Seismic and electric combined tunnel advanced geological prediction method based on double convolution neural network

Abstract

The application discloses a seismic and electric combined tunnel advanced geological prediction method based on a double convolutional neural network, relates to the field of tunnel engineering geophysical exploration, and comprises the following steps: collecting seismic wave reflection data and transient electromagnetic response data in front of a tunnel face; pre-processing and spatial grid aligning the heterogeneous data, constructing a double branch convolutional neural network, extracting structural form features of seismic data and electrical property evolution features of electromagnetic data, introducing a cross attention mechanism to dynamically weight fuse the double branch features, realizing mutual checking of seismic and electric physical fields, and outputting a three-dimensional geological risk probability graph. Through an end-to-end feature level fusion framework, the application significantly improves the identification accuracy and spatial positioning precision of adverse geological bodies such as faults, karst caves and water-rich zones, gets rid of the dependence on expert experience, and achieves a processing speed of minutes. The introduction of a physical constraint term effectively eliminates false predictions of reverse logic, and the application has the outstanding advantages of high automation and strong dynamic adaptability.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration in tunnel engineering, and more specifically, to a method for advanced geological prediction of tunnels based on a dual convolutional neural network combining seismic and electrical methods. Background Technology

[0002] When tunnel engineering traverses complex geological conditions, adverse geological formations ahead of the tunnel face (such as faults, karst caves, and water-rich zones) can easily trigger disasters such as water inrush and landslides. Therefore, advanced geological prediction is a crucial link in ensuring construction safety. Currently, long-distance advanced geological prediction mainly relies on geophysical exploration methods, among which seismic wave methods (such as TSP and TGP) and transient electromagnetic methods (TEM) are the two most widely used technologies. Seismic wave methods are sensitive to structural interfaces with large differences in wave impedance (such as faults and fracture zones) and have high spatial resolution; transient electromagnetic methods are extremely sensitive to low-resistivity bodies (water-bearing structures) and can effectively identify water-rich areas. However, each method has inherent limitations and cannot independently complete the accurate interpretation of complex geological bodies.

[0003] In existing technologies, to achieve more reliable forecasts, a combined seismic and electromagnetic interpretation strategy is typically employed. This involves first independently inverting seismic and transient electromagnetic data, then manually performing qualitative comparisons or simple weighted superpositions based on experience to comprehensively assess the geological conditions ahead. However, this traditional combined interpretation method has the following drawbacks:

[0004] Data dimensional mismatch: Seismic data is a time-amplitude sequence, reflecting differences in elastic wave impedance; transient electromagnetic data is a time-decay voltage sequence, reflecting differences in electrical properties. The sampling rate, physical dimension, and resolution of the two differ significantly, making effective spatial alignment and quantitative fusion difficult using traditional methods. Different physical field response mechanisms: Seismic waves are sensitive to tectonic boundaries but not to water content, while transient electromagnetic data are sensitive to low-resistivity bodies but have ambiguous spatial positioning. The respective advantages and disadvantages of both methods have not been deeply complementary in traditional joint interpretation, and the problem of multiple solutions remains prominent. Reliance on human experience: Existing technologies mostly adopt a process of "independent inversion first, then manual synthesis," with the final judgment heavily dependent on the professional experience of interpreters. This results in strong subjectivity, large fluctuations in prediction accuracy, and long processing times (on the order of hours), making it difficult to meet the needs of rapid tunnel excavation.

[0005] In summary, the existing combined seismic and electrical tunnel advanced geological prediction technology mainly has the following technical problems:

[0006] High ambiguity: Due to the lack of in-depth quantitative coupling of seismoelectric heterogeneous data, it is impossible to effectively distinguish between the authenticity of tectonic anomalies and water-bearing anomalies, resulting in prominent ambiguity in the forecast results and a high risk of false alarms and missed alarms;

[0007] Low positioning accuracy: The transient electromagnetic method itself has insufficient spatial resolution, and the traditional combined method fails to utilize the high resolution of seismic waves to constrain the low-resistivity anomaly at the boundary, resulting in a large spatial positioning error of the anomaly body, which is difficult to meet the requirements of refined construction.

[0008] Low level of automation: The entire process relies on manual intervention and experience-based judgment, resulting in low processing efficiency and poor consistency of results, making it impossible to achieve automated and standardized intelligent forecasting.

[0009] Therefore, the present invention aims to provide a method for advanced geological prediction of tunnels based on a dual convolutional neural network and combined seismic and electrical signals, in order to solve the above-mentioned problems. Summary of the Invention

[0010] The purpose of this invention is to provide a combined seismic and electrical geological prediction method for tunnels based on dual convolutional neural networks. This invention significantly improves the accuracy of identifying adverse geological bodies such as faults, karst caves, and water-rich zones, as well as the accuracy of spatial positioning, through an end-to-end feature-level fusion framework. The positioning error is controlled within 3 meters, eliminating the reliance on expert experience and achieving a processing speed of minutes. The introduction of physical constraints effectively eliminates false predictions with anti-logic. It has outstanding advantages such as high automation and strong dynamic adaptability, and can be widely applied in the field of tunnel advanced geological prediction.

[0011] The above-mentioned technical objective of this invention is achieved through the following technical solution: a method for advanced geological prediction of tunnels based on a dual convolutional neural network, comprising the following steps:

[0012] S1. Data Acquisition: Acquire seismic wave reflection data and transient electromagnetic response data in front of the tunnel face;

[0013] S2. Data Preprocessing and Spatial Grid Alignment: The seismic wave reflection data and transient electromagnetic response data are preprocessed separately, and both are projected onto the same tunnel axial coordinate system to establish a spatial grid model, forming a spatially aligned seismic input tensor. and electromagnetic input tensor ;in: For the location of the measuring point, For depth;

[0014] S3. Dual-branch feature extraction: Construct parallel first and second convolutional neural network branches to extract the seismic input tensor. The first convolutional neural network branch is input to extract earthquake feature vectors. The electromagnetic input tensor The electromagnetic feature vector is extracted by inputting the second convolutional neural network branch. ;

[0015] S4. Heterogeneous Feature Fusion: The seismic feature vectors are fused... With the electromagnetic characteristic vector After channel stitching, a spatial attention weight map is calculated using an attention mechanism. The attention weight map is then multiplied element-wise with the concatenated feature map to obtain the fused feature. ;

[0016] S5. Output prediction results: The fused features... Input a classifier or regressor, and output the prediction results of the geological bodies in front of the tunnel face;

[0017] S6. Training with a composite loss function: Constructing a function that includes classification loss. Location loss and physical constraint loss Composite loss function The first convolutional neural network branch and the second convolutional neural network branch are jointly trained using this composite loss function; wherein: This is the geological type weighting coefficient, with a value range of [0,1]. Positioning and size weighting coefficients, with a value range of [0,1]; This represents the physical constraint weighting coefficient, with a value range of [0,1].

[0018] The physical constraint loss This is achieved by calculating the cross-gradient constraint between seismic characteristic gradients and electromagnetic characteristic gradients. The calculation formula is as follows:

[0019] .

[0020] The present invention is further configured such that: the preprocessing of the seismic wave reflection data in step S2 specifically includes: denoising, band filtering, migration imaging, amplitude normalization, and conversion into a two-dimensional spatial-depth migration matrix; the preprocessing of the transient electromagnetic response data specifically includes: removing early noise, calculating apparent resistivity, inverting to obtain resistivity profiles, normalization, and conversion into a two-dimensional spatial-depth resistivity matrix.

[0021] The present invention is further configured such that: the first convolutional neural network branch adopts a deep residual network structure to capture the continuity, amplitude anomalies, and wave velocity variation characteristics of the reflection interface; its second... The feature transfer formula for a layer is:

[0022]

[0023] in: Represents the convolution weights. Represents the ReLU activation function. For the first Seismic characteristic map of the layer.

[0024] The invention is further configured such that: the second convolutional neural network branch employs multi-scale convolutional kernels, using 3×3 and 5×5 convolutional kernels in parallel, to capture the longitudinal evolution trend of apparent resistivity and the morphological characteristics of low-resistivity anomalies; its... The feature transfer formula for a layer is:

[0025]

[0026] in: This indicates splicing along the channel dimension.

[0027] The present invention is further configured such that: the attention mechanism in step S4 specifically includes: calculating a spatial attention weight map through a 1×1 convolutional layer and a sigmoid activation function. The calculation formula is as follows:

[0028]

[0029] in: Earthquake eigenvectors With electromagnetic eigenvectors The channel splicing result, where b is the bias term; then the attention weight map is... With the stitched feature map Element-wise multiplication is performed to obtain the fusion feature. .

[0030] The present invention is further configured such that the prediction results in step S5 include: geological body classification results and / or regression prediction results; the geological body classification results include normal surrounding rock, fault fracture zone, water-rich area, karst and weak interlayer; the regression prediction results include water-rich probability value, risk level index and anomaly location coordinates.

[0031] The present invention is further configured such that: the classification loss The cross-entropy loss function is used; the localization loss The mean squared error loss function or the IoU loss function is used; the weighting coefficients , , The value range is [0, 1], and it can be dynamically adjusted according to the engineering stage and geological conditions.

[0032] The present invention is further configured such that: a multi-stage optimization strategy is adopted during the training process in step S6: the first stage is set , , First, perform basic classification feature extraction training; second stage: increase... The third stage involves training spatial positioning accuracy; physical constraints are then introduced and set... Physical consistency calibration training was conducted.

[0033] The present invention is further configured such that: the prediction result output in step S5 also includes: a three-dimensional geological risk probability map within a range of 100-120 meters in front of the tunnel face; the method also includes a step of post-processing the prediction result, the post-processing including spatial interpolation and three-dimensional visualization.

[0034] The present invention is further configured such that the method also includes: incrementally optimizing the trained model through transfer learning as real geological data is revealed during tunnel excavation.

[0035] In summary, the present invention has the following beneficial effects:

[0036] 1. This invention adopts an end-to-end feature-level fusion framework, extracts deep features from seismic and electromagnetic data through a dual-branch convolutional neural network, and introduces a cross-attention mechanism for dynamic weight fusion, realizing deep quantitative coupling of seismoelectric data. It effectively solves the dilemma of "seeing artifacts in seismic data and fuzziness in electromagnetic data" in traditional methods. It achieves complementary verification of seismoelectric physical fields through a data-driven approach, significantly reduces the ambiguity of single methods, and greatly improves the accuracy of identifying water-rich soft strata (such as faults, karst caves, and fracture zones).

[0037] 2. This invention utilizes the high spatial resolution of the seismic wave method to define precise physical boundaries for the low-resistivity anomaly provided by the transient electromagnetic method, by introducing a location loss term ( ) and physical constraint loss term ( This method forces the model to predict anomaly boundaries that are consistent with the seismic reflection interface. In practical engineering applications, this method can control the accuracy deviation of long-distance predictions of 30-120 meters to within 3 meters and the accuracy deviation of predictions of 0-30 meters to within 1 meter, meeting the needs of refined construction and risk avoidance.

[0038] 3. This invention breaks away from the subjective reliance on expert experience in traditional seismic and electrical joint interpretation. It uses a deep learning model to automatically complete feature extraction, fusion, and interpretation, reducing data processing speed from hours to minutes. This can meet the real-time forecasting needs of rapid tunnel excavation, while reducing the subjectivity of manual interpretation and improving the objectivity and consistency of forecast results.

[0039] 4. This invention uses multi-scale convolution kernels (3×3 and 5×5 in parallel) and deep residual network structure to simultaneously capture geological features at different scales. It has stronger sensitivity and recognition ability for weak anomalies such as small karst caves and narrow fracture zones, avoiding the risk of missed detection due to insufficient resolution in traditional methods.

[0040] 5. This invention innovatively incorporates a physical constraint term into the loss function. By calculating the cross-gradient constraint between seismic characteristic gradient and electromagnetic characteristic gradient, the model output is forced to conform to the basic laws of geophysics, effectively eliminating false predictions that contradict physical logic and significantly improving the geological reliability of the forecast results.

[0041] 6. This invention supports transfer learning. As real geological data revealed during tunnel excavation accumulates, the model can continuously optimize its performance through incremental learning, achieving adaptive dynamic improvement. At the same time, the technical framework of this invention has good scalability and can be easily extended to three-physics field fusion (such as adding ground-penetrating radar), constructing 3D convolutional networks, introducing Transformer structures to enhance long-distance dependency modeling, or constructing a real-time online prediction system. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall process of combined seismic and electrical advanced geological prediction based on Dual-CNN in Embodiment 1 of the present invention;

[0043] Figure 2 This is a detailed schematic diagram of the dual convolutional network and cross-attention fusion structure in Embodiment 1 of the present invention;

[0044] Figure 3 This is a schematic diagram of the dynamic training strategy of the composite loss function in Embodiment 1 of the present invention. Detailed Implementation

[0045] The following is in conjunction with the appendix Figures 1-3 The present invention will be described in further detail below.

[0046] Example 1: A combined seismic and electrical tunnel advanced geological prediction method based on dual convolutional neural networks, comprising the following steps:

[0047] Step S1: Multi-source data acquisition.

[0048] Seismic wave reflection data: Collect seismic wave reflection data in front of the tunnel face. The seismic data includes, but is not limited to, parameters such as seismic wave velocity, amplitude, and frequency.

[0049] Transient electromagnetic response data: Collect transient electromagnetic response data for the same area. The transient electromagnetic data includes information such as the transient response and depth distribution of the electromagnetic field.

[0050] Step S2: Data preprocessing and spatial grid alignment.

[0051] Because seismic data (time-amplitude sequence) and transient electromagnetic data (time-decaying voltage sequence) have different sampling rates and physical dimensions, the following processing is required in this embodiment:

[0052] Step S2.1: Seismic data processing;

[0053] Denoising (FK filtering, wavelet denoising);

[0054] Band filtering;

[0055] Offset imaging;

[0056] Amplitude normalization, Min-Max normalization, eliminate magnitude differences;

[0057] Convert to a two-dimensional spatial depth offset matrix;

[0058] The input tensor is formed as S(x, z), where x is the measurement point position and z is the depth; each record is a time series.

[0059] Step S2.2: Transient electromagnetic data processing;

[0060] Remove early noise;

[0061] Apparent resistivity calculation;

[0062] Resistivity profiles are obtained through inversion;

[0063] Min-Max normalization eliminates magnitude differences;

[0064] Time-frequency analysis was performed on transient electromagnetic data to extract key geological information features;

[0065] Convert to a two-dimensional space-depth resistivity matrix;

[0066] The input tensor is formed as: E(x, z), where x is the position of the measurement point and z is the depth.

[0067] Step S2.3, Spatial Coordinate System 1;

[0068] The apparent resistivity profile of the TEM and the seismic reflection interface are projected onto the same tunnel axial coordinate system (ZX plane).

[0069] Step S3: Construct a dual-branch convolutional neural network.

[0070] Construct a dual-branch convolutional neural network architecture, which contains two convolutional neural network modules, one for processing seismic data and the other for processing transient electromagnetic data.

[0071] Step S3.1, Seismic Stream:

[0072] A deep residual network (ResNet structure) is employed to capture the continuity, amplitude anomalies, wave velocity variation characteristics, and frequency attenuation characteristics of the reflection interface. Its convolutional kernels tend to extract "linear" features (constructing boundaries).

[0073] Output earthquake feature vector:

[0074]

[0075] To highlight the capture of the continuity of the reflection interface and the changes in wave velocity, the residual block (ResBlock) is... The feature transfer of a layer can be represented as:

[0076]

[0077] Where W represents the convolution weights, Represents the ReLU activation function. For the first Seismic characteristic map of the layer.

[0078] Step S3.2, Electromagnetic Branch (TEM-Stream):

[0079] Multi-scale convolutional kernels (3×3, 5×5) are used to capture the longitudinal evolution trend of apparent resistivity in front of the tunnel face (morphology of low-resistivity anomalies, changes in electrical gradients, and boundaries of conductive anomalies). The convolutional kernels tend to extract "clump-like" features (including water body areas).

[0080] Output electromagnetic eigenvectors:

[0081]

[0082] To capture the clumping features of low-resistivity anomalies at different scales, multi-scale convolutional kernels (3×3, 5×5) are used, mathematically represented as the concatenation of feature channels:

[0083]

[0084] This indicates concatenation along the channel dimension.

[0085] Step S4: Dynamic weight feature fusion mechanism.

[0086] In this embodiment, the feature data processed by their respective convolutional networks are fused. A weighted fusion strategy is adopted, which weights the features from the two data sources according to their information importance;

[0087] Multilayer perceptron (MLP) or attention mechanism modules can be used to further integrate features from seismic and electromagnetic data. Through the fused features, the model can learn more accurate geological prediction information.

[0088] The extracted fusion features include, but are not limited to, the structure, tectonic anomalies, fissures, and hydrogeological conditions of underground soil and rock;

[0089] The fusion layer is not simply an addition, but introduces a cross-attention mechanism:

[0090] Mutual verification: When the seismic branch detects a strong reflection interface and the electromagnetic branch simultaneously detects a low-resistivity anomaly, the weight of the "water-rich fault" in this area is amplified.

[0091] Masking filtering: If the earthquake shows that the rock mass is intact, but the electromagnetic field shows low resistance (which may be due to interference from metal pipelines), the system will reduce its false alarm rate by comparing spatial features.

[0092] The specific formula is as follows:

[0093] First, output the earthquake feature vector. and electromagnetic output eigenvector Perform channel splicing:

[0094]

[0095] Then, the spatial attention weight map is computed using a 1x1 convolution:

[0096]

[0097] Here, b is the bias term, which allows the model to be independent of the input data and comes with a basic "propensity" or "threshold", thereby greatly enhancing the network's ability to fit complex geological data.

[0098] Sigmoid is a non-linear activation function. Its core mathematical property is: its mathematical formula is... No matter how large or small the input number is (from negative infinity to positive infinity), it will "squeeze" the output between 0 and 1.

[0099] The output of Sigmoid is always between (0,1), so it perfectly acts as a "weight allocator" or "gating switch":

[0100] In this embodiment, if the output is close to 1 (e.g., 0.95), it's equivalent to the model saying, "Attention! The seismic reflection here is extremely strong and the resistivity drops sharply; it's 100% a water-rich fault. Amplify this feature!"

[0101] In this embodiment, if the output is close to 0 (e.g., 0.05), it is equivalent to the model saying, "Attention! Although there is a low impedance anomaly in the electromagnetic field, the seismic wave is completely undulating. This may be due to interference from the metal pipeline. Filter out this feature!"

[0102] Finally, the attention weights are multiplied element-wise with the original feature map (Hadamard product) to obtain the fused features:

[0103]

[0104] Step S5: Output layer design.

[0105] Classification pattern

[0106] Output categories include normal surrounding rock, fault fracture zone, water-rich area, karst, and weak interlayers, using Softmax output.

[0107]

[0108] Regression mode: Output: water abundance probability value, risk level index, and coordinates of anomaly location.

[0109] Step S6: Training the composite loss function.

[0110] In this embodiment, the loss function in deep learning acts as the model's "command stick," representing how much the model's predictions differ from reality. The goal of this loss function design is to compress the traditional 5-10m error into the engineering requirements (1-3m) you mentioned through AI training. Earthquake-electricity combined tunnel early warning is a complex task, and relying on a single loss function (such as only considering classification) is far from sufficient. This is similar to evaluating a forecasting engineer; one cannot only look at whether they have detected faults, but also at the accuracy of their location determination and whether their water content interpretation conforms to physical principles. Therefore, this embodiment defines a composite loss function L:

[0111]

[0112] The classification loss for geological types (faults, caves, intact rock masses) is usually achieved through cross-entropy loss. Its function is to force the model to accurately determine the geological structure ahead based on the reflection disturbances of seismic waves and the low-resistivity characteristics of electromagnetic waves, such as identifying a "fault," "intact limestone," or "isolated cave" 20-40m ahead.

[0113] Boundary localization and size deviation loss, typically using mean squared error (MSE) or IoU (intersection over union) loss. Its function is to limit the spatial deviation of the prediction. For example, if the actual width of a fault is 20m, but the model predicts a width of 10m or 30m, this loss term will impose a significant "penalty" on the model.

[0114] Physical constraint loss (ensuring predictions conform to fundamental geophysical laws, such as the diffusion of electromagnetic fields). Neural networks sometimes produce "anti-physics errors" to reduce the losses in the first two categories (for example, providing a physical combination of extremely high wave velocity but extremely low resistivity, which is almost impossible in actual geology).

[0115]

[0116] This formula represents the L2 norm of the cross product of the seismic feature gradient vector and the electromagnetic feature gradient vector in space. It forces that the gradient directions of the two in space be consistent or opposite, thereby eliminating false predictions that contradict physical logic.

[0117] Geological type weighting coefficient, with a value range of [0, 1].

[0118] Positioning and size weighting coefficient, with a value range of [0, 1];

[0119] Physical constraint weight coefficient, with a value range of [0, 1].

[0120] Weighting coefficient adjustment: Initial stage of the project (focusing on discovery): Increase the weighting coefficient. As long as we can identify the abnormal entity, that's fine.

[0121] During the meticulous construction phase (with an emphasis on risk avoidance): increase the... The boundary of the anomaly must be precisely locked. Generally, engineering requirements stipulate that the prediction accuracy deviation for a length of 30-120m should be within 3m, and the accuracy deviation for a length of 0-30m should be within 1m.

[0122] Complex strata (focusing on identifying and verifying authenticity): Increase This is to prevent false electromagnetic anomalies caused by underground metal pipelines or construction interference.

[0123] Step S7: Model training and optimization.

[0124] The dual convolutional neural network model was trained using geological data revealed by actual tunnel excavation. The model's weight parameters were optimized using the backpropagation algorithm to minimize prediction error.

[0125] K-fold cross-validation was used to evaluate the model to ensure its generalization ability and accuracy.

[0126] Optimize using appropriate loss functions (such as mean squared error, cross-entropy, etc.) according to different task requirements;

[0127] Overfitting prevention techniques, such as Dropout or L2 regularization, are employed to improve the stability and reliability of the model.

[0128] Step S8: Forecast Result Generation and Geological Analysis

[0129] The trained model was applied to actual tunnel geological prediction. The geological environment ahead of the tunnel was analyzed based on the model's prediction results.

[0130] Based on the prediction results, a structural map of the underground geology is generated, and sections with different risks are divided to help tunnel engineers grasp the geological conditions of the construction area in real time and formulate reasonable construction plans.

[0131] Further post-processing of the prediction results, such as spatial interpolation and 3D visualization, can improve operability and understandability.

[0132] Example 2: Specific application for a 20-meter normal fault.

[0133] This embodiment takes the identification of a normal fault 20 meters in front of the tunnel face as an example to provide a detailed explanation of the method in Embodiment 1.

[0134] 1. Network design considerations for a 20m normal fault

[0135] Receptive field matching: 20 meters accounts for about 1 / 6 of the prediction range of 100-120 meters. The size and stride of the convolution kernel need to ensure that the deep feature layer can cover a range of at least 30-40 meters so that the model can "see" the contact relationship between the upper and lower plates of the fault.

[0136] Multi-scale extraction: Normal faults are usually accompanied by secondary fractures. Seismic branches require high-frequency convolution to capture the interface, while electromagnetic branches require wide convolution to capture the overall low-resistivity trend.

[0137] 2. Detailed Dual-CNN Architecture Design

[0138] Seismic feature extraction branch (Seismic-Stream)

[0139] Input: Preprocessed 2D offset image or set of common reflection points (128x128 dimensions);

[0140] Layer design: 5 layers of residual blocks;

[0141] Layer 1-2: Use 3×3 convolution to extract edge features of the fracture surface (sloping straight line);

[0142] Layer 3-5: Introducing dilated convolution to increase the receptive field and capture waveform disorder areas caused by fault fracture zones.

[0143] Electromagnetic Feature Extraction Branch (TEM-Stream)

[0144] Input: Apparent resistivity profile or multichannel attenuation curve (128×128 dimensions).

[0145] Layer design: 4 layers of convolution + global pooling.

[0146] Layer 1-2: Use a 5×5 large convolution kernel to smooth out high-frequency noise and extract the spatial distribution of the low-resistivity core region.

[0147] Layer 3-4: 1×1 convolution, used for feature compression between channels to extract the nonlinear trend of resistivity evolution.

[0148] Cross-Attention Fusion: For a 20m normal fault, the attention mechanism will concentrate the weights on the boundary region where seismic reflection is strong and resistivity drops sharply.

[0149] 3. Core Code Implementation

[0150] In this embodiment, the core PyTorch logic code is implemented as follows:

[0151] The key points of feature interpretation in this embodiment are as follows:

[0152] Fault dip identification logic: The seismic branch can identify the waveform travel time difference caused by the dip angle of the normal fault plane through a multi-scale residual structure and convert it into a feature vector; at the same time, the electromagnetic branch can achieve joint determination of fault attitude by synchronously identifying the low-resistivity filling material (such as water and wet mud) below the dip interface.

[0153] Width quantification logic: The fusion model calculates the overlap between the axial length of the seismic wave amplitude disturbance zone and the electromagnetic gradient change zone, and quantitatively outputs the width of the fault fracture zone (such as the 20m anomaly zone in this embodiment) through fully connected layer mapping.

[0154] (Distance constraint in the loss function: Structural similarity (SSIM) constraint is introduced into the loss function, which forces that the predicted water-rich area boundary must fall within the fault boundary identified by the earthquake, and the error threshold is set to 10% of the fault width.)

[0155] 4. Application of the composite loss function

[0156] In this embodiment, the significance of the composite loss function for a 20m fault is as follows:

[0157] (Classification loss): Enables the model to generate strong neural responses to the unique attitude (dipping angle) and fracture zone features of normal faults during the feature extraction stage.

[0158] (Boundary location and size deviation loss): The high spatial resolution of seismic waves is used to constrain the "diffuseness" of the electromagnetic method. Without this, the low-resistivity "clumps" of the aquifer predicted by the electromagnetic method may extend far beyond the fault boundary. Lpos can trim these "clumps" to a precise range of 20m that conforms to the seismic reflection interface.

[0159] (Physical constraint loss): If seismic waves show extreme fracturing at this location (wave velocity decrease), then the resistivity will inevitably show a corresponding anomalous trend (decreased water content, increased air content). Simplified constraints from the electromagnetic field diffusion equation or the seismic wave equation are introduced to ensure that the AI-generated "prediction map" conforms to geological logic, rather than being a random patchwork of images.

[0160] The specific application steps of this embodiment include:

[0161] Preparing the "Ground Truth" (GT): Before applying the loss function, the geological expert's experience must be transformed into "labels" that the machine can understand.

[0162] Category tags (y cls ): A one-hot code, for example [1, 0, 0] represents a fault.

[0163] Location tag (y) pos ): A mask the same size as the input image, with pixel values ​​of 1 in the 20m region where the fault exists, and 0 elsewhere.

[0164] Physical tag (y phy ): Here, "labels" are usually prior knowledge, such as the known background values ​​of normal wave velocity and resistivity of the surrounding rock in the area.

[0165] The specific calculations for each loss item are as follows:

[0166] Category Items To address the "qualitative" issue, Cross Entropy Loss is typically used in PyTorch.

[0167] If the model misclassifies a "normal fault" as a "complete rock mass", the gradient generated by this term will be very large, forcing the model weights to be adjusted in the direction that can extract the features of "interface reflection" and "low-resistivity core".

[0168] Location item To address "quantitative" (accuracy over a 20m width): use Dice Loss (for overlap) or Smooth L1 Loss (for boundary coordinates).

[0169] The model predicts a 20m anomaly boundary. The intersection-union ratio (LOU) between the predicted and actual regions is calculated.

[0170] For normal faults, the seismic branch is very accurate to the boundary (dipping surface). This will force the “fuzzy low-resistivity clusters” generated by the electromagnetic branch to adapt to the “clear boundaries” generated by the seismic branch.

[0171] Physical constraints Solving the "consistency" problem:

[0172] This is the core application, which is usually achieved by calculating the Pearson correlation coefficient or cross-gradient constraints.

[0173] Mathematical Logic: That is, the gradient of seismic wave velocity change and the gradient of resistivity change should be in the same or opposite direction in space.

[0174] If the model predicts that there is a fault (seismic wave disturbance) and the geological background is aquifer, but the resistivity shows high resistivity (dry and dense rock). This will cause the value to spike, telling the model, "This doesn't make geological sense, please retrain!"

[0175] 5. Multi-stage optimization strategies during training

[0176] The dynamic application logic in this embodiment during the training process: the composite loss function is not static, and a multi-stage optimization strategy is usually adopted.

[0177] Phase 1 (Warm-up Period): Setting up , , .

[0178] Objective: To first enable the model to distinguish between faults and rocks, and to establish basic feature extraction capabilities.

[0179] Phase Two (Refinement Period): Gradually increase β, for example... , , .

[0180] Objective: To focus on training the model's sensitivity to the width of "20 meters" and to use the high spatial resolution of seismic data to constrain positioning.

[0181] Phase Three: (Physical Calibration Period): Incorporating physical constraints. , , .

[0182] Objective: To eliminate "false predictions" that, while conforming to image characteristics, do not conform to geophysical laws.

[0183] The pseudocode logic implementation in this embodiment is as follows:

[0184]

[0185] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for advanced geological prediction of tunnels based on a dual convolutional neural network and combined seismic and electrical signals, characterized in that: Includes the following steps: S1. Data Acquisition: Acquire seismic wave reflection data and transient electromagnetic response data in front of the tunnel face; S2. Data Preprocessing and Spatial Grid Alignment: The seismic wave reflection data and transient electromagnetic response data are preprocessed separately, and both are projected onto the same tunnel axial coordinate system to establish a spatial grid model, forming a spatially aligned seismic input tensor. and electromagnetic input tensor ;in: For the location of the measuring point, For depth; S3. Dual-branch feature extraction: Construct parallel first and second convolutional neural network branches to extract the seismic input tensor. The first convolutional neural network branch is input to extract earthquake feature vectors. The electromagnetic input tensor The electromagnetic feature vector is extracted by inputting the second convolutional neural network branch. ; S4. Heterogeneous Feature Fusion: The seismic feature vectors are fused... With the electromagnetic characteristic vector After channel stitching, a spatial attention weight map is calculated using an attention mechanism. The attention weight map is then multiplied element-wise with the concatenated feature map to obtain the fused feature. ; S5. Output prediction results: The fused features... Input a classifier or regressor, and output the prediction results of the geological bodies in front of the tunnel face; S6. Training with a composite loss function: Constructing a function that includes classification loss. Location loss and physical constraint loss Composite loss function The first convolutional neural network branch and the second convolutional neural network branch are jointly trained using this composite loss function; wherein: This is the geological type weighting coefficient, with a value range of [0,1]. Positioning and size weighting coefficients, with a value range of [0,1]; This represents the physical constraint weighting coefficient, with a value range of [0,1]. The physical constraint loss This is achieved by calculating the cross-gradient constraint between seismic characteristic gradients and electromagnetic characteristic gradients. The calculation formula is as follows: 。 2. The method for advanced geological prediction of tunnels based on combined seismic and electrical signals according to claim 1, characterized in that: The preprocessing of seismic wave reflection data in step S2 specifically includes: denoising, band filtering, migration imaging, amplitude normalization, and conversion into a two-dimensional spatial-depth migration matrix; the preprocessing of transient electromagnetic response data specifically includes: removing early noise, calculating apparent resistivity, inverting to obtain resistivity profiles, normalization, and conversion into a two-dimensional spatial-depth resistivity matrix.

3. The method for advanced geological prediction of tunnels based on combined seismic and electrical signals according to claim 1, characterized in that: The first convolutional neural network branch adopts a deep residual network structure to capture the continuity, amplitude anomaly, and wave velocity variation characteristics of the reflection interface; its first The feature transfer formula for a layer is: in: Represents the convolution weights. Represents the ReLU activation function. For the first Seismic characteristic map of the layer.

4. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The second convolutional neural network branch uses multi-scale convolutional kernels, and uses 3×3 convolutional kernels and 5×5 convolutional kernels in parallel to capture the longitudinal evolution trend of apparent resistivity and the morphological characteristics of low-resistivity anomalies. its first The feature transfer formula for a layer is: in: This indicates splicing along the channel dimension.

5. The method for advanced geological prediction of tunnels based on combined seismic and electrical signals according to claim 1, characterized in that: The attention mechanism in step S4 specifically includes: calculating a spatial attention weight map using a 1×1 convolutional layer and a Sigmoid activation function. The calculation formula is as follows: in: Earthquake eigenvectors With electromagnetic eigenvectors The channel splicing result, where b is the bias term; then the attention weight map is... With the stitched feature map Element-wise multiplication is performed to obtain the fusion feature. .

6. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The prediction results in step S5 include: geological body classification results and / or regression prediction results; the geological body classification results include normal surrounding rock, fault fracture zone, water-rich area, karst and weak interlayer; the regression prediction results include water-rich probability value, risk level index and anomaly location coordinates.

7. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The classification loss The cross-entropy loss function is used; the localization loss The mean squared error loss function or the IoU loss function is used; the weighting coefficients , , The value range is [0, 1], and it can be dynamically adjusted according to the engineering stage and geological conditions.

8. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The training process in step S6 employs a multi-stage optimization strategy: the first stage is set... , , Basic classification feature extraction training is performed; The second phase of adjustment The third stage involves training spatial positioning accuracy; physical constraints are then introduced and set... Physical consistency calibration training was conducted.

9. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The prediction results output in step S5 also include a three-dimensional geological risk probability map within a range of 100-120 meters in front of the tunnel face; the method also includes a post-processing step for the prediction results, which includes spatial interpolation and three-dimensional visualization.

10. The method for advanced geological prediction of tunnels based on dual convolutional neural networks according to claim 1, characterized in that: The method also includes incrementally optimizing the trained model through transfer learning as real geological data is revealed during tunnel excavation.

Images