A deep learning tunnel advance prediction method based on transient electromagnetic images

Abstract

The present invention proposes a deep learning tunnel advance prediction method based on transient electromagnetic images, comprising: inputting the TEM detection image to be predicted into a segmentation model to obtain continuous semantic segmentation result data; wherein the segmentation model is obtained by training the Swin‑Unet network based on a training set, and the training set includes: randomly selected TEM detection images and annotated feature label data; three-dimensional reconstruction and image data analysis are performed on the semantic segmentation result data respectively to obtain a three-dimensional reconstructed image and a TEM synthetic image; based on the three-dimensional reconstructed image and the TEM synthetic image, the drilling positions and the amount of explosives used are distributed. The present invention uses three-dimensional images to provide more accurate three-dimensional information for tunnel drilling and blasting, and compresses the three-dimensional information into two-dimensional information on the face image, without the need for manual calculation and analysis, which helps to improve the drilling efficiency and blasting efficiency. The present invention saves a lot of labor costs and has a good application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a deep learning tunnel advance prediction method based on transient electromagnetic images. Background Art

[0002] In recent years, as the terrain and geological conditions faced by engineering construction have become increasingly complex, the requirements for the accuracy of geophysical exploration have become increasingly higher. Due to the complex underlying data, it is sometimes difficult to ascertain the accurate spatial information of boulders, fault zones, and underground rivers in front of the tunnel face during tunnel blasting and excavation, resulting in unreasonable layout and use of explosives, which ultimately leads to major geological disasters, causing casualties, economic losses, construction stagnation, and adverse social impacts.

[0003] How to more accurately obtain geological information in front of the tunnel face during blasting is crucial to the location and amount of explosives, and transient electromagnetic detection technology (TEM) can effectively provide accurate geological information in front of the tunnel face. However, tunnel face information is based on continuous two-dimensional image slices, which requires professional manual analysis and processing to accurately control the holes and explosives used in tunnel face blasting, which is time-consuming and labor-intensive.

[0004] TEM (Transient Electromagnetic Method) is an active source electromagnetic exploration method that emits a pulsed electromagnetic field in front of a plane in a vertical direction to induce eddy currents in the conductor in the front area, and measures and analyzes the time decay characteristics of the secondary magnetic field generated during the eddy current decay process, thereby inferring the spatial distribution of the conductive structure in front of a plane in a vertical direction. Through TEM scanning imaging, the spatial distribution in front of the tunnel face can be obtained, including boulders, fault zones, underground rivers, etc., providing important data support for face drilling and blasting.

[0005] Since the images obtained by TEM are based on two-dimensional image slices, it is necessary to manually observe each image slice for analysis, and it is impossible to intuitively and quickly obtain accurate three-dimensional spatial information, which limits the accuracy of face drilling and explosive quantity control. Therefore, it is urgent to propose a deep learning tunnel advance prediction method based on transient electromagnetic images. Summary of the invention

[0006] The purpose of the present invention is to propose a deep learning tunnel advance prediction method based on transient electromagnetic images, which mainly involves deep learning segmentation of continuous image sets obtained by TEM detection technology, and provides the front three-dimensional structure distribution information for tunnel face blasting based on the analysis data of the segmentation results.

[0007] To achieve the above object, the present invention provides a deep learning tunnel advance prediction method based on transient electromagnetic images, comprising:

[0008] Inputting the TEM detection image to be predicted into the segmentation model to obtain continuous semantic segmentation result data; wherein the segmentation model is obtained by training the Swin-Unet network based on a training set, and the training set includes: randomly selected TEM detection images and annotated feature label data;

[0009] Respectively performing three-dimensional reconstruction and image data analysis on the semantic segmentation result data to obtain a three-dimensional reconstructed image and a TEM synthetic image;

[0010] Based on the three-dimensional reconstructed image and the TEM synthetic image, the drilling positions and the amount of explosive used are distributed.

[0011] Optionally, generating the training set includes:

[0012] Based on TEM detection imaging, continuous TEM images are obtained;

[0013] Randomly selecting the TEM images to form a TEM image set;

[0014] Performing feature extraction on the TEM image set to obtain a feature label set;

[0015] The training set is generated based on the TEM image set and the feature label set.

[0016] Optionally, extracting features from the TEM image set includes:

[0017] Segmenting the features in the TEM image set using a segmentation algorithm; wherein the features include: boulders, fault zones, and underground rivers;

[0018] Smear and patch the segmented features to generate a set of labeled images;

[0019] Converting the images in the labeled image set into grayscale images;

[0020] Extracting a preset grayscale value in the grayscale image;

[0021] Use OpenCV region growing denoising algorithm to filter out non-feature noise points outside the target features;

[0022] The OpenCV erosion and expansion algorithm is used to smooth the feature edges and arrange the grayscale values ​​of different features in sequence starting from 1.

[0023] Optionally, the Swin-Unet network includes: an encoder, a bottleneck module, a decoder and a skip connection unit;

[0024] The encoder is used to gradually downsample the input image into a feature map;

[0025] The bottleneck module is used to enhance the expressiveness of the feature map output by the encoder;

[0026] The decoder is used to upsample the enhanced feature map into a prediction map of the same scale as the input image;

[0027] The skip connection unit is used to connect the output of the encoder and the input of the decoder in the decoder.

[0028] Optionally, performing image data analysis on the semantic segmentation result data includes:

[0029] Analyze the semantic segmentation result data by an image data analysis algorithm to obtain preset information in the image; wherein the preset information includes: feature position data, feature volume data and feature three-dimensional direction information;

[0030] The preset information is projected onto a TEM two-dimensional tunnel face image to obtain the TEM synthetic image.

[0031] Optionally, acquiring the preset information in the image includes:

[0032] Calculate the position of the feature in front of the palm face according to the resolution between images in the continuous semantic segmentation result data, that is, the feature position; wherein the feature position includes: the distance when the feature appears in the image and the distance when it disappears;

[0033] According to the number of pixels in the region where the features exist on the continuous semantic segmentation result data, multiply it by the square of the resolution between pixels to obtain the area of ​​the feature region on a single image, and then multiply it by the resolution between images to obtain the volume of the feature region between single images, and finally obtain the feature volume data by accumulating the feature volumes of multiple images;

[0034] By analyzing the three-dimensional minimum circumscribed cuboid of the area where the feature exists in the semantic segmentation result data, the longest side is calculated to obtain the three-dimensional direction information of the feature.

[0035] Optionally, projecting the preset information onto the TEM two-dimensional tunnel face image includes:

[0036] The three-dimensional spatial volume data of the preset information is projected onto the x, y plane, i.e., the TEM two-dimensional tunnel face, with the z axis as the projection direction, and a circumscribed rectangle is drawn.

[0037] Drawing the feature position data and feature volume data on the edge of the circumscribed rectangle of the corresponding feature;

[0038] The three-dimensional trend information of the feature is projected onto the x- and y-plane angles θ and the angle α with the z-axis, wherein θ represents the angle between the longest side of the smallest circumscribed cuboid of the feature and the x-axis, and α represents the angle between the longest side of the smallest circumscribed cuboid of the feature and the z-axis. θ and α are used to represent the relative trend information of the feature based on the palm face, and θ and α are drawn on the edges of the circumscribed rectangle of the corresponding feature.

[0039] Optionally, distributing the drilling positions and the amount of explosives used based on the three-dimensional reconstructed image and the TEM synthetic image includes:

[0040] Based on the three-dimensional reconstructed image and the TEM synthetic image, adding or reducing blasting holes for different density areas;

[0041] The amount of explosives used in each blasting hole is arranged according to the characteristic position data, characteristic volume data and characteristic three-dimensional trend information in the TEM synthetic image.

[0042] The present invention has the following beneficial effects:

[0043] The present invention uses three-dimensional images to provide more accurate three-dimensional information for tunnel drilling and blasting, and compresses the three-dimensional information into two-dimensional information on the tunnel face image, without the need for manual calculation and analysis, which helps to improve the drilling efficiency and blasting efficiency. The present invention saves a lot of labor costs and has a good application prospect. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0045] Figure 1 A schematic diagram of a process flow of a deep learning tunnel advance prediction method based on transient electromagnetic images according to an embodiment of the present invention;

[0046] Figure 2 Schematic diagram of the data processing stage of an embodiment of the present invention; wherein (a) is a schematic diagram of a three-dimensional model of tunnel blasting excavation, (b) is a schematic diagram of TEM detection imaging, and (c) is a schematic diagram of deep learning segmentation results;

[0047] Figure 3 Schematic diagrams showing the results of an embodiment of the present invention; wherein (a) is a schematic diagram of a three-dimensional reconstruction model of semantic segmentation results based on deep learning image segmentation, (b) is a two-dimensional polar coordinate projection method of a three-dimensional orientation, and (c) is a schematic diagram of the effect of a TEM synthetic image;

[0048] Figure 4 Schematic diagram of hole distribution according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0051] This embodiment proposes a deep learning tunnel advance prediction method based on transient electromagnetic images, including:

[0052] The TEM detection image to be predicted is input into the segmentation model to obtain continuous semantic segmentation result data; wherein the segmentation model is obtained by training the Swin-Unet network based on the training set, and the training set includes: randomly selected TEM detection images and annotated feature label data;

[0053] Perform three-dimensional reconstruction and image data analysis on the semantic segmentation result data respectively to obtain a three-dimensional reconstructed image and a TEM synthetic image;

[0054] Based on the 3D reconstruction images and TEM synthetic images, the drilling locations and the amount of explosives used were distributed.

[0055] Furthermore, generating a training set includes:

[0056] Based on TEM detection imaging, continuous TEM images are obtained;

[0057] TEM images are randomly selected to form a TEM image set;

[0058] Perform feature extraction on the TEM image set to obtain a feature label set;

[0059] A training set is generated based on the TEM image set and the feature label set.

[0060] Furthermore, feature extraction of the TEM image set includes:

[0061] The segmentation algorithm is used to segment the features in the TEM image set; the features include: boulders, fault zones, and underground rivers;

[0062] Smear and patch the segmented features to generate a set of labeled images;

[0063] Convert images in the labeled image set to grayscale;

[0064] Extract the preset grayscale value in the grayscale image;

[0065] Use OpenCV region growing denoising algorithm to filter out non-feature noise points outside the target features;

[0066] The OpenCV erosion and expansion algorithm is used to smooth the feature edges and arrange the grayscale values ​​of different features in sequence starting from 1.

[0067] Furthermore, the Swin-Unet network includes: an encoder, a bottleneck module, a decoder, and a skip connection unit;

[0068] The encoder is used to gradually downsample the input image into feature maps;

[0069] The bottleneck module is used to enhance the expressiveness of the feature map output by the encoder;

[0070] The decoder is used to upsample the enhanced feature map to a prediction map of the same scale as the input image;

[0071] Skip connection units are used in the decoder to connect the encoder output and the decoder input.

[0072] Furthermore, performing image data analysis on the semantic segmentation result data includes:

[0073] Analyze the semantic segmentation result data through the image data analysis algorithm to obtain the preset information in the image; the preset information includes: feature position data, feature volume data and feature three-dimensional direction information;

[0074] The preset information is projected onto the TEM two-dimensional tunnel face image to obtain a TEM synthetic image.

[0075] Furthermore, obtaining preset information in the image includes:

[0076] The position of the feature in front of the palm face, i.e., the feature position, is calculated according to the resolution between images in the continuous semantic segmentation result data; wherein the feature position includes: the distance when the feature appears in the image and the distance when it disappears;

[0077] According to the number of pixels in the region where the features exist on the continuous semantic segmentation result data, multiply it by the square of the resolution between pixels to get the area of ​​the feature region on a single image, and then multiply it by the resolution between images to get the volume of the feature region between single images. By accumulating the feature volumes of multiple images, the feature volume data is finally obtained;

[0078] By analyzing the three-dimensional minimum circumscribed cuboid of the area where the features exist in the semantic segmentation result data, the longest side is found and the three-dimensional direction information of the features is obtained.

[0079] Furthermore, projecting the preset information onto the TEM two-dimensional tunnel face image includes:

[0080] The three-dimensional spatial volume data of the preset information is projected onto the x, y plane, i.e., the TEM two-dimensional face, with the z axis as the projection direction, and a circumscribed rectangle is drawn.

[0081] The feature position data and the feature volume data are drawn on the edge of the circumscribed rectangle of the corresponding feature;

[0082] The three-dimensional trend information of the feature is projected onto the x- and y-plane angles θ and α with the z-axis, where θ represents the angle between the longest side of the feature’s smallest circumscribed cuboid and the x-axis, and α represents the angle between the longest side of the feature’s smallest circumscribed cuboid and the z-axis. θ and α are used to represent the relative trend information of the feature based on the face, and θ and α are drawn on the edges of the circumscribed rectangle of the corresponding feature.

[0083] Furthermore, based on the 3D reconstruction image and the TEM synthetic image, the distribution of the drilling positions and the amount of explosives used includes:

[0084] Based on 3D reconstruction images and TEM synthetic images, blasting holes are added or reduced for different density areas;

[0085] The amount of explosives to be used in each blasting hole is arranged according to the characteristic position data, characteristic volume data and characteristic three-dimensional trend information in the TEM synthetic image.

[0086] In this embodiment, if Figure 1-Figure 4 As shown, a deep learning tunnel advance prediction method based on transient electromagnetic images specifically includes the following steps:

[0087] Step 1: Use manual annotation method to perform algorithm segmentation and manual assisted segmentation on randomly selected TEM detection images, construct image sets, mark image sets, and extract label sets through feature extraction algorithm;

[0088] Step 2: The image set and the label set are combined into a training set and input into the Swin-Unet network for training to obtain the network weights. The network weights are used to test the continuous TEM images to obtain the semantic segmentation results.

[0089] Step 3: Perform three-dimensional reconstruction on the continuous semantic segmentation results and perform image data analysis, including the location, volume, and direction of features in the image. This information is annotated in the TEM face image to obtain a TEM synthetic image.

[0090] Step 4: Based on the TEM synthetic image and the 3D reconstructed image in step 3, the drilling positions and the amount of explosives used are reasonably distributed.

[0091] In this embodiment, step 1 specifically includes:

[0092] Step 1.1, obtaining a continuous set of TEM images through TEM detection imaging;

[0093] Step 1.2: randomly select TEM images to form an image set, manually mark the image set, paint the same feature area with a uniform grayscale value to obtain a marked image set, and extract the label set through a feature extraction algorithm.

[0094] In this embodiment, step 1.2 specifically includes:

[0095] Step 1.2.1, use the brush tool to segment the features (boulders, fault zones, underground rivers, etc.) in the image set using the algorithm and manually use different grayscale brushes to smear and repair to form a labeled image set;

[0096] Step 1.2.2, convert the real labeled image into a grayscale image;

[0097] Step 1.2.3, extract specific grayscale values, filter other grayscale values ​​and set the grayscale value to 0;

[0098] Step 1.2.4: Since some areas that do not belong to the smeared area also have specific grayscale values, it is necessary to use the OpenCV region growing denoising algorithm to filter out the smaller noise points in the pixel connected area;

[0099] Step 1.2.5: Use OpenCV corrosion and expansion algorithms to smooth the feature edges and arrange the grayscale values ​​of different features in sequence starting from 1 (required by Swin-Unet semantic segmentation network).

[0100] In this embodiment, step 2 specifically includes:

[0101] The image set and the label set form a training set and are input into the Swin-UNet network for training. After the training is completed, the TEM continuous images are tested to obtain continuous semantic segmentation result images.

[0102] The architecture of Swin-UNet consists of an encoder, a bottleneck, and a decoder, where both the encoder and the decoder are constructed based on the Swin Transformer block. The encoder is responsible for gradually downsampling the input image into a feature map to capture semantic information of different scales; the bottleneck module further processes the output of the encoder to enhance the expressiveness of the features; the decoder upsamples the processed feature map to a prediction map of the same scale as the input image. In addition, skip connections are used to connect the output of the encoder and the input of the decoder in the decoder to help retain more high-level semantic information and improve the accuracy of the prediction. In summary, Swin-UNet combines the advantages of Swin Transformer with the structure of UNet, and can effectively handle image semantic segmentation tasks.

[0103] In this embodiment, step 3 specifically includes:

[0104] Step 3.1: Input the continuous semantic segmentation result images into three-dimensional imaging software such as Avizo and 3D Slicer for three-dimensional reconstruction to obtain the TEM three-dimensional deep learning segmentation result images.

[0105] Step 3.2: Analyze the continuous semantic segmentation result image through an image data analysis algorithm to obtain information such as feature position, volume, and three-dimensional direction in the image.

[0106] Step 3.3: Project the obtained three-dimensional information onto the TEM two-dimensional tunnel face image to obtain a TEM composite image.

[0107] In this embodiment, step 3.2 specifically includes:

[0108] Step 3.2.1. Based on the range of features on the continuous slice images, such as boulders, calculate the position of the boulder in front of the palm face based on the resolution (spacing) between slices, including the distance when it appears and the distance when it disappears.

[0109] Step 3.2.2: Based on the area where the features exist on the continuous slice images, such as the number of pixels of the boulder on each slice, the volume data of the boulder can be obtained by multiplying the image resolution and the resolution between slices. The other features are similar.

[0110] Step 3.2.3: Based on the areas where features exist on the continuous slice images, such as underground rivers, the three-dimensional direction of the underground river can be obtained by analyzing its three-dimensional minimum circumscribed cuboid and finding the longest side. Other features are similar.

[0111] In this embodiment, step 3.3 specifically includes:

[0112] Step 3.3.1, project the three-dimensional data of multiple features onto the x, y plane with the z axis as the baseline, and draw a circumscribed rectangle; draw the distance data and volume data obtained in steps 3.2.1 and 3.2.2 on the edges of the circumscribed rectangles of the corresponding features.

[0113] Step 3.3.2: Project the three-dimensional direction information of the feature obtained in step 3.2.3 to the x- and y-plane angles θ and the angle α with the z-axis, and draw the direction information θ and α on the edges of the corresponding feature circumscribed rectangle.

[0114] In this embodiment, step 4 specifically includes:

[0115] Step 4.1: Based on the TEM composite image and three-dimensional reconstructed image obtained in step 3, increase blasting holes in the high-density area, reduce blasting holes in the low-density area, arrange them reasonably, and improve blasting efficiency.

[0116] Step 4.2: According to the spatial distribution information, volume information, and direction information of each feature in the TEM synthetic image, the amount of explosives used in each blasting hole is reasonably arranged, and then the drilling and blasting work is completed.

[0117] Figure 2 Schematic diagram of the data processing stage of an embodiment of the present invention; wherein, Figure 2 (a) is a three-dimensional model diagram of tunnel blasting excavation. Figure 2 (b) is a schematic diagram of TEM detection imaging. Figure 2 (c) is a schematic diagram of deep learning segmentation results; Figure 3 The result of the embodiment of the present invention is presented in a schematic diagram; wherein, Figure 3 (a) is a schematic diagram of the 3D reconstruction model based on the semantic segmentation result of deep learning image segmentation. Figure 3 (b) is the two-dimensional polar coordinate projection method of the three-dimensional direction. Figure 3 (c) is a schematic diagram of the effect of TEM composite image; Figure 4 Schematic diagram of hole distribution according to an embodiment of the present invention.

[0118] This embodiment provides a deep learning tunnel advance prediction method based on transient electromagnetic images. The image in front of the relevant tunnel face is obtained by transient electromagnetic detection technology (TEM), the original data set is obtained, and the features in the image are manually marked to generate a marked image set, and the corresponding labels are extracted by a feature extraction algorithm; then the marked image is input into the Swin-Unet network for training; then the continuous TEM images are input into the trained Swin-Unet network to obtain the segmentation results and reconstruct them in three dimensions; finally, the image data is analyzed by the algorithm, including the spatial distribution, volume, direction and other information of the features in the image, and these information are marked in the TEM original face image to obtain a TEM synthetic face image, and the hole distribution and explosive dosage are arranged according to the information of the three-dimensional model in front of the face and the TEM synthetic image to complete the drilling and blasting work. This embodiment uses three-dimensional images to provide more accurate three-dimensional information for tunnel drilling and blasting, and compresses the three-dimensional information into two-dimensional information on the face image, without the need for manual calculation and analysis, which helps to improve the drilling efficiency and blasting efficiency. This method saves a lot of labor costs and has a good application prospect.

[0119] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A deep learning tunnel advance prediction method based on transient electromagnetic images, characterized in that: include: Inputting the TEM detection image to be predicted into the segmentation model to obtain continuous semantic segmentation result data; wherein the segmentation model is obtained by training the Swin-Unet network based on a training set, and the training set includes: randomly selected TEM detection images and annotated feature label data; Respectively performing three-dimensional reconstruction and image data analysis on the semantic segmentation result data to obtain a three-dimensional reconstructed image and a TEM synthetic image; Based on the three-dimensional reconstructed image and the TEM synthetic image, distribute the drilling positions and the amount of explosives used; Generating the training set includes: Based on TEM detection imaging, continuous TEM images are obtained; Randomly selecting the TEM images to form a TEM image set; Performing feature extraction on the TEM image set to obtain a feature label set; Based on the TEM image set and the feature label set, generating the training set; Extracting features from the TEM image set includes: Segmenting the features in the TEM image set using a segmentation algorithm; wherein the features include: boulders, fault zones, and underground rivers; Smear and patch the segmented features to generate a set of labeled images; Converting the images in the labeled image set into grayscale images; Extracting a preset grayscale value in the grayscale image; Use OpenCV region growing denoising algorithm to filter out non-feature noise points outside the target features; The OpenCV erosion and expansion algorithm is used to smooth the feature edges and arrange the grayscale values ​​of different features in sequence starting from 1.

2. The deep learning tunnel advance prediction method based on transient electromagnetic images according to claim 1 is characterized in that: The Swin-Unet network includes: an encoder, a bottleneck module, a decoder and a skip connection unit; The encoder is used to gradually downsample the input image into a feature map; The bottleneck module is used to enhance the expressiveness of the feature map output by the encoder; The decoder is used to upsample the enhanced feature map into a prediction map of the same scale as the input image; The skip connection unit is used to connect the output of the encoder and the input of the decoder in the decoder.

3. The deep learning tunnel advance prediction method based on transient electromagnetic images according to claim 1 is characterized in that: Performing image data analysis on the semantic segmentation result data includes: Analyze the semantic segmentation result data by an image data analysis algorithm to obtain preset information in the image; wherein the preset information includes: feature position data, feature volume data and feature three-dimensional direction information; The preset information is projected onto a TEM two-dimensional tunnel face image to obtain the TEM synthetic image.

4. The deep learning tunnel advance prediction method based on transient electromagnetic images according to claim 3 is characterized in that: Acquiring the preset information in the image includes: Calculate the position of the feature in front of the palm face according to the resolution between images in the continuous semantic segmentation result data, that is, the feature position; wherein the feature position includes: the distance when the feature appears in the image and the distance when it disappears; According to the number of pixels in the region where the features exist on the continuous semantic segmentation result data, multiply it by the square of the resolution between pixels to obtain the area of ​​the feature region on a single image, and then multiply it by the resolution between images to obtain the volume of the feature region between single images, and finally obtain the feature volume data by accumulating the feature volumes of multiple images; By analyzing the three-dimensional minimum circumscribed cuboid of the area where the feature exists in the semantic segmentation result data, the longest side is calculated to obtain the three-dimensional direction information of the feature.

5. The deep learning tunnel advance prediction method based on transient electromagnetic images according to claim 3 is characterized in that: Projecting the preset information onto the TEM two-dimensional tunnel face image includes: Projecting the three-dimensional spatial volume data of the preset information onto the x, y plane, i.e., the TEM two-dimensional tunnel face, with the z axis as the projection direction, and drawing a circumscribed rectangle; Drawing the feature position data and feature volume data on the edge of the circumscribed rectangle of the corresponding feature; The three-dimensional trend information of the feature is projected onto the x- and y-plane angles θ and the angle α with the z-axis, wherein θ represents the angle between the longest side of the smallest circumscribed cuboid of the feature and the x-axis, and α represents the angle between the longest side of the smallest circumscribed cuboid of the feature and the z-axis. θ and α are used to represent the relative trend information of the feature based on the palm face, and θ and α are drawn on the edges of the circumscribed rectangle of the corresponding feature.

6. The deep learning tunnel advance prediction method based on transient electromagnetic images according to claim 3 is characterized in that: Based on the 3D reconstructed image and the TEM synthetic image, the distribution of the drilling positions and the amount of explosives used includes: Based on the three-dimensional reconstructed image and the TEM synthetic image, adding or reducing blasting holes for different density areas; The amount of explosives used in each blasting hole is arranged according to the characteristic position data, characteristic volume data and characteristic three-dimensional trend information in the TEM synthetic image.

Images