A tunnel karst advanced geological prediction automatic interpretation method, device and medium

Abstract

The application discloses a tunnel karst advanced geological prediction automatic interpretation method and device and medium, relates to the field of image processing, obtains tunnel karst advanced geological prediction data on different scales, preprocesses the tunnel karst advanced geological prediction data, reconstructs unsupervised images, constructs a three-dimensional reconstruction model, constructs a neural network model to extract a two-dimensional distribution feature map of the three-dimensional reconstruction model, constructs a data automatic interpretation model based on the two-dimensional distribution feature map, and after the training of the data automatic interpretation model is completed, inversion is carried out based on the trained data automatic interpretation model, multi-scale data fusion, unsupervised image reconstruction and a three-dimensional reconstruction model are used, the interpretation accuracy and efficiency of complex images in prediction data are significantly improved, the model firmware method enhances the generalization ability of the model, improves the utilization efficiency of image data, ensures the reliability and accuracy of the interpretation result, and further improves the construction safety and efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to a tunnel karst advanced geological prediction automatic interpretation method, device and medium. BACKGROUND

[0002] Tunnel advanced geological prediction is a report that predicts the geological characteristics, structural characteristics and integrity of the surrounding rock of a certain section or part of the tunnel and its front range, the surrounding rock grade and the stability after tunnel excavation, and proposes excavation and support suggestions in front of the tunnel. It aims to master the structure, nature and state of the rock-soil body in front of the construction, as well as the occurrence of groundwater, gas and geostress, etc. geological information, to guide the tunnel construction, so as to avoid geological disasters such as water gushing, gas outburst, rock burst, large deformation, etc. in the construction and operation process, and ensure the safety of construction. After obtaining the tunnel advanced geological prediction, data interpretation needs to be carried out in time to ensure that the geological conditions can be found out in time, the adverse geology and major geological problems can be predicted, the geological disaster risk can be reduced, and the construction difficulty and cost increase caused by unknown geological conditions can be reduced.

[0003] However, in the interpretation process of tunnel advanced geological prediction data, the advanced geological tunnel karst advanced geological prediction data comes from various sources, and may have problems such as noise and missing values, which makes it difficult to interpret the data images obtained, and on this basis, due to the lack of a large amount of annotation of image data, it is difficult to effectively extract representative data in a complex three-dimensional geological model. Based on the complexity of tunnel karst advanced geological prediction data, a large amount of labeled data and feature engineering are needed for interpretation through machine learning, so it is easy to overfit; through the deep learning method, the demand for computing resources is high, the demand for data is large, and the model complexity is high, resulting in low engineering efficiency, few available conclusions, and facing image data obtained by multi-source and multi-geological exploration, how to quickly and accurately invert the three-dimensional distribution of the exploration area is a major theoretical problem. SUMMARY

[0004] The technical problem to be solved by the present application is that the existing detection image processing process is complex and the interpretation method is low in efficiency. The present application provides a tunnel karst advanced geological prediction automatic interpretation method, device and medium, which significantly improves the interpretation accuracy and efficiency of complex images in the prediction data through multi-scale data fusion, unsupervised image reconstruction and three-dimensional reconstruction model. The model firmware method of the present application enhances the generalization ability of the model, improves the utilization efficiency of image processing data, ensures the reliability and accuracy of the interpretation results, and further improves the construction safety and efficiency.

[0005] The present application is realized by the following technical scheme:

[0006] The first aspect of the present application provides an automatic interpretation method for tunnel karst advanced geological prediction, comprising the following specific steps:

[0007] Obtaining tunnel karst advanced geological prediction data at different scales, preprocessing the tunnel karst advanced geological prediction data;

[0008] Reconstructing an unsupervised image for the preprocessed tunnel karst advanced geological prediction data;

[0009] Building a three-dimensional reconstruction model according to the unsupervised image;

[0010] Building a neural network model, combining a two-dimensional encoder to extract a two-dimensional distribution feature map of the three-dimensional reconstruction model;

[0011] Building a three-dimensional decoder to decode the two-dimensional feature map layer by layer, and building a data automatic interpretation model based on the decoding result; the three-dimensional decoder comprises a plurality of deconvolution layers and up-sampling layers, the deconvolution layers are used to increase the feature dimension, and the up-sampling layers are used to restore the spatial resolution of the data;

[0012] Building a generative adversarial network using a generator and a discriminator, in which the generator and the discriminator perform an antagonistic game until a Nash equilibrium is reached, the training of the data automatic interpretation model is completed, and a trained data automatic interpretation model is obtained;

[0013] Based on the trained data automatic interpretation model, inversion is carried out, the inversion result is compared and verified with the actual geological condition or other data, and the inversion result is corrected and improved.

[0014] Further, the preprocessing of the tunnel karst advanced geological prediction data specifically comprises:

[0015] Obtaining tunnel karst advanced geological prediction data at different scales, and standardizing the tunnel karst advanced geological prediction data at different scales;

[0016] Position matching and correction are performed on the standardized tunnel karst advanced geological prediction data at different scales;

[0017] Color synthesis band selection is performed on the corrected data for data fusion;

[0018] The histogram matching method is used to coordinate the brightness and contrast of the spliced image, and the preprocessing of the tunnel karst advanced geological prediction data is completed.

[0019] Further, the unsupervised image reconstruction of the preprocessed tunnel karst advanced geological prediction data specifically comprises:

[0020] Extract image data in the tunnel karst advanced geological prediction data after pretreatment;

[0021] Extract the mask matrix of the seed point of the incomplete image in the image data, which is used to mark the position of the known point;

[0022] Random white noise is input into the variational autoencoder as a blank image to be generated, and a reconstructed image is obtained;

[0023] The reconstructed image is sampled by the mask matrix;

[0024] According to the pixel point value and the seed point value of the incomplete image, the reconstruction loss is calculated;

[0025] Through the reconstruction loss, the network parameters are updated, and one round of training is completed;

[0026] Random white noise is input into the trained variational autoencoder;

[0027] The reconstructed image generated by the variational autoencoder is obtained.

[0028] Further, the calculation of the reconstruction loss specifically includes:

[0029] A loss function is constructed in the variational autoencoder, and the reconstruction loss is obtained based on the loss function;

[0030] The loss function includes reconstruction error and KL divergence;

[0031] The reconstruction error is used to measure the difference between the reconstructed data and the corresponding input data, and is obtained according to the root mean square error of the known seed points in the original data and the pixels at the same position of the reconstructed image;

[0032] The KL divergence is used to measure the difference between the latent distribution output by the encoder and the prior distribution, and is obtained according to the mean and variance output by the variational autoencoder.

[0033] Further, the construction of a three-dimensional reconstruction model according to unsupervised images specifically includes:

[0034] Based on the unsupervised image data, a plane reflection model is constructed, and each layer of the plane reflection model has an attribute value and increases from bottom to top;

[0035] Add wrinkles and tilts to the plane reflection model to simulate the layer fluctuation characteristics in the real data;

[0036] An elliptical surface is used to simulate a fault surface, a fault plane is added, and then the plane is disturbed;

[0037] Based on the disturbed data, the labels and fault data of the initial three-dimensional model are obtained;

[0038] Extract isosurfaces of the initial 3D model labels from the initial 3D model labels and tomographic data;

[0039] Isosurfaces are embedded into a 3D mesh to obtain layer data;

[0040] Based on the layer data, the initial 3D model is meshed into layers to obtain the spatial 3D coordinates of each grid.

[0041] Based on the spatial three-dimensional coordinates of each grid, the boundary and depth of the initial three-dimensional model are determined;

[0042] Spatial interpolation is performed on the boundaries and depth of the initial 3D model to obtain the top and bottom surfaces;

[0043] By combining the 3D mesh, the boundaries and depth of the initial 3D model, and the top and bottom surfaces of the initial 3D model, a 3D reconstruction model is generated.

[0044] Furthermore, the extraction of the two-dimensional distribution feature map of the three-dimensional reconstruction model specifically includes:

[0045] Construct a neural network model and input the 3D reconstruction model into the neural network model:

[0046] The input data is mapped to a low-dimensional latent space through a fully connected layer, which enables dimensionality reduction and preliminary fusion of the data.

[0047] A two-dimensional encoder is used to encode the fused latent space data to extract the spatial features of the input data;

[0048] The two-dimensional encoder includes multiple convolutional layers and pooling layers. The convolutional layers are used to extract local features, and the pooling layers are used to reduce the feature dimensionality and enhance the invariance of the features.

[0049] Based on the spatial characteristics of the input data, a two-dimensional distribution feature map reflecting the advanced geological prediction data of tunnel karst is constructed in two-dimensional space.

[0050] Furthermore, after mapping the input data to a low-dimensional latent space through a fully connected layer, an attention mechanism is added. Utilizing the weight allocation capability of the attention mechanism, data from different sources are dynamically selected and weighted to achieve high-dimensional fusion of multi-source, multi-field geological tunnel karst advanced geological prediction data, including:

[0051] The attention weights between different data modalities are calculated using the attention module.

[0052] A similarity metric between data is calculated based on attention weights;

[0053] Based on similarity metrics, weighted summation of each data modality is performed to obtain the fused latent space data.

[0054] Furthermore, the training process for the 3D model specifically includes:

[0055] A discriminator is constructed using a convolutional neural network, and the model parameters of the generator and discriminator are initialized.

[0056] The 3D model generated by the generator and the real 3D model are input into the discriminator;

[0057] The discriminator extracts spatial and semantic features from the input data through multiple convolutional and pooling layers;

[0058] Determine whether the input data is a simulated sample or a real sample based on its features;

[0059] A loss function discriminator is constructed to score simulated or real samples based on the loss function. The parameters of the discriminator are updated by constructing the loss function through the backpropagation algorithm, thereby optimizing the sample scoring of the discriminator.

[0060] When the output probabilities of real samples and generated samples are within a set threshold, the output probabilities of real samples and generated samples should be close to reaching a balance, and training should end.

[0061] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement an automatic interpretation method for advanced geological prediction of tunnel karst.

[0062] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an automatic interpretation method for advanced geological prediction of tunnel karst.

[0063] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0064] By acquiring and preprocessing advanced geological prediction data of tunnel karst at different scales, we can more comprehensively capture the details and macroscopic features of geological structures, thereby improving the accuracy of interpretation. Unsupervised image reconstruction of the preprocessed data can automatically extract useful information from the data, reduce manual intervention, and improve interpretation efficiency. Constructing a three-dimensional reconstruction model can more intuitively display the geological structure, help engineers better understand the geological conditions, and improve the accuracy of decision-making.

[0065] By combining a 2D encoder to extract the 2D distribution feature map of the 3D reconstruction model, the high-dimensional features of the data can be learned automatically, improving the generalization ability of the model. Through the adversarial game between the generator and the discriminator, the model can be continuously optimized to generate samples that are closer to real data, thereby improving the robustness and accuracy of the interpretation model.

[0066] The deconvolutional and upsampling layers in the 3D decoder can gradually restore the spatial resolution of the data, ensuring the spatial consistency between the decoded data and the original data, and improving the efficiency of data utilization.

[0067] Inversion is performed using an automatic interpretation model based on trained data. The inversion results are then compared and verified with actual geological conditions or other data. This allows for the timely identification and correction of model deficiencies, thereby improving the reliability of the interpretation results.

[0068] Through comparative verification, model parameters can be continuously adjusted and optimized to form an optimization loop between the model and the data, thereby further improving the accuracy of intelligent interpretation.

[0069] Accurate geological forecasting can help detect unfavorable geological features ahead of the tunnel, such as fault fracture zones, karst, and weak layers, thereby preventing catastrophic accidents such as water inrush, mudslides, and gas inrushes, and ensuring the safety of construction personnel and equipment. The interpretation results can provide detailed geological information for tunnel construction, helping engineering units to rationally arrange the construction schedule, optimize engineering design, and reduce construction costs. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0071] Figure 1 This is a flowchart illustrating an automatic interpretation method for advanced geological prediction of tunnel karst formations, provided in an embodiment of the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0073] like Figure 1 As shown, in order to improve data processing capabilities in the direction of image data acquisition, this embodiment provides an automatic interpretation method for tunnel karst advanced geological prediction, which improves the processing capabilities of image data, including the following specific steps:

[0074] S101, Obtain advanced geological prediction data of tunnel karst at different scales, and preprocess the advanced geological prediction data of tunnel karst;

[0075] S102, Unsupervised image reconstruction of preprocessed tunnel karst advanced geological prediction data;

[0076] S103, Construct a 3D reconstruction model based on unsupervised images;

[0077] S104, Construct a neural network model and combine it with a two-dimensional encoder to extract the two-dimensional distribution feature map of the three-dimensional reconstruction model;

[0078] S105, a 3D decoder is constructed to decode the 2D feature map layer by layer, and an automatic data interpretation model is constructed based on the decoding results; the 3D decoder includes multiple deconvolution layers and upsampling layers, the deconvolution layers are used to increase the feature dimension, and the upsampling layers are used to restore the spatial resolution of the data.

[0079] S106 uses a generator and a discriminator to construct a generative adversarial network. In the generative adversarial network, the generator and the discriminator engage in adversarial games until a Nash equilibrium is reached, thus completing the training of the automatic data interpretation model and obtaining the trained automatic data interpretation model.

[0080] S107, based on the trained data, automatically interprets the model to perform inversion, compares and verifies the inversion results with actual geological conditions or other data, and corrects and improves the inversion results.

[0081] In some possible implementations, after obtaining tunnel karst advanced geological prediction data using geological analysis and geophysical methods, preprocessing is required, including data cleaning and standardization, to ensure the quality of the input data. Then, unsupervised image reconstruction is performed on the preprocessed tunnel karst advanced geological prediction data. A 3D reconstruction model is constructed based on the unsupervised images. When constructing the 3D reconstruction model, settings such as multi-index loss function weight parameters need to be considered to optimize model performance. A neural network model is constructed, combined with a 2D encoder to extract the 2D distribution feature map of the 3D reconstruction model. The encoder is used to encode the input data into a low-dimensional feature representation. When constructing the neural network model, an appropriate model size needs to be selected to prevent overfitting, and attention mechanisms can be used to help the model better capture long-distance dependencies. A 3D decoder is constructed to decode the 2D feature map layer by layer. Based on the decoding results, an automatic data interpretation model is built. The decoder typically consists of multiple deconvolutional layers (transposed convolutional layers), responsible for restoring the feature vectors or feature maps output by the encoder to an output of the same size as the input image. As the decoder progresses, the spatial resolution of the feature map gradually increases. When constructing an automatic data interpretation model, the automatic interpretation technology for tunnel rock mass integrity based on the fusion of digital drilling and multi-scale models can be referenced. This technology, while meeting the basic condition of high prediction accuracy, can efficiently and conveniently provide more detailed geological prediction information. A generative adversarial network (GAN) is constructed using a generator and a discriminator. In this GAN, the generator and discriminator engage in adversarial game until a Nash equilibrium is reached, completing the training of the automatic data interpretation model and obtaining the trained model. The GAN consists of a generator and a discriminator; the generator generates data, and the discriminator judges the authenticity of the data, maintaining stability and accelerating the training process. Inversion is performed based on the trained automatic data interpretation model. The inversion results are compared and verified with actual geological conditions or other data. The inversion results are then corrected and improved. A multi-level, multi-scale accuracy verification strategy can be adopted, comprehensively considering multi-source surface ground truth datasets obtained through field surveys, manual interpretation, and historical data collection. The variation characteristics of surface ground truth with observation scale are analyzed, and a surface relative ground truth probability expression model and multi-scale ground truth sample reconstruction technology are developed. In addition, studies on the advanced geological prediction effects of different geophysical methods can be referenced, and the prediction effect can be improved by combining multiple comprehensive geophysical methods.

[0082] In some possible implementations, standardizing tunnel karst advanced geological prediction data at different scales can eliminate dimensional and magnitude differences between data, giving the data a unified scale and range, improving data consistency and comparability. Position matching and correction of the standardized data ensures spatial alignment of data at different scales, avoiding misalignment caused by data acquisition or processing errors, thus improving data accuracy and reliability. Selecting appropriate color composite bands highlights geological structures and features, making geological information more intuitive and clear. Color composite enhances the visual effect of the data, enabling better identification and analysis of geological features. By fusing data at different scales and integrating the advantages of different data sources, more geological information can be extracted, improving the accuracy and comprehensiveness of interpretation. Histogram matching methods are used to coordinate the brightness and contrast of the stitched images, making the fused image more visually consistent and natural. Histogram matching can eliminate brightness and contrast differences caused by different data sources or processing methods, improving the overall image quality. Multi-scale data fusion: The fused data contains geological information at different scales, providing a more comprehensive geological background and helping geological engineers to interpret geological data more accurately, improving interpretation efficiency and accuracy. Automated processing reduces human intervention and subjective judgment, improving data processing efficiency and objectivity, and reducing errors caused by human factors. Data processed through standardization, correction, fusion, and histogram matching is of higher quality and more suitable as input data for training machine learning or deep learning models. High-quality data can improve the model's generalization ability, enabling it to maintain high performance under different geological environments and data conditions. Therefore, the steps of multi-scale data standardization, location matching and correction, color composite band selection, data fusion, and histogram matching significantly improve the consistency, fusion effect, and image quality of tunnel karst advanced geological prediction data. These processing steps not only improve the accuracy and reliability of the data but also enhance the model's generalization ability, ultimately improving the efficiency and accuracy of geological interpretation and providing more reliable geological information support for tunnel construction.

[0083] In some possible implementations, unsupervised image reconstruction is performed on the preprocessed tunnel karst advanced geological prediction data, specifically including:

[0084] Image data is extracted from the preprocessed data. This image data may contain information such as geological structure and rock layer distribution. Through data filtering and extraction techniques, the image data is separated from the multi-source data.

[0085] Identify and label known points in the image, which will be used for subsequent image reconstruction and loss calculation, i.e., extract the mask matrix of seed points of the incomplete image in the image data;

[0086] Random white noise is input as a blank image to be generated into the variational autoencoder to obtain a reconstructed image. The variational autoencoder can generate a reconstructed image similar to the original image.

[0087] The reconstructed image is sampled using a mask matrix. The mask matrix only samples the pixel values ​​of known points and ignores unknown points.

[0088] The reconstruction loss is calculated based on the sampled pixel values ​​and the seed point values ​​of the incomplete image;

[0089] By reconstructing the loss and updating the network parameters, one round of training is completed. That is, the network parameters are updated by backpropagation based on the reconstruction loss, thus optimizing the performance of the variational autoencoder.

[0090] Input random white noise into the trained variational autoencoder;

[0091] The reconstructed image generated by the variational autoencoder is obtained, producing a high-quality reconstructed image for geological analysis and prediction.

[0092] The reconstructed images generated by the variational autoencoder can more accurately restore the details of incomplete images, improving image integrity and accuracy. Through the optimization of random white noise input and reconstruction loss, the model can better handle different types of geological images, enhancing its generalization ability. Under different geological environments and data conditions, the model can maintain high reconstruction accuracy and is suitable for various geological exploration scenarios. Automated generation of high-quality reconstructed images reduces the workload of manual labeling and correction, improving work efficiency. More accurate image reconstruction can provide a more reliable basis for geological forecasting, improving the accuracy and reliability of forecasts. In complex geological environments, multi-scale analysis can provide a more comprehensive understanding of geological structures, supporting more refined geological interpretation and prediction, significantly improving the processing and analysis capabilities of tunnel karst advanced geological forecasting data, and providing stronger support for geological engineering.

[0093] In some possible implementations, the calculation of the reconstruction loss specifically includes:

[0094] In the variational autoencoder, a loss function is constructed, and the reconstruction loss is obtained based on the loss function. That is, the total loss function of the variational autoencoder is the weighted sum of the reconstruction error and the KL divergence.

[0095] The loss function includes reconstruction error and KL divergence. Reconstruction error is a measure of the difference between the reconstructed data and the corresponding input data, and is obtained by calculating the root mean square error of the known seed points in the original data and the pixels at the same position in the reconstructed image. Depending on the type of data, different loss functions can be selected to calculate the reconstruction error, including mean square error (MSE) and cross-entropy loss. KL divergence is used to measure the difference between the latent distribution of the encoder output and the prior distribution, and is obtained by calculating the mean and variance of the variational autoencoder output.

[0096] In some possible implementations, a 3D reconstruction model is constructed based on the unsupervised image, specifically including:

[0097] Based on unsupervised image data, unsupervised learning algorithms, such as autoencoders or variational autoencoders, are used to extract features from the image data and construct a planar reflection model. The attribute values ​​of each layer can be determined through feature extraction and clustering methods, and are hierarchically increasing; that is, each layer of the planar reflection model has one attribute value that increases from bottom to top. To simulate the undulation characteristics of the stratigraphy in real data, folds and tilts are added to the planar reflection model. This can be achieved by simulating geological structural movements, such as folds and faults. Fault planes are simulated using elliptical surfaces, folds can be simulated using sine wave functions, and tilts can be simulated using rotation and translation operations. That is, fault planes are added, and then the planes are perturbed. Through data labeling and feature extraction methods, the labels of the initial 3D model and fault data are extracted from the perturbed data. Using an isosurface extraction algorithm, isosurfaces of the initial 3D model labels are extracted from the initial 3D model labels and fault data. The extracted isosurfaces are embedded into the 3D mesh to generate stratigraphic data. A 3D mesh can be constructed using rasterization or finite element methods, and isosurface data can be embedded into the mesh. Layered meshing techniques are used to mesh the initial 3D model according to layer data, generating spatial 3D coordinates for each mesh. Through spatial analysis and geometric calculations, the boundaries and depths of each mesh are determined. Boundary detection algorithms and depth calculation methods, such as least squares or interpolation methods, can be used to determine the boundaries and depths. Spatial interpolation methods, such as Kriging or spline interpolation, are used to interpolate the boundary and depth data, generating top and bottom surfaces. Combining the 3D mesh, the boundaries and depths of the initial 3D model, and the top and bottom surfaces of the initial 3D model, a 3D reconstructed model is generated. In other words, all the above data and results are combined to generate the final 3D reconstructed model.

[0098] In some possible implementations, by simulating real geological features such as folds, dips, and faults, the generated 3D model is closer to the actual geological conditions, improving the accuracy and reliability of the model. Isosurface extraction and hierarchical meshing techniques can capture more detailed information, making the model more refined. By automating data extraction and processing, manual intervention is reduced, improving work efficiency. Furthermore, the generated 3D model supports multi-scale analysis, can provide geological information at different levels, and is suitable for various geological analysis and prediction tasks.

[0099] In some possible implementations, the two-dimensional distribution feature map of the three-dimensional reconstruction model is extracted, specifically including:

[0100] Construct a neural network model by inputting the 3D reconstruction model into the neural network model. The neural network model includes:

[0101] Convolutional layers: Multiple convolutional layers are used to extract local features. Convolutional layers extract local features by sliding convolutional kernels (filters) across the input data.

[0102] Pooling layers: Pooling layers reduce feature dimensionality and enhance feature invariance. Through max pooling or average pooling operations, pooling layers reduce the size of feature maps while preserving important information.

[0103] Structure: Two-dimensional encoders typically consist of a combination of multiple convolutional and pooling layers to progressively extract and compress features.

[0104] The input data is mapped to a low-dimensional latent space through a fully connected layer, which enables dimensionality reduction and preliminary fusion of the data.

[0105] A two-dimensional encoder is used to encode the fused latent space data to extract the spatial features of the input data;

[0106] The two-dimensional encoder includes multiple convolutional layers and pooling layers. The convolutional layers are used to extract local features, and the pooling layers are used to reduce the feature dimensionality and enhance the invariance of the features.

[0107] Based on the spatial characteristics of the input data, a two-dimensional distribution feature map reflecting the advanced geological prediction data of tunnel karst is constructed in two-dimensional space.

[0108] In some possible implementations, a two-dimensional distribution feature map is extracted, specifically including:

[0109] The role of fully connected layers: Fully connected layers map input data to a low-dimensional latent space, achieving dimensionality reduction and initial data fusion. In neural networks, fully connected layers typically follow convolutional layers or encoders. They globally integrate the local features extracted by previous layers to obtain a more representative feature representation. For example, in convolutional neural networks, feature maps extracted by convolutional and pooling layers can be weighted and summed by fully connected layers to obtain a fixed-length feature vector, thereby reducing computation and improving model efficiency.

[0110] The role of a 2D encoder: A 2D encoder can further encode the features output by the fully connected layer, extracting more effective spatial features. It can learn the distribution patterns and inherent relationships of data in two-dimensional space, transforming complex input data into a 2D distribution feature map with certain patterns and interpretability. For example, in image processing tasks, a 2D encoder can encode the pixel information of an image into a 2D feature map, where the value of each point represents the feature strength or importance at that location, thus intuitively displaying the spatial distribution characteristics of the image.

[0111] The role of pooling layers: Pooling layers enhance feature invariance by reducing feature dimensionality. In convolutional neural networks, pooling layers typically follow convolutional layers. They downsample the feature maps output by the convolutional layers, retaining the main feature information while removing some details. This makes the model more robust to small changes in the input data. For example, in image recognition tasks, even if the input image undergoes some degree of translation, rotation, or scaling, the feature maps processed by the pooling layer can still remain relatively stable, thereby improving the model's ability to recognize images of different shapes.

[0112] High-dimensional fusion of multi-source, multi-field geological tunnel karst advanced geological prediction data:

[0113] The introduction of attention mechanisms: After mapping input data to a low-dimensional latent space through fully connected layers, adding an attention mechanism leverages its weight allocation capabilities to dynamically select and weight data from different sources. The attention mechanism automatically learns the importance of different parts of the data, allocating more weight to key information, thereby achieving effective fusion of multi-source, multi-field geological tunnel and karst advanced geological prediction data. For example, in geological exploration tasks involving data from multiple sensors, the data from different sensors may contain different important information. The attention mechanism can automatically identify and highlight this key information, improving the quality and representativeness of the fused data.

[0114] Calculation and application of attention weights:

[0115] Calculating Attention Weights: The attention module calculates the attention weights between each data modality. The attention module typically assigns a weight value to each data modality based on factors such as similarity and correlation between the data. These weight values ​​represent the importance of that data modality in the overall fusion process. For example, methods such as dot product attention and scaled dot product attention can be used to calculate the weights. Dot product attention obtains the weights by calculating the dot product of the feature vectors of two data modalities, while scaled dot product attention introduces a scaling factor to prevent problems such as vanishing gradients caused by excessively large dot product results.

[0116] Calculating similarity metrics: Similarity metrics are calculated between data based on attention weights. These metrics further reflect the degree of association between different data modalities, providing a more accurate basis for subsequent weighted summation. Commonly used similarity metrics include cosine similarity and Euclidean distance. Cosine similarity measures the angle between two vectors, reflecting their directional similarity; Euclidean distance represents the spatial distance between two vectors, with smaller distances indicating higher similarity.

[0117] Weighted summation yields the fused latent space data: Based on a similarity metric, the feature vectors of each data modality are weighted and summed to obtain the fused latent space data. Specifically, the feature vector of each data modality is multiplied by its corresponding attention weight, and then all weighted feature vectors are summed to obtain a comprehensive feature vector, which represents the fused latent space data. This approach comprehensively considers information from different data modalities, providing a more comprehensive and accurate feature representation for subsequent geological engineering predictions and analyses.

[0118] In some possible implementations, a 3D decoder is constructed to decode the 2D feature map layer by layer, and an automatic data interpretation model is constructed based on the decoding results; the 3D decoder includes multiple deconvolution layers and upsampling layers, the deconvolution layers are used to increase the feature dimension, and the upsampling layers are used to restore the spatial resolution of the data;

[0119] Deconvolutional layers are used to increase feature dimension. In deep learning, deconvolutional layers (also known as transposed convolutional layers) map low-resolution feature maps to high-resolution feature maps by learning the inverse operation of the convolution kernel. This helps restore the spatial resolution of the feature maps, gradually bringing them closer to the size of the original input data. For example, in image segmentation tasks, deconvolutional layers can progressively restore the low-resolution feature map output by the encoder to an output of the same size as the input image.

[0120] Upsampling layers are used to restore the spatial resolution of data. Upsampling operations can be implemented in various ways, such as nearest neighbor interpolation and bilinear interpolation. These methods increase the size of the feature map by inserting new pixels, thereby restoring the spatial resolution of the data. For example, nearest neighbor interpolation copies each pixel to a new location, while bilinear interpolation interpolates based on the values ​​of surrounding pixels.

[0121] 3D decoders typically consist of multiple deconvolutional layers and upsampling layers. These layers work together to progressively restore the low-resolution feature maps output by the encoder to high-resolution feature maps. After each upsampling step, skip connections can be used to concatenate the feature maps of the corresponding encoder layers with those of the decoder to preserve more detailed information. This structure is widely used in networks such as U-Net.

[0122] The 3D decoder starts with the feature map received from the encoder and progressively increases the size of the feature map through deconvolutional layers and upsampling layers. After each upsampling step, the feature map of the corresponding layer from the encoder is concatenated with the feature map from the decoder through skip connections to retain more detailed information. Finally, the high-resolution feature map output by the decoder can be used for subsequent interpretation tasks.

[0123] A data automatic interpretation model is constructed based on the high-resolution feature map output by the decoder.

[0124] In some possible implementations, the training process of the 3D model is as follows:

[0125] Constructing the discriminator: A convolutional neural network (CNN) is used to construct the discriminator. CNNs have powerful feature extraction capabilities and can handle spatially correlated data well, such as 3D model data. In the discriminator, convolutional layers can capture the local spatial features of the 3D model, such as details like edges and textures. By stacking multiple convolutional layers, higher-level semantic features, such as the model's shape and structure, can be extracted progressively.

[0126] Initializing the model parameters of the generator and discriminator: Initializing the model parameters of the generator and discriminator is crucial. Random initialization, such as Gaussian distribution initialization, is typically used. This is because if all parameters are initialized to the same value (e.g., all 0), then during training, each neuron in the network will behave identically during forward and backward propagation, preventing the network from learning effectively. Random initialization breaks this symmetry, allowing each neuron to learn different features early in training, thus promoting network convergence. For example, for a generator with multiple convolutional and fully connected layers, the weights and biases of the convolutional kernels, as well as the weights and biases of the fully connected layers, are initialized according to a certain random distribution.

[0127] The generator produces 3D models: During training, the generator generates 3D models based on the input noise vector (usually a random vector). In the early stages of training, these generated 3D models often differ significantly from the real 3D model, exhibiting noticeable unnatural features. As training progresses, the generator gradually learns how to generate samples that more closely resemble the real 3D model. The generated 3D model data format needs to be consistent with the real 3D model so that it can be input into the discriminator for processing.

[0128] Realistic 3D Models: Realistic 3D models are high-quality 3D data acquired from actual scenes. This data possesses rich spatial features and semantic information, serving as a crucial basis for the discriminator to learn and distinguish between real and fake samples during training. The data sources for realistic 3D models can include object models, architectural models, etc., acquired by 3D scanning equipment, or precise models exported from 3D modeling software. Before being input into the discriminator, realistic 3D models may require preprocessing, such as normalizing dimensions and adjusting data formats, to ensure consistency with the 3D models generated by the generator at the data level.

[0129] The discriminator's feature extraction process: After the 3D model generated by the generator and the real 3D model are input into the discriminator, the discriminator first extracts spatial features from the input data through multiple convolutional layers. The convolutional kernels in the convolutional layers slide across the 3D model data, extracting local feature information. For example, when processing a 3D building model, the convolutional layer can extract local features such as window edges and door outlines. Next, pooling layers downsample the extracted features, reducing the dimensionality of the feature data while retaining important feature information. Pooling layers can use max pooling or average pooling. Max pooling highlights the maximum value in the features, emphasizing salient features; average pooling is relatively smooth and retains the average information of the features. After multiple convolutional and pooling operations, the discriminator can extract high-level semantic features of the input 3D model, such as the overall style of the building and the category of objects.

[0130] Determining the authenticity of input data: Based on the extracted features, the discriminator judges the input data to determine whether it is a simulated sample (generated by the generator) or a real sample. This is typically achieved through a fully connected layer or an output layer, whose output can be a probability value representing the probability that the input data is a real sample. For example, an output probability of 0.9 means the discriminator considers the input data to be 90% likely to be a real sample, while an output probability of 0.1 means the discriminator considers the input data to be 90% likely to be a simulated sample generated by the generator. During training, the discriminator continuously learns how to accurately make this authenticity judgment based on the features.

[0131] Constructing the Loss Function: The loss function is a key metric for measuring the accuracy of the discriminator's judgments. For a discriminator, its loss function typically consists of two parts: one part is the loss for real samples, where the discriminator aims to output a probability as close to 1 as possible for real samples (i.e., correctly classifying them as real); the other part is the loss for generated samples, where the discriminator aims to output a probability as close to 0 as possible for generated samples (i.e., correctly classifying them as simulated samples). For example, a binary cross-entropy loss function can be used to construct the discriminator's loss function. For real samples, the loss function is y... ture ×log(y pred ), where y ture For real labels, y pred Let be the predicted output probability of the discriminator for real samples; for generated samples, the loss function is (1-y ture )×log(1-y pred ), where y ture To generate labels for the samples, y pred Let be the predicted output probability of the discriminator for the generated sample. Adding these two loss components together yields the total loss function of the discriminator.

[0132] Backpropagation and Parameter Update: Based on the constructed loss function, the discriminator's parameters are updated using the backpropagation algorithm. During backpropagation, the gradient of the loss function with respect to the discriminator's output layer is first calculated, and then this gradient is backpropagated to the preceding convolutional layers, pooling layers, and other network layers. At each network layer, the gradient of the layer's parameters is calculated according to the chain rule; for example, for convolutional layers, the gradients of the convolutional kernel weights and biases are calculated. After obtaining the gradients of each layer's parameters, optimization algorithms (such as stochastic gradient descent, Adam optimization, etc.) are used to update the parameters. The optimization algorithm adjusts the parameter values ​​according to the direction and magnitude of the gradients, gradually reducing the value of the loss function, thereby optimizing the discriminator's scoring of samples and improving its accuracy in distinguishing between true and false samples.

[0133] Output probability balance: When the output probabilities of real samples and generated samples are within a set threshold range, it means that the discriminator has difficulty distinguishing between real and generated samples. At this point, training can be considered to have reached a balance, and training can be terminated. For example, setting the threshold to 0.05, when the output probability of real samples is between 0.475 and 0.525, and the output probability of generated samples is also between 0.475 and 0.525, it indicates that the discriminator's judgment of the two is very close, and the 3D model generated by the generator is realistic enough to "deceive" the discriminator, thus achieving the training goal. The setting of this threshold needs to be determined based on the specific training task and model performance requirements. Different application scenarios may have different requirements for the realism of the generated model, so the threshold will also vary.

[0134] As one possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements an automatic interpretation method for advanced geological prediction of tunnel karst.

[0135] As one possible implementation, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements an automatic interpretation method for advanced geological prediction of tunnel karst.

[0136] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic interpretation method for advanced geological prediction of tunnel karst formations, characterized in that, The specific steps include the following: Acquire advanced geological prediction data of tunnel karst at different scales, and preprocess the advanced geological prediction data of tunnel karst; Unsupervised image reconstruction was performed on the preprocessed tunnel karst advanced geological prediction data; A 3D reconstruction model is constructed based on unsupervised images. Specifically, this construction includes: constructing a planar reflection model based on unsupervised image data, where each layer of the planar reflection model has an attribute value that increases from bottom to top; adding wrinkles and tilts to the planar reflection model to simulate the layer undulations in real data; simulating fault planes with elliptical surfaces, adding fault planes, and then perturbing the planes; obtaining the initial 3D model's labels and fault data based on the perturbed data; extracting isosurfaces of the initial 3D model labels from the initial 3D model's labels and fault data; embedding the isosurfaces into a 3D mesh to obtain layer data; performing layered meshing on the initial 3D model based on the layer data to obtain the spatial 3D coordinates of each mesh; determining the boundary and depth of the initial 3D model based on the spatial 3D coordinates of each mesh; performing spatial interpolation on the boundary and depth of the initial 3D model to obtain the top and bottom surfaces; and combining the 3D mesh, the boundary and depth of the initial 3D model, and the top and bottom surfaces of the initial 3D model to generate a 3D reconstruction model. A neural network model is constructed, and a two-dimensional encoder is used to extract the two-dimensional distribution feature map of the three-dimensional reconstruction model; A 3D decoder is constructed to decode the 2D feature map layer by layer, and an automatic data interpretation model is built based on the decoding results. The 3D decoder includes multiple deconvolution layers and upsampling layers. The deconvolution layers are used to increase the feature dimension, and the upsampling layers are used to restore the spatial resolution of the data. A generative adversarial network is constructed using a generator and a discriminator. In the generative adversarial network, the generator and the discriminator engage in adversarial games until a Nash equilibrium is reached, thereby completing the training of the automatic data interpretation model and obtaining the trained automatic data interpretation model. The inversion is performed based on the trained data automatic interpretation model to obtain the interpretation results of tunnel karst advanced geological prediction.

2. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 1, characterized in that, The preprocessing of tunnel karst advanced geological prediction data specifically includes: Acquire advanced geological prediction data of tunnel karst at different scales, and standardize the advanced geological prediction data of tunnel karst at different scales; Location matching and correction were performed on tunnel karst advanced geological prediction data at different scales after standardization. The corrected data is then subjected to color composite band selection and data fusion. Histogram matching was used to coordinate the brightness and contrast of the stitched images, thus completing the preprocessing of tunnel karst advanced geological prediction data.

3. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 1, characterized in that, Unsupervised image reconstruction was performed on the preprocessed tunnel karst advanced geological prediction data, specifically including: Extract image data from the preprocessed tunnel karst advanced geological prediction data; Extract the mask matrix of seed points from the incomplete image data to mark the positions of known points; Random white noise is input as a blank image to be generated into a variational autoencoder to obtain a reconstructed image; The reconstructed image is sampled using a mask matrix; The reconstruction loss is calculated based on the sampled pixel values ​​and the seed point values ​​of the incomplete image; By reconstructing the loss and updating the network parameters, one round of training is completed. Input random white noise into the trained variational autoencoder; The reconstructed image generated by the variational autoencoder is obtained.

4. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 3, characterized in that, The calculation of the reconstruction loss specifically includes: A loss function is constructed in the variational autoencoder, and the reconstruction loss is obtained based on the loss function. The loss function includes reconstruction error and KL divergence; The reconstruction error is a measure of the difference between the reconstructed data and the corresponding input data, and is obtained by calculating the root mean square error of the known seed points in the original data and the pixels at the same position in the reconstructed image. The KL divergence is used to measure the difference between the latent distribution and the prior distribution of the encoder output, and is obtained from the mean and variance of the variational autoencoder output.

5. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 1, characterized in that, The extraction of the two-dimensional distribution feature map of the three-dimensional reconstruction model specifically includes: Construct a neural network model and input the 3D reconstruction model into the neural network model: The input data is mapped to a low-dimensional latent space through a fully connected layer, which enables dimensionality reduction and preliminary fusion of the data. A two-dimensional encoder is used to encode the fused latent space data to extract the spatial features of the input data; The two-dimensional encoder includes multiple convolutional layers and pooling layers. The convolutional layers are used to extract local features, and the pooling layers are used to reduce the feature dimensionality and enhance the invariance of the features. Based on the spatial characteristics of the input data, a two-dimensional distribution feature map reflecting the advanced geological prediction data of tunnel karst is constructed in two-dimensional space.

6. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 5, characterized in that, After mapping the input data to a low-dimensional latent space through a fully connected layer, the system also includes an attention mechanism. Utilizing the weight allocation capability of the attention mechanism, it dynamically selects and weights data from different sources to perform high-dimensional fusion of multi-source, multi-field geological tunnel karst advanced geological prediction data, including: The attention weights between different data modalities are calculated using the attention module. A similarity metric between data is calculated based on attention weights; Based on similarity metrics, weighted summation of each data modality is performed to obtain the fused latent space data.

7. The automatic interpretation method for tunnel karst advanced geological prediction according to claim 1, characterized in that, The training process for a 3D model specifically includes: A discriminator is constructed using a convolutional neural network, and the model parameters of the generator and discriminator are initialized. The 3D model generated by the generator and the real 3D model are input into the discriminator; The discriminator extracts spatial and semantic features from the input data through multiple convolutional and pooling layers; Determine whether the input data is a simulated sample or a real sample based on its features; A loss function discriminator is constructed to score simulated or real samples based on the loss function. The parameters of the discriminator are updated by constructing the loss function through the backpropagation algorithm, thereby optimizing the sample scoring of the discriminator. When the output probabilities of real samples and generated samples reach a balance at a set threshold, training should end.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the automatic interpretation method for advanced geological prediction of tunnel karst as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the automatic interpretation method for advanced geological prediction of tunnel karst as described in any one of claims 1 to 7.

Images