching zero. When real-time data causes significant changes in composite anomaly indicators, Increasing the value makes the fusion results more biased towards real-time new data; when historical stability is strong and the changes in composite anomaly indicators are small. This reduces the number of elements, allowing the fusion results to retain more historically stable characteristics.

[0049] Furthermore, the new feature vector in real time Characteristics of historical nodes Incremental fusion is represented as: in, Indicates the fused voxels Node feature vectors This represents the voxels obtained after real-time data has undergone preprocessing and feature encoding. The new feature vector. Furthermore, based on the anomaly propagation domain... The real-time composite anomaly index changes of each voxel are analyzed, and the proportion of new data anomalies within the anomaly propagation domain is calculated, as expressed as: in, Indicates the abnormal propagation domain The proportion of new anomalies caused by real-time data. This indicates the number of voxels in the anomaly propagation domain. This represents the threshold for changes in composite anomaly indicators, used to determine whether anomalies caused by real-time data have reached a level that triggers incremental fine-tuning of the anomaly propagation domain. The preferred value is 0.05 to 0.15; when the geological conditions in the construction area are complex and the loose mass is sensitive to changes, The preferred value is 0.05 to 0.10; when the geological conditions in the construction area are relatively stable or the real-time data noise is high, The preferred value is 0.10 to 0.15. When or When this occurs, it indicates that real-time data causes local anomaly enhancement or anomaly propagation domain expansion. In this case, the anomaly propagation domain... Incremental fine-tuning of the corresponding local subgraph is represented as follows: in, This represents the incremental loss function value for the anomaly propagation domain, calculated based on real-time data. The threshold value representing the proportion of new anomalies within the anomaly propagation domain is preferably 0.1 to 0.2, indicating that optimization is triggered when 10% to 20% of the anomalies are present. By performing incremental fine-tuning within the local subgraph corresponding to the anomaly propagation domain, the model focuses on responding to the anomaly expansion region caused by real-time data, rather than indiscriminately updating the entire 3D model.

[0050] After incremental learning and fusing real-time data, to prevent feature mutations or noise interference caused by local anomaly fine-tuning, the globally fused features are further smoothed. This ensures the overall spatial continuity and physical consistency of the 3D geological model while maintaining real-time responsiveness, achieving a balance between local anomaly response and global stability. and When the real-time data does not cause significant abnormal expansion, a smooth update of the global fusion features is performed, represented as: in, This represents the mean of the global voxel fusion features. This global smooth update can suppress small noise disturbances in real-time data, preventing new data from over-correcting historically stable models.

[0051] After incremental fusion, incremental fine-tuning of the anomaly propagation domain, and smoothing update of global fusion features, a dynamically updated 3D geological model is output. , is represented as: It should also be noted that the newly acquired multimodal data in real time is first encoded into new feature vectors for each voxel, then the real-time composite anomaly index is recalculated and compared with the historical composite anomaly index to obtain the change in the composite anomaly index. This allows for the differentiation between real geological changes and sensor noise or short-term disturbances. Based on this, a dynamic weight adjustment coefficient calculated from the real-time composite anomaly index and historical feature stability is used to fuse the new feature vectors and historical node features. This allows voxels with significant anomaly changes to absorb more real-time data, while voxels with high historical stability retain more of the original model features, thus solving the problem of new data updates easily covering historically stable geological structures. Furthermore, this invention determines whether anomalies have spatially expanded based on the proportion of voxels whose composite anomaly index changes exceed a threshold within the anomaly propagation domain. Only when the proportion of new anomalies within the anomaly propagation domain reaches a preset condition is the corresponding local subgraph incrementally fine-tuned; otherwise, the global fused features are smoothly updated. The model can focus on updating local areas where loose body flow, pore expansion, or anomaly diffusion may occur, while suppressing invalid fluctuations in stable areas caused by construction vibration, equipment errors, or local noise. Achieving a balance between real-time data import, historical structure preservation, local response of anomaly propagation domain, and global background stability improves the dynamic representation accuracy and update reliability of 3D geological models during continuous construction.

[0052] S6: Visualize the results of the three-dimensional geological model analysis to generate construction support decisions.

[0053] Furthermore, the construction support decision generation includes visualizing the output of the three-dimensional geological model as anomaly heatmaps, loose body density distribution profiles, and potential sliding zone maps; and generating construction early warnings by combining logical judgments and threshold triggering conditions.

[0054] It should be noted that in this embodiment, the updated three-dimensional geological model is used. Visualize the construction geological environment and generate construction decision support. For each voxel... Recalculate the updated anomaly representation values and normalized to thermal intensity , is represented as: in, Voxel representation The degree of abnormal deviation relative to the background field, and These represent the minimum and maximum values ​​among all voxel anomaly representation values, respectively. This represents the normalization smoothing constant, used to avoid the denominator being zero. Voxel representation The normalized outlier deviation is determined by Values ​​ranging from 0 to 1 can be mapped to color intensity and generate anomaly heatmaps. Higher heatmap values ​​indicate a greater likelihood of local anomalies or structural mutations in the corresponding voxel region. After generating the anomaly heatmap, a loose body density distribution profile is further output, assembled along a preset cutting plane. The three-dimensional geological model is cropped, and the profile density field is calculated on each profile. , is represented as: in, Indicates the first A cutting plane in two-dimensional coordinates The profile density value, Indicates mapping to a profile The following set of voxels, Indicates the number of cutting planes. Indicates the first A cutting plane, Indicates the cutting plane index. Indicates the first Two-dimensional coordinates on a cutting plane. Since the profile map and the anomalous thermal map share the same three-dimensional geological model source, they have a natural spatial correspondence. This allows for the overlay and comparison of thermal anomaly areas and density abrupt change areas, improving the accuracy of anomaly assessment. Based on this, anomaly information, density gradient information, and porosity information are jointly transformed into a sliding risk indicator. , is represented as: in, Voxel representation Potential slip risk indicators This represents the risk-weighted coefficient for adjusting abnormal thermal intensity. This represents the risk weighting coefficient of the density gradient. The risk-weighted coefficient representing porosity. Voxel representation The spatial density gradient reflects the degree of drastic change in local geological parameters. Voxel representation The porosity. Since loose material regions typically exhibit abnormal enhancement, sudden increases in density, and high porosity before local instability, the sliding risk index can further transform identified structural changes into a potential sliding risk distribution. In practical applications, it can be... Exceeding the preset risk threshold The voxel regions are colored to create a potential sliding region map.

[0055] After completing the visualization, the 3D geological model analysis results are further transformed into construction early warning and construction decision support. The construction area is divided into several construction decision zones. , Indicates the total number of partitions. Indicates the first For each construction zone, calculate the comprehensive risk score. , is represented as: in, Indicates the first Comprehensive risk score for each construction zone Indicates the first The number of voxels within each construction zone. Based on the comprehensive risk score, a warning level is further constructed. , is represented as: in, This indicates a preset low warning threshold. , This indicates a preset high warning threshold, used to distinguish between low, medium, and high risk levels. When the warning level is low, it indicates that the geological environment of the corresponding area is relatively stable, the voxel anomaly value is low, and the risk of local loose bodies or landslides is small; when At this time, a moderate warning is issued, indicating the presence of moderately abnormal bodies or loose bodies in the corresponding area, with the possibility of small-scale displacement or soil loosening in some areas; when When a high level of alert is triggered, it indicates that there is a significant risk of sliding or instability in the corresponding area, with high voxel anomalies and the possibility of local soil collapse or landslides.

[0056] It should also be noted that by calculating anomaly indicators, generating heat maps, profile density distribution, and potential sliding risk distribution from the updated 3D model, and combining this with zonal comprehensive risk scoring to form construction early warning levels, 3D geological environment visualization and construction auxiliary decision-making are achieved. This solves the problem that traditional construction decision-making relies on 2D maps or abstract data, which cannot intuitively reflect risks. It enables construction personnel to intuitively identify anomalies and high-risk areas in space. Complex multimodal 3D geological information is transformed into visualized and quantifiable engineering decision-making basis, allowing construction sequence, support design, and risk prevention and control plans to be directly formulated, improving construction safety and geological risk management efficiency, and enhancing the operability and scientific nature of decision-making.

[0057] One embodiment of the present invention provides a three-dimensional analysis system for multimodal construction geological environment, including a multi-dimensional data acquisition module, a small sample enhancement module, a three-dimensional mapping modeling module, an adaptive analysis module, a dynamic update module, and a visualization decision module.

[0058] The system comprises the following modules: a multidimensional data acquisition module for collecting multidimensional data from the construction area and preprocessing it to form a multimodal dataset; a small-sample augmentation module for performing small-sample augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors; a 3D mapping and modeling module for mapping the multimodal feature vectors to a unified 3D mesh space, constructing a graph structure, and performing cross-modal feature fusion to generate a 3D geological model; an adaptive analysis module for adaptively analyzing the 3D geological model, generating voxel-level dynamic adjustment vectors based on the loose body conditions, adaptively adjusting graph structure edge weights, neighborhood ranges, and node features, and identifying local anomalies and small-scale loose body structures; a dynamic update module for dynamically updating the 3D geological model by combining real-time data and historical geological information, adjusting node features and global structure through incremental learning; and a visualization decision-making module for visualizing the analysis results of the 3D geological model and generating construction auxiliary decisions.

[0059] Reference Figure 3 This embodiment also provides a computer device applicable to the three-dimensional analysis method of multimodal construction geological environment, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the three-dimensional analysis method of multimodal construction geological environment as proposed in the above embodiment.

[0060] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0061] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the three-dimensional analysis method for multimodal construction geological environments as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A method of multi-modal three-dimensional analysis of a construction geologic environment, characterized by, include: Collect multidimensional data from the construction area and preprocess the multidimensional data to form a multimodal dataset; Perform few-sample augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors; Multimodal feature vectors are mapped to a unified 3D mesh space, a graph structure is constructed, and cross-modal feature fusion is performed to generate a 3D geological model; Adaptive analysis of the three-dimensional geological model is performed. Based on the loose body conditions, a voxel-level dynamic adjustment vector is generated to adaptively adjust the graph structure edge weights, neighborhood ranges, and node features, and to identify local anomalies and small-scale loose body structures. The three-dimensional geological model is dynamically updated by combining real-time data and historical geological information, and node features and global structure are adjusted through incremental learning. Visualize the results of the three-dimensional geological model analysis to generate construction support decisions.

2. The method of multi-modal three-dimensional analysis of a construction geologic environment of claim 1, wherein: The formation of the multimodal dataset includes deploying multiple sensor acquisition nodes in the construction area to collect image data, radar data, acoustic data, vibration data, borehole and soil survey information; The spatial coordinates of borehole and soil survey information are standardized and used as a global coordinate reference. Image, radar, and acoustic data are aligned to a global coordinate reference using a coordinate transformation matrix and an iterative nearest-point algorithm. The system performs noise reduction and format unification on various types of data, and outputs a multimodal dataset that has undergone spatiotemporal alignment and preliminary registration.

3. The method of multi-modal three-dimensional analysis of a construction geologic environment of claim 2, wherein: The generation of multimodal feature vectors includes analyzing and filtering multimodal data of the construction area using historical survey records in the multimodal dataset, and identifying and obtaining sparse anomalous areas and small-scale loose structures within the construction area. Small sample augmented data is generated through rotation, mirroring, noise injection, and spectral transformation. Small sample augmented data is input into a feature encoding network for mapping, and modal information is encoded and fused in a unified high-dimensional feature space to generate multimodal feature vectors that can characterize the heterogeneity and physical properties of loose bodies. The porosity, particle distribution, and compressibility modulus of loose materials are quantized into high-dimensional tensors and concatenated with multimodal feature vectors along the channel dimension. The contribution of each modality feature in the feature vector is dynamically adjusted through a modal weight adaptive mechanism.

4. The method of multi-modal three-dimensional analysis of a construction geologic environment of claim 3, wherein: The generation of the three-dimensional geological model includes projecting multimodal feature vectors onto three-dimensional mesh nodes and constructing a graph structure in which node attributes include physical parameters, amplitude and texture features; The spatial dependencies between nodes are captured by a multi-scale graph convolutional network. At the same time, additional edge connections based on physical spatial distance and topological neighborhood are introduced into the graph structure. The extended connections are weighted in the attention calculation to handle cross-modal and long-distance dependencies. Based on the connectivity between nodes and attention weights, the feature values ​​of each node are accumulated and summed in the spatial neighborhood to obtain the aggregated feature vector of each node in the current layer; The aggregated feature vectors are mapped from node features to the voxel grid according to the voxel resolution of the 3D grid to generate a 3D geological model.

5. The method for multi-modal three-dimensional analysis of a construction geologic environment of claim 4, wherein: The identification of local anomalies and small-scale loose body structures includes generating a voxel-level dynamic adjustment vector for any voxel in the three-dimensional geological model based on the voxel's loose body density, porosity, density spatial gradient, and historical anomaly continuity index. The node features and channel features of the current layer are weighted and adjusted using a dynamic adjustment vector, while the edge weights and neighborhood range of the graph structure are adaptively adjusted at the same time. Multidimensional features are quantified and integrated to construct composite anomaly indicators; The multidimensional features include adjusted node features, density gradient, porosity, and historical anomaly continuity; The composite anomaly index of each voxel in the three-dimensional geological model is evaluated, and the adaptive learning rate of the local subgraph within the anomaly propagation domain of the voxel is calculated based on the magnitude of the index. The background field smoothing coefficient is dynamically determined based on the magnitude of the voxel composite anomaly index relative to the global maximum value. Perform weight optimization or apply smoothing constraints to the background field within the local subgraph corresponding to the anomaly propagation domain; When the composite anomaly index exceeds the preset composite anomaly threshold, weight optimization is performed in the local subgraph corresponding to the anomaly propagation domain. The weight optimization process includes determining the weight update step size based on the correlation between the features of each voxel in the local subgraph and the anomaly propagation increment, dynamically adjusting the weight amplitude according to the anomaly degree of each voxel, and strengthening the expression of relevant voxel features along the anomaly propagation direction. When the composite anomaly index does not exceed the preset composite anomaly threshold, a smoothing constraint is applied to the background field. The implementation of smoothing constraints includes weighted averaging of each voxel feature in the spatial neighborhood based on voxel-level dynamic adjustment vectors, using stable voxel features in the neighborhood as constraint benchmarks, applying strong constraints to voxels with low anomaly levels, and combining density gradient, porosity and historical anomaly continuity information to weight and adjust the global fused features to suppress the influence of local noise. The updated 3D geological model is output through dynamic adjustment and local weight optimization or smoothing constraints.

6. The method of multi-modal three-dimensional analysis of a construction geologic environment of claim 5, wherein: The step of adjusting node features and global structure through incremental learning includes inputting newly acquired multimodal data in real time into the updated three-dimensional geological model to obtain new feature vectors for each voxel. Real-time composite anomaly index is calculated based on the new feature vector and compared with historical composite anomaly index to obtain the change in composite anomaly index. The dynamic weight adjustment coefficient is calculated based on the real-time composite anomaly index and the stability of historical features. The new feature vector and historical node features are then fused using the dynamic weight adjustment coefficient to obtain the fused voxel node features. The decision to perform incremental fine-tuning on the local subgraph corresponding to the anomaly propagation domain is based on the change in the composite anomaly index. If incremental fine-tuning is not triggered, the global fusion features are updated smoothly. Based on the incremental fine-tuning results or the global smoothing update results, output the dynamically updated 3D geological model.

7. The method of multi-modal three-dimensional analysis of a construction geologic environment of claim 6, wherein: The generated construction auxiliary decision includes visualizing the output of the three-dimensional geological model as anomaly heat map, loose body density distribution profile and potential sliding area map; Construction warnings are generated by combining logical judgments and threshold triggering conditions.

8. A multi-modal three-dimensional analysis system of a construction geological environment, which adopts the multi-modal three-dimensional analysis method of a construction geological environment according to any one of claims 1 to 7, characterized by: It includes a multi-dimensional data acquisition module, a small sample enhancement module, a 3D mapping modeling module, an adaptive analysis module, a dynamic update module, and a visualization decision-making module; The multidimensional data acquisition module is used to collect multidimensional data from the construction area and preprocess the multidimensional data to form a multimodal dataset; The few-shot augmentation module is used to perform few-shot augmentation and physical perception feature learning on the multimodal dataset to generate multimodal feature vectors; The three-dimensional mapping modeling module is used to map multimodal feature vectors to a unified three-dimensional mesh space, construct a graph structure, and perform cross-modal feature fusion to generate a three-dimensional geological model. The adaptive analysis module is used to perform adaptive analysis on the three-dimensional geological model, generate voxel-level dynamic adjustment vectors based on the loose body conditions, adaptively adjust the graph structure edge weights, neighborhood ranges and node features, and identify local anomalies and small-scale loose body structures. The dynamic update module is used to dynamically update the three-dimensional geological model by combining real-time data and historical geological information, and to adjust node features and global structure through incremental learning. The visualization decision module is used to visualize the analysis results of the three-dimensional geological model and generate construction auxiliary decisions. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the three-dimensional analysis method for the multimodal construction geological environment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the three-dimensional analysis method for the multimodal construction geological environment as described in any one of claims 1 to 7.

Images