Advanced prediction method and system for unfavorable geology of tunnel based on deep learning

Abstract

The invention provides a tunnel unfavorable geology advanced prediction method and system based on deep learning, and relates to the technical field of tunnel engineering, and the method comprises the steps: obtaining unfavorable geology radar image data, and carrying out the preprocessing; inputting the preprocessed image data into a geological radar interpretation model GPR-YOLOv5, and outputting an unfavorable geological prediction recognition result; wherein in the geological radar interpretation model GPR-YOLOv5, a CA attention mechanism module pays attention to a salient region in image data, pooling is carried out in the horizontal direction and the vertical direction by utilizing a pooling kernel, after image features are extracted, feature maps are fused and spliced, and a transition feature map is generated; a GhostC3 structure is used for carrying out channel expansion, low-level features and high-level features in transition feature maps are combined, rich image context information is obtained, then a GSConv module convolution module is used for carrying out weighting and aggregation on the transition feature maps of neighbor nodes of each node, feature representation of the nodes is updated, and an unfavorable geological prediction recognition result is output.

Description

Technical Field

[0001] The present disclosure relates to the field of tunnel engineering technology, and in particular to a method and system for advanced prediction of poor tunnel geology based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Tunnel engineering is a hidden project and is often affected by various unfavorable geological bodies, leading to catastrophic accidents such as surrounding rock collapse. Therefore, it is particularly important to conduct advanced geological forecasting. Geological radar is the engineering geophysical method with the highest resolution at present. It is widely used in engineering quality inspection and site investigation. In recent years, it has also been used for tunnel advance forecasting.

[0004] Traditional geological radar image interpretation mainly relies on professional technicians, which is inefficient and subjective. Image processing technology based on ground penetrating radar image interpretation has been widely used in detecting hidden targets. The most commonly used methods are edge detection, Hough transform and mitigation technology.

[0005] With the advancement of technology, machine learning methods have been used to analyze the features of GPR b-scans. Todkar et al. used the finite difference time domain (FDTD) forward modeling algorithm to build a simulated GPR b-scan and introduced a support vector machine method to identify the roadbed with detachment. The simulation and field test results showed that this method can identify millimeter-level interlayer debonding in ground penetrating radar b-scans. Maas et al. used the VJ algorithm to extract the target area in GPR data. Pasolli et al. designed a genetic algorithm-based technique to identify linear and hyperbolic features of underground objects. By using these extracted features, they were classified into linear or hyperbolic shapes using a support vector machine classifier.

[0006] However, there are still some problems with the above methods: these methods mainly rely on empirical parameters or hand-crafted features, and it is difficult to accurately identify the features of various complex-shaped targets from geological radar images with noisy background interference. The existing geological radar image interpretation model has an unbalanced ability to extract features. Manual feature engineering is usually used to extract relevant features from radar data, and these features are used to train models for tasks such as classification or regression. The degree of intelligence and automation is relatively low, and the detection accuracy is also relatively low. Summary of the invention

[0007] In order to solve the above problems, this paper proposes a method and system for advanced prediction of poor tunnel geology based on deep learning, improves the target detection model YOLOv5s, highlights the target features by adding attention mechanism and improving convolution module, improves detection accuracy, and enhances the feature extraction and fusion capability of the model.

[0008] The deep learning-based advanced prediction method for poor tunnel geology includes:

[0009] Obtain poor geological radar image data and perform preprocessing;

[0010] The preprocessed image data is input into the geological radar interpretation model GPR-YOLOv5 to output the adverse geological prediction and identification results;

[0011] Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts the image features, and fuses and splices the feature maps to generate a transition feature map; uses the GhostC3 structure to expand the channel, combines the low-level features in the transition feature map with the high-level features to obtain rich image context information, and then uses the GSConv module convolution module to weight and aggregate the transition feature maps of the neighboring nodes of each node, updates the feature representation of the node, and outputs the poor geological prediction recognition results.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] The deep learning-based tunnel adverse geological advance prediction system includes:

[0014] Data acquisition module, used to acquire poor geological radar image data and perform preprocessing;

[0015] The prediction module is used to input the preprocessed image data into the geological radar interpretation model GPR-YOLOv5 and output the adverse geological prediction identification results;

[0016] Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts the image features, and fuses and splices the feature maps to generate a transition feature map; uses the GhostC3 structure to expand the channel, combines the low-level features in the transition feature map with the high-level features to obtain rich image context information, and then uses the GSConv module convolution module to weight and aggregate the transition feature maps of the neighboring nodes of each node, updates the feature representation of the node, and outputs the poor geological prediction recognition results.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] A non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the method for advanced prediction of poor tunnel geology based on deep learning is implemented.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the method for advanced prediction of poor tunnel geology based on deep learning.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] This paper proposes a method for advanced prediction of poor geology in tunnels based on deep learning. It uses the geological radar interpretation model GPR-YOLOv5, adds a CA attention mechanism module to the backbone network, adaptively learns the correlation between different feature channels, and applies attention weights to different channels of the feature map, so that the model can pay more attention to useful features, improving the detection performance and expression ability of the model. Secondly, it also integrates the GhostC3 module and the GSConv module, so that the model can better fit the task of poor geological detection, and uses the GSConv module convolution to replace the CBS module in YOLOv5s. The convolution in YOLOv5s is mostly standard convolution, but if the model uses this convolution method in all stages, it will cause the network layer of the model to be deeper, significantly increasing the inference time. In the Neck part, the spatial dimension of the feature map reaches the minimum and the channel dimension reaches the maximum. The feature map becomes "small" and no longer needs channel transformation. At this time, the use of GSConv convolution method not only reduces redundant and repeated information, ensures the full extraction of target feature information, but also reduces the number of parameters, achieves the purpose of lightweight network, makes the model easier to be installed on mobile devices, and is more in line with the needs of real-time processing of radar data; improves the accuracy of detection. At the same time, the calculation amount of these two modules is relatively small, which also improves the calculation efficiency of the model. In addition, the introduction of Varifocal loss as confidence and classification loss helps to focus the training on high-quality samples and make the ratio of positive and negative samples more reasonable. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0024] Figure 1 A schematic diagram of a CA attention mechanism module according to an embodiment of the present disclosure;

[0025] Figure 2 It is a schematic diagram of the GhostC3 structural module of an embodiment of the present disclosure;

[0026] Figure 3 Schematic diagram of a GSConv module according to an embodiment of the present disclosure;

[0027] Figure 4 This is a network architecture diagram of the geological radar interpretation model GPR-YOLOv5 of an embodiment of the present disclosure;

[0028] Figure 5 This is a diagram of the model training architecture of an embodiment of the present disclosure;

[0029] Figure 6 It is a schematic diagram of the verification result of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0031] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0033] Example 1

[0034] In one embodiment of the present disclosure, a method for advanced prediction of poor tunnel geology based on deep learning is provided, comprising:

[0035] Obtain poor geological radar image data and perform preprocessing;

[0036] The preprocessed image data is input into the geological radar interpretation model GPR-YOLOv5 to output the adverse geological prediction and identification results;

[0037] Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts the image features, and fuses and splices the feature maps to generate a transition feature map; uses the GhostC3 structure to expand the channel, combines the low-level features in the transition feature map with the high-level features to obtain rich image context information, and then uses the GSConv module convolution module to weight and aggregate the transition feature maps of the neighboring nodes of each node, updates the feature representation of the node, and outputs the poor geological prediction recognition results.

[0038] As an embodiment, the geological radar interpretation model GPR-YOLOv5 selects a lightweight YOLOv5s model. In the backbone network, a CA attention mechanism module is added before the backbone network SPPF module to adaptively learn the correlation between different feature channels and apply the attention weights to different channels of the feature map, so that the model can pay more attention to useful features and improve the detection performance and expression ability of the model. Secondly, the GhostC3 module and the GSConv module are also integrated to enable the model to better fit the task of poor geological detection and improve the detection accuracy. At the same time, the calculation amount of these two modules is relatively small, which also improves the calculation efficiency of the model. In addition, the Varifocal loss is introduced as the confidence and classification loss, which helps to focus the training on high-quality samples and make the ratio of positive and negative samples more reasonable. Finally, the geological radar interpretation model GPR-YOLOv5 is generated.

[0039] As an embodiment, in the YOLOv5s model, a CA attention module is added before the backbone network SPPF module. Figure 1 As shown in the figure, when YOLOv5s interprets geological radar images, the detected objects are often messy and difficult to identify, and key feature information is easily lost during the detection process, resulting in poor detection results. Therefore, the CA attention mechanism is added to the present disclosure, which allows the algorithm model to focus on the area that needs attention and effectively improves the accuracy of small target detection. The CA attention mechanism uses two pooling kernels at the input end to perform pooling in the horizontal and vertical directions, thereby obtaining an output with a height of h and a width of w. The output feature maps are then fused and spliced ​​and then convolved with a 1×1 convolution kernel to generate a transition feature map. Secondly, the transition feature map is decomposed into separate feature maps, and then convolved with a 1×1 convolution kernel to obtain a number of channels consistent with the input.

[0040] The specific process of the CA attention mechanism is as follows:

[0041] 1. Input: Assume that the dimension of the input feature map is C×H×W, where C represents the number of channels, H and W represent the height and width respectively.

[0042] 2. Global average pooling: First, perform global average pooling on the input feature map to reduce its dimension to a vector of length C. The vector obtained in this way can represent the importance of each channel in the entire feature map.

[0043] 3. Fully connected layer: The vector obtained by global average pooling is mapped through a fully connected layer and converted into a vector of length C. This vector represents the weight of each channel.

[0044] 4. Activation function: In order to ensure that the channel weight is positive and can express the relative importance between different channels, an activation function such as sigmoid or softmax is usually used on the output of the fully connected layer.

[0045] 5. Reshape: Resize the channel weight vector processed by the activation function to the same dimension as the input feature map, that is, C×1×1.

[0046] 6. Feature scaling: Multiply the adjusted channel weights by the input feature map element by element to obtain a weighted feature map. This can amplify or suppress the corresponding feature response according to the importance of the channel.

[0047] 7. Output: The feature map after feature scaling is used as the output of the CA attention mechanism and can be directly used for subsequent tasks such as classification and target detection.

[0048] In general, the process of CA attention mechanism includes global average pooling, fully connected layer mapping, activation function processing, reshaping, feature scaling and other steps. It improves the expression ability and performance of the network by learning the correlation between channels and strengthening important feature channels.

[0049] In GPR-YOLOv5, the input shape of the CA layer should be [512, 20, 20], representing 512 channels and a 20x20 feature map. The output of the CA layer will have the same shape, which is [512, 20, 20]. This module extracts image features by weighting each channel with attention.

[0050] In the YOLOv5s model, the GSConv module convolution is used to replace the CBS module in YOLOv5s. The convolutions in YOLOv5s are mostly standard convolutions, but if the model uses this convolution method in all stages, the network layer of the model will be deeper, which will significantly increase the inference time. In the Neck part, the spatial dimension of the feature map reaches the minimum and the channel dimension reaches the maximum. The feature map becomes "small" and no longer needs channel transformation. At this time, the use of the GSConv convolution method not only reduces redundant and repeated information, ensuring the full extraction of target feature information, but also reduces the number of parameters, achieving the purpose of lightweight network, making the model easier to carry on mobile devices, and more in line with the needs of real-time processing of radar data. Figure 3 As shown in the figure, the GSConv module has 4 layers in GPR-YOLOv5. GSConv updates the feature representation of the node by adaptively weighting and aggregating the features of the neighboring nodes of each node.

[0051] Furthermore, the specific process in the GSConv module is:

[0052] 1. Input: Assume that the dimension of the input data is N×C×D, where N represents the number of samples, C represents the number of channels, and D represents the dimension of the data.

[0053] 2. Graph construction: In order to represent the data as a graph structure, you need to first construct an adjacency matrix A. The adjacency matrix describes the relationship between data points and is usually constructed using the k-nearest neighbor or radius neighborhood method. After the construction is completed, the adjacency matrix A can be represented as a sparse matrix with a dimension of N×N.

[0054] 3. Feature transformation: The input data is transformed through a linear transformation to obtain a new feature representation. This linear transformation is usually implemented by a fully connected layer, and the output dimension is C'.

[0055] 4. Neighbor aggregation: Aggregate the neighbor nodes of each data point with their corresponding features to obtain the neighbor feature representation of each node. This step can be expressed as X'=AXW, where X' represents the neighbor feature matrix and W represents the weight matrix.

[0056] 5. Information exchange: The neighbor feature representation is exchanged with the input feature representation to obtain a new feature representation for each node. This operation is usually implemented using convolution, such as 1×1 convolution or 3×3 convolution in image processing.

[0057] 6. Output: After information interaction, a new feature representation of each data point is obtained as the output of the GSConv module, which can be used for subsequent tasks such as classification and segmentation.

[0058] In general, the process of the GSConv module includes steps such as graph construction, feature conversion, neighbor aggregation, and information interaction, and has strong expressiveness and adaptability.

[0059] In the YOLOv5s model, the GhostC3 structure is used to replace the original C3 structure in the backbone network, such as Figure 2 As shown in Figure 2, the Ghostbottleneck process in the GhostC3 structure seems similar to the basic residual block in ResNet, in which several convolutional layers and shortcuts are integrated. The proposed Ghostbottleneck mainly consists of two stacked ghost modules. The first Ghost module acts as an expansion layer to increase the number of channels. We call the ratio of the number of output channels to the number of input channels the expansion ratio. The second Ghost module reduces the number of channels matching the shortcut path. Then a shortcut is used to connect the input and output of these two Ghost modules.

[0060] The GhostC3 module has 4 layers in GPR-YOLOv5. Through the GhostC3 module, low-level features can be combined with high-level features to capture richer contextual information while maintaining detailed information.

[0061] Furthermore, the specific process in the GhostC3 module is:

[0062] 1. Input: Assume that the dimension of the input feature map is N×C×H×W, where N represents the number of samples, C represents the number of channels, and H and W represent the height and width respectively.

[0063] 2. Split features: The input feature map is divided into two parts, one part is called the main feature (Main Feature), and the other part is called the ghost feature (Ghost Feature). The split method can be selected according to the actual situation, such as equal division, according to the proportion of the number of channels, etc.

[0064] 3. Main feature processing:

[0065] a. Backbone feature convolution: Perform convolution operation on the backbone features. You can use different convolution kernel sizes and step sizes to obtain the convolution results of the backbone features.

[0066] b. Upsampling of backbone features: In order to maintain the same size after adding the ghost features, the backbone features can be upsampled to make them consistent with the size of the ghost features.

[0067] 4. Ghost feature processing:

[0068] a. Ghost feature convolution: Perform convolution operation on ghost features, where the number of channels of the convolution kernel is a fraction of the original number of channels, i.e., Ghost Ratio. This can reduce the number of parameters and calculations and improve efficiency.

[0069] b. Downsampling of ghost features: In order to keep the same size after adding to the backbone features, the ghost features can be downsampled to make them consistent with the size of the backbone features.

[0070] 5. Feature fusion: Add the main features and ghost features to get the final feature map. This can retain the rich information of the main features and incorporate the local details of the ghost features.

[0071] 6. Output: After feature fusion, the final feature map is obtained as the output of the GhostC3 module, which can be used for subsequent tasks such as classification and target detection.

[0072] In general, the process of the GhostC3 module includes the steps of segmentation feature, backbone feature processing, ghost feature processing, feature fusion, etc. By using the combination of ghost features and backbone features, the GhostC3 module can maintain good performance and effects while reducing the amount of parameters and calculations.

[0073] The input and output of each module are shown in Table 1.

[0074] Table 1 Input and output of each module

[0075]

[0076]

[0077]

[0078]

[0079] In Table 1, the “-1” in the previous layer column indicates that the

[0080] The Detect output layer [3, 80, 80, 85] indicates that the output size of the model is 80x80 pixels, and the model predicts 3 bounding boxes, each containing 80 feature maps and 85 channels.

[0081] Furthermore, Varifocal loss is used as the confidence and classification loss function. The original confidence and classification loss function of the YOLOv5s model is BCEWithLogitsLoss, but it is usually used to detect more obvious samples. For poor geological radar samples with more complex backgrounds, there are a large number of negative samples, and the large proportion is easy to affect the training results, thereby reducing the accuracy of the model detection target. Therefore, the present disclosure introduces Varifocal loss, and its formula is as follows:

[0082]

[0083] Where p is the predicted probability of the foreground class, and q is the target IoU score. In the positive training samples, p is set to the IoU between the generated predicted box and the true box; while in the negative training samples, the training target q of all categories is 0, α is the positive sample weight, and γ is the difficult sample weight.

[0084] The above is as follows: Using the original YOLOv5s model as the baseline network, a new GPR-YOLOv5 network structure is constructed. Each module is improved to generate a new GPR-YOLOv5 network. Figure 4 As shown,

[0085] The present invention uses the obtained GPR-YOLOv5 network optimal model to perform target detection and recognition on the verification set, and obtains recognition results, wherein the recognition accuracy rate is 92%, the recall rate is 87.1%, and the detection effect is good.

[0086] In order to find a model that is more suitable for geological radar image interpretation, a series of experiments were conducted on the YOLOv5 basic model, introducing the precision rate, recall rate, and mean average precision (mAP) of multiple categories to evaluate the performance of the detection algorithm. The precision rate P, recall rate R, and mAP are expressed as:

[0087]

[0088] Among them, TP (True Positives) indicates the number of correctly detected targets, FP (False Positives) indicates the number of incorrectly detected targets, FN (False Negatives) indicates the number of undetected targets, n indicates the total number of categories to be classified, and AP (Average Precision) indicates the average precision of a target class. mAP@.5:.95 indicates the average mAP at different IoU thresholds (from 0.5 to 0.95, with a step size of 0.05).

[0089] The batch-size is set to 16 and the total number of iterations is set to 200. The hardware parameters of the experiment are AMD EPYC 7543 32-Core Processor and NVIDIA GeForce RTX 4090, and the Pytorch deep learning framework and Python programming are used to implement

[0090] The experiment made some module adjustments and loss function replacements and found the currently optimal model combination, which is the current GPR-YOLOv5s architecture. The experimental data is shown in the following table.

[0091] As can be seen from the table, the disclosed model is effective for interpreting radar images. The precision is 92%, the recall is 87.1%, and the mAP@.5:.95 is 57.9%. Compared with the original YOLOv5s algorithm, the precision is improved by 5%, the recall is improved by 1.1%, and the mAP@.5:.95 is improved by 1.6%. This is a model in which all indicators are improved in the above improvement measures.

[0092] Example 2

[0093] In one embodiment of the present disclosure, a system for advanced prediction of poor tunnel geology based on deep learning is provided, comprising:

[0094] Data acquisition module, used to acquire poor geological radar image data and perform preprocessing;

[0095] The prediction module is used to input the preprocessed image data into the geological radar interpretation model GPR-YOLOv5 and output the adverse geological prediction identification results;

[0096] Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts the image features, and fuses and splices the feature maps to generate a transition feature map; uses the GhostC3 structure to expand the channel, combines the low-level features in the transition feature map with the high-level features to obtain rich image context information, and then uses the GSConv module convolution module to weight and aggregate the transition feature maps of the neighboring nodes of each node, updates the feature representation of the node, and outputs the poor geological prediction recognition results.

[0097] Example 3

[0098] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method for advanced prediction of poor tunnel geology based on deep learning is implemented.

[0099] Example 4

[0100] An electronic device is provided in an embodiment of the present disclosure, characterized in that it includes: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the method for advanced prediction of poor tunnel geology based on deep learning.

[0101] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0103] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A method for predicting adverse tunnel geology based on deep learning, characterized in that: include: Obtain poor geological radar image data and perform preprocessing; The preprocessed image data is input into the geological radar interpretation model GPR-YOLOv5 to output the adverse geological prediction and identification results; Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts image features, and then fuses and splices the feature maps to generate transition feature maps; The GhostC3 structure is used for channel expansion, and the low-level features in the transition feature map are combined with the high-level features to obtain rich image context information. The GSConv module convolution module is then used to weight and aggregate the transition feature maps of each node's neighbor nodes, update the node's feature representation, and output the poor geological prediction recognition results.

2. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 1, characterized in that: Preprocessing includes: manually labeling the defect shapes and types in radar images using the LabelImg labeling tool, labeling the poor geology on each image with the smallest rectangular box, manually labeling the defect range and type in the GPRb scan, generating a txt file as a label, and using the ImgAug library to perform data enhancement on the labeled data set to convert a set of input images into a new set of images.

3. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 1, characterized in that: A CA attention mechanism module is added before the backbone network SPPF module. The CA attention mechanism module uses two pooling kernels to perform pooling in the horizontal and vertical directions to obtain a feature map with a height of h and a width of w. The output feature map is then fused and convolved with a convolution kernel to generate a transition feature map. The transition feature map is decomposed into separate feature maps, which are then convolved with a convolution kernel to obtain a number of channels consistent with the input.

4. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 1, characterized in that: In the backbone network, the GhostC3 structure is used to replace the C3 structure. Several convolutional layers and shortcuts are integrated in the GhostC3 structure. The Ghostbottleneck is mainly composed of two stacked ghost modules. The first ghost module acts as an expansion layer to increase the number of channels. The ratio of the number of output channels to the number of input channels is called the expansion ratio. The second ghost module reduces the number of channels matching the shortcut path, and then a shortcut is used to connect the input and output of the two ghost modules.

5. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 4, characterized in that: The GhostC3 module has 4 layers in the geological radar interpretation model GPR-YOLOv5. Through the GhostC3 module, low-level features are combined with high-level features to capture richer contextual information while maintaining detailed information.

6. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 1, characterized in that: The GSConv module convolution is used to replace the CBS module in YOLOv5s. The GSConv convolution method has less redundant and repeated information, fully extracts the target feature information, reduces the number of parameters, and achieves the purpose of lightweight network.

7. The method for advanced prediction of tunnel adverse geology based on deep learning according to claim 6, characterized in that: The GSConv module has 4 layers in the geological radar interpretation model GPR-YOLOv5. The GSConv module updates the feature representation of the node by adaptively weighting and aggregating the features of each node's neighbor nodes.

8. The tunnel adverse geological advance prediction system based on deep learning is characterized by: include: Data acquisition module, used to acquire poor geological radar image data and perform preprocessing; The prediction module is used to input the preprocessed image data into the geological radar interpretation model GPR-YOLOv5 and output the adverse geological prediction identification results; Among them, in the geological radar interpretation model GPR-YOLOv5, the CA attention mechanism module focuses on the salient areas in the image data, uses the pooling kernel to perform pooling in the horizontal and vertical directions, extracts image features, and then fuses and splices the feature maps to generate transition feature maps; The GhostC3 structure is used for channel expansion, and the low-level features in the transition feature map are combined with the high-level features to obtain rich image context information. The GSConv module convolution module is then used to weight and aggregate the transition feature maps of each node's neighbor nodes, update the node's feature representation, and output the poor geological prediction recognition results.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the method for advanced prediction of poor tunnel geology based on deep learning as described in any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the method for advanced prediction of poor tunnel geology based on deep learning as described in any one of claims 1-7.

Images