Method for intelligently determining grade of surrounding rock of tunnel face

Through the deep learning model, intelligently identifying the photos of the palm face of the tunnel, the problems of low efficiency and safety risks in the traditional method are solved, and the rapid and accurate evaluation of the surrounding rock level during tunnel construction is achieved, and reliable geological information support is provided. .

CN119068236APending Publication Date: 2024-12-03CHINA RAILWAY SIYUAN SURVEY & DESIGN GRP CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411001474.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In mining tunnel construction, the traditional method used to judge the level of surrounding rock in the palm surface of the tunnel requires a large amount of operating time, which poses safety risks, and is affected by the technical level of on-site personnel, and the judgment standards are not unified, which affects the construction period and quality. .

Method used

The deep learning model is used to combine multi-level judgment logic, and intelligently identify the photos of the tunnel palm surface, and a deep learning model of ResNet50 backbone network, feature pyramid network and multi-classification head are built, and image data is processed automatically to obtain comprehensive evaluation of surrounding rock level.

Benefits of technology

It realizes intelligent and rapid judgment of the surrounding rock level of the tunnel palm surface, improves efficiency and accuracy, provides more comprehensive and objective geological information support, and reduces construction risks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068236A_ABST
    Figure CN119068236A_ABST
Patent Text Reader

Abstract

The invention provides a tunnel face surrounding rock level intelligent determination method, and relates to the technical field of tunnel engineering investigation and construction, and the method comprises the steps: S1, constructing a deep learning model, carrying out the iterative training of the deep learning model through a training set till the model converges, and obtaining a trained deep learning model; s2, obtaining a to-be-recognized image, inputting the to-be-recognized image into the trained deep learning model, and outputting to obtain five classification results; s3, carrying out standard value domain processing on the five classification results, and then carrying out comprehensive calculation on the five classification results subjected to standard value domain processing by adopting a comprehensive scoring method to obtain a comprehensive score; and S4, judging the comprehensive score according to a surrounding rock level rule to obtain the surrounding rock level of the to-be-recognized image. According to the method, the surrounding rock level of the tunnel face can be intelligently and quickly judged and evaluated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel engineering investigation and construction, and in particular to a method for intelligently determining the surrounding rock grade of a tunnel face. Background Art

[0002] At present, in the process of tunnel construction using the mining method, traditional methods are often used to determine the surrounding rock level of the face: after tunnel blasting and excavation, the quality of the surrounding rock in the face and a certain range ahead is comprehensively evaluated through traditional geological means such as face geological sketching, advanced horizontal drilling, deepening blastholes, and engineering geophysical exploration. The surrounding rock level is identified based on the evaluation results, and suggestions are made for the next construction process. Frequent face observation and exploration methods require a large amount of face operation time, which seriously affects the on-site construction organization and the tunnel construction period, and there are many safety risks. In addition, the comprehensive technical level of on-site personnel has a great influence on the determination of surrounding rock quality through face geological inspection, and there are many subjective factors, and the judgment standards and basis are not unified. Summary of the invention

[0003] In view of this, the present invention proposes a method for intelligently determining the surrounding rock grade of the tunnel face, which can realize intelligent and rapid determination and evaluation of the surrounding rock grade of the tunnel face. By deep learning the face photos and establishing a complete deep learning model, the multi-level judgment logic is trained to obtain the surrounding rock grade of the tunnel face and guide the next construction cycle sequence.

[0004] The technical solution of the present invention is implemented as follows: The present invention provides a method for intelligently determining the surrounding rock level of a tunnel face, comprising:

[0005] S1 builds a deep learning model and iteratively trains the deep learning model using the training set until the model converges to obtain a trained deep learning model;

[0006] S2 obtains the image to be identified, inputs the image to be identified into the trained deep learning model, and outputs five classification results;

[0007] S3 standardizes the five classification results, and then uses a comprehensive scoring method to comprehensively calculate the five classification results of the standard value range to obtain a comprehensive score;

[0008] S4 determines the comprehensive score according to the surrounding rock grade rule to obtain the surrounding rock grade of the image to be identified.

[0009] Based on the above technical solution, preferably, the deep learning model includes a ResNet50 backbone network, a feature pyramid network and a classification network, wherein:

[0010] The ResNet50 backbone network includes a one-dimensional convolutional layer, a BN layer, a maximum pooling layer, and 4 residual blocks. The first residual block includes 3 residual units, the second residual block includes 4 residual units, the third residual block includes 6 residual units, and the fourth residual block includes 3 residual units.

[0011] The feature pyramid network adopts a top-down path, including 4 convolutional layers, and each convolutional layer uses lateral connections;

[0012] The classification network includes 5 classification heads. The first classification head includes 3 convolutional layers, a global average pooling layer and a fully connected layer. The second classification head includes 3 consecutive convolutional layers, a global average pooling layer and a fully connected layer. The third classification head includes a convolutional layer, a self-attention layer, an average pooling layer and a fully connected layer. The fourth classification head includes a convolutional layer, a channel attention block, a global average pooling layer and a fully connected layer. The fifth classification head includes 2 convolutional layers, an adaptive maximum pooling layer, a bidirectional GRU layer and a fully connected layer.

[0013] On the basis of the above technical solution, preferably, the first classification head is used for surrounding rock color classification; the second classification head is used for crack width classification; the third classification head is used for crack length classification; the fourth classification head is used for crack group number classification; and the fifth classification head is used for groundwater development degree classification.

[0014] Based on the above technical solution, preferably, the process of the deep learning model processing the image to be recognized is:

[0015] The image to be recognized is input into the ResNet50 backbone network, and four residual blocks are used to extract features of four scales in sequence, namely the first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4;

[0016] The first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4 are input into the feature pyramid network, and the four scale features C i Both apply 1*1 convolution and adjust the number of channels to the same value. i =Conv1*1(C i ); Following the top-down path, starting from the top layer, the high-level scale features are upsampled and added to the next layer to form a four-layer pyramid feature A i , A i =B i +Upsample(A i+1 ); for each pyramid feature A i Apply 3*3 convolution for smoothing and get 4 smooth features P i , P i =Conv3*3(A i ); where i = 1, 2, 3, 4;

[0017] Input P1 into the first classification head to obtain the first classification result, input P1 and P2 into the second classification head to obtain the second classification result, input P2, P3 and P4 into the third classification head to obtain the third classification result, input P1, P2, P3 and P4 into the fourth classification head to obtain the fourth classification result, and input P3 and P4 into the fifth classification head to obtain the fifth classification result.

[0018] On the basis of the above technical solution, preferably, P1 is input into the first classification head, and color features are gradually extracted through 3 consecutive convolutional layers. Batch normalization and ReLU activation are performed after each convolution. During the convolution process, the number of channels is gradually reduced, and the features are compressed into a one-dimensional vector using global average pooling, and then mapped to the output space of 5 categories through a fully connected layer to obtain the first classification result;

[0019] P1 and P2 are input into the second classification head, and width-related features are extracted through one 3*3 convolution layer and two 3*3 hole convolution layers. The ReLU activation function is applied after each convolution layer. After that, the processed P1 and P2 are upsampled or downsampled to the same size, and then spliced ​​in the channel dimension. The spliced ​​features are fused through a 1*1 convolution, and the features are compressed into a one-dimensional vector using global average pooling. Then, they are mapped to the output space of 5 categories through a fully connected layer to obtain the second classification result;

[0020] P2, P3, and P4 are input into the third classification head. Each feature is independently subjected to 1*1 convolution and self-attention mechanism. The three processed features are upsampled to the same size and spliced ​​in the channel dimension. The spliced ​​features are passed through a self-attention layer to fuse multi-scale information. Finally, they are mapped to the output space of 5 categories through a fully connected layer to obtain the third classification result.

[0021] P1, P2, P3 and P4 are input into the fourth classification head. Each feature is independently processed by 1*1 convolution and channel attention block. The processed 4 features are adjusted to the same size and spliced ​​in the channel dimension. The spliced ​​features are passed through an additional channel attention block to further fuse multi-scale information. The features are compressed into a one-dimensional vector through global average pooling and then mapped to the output space of 5 categories through a fully connected layer to obtain the fourth classification result.

[0022] P3 and P4 are input into the fifth classification head, and each feature is subjected to two 3*3 convolutional layers to extract water development-related features. Batch normalization and ReLU activation are performed after each convolution. Then, an adaptive maximum pooling layer is used to obtain two feature sequences. These two feature sequences are spliced ​​in the time dimension to form a spliced ​​sequence, and then input into the bidirectional GRU for processing. The output of the GRU is average pooled and then mapped to the output space of 5 categories through the fully connected layer to obtain the fifth classification result.

[0023] Based on the above technical scheme, preferably, the first classification results include single, relatively single, relatively turbid, turbid, and extremely turbid; the second classification results include extremely narrow, narrow, relatively narrow, relatively wide, and wide; the third classification results include less than extremely short, short, relatively short, relatively long, and long; the fourth classification results include very little, few, fewer, more, and much; the fifth classification results include dry, dot-shaped, drop-shaped, line-shaped, and strand-shaped.

[0024] Based on the above technical solution, preferably, step S3 includes:

[0025] S31 sets a standard value range, divides the standard value range into five sub-ranges, and corresponds five categories of five classification results to the five sub-ranges one by one, wherein the five classification results are respectively recorded as a first classification result, a second classification result, a third classification result, a fourth classification result, and a fifth classification result;

[0026] S32 obtains the predicted probability values ​​of the first classification result, the second classification result, the third classification result, the fourth classification result, and the fifth classification result, and calculates the individual scores of the five classification results according to the conversion relationship between the predicted probability values ​​and the corresponding sub-ranges;

[0027] S33 sets weights for the five classification results and calculates a comprehensive score based on the weights and individual scores.

[0028] Based on the above technical solution, preferably, the calculation formula for a single score is:

[0029]

[0030] In the formula, Represents the single score of the j1-th classification result, j1 = 1, 2, 3, 4, 5; Indicates the lower limit of the value range of the sub-range j2 corresponding to the j1-th classification result, j2 = 1, 2, 3, 4, 5; Represents the predicted probability value of the j1th classification result; Represents the prediction threshold of the j1th classification result; Represents the adjustment factor for sub-range j2.

[0031] Based on the above technical solution, preferably, the calculation formula of the comprehensive score is:

[0032]

[0033] In the formula, S total is the comprehensive score, Represents the weight of the j1th classification result.

[0034] On the basis of the above technical solution, preferably, step S4 includes:

[0035] The value range of the comprehensive score is divided into 6 level ranges, and the 6 level ranges are set as surrounding rock level 1, surrounding rock level 2, surrounding rock level 3, surrounding rock level 4, surrounding rock level 5, and surrounding rock level 6 according to the comprehensive scores from high to low. According to the value of the comprehensive score calculated in step S3, it is assigned to the corresponding surrounding rock level to obtain the surrounding rock level of the image to be identified.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The present invention realizes the intelligent determination of the surrounding rock grade of the tunnel face by constructing a deep learning model and combining it with a comprehensive scoring method. This method can not only automatically process a large amount of image data, thereby improving the efficiency and accuracy of surrounding rock grade determination, but also comprehensively consider multiple geological characteristics, including surrounding rock color, crack width, crack length, number of crack groups, and groundwater development, thereby obtaining a more comprehensive and objective evaluation result, providing reliable geological information support for tunnel construction;

[0038] (2) The present invention constructs a deep learning model consisting of a ResNet50 backbone network, a feature pyramid network, and a multi-classification head, which can extract image features from different scales and levels and perform comprehensive analysis. This structure not only improves the depth and breadth of feature extraction, but also demonstrates strong adaptability and efficiency when processing various geological features, greatly improving the recognition accuracy of the model after training;

[0039] (3) Five special classification heads are designed in the classification network of the present invention, which classify the surrounding rock color, crack width, crack length, number of crack groups and groundwater development degree respectively. This design can not only effectively separate the influence of different geological characteristics and ensure that each feature can be fully analyzed and identified, but also more comprehensively evaluate the overall stability and quality of the surrounding rock by integrating these classification results;

[0040] (4) The standard value domain and comprehensive scoring method introduced in the present invention unifies the evaluation results of different types of geological features into a standard scale, and realizes the comprehensive consideration of multiple features through weight allocation. This method not only solves the problem of inconsistent quantitative standards of different geological features, but also can flexibly adjust the importance of each feature, making the final surrounding rock grade determination more objective and comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0043] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] like Figure 1 As shown, the present invention provides a method for intelligently determining the surrounding rock level of a tunnel face, comprising:

[0045] S1 builds a deep learning model and iteratively trains the deep learning model using the training set until the model converges to obtain a trained deep learning model;

[0046] S2 obtains the image to be identified, inputs the image to be identified into the trained deep learning model, and outputs five classification results;

[0047] S3 standardizes the five classification results, and then uses a comprehensive scoring method to comprehensively calculate the five classification results of the standard value range to obtain a comprehensive score;

[0048] S4 determines the comprehensive score according to the surrounding rock grade rule to obtain the surrounding rock grade of the image to be identified.

[0049] Specifically, the present invention constructs a deep learning model, uses the images of the tunnel face as a training set, and iteratively trains the model. The specific training process is as follows:

[0050] Firstly, a large number of tunnel face images from home and abroad were collected, and the surrounding rock color, crack width, crack length, number of crack groups, and groundwater development degree in the images were annotated. Specifically, the surrounding rock color was divided into single, relatively single, relatively turbid, turbid, and extremely turbid; the crack width was divided into extremely narrow, narrow, relatively narrow, relatively wide, and wide; the crack length was divided into extremely short, short, relatively short, relatively long, and long; the number of crack groups was divided into very few, few, relatively few, relatively many, and a lot; the groundwater development degree was divided into dry, point-like, drop-like, linear, and strand-like. Image annotation tools such as LabelImg can be used for annotation. Among the five categories, the crack width, crack length and number of crack groups can be converted into numerical labels. For example, the crack width is divided into extremely narrow, narrow, relatively narrow, relatively wide, and wide, corresponding to crack width <5mm, 5mm<crack width ≤10mm, 10mm<crack width ≤20mm, 20mm<crack width ≤30mm, and 30mm<crack width; the crack length is divided into extremely short, short, relatively short, relatively long, and long, corresponding to crack length <1m, 1m<crack length ≤2m, 2m<crack length ≤3m, 3m<crack length ≤5m, and 5m<crack length; the number of crack groups is divided into extremely few, few, few, more, and many, corresponding to number of crack groups ≤1 group, 1 group <number of crack groups ≤2 groups, 2 groups <number of crack groups ≤3 groups, 3 groups <number of crack groups ≤5 groups, and 5 groups <number of crack groups. Create a corresponding annotation file for each image, which can be in JSON format. After annotation, a dataset is formed.

[0051] The dataset is preprocessed, including resizing to a uniform size, performing data enhancement (such as random cropping, flipping, rotating, adjusting brightness and contrast, etc.), and finally standardizing the image.

[0052] Divide the dataset into training set, validation set, and test set in a ratio of 70%:15%:15%.

[0053] Next, a deep learning model is built, including the ResNet50 backbone network, feature pyramid network, and classification network.

[0054] ResNet50 backbone network: includes a one-dimensional convolution layer, a BN layer, a maximum pooling layer and 4 residual blocks, the first residual block includes 3 residual units, the second residual block includes 4 residual units, the third residual block includes 6 residual units, and the fourth residual block includes 3 residual units. In this embodiment, a pre-trained ResNet50 network is used, the first layer is modified to a one-dimensional convolution, and the outputs of the four residual blocks are retained: C1, C2, C3, C4.

[0055] Feature Pyramid Network: It adopts a top-down path, including 4 convolutional layers, and lateral connections are used between each convolutional layer.

[0056] Classification network: includes 5 classification heads. The first classification head includes 3 convolutional layers, a global average pooling layer and a fully connected layer. The second classification head includes 3 consecutive convolutional layers, a global average pooling layer and a fully connected layer. The third classification head includes a convolutional layer, a self-attention layer, an average pooling layer and a fully connected layer. The fourth classification head includes a convolutional layer, a channel attention block, a global average pooling layer and a fully connected layer. The fifth classification head includes 2 convolutional layers, an adaptive maximum pooling layer, a bidirectional GRU layer and a fully connected layer.

[0057] The first classification head is used for surrounding rock color classification; the second classification head is used for crack width classification; the third classification head is used for crack length classification; the fourth classification head is used for crack group number classification; the fifth classification head is used for groundwater development degree classification.

[0058] According to the five classification tasks, the cross entropy loss is defined, and the total loss is the weighted sum of the losses of the five classification tasks.

[0059] Select Adam optimizer, set learning rate Lr to 0.001, and weight decay to 10 -4 .

[0060] Set the learning rate scheduling strategy, use StepLR, set the step size to 30, and set the scheduling factor to 0.1.

[0061] Start the training loop, set epoch to 100, and iterate for a predefined number of rounds:

[0062] a: Set the model to training mode at the beginning of each epoch.

[0063] b: Iterate over each batch of the training dataset:

[0064] The training images are fed into the model to obtain prediction results for the five classification tasks.

[0065] Calculate the loss for each task and sum them up to get the total loss.

[0066] Backpropagation calculates the gradients and uses the optimizer to update the model parameters.

[0067] c: After each epoch, evaluate the model performance on the validation set:

[0068] Set the model to evaluation mode.

[0069] Calculate the loss on the validation set and the accuracy of each task.

[0070] d: Record and print the training loss, validation loss, and accuracy of each task of the current epoch.

[0071] e: If the current model performance is better than the previous best performance, save the model parameters.

[0072] f: Adjust the learning rate according to the learning rate scheduling strategy.

[0073] After training is complete, the best performing model parameters are loaded and a final evaluation is performed on the test set. The accuracy, precision, recall, and F1 score for each classification task are calculated and reported.

[0074] Analyze the performance of the model on different tasks and identify tasks with poor performance. Based on the analysis results, modify the loss function or use Dropout to prevent overfitting.

[0075] According to the test results, repeat the above process several times until the requirements are met.

[0076] After the deep learning model is trained, the image to be recognized is obtained and input into the trained deep learning model. The specific processing process is as follows:

[0077] The image to be recognized is input into the ResNet50 backbone network, and four residual blocks are used to extract features of four scales in sequence, namely the first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4.

[0078] Specifically, the image to be recognized is 224*224*3. The image first passes through a 7*7 convolution layer with a step size of 2, then passes through BN, ReLU activation and a 3*3 maximum pooling layer (with a step size of 2), and then the image passes through 4 main residual blocks in sequence:

[0079] The first residual block (3 residual units): outputs the first scale feature C1 (56*56*256).

[0080] The second residual block (4 residual units): outputs the second scale feature C2 (28*28*512).

[0081] The third residual block (6 residual units): outputs the third scale feature C3 (14*14*1024).

[0082] The fourth residual block (3 residual units): outputs the fourth scale feature C4 (7*7*2048).

[0083] Each residual block contains multiple bottleneck structures, each bottleneck includes 1*1 convolution dimensionality reduction, 3*3 convolution processing, and 1*1 convolution dimensionality increase.

[0084] The first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4 are input into the feature pyramid network, and the four scale features C i Both apply 1*1 convolution and adjust the number of channels to the same value.i =Conv1*1(C i ); Following the top-down path, starting from the top layer, the high-level scale features are upsampled and added to the next layer to form a four-layer pyramid feature A i , A i =B i +Upsample(A i+1 ); for each pyramid feature A i Apply 3*3 convolution for smoothing and get 4 smooth features P i , P i =Conv3*3(A i ); where i=1,2,3,4.

[0085] Specifically, first use 1*1 convolution to adjust the number of channels of the first scale feature C1, the second scale feature C2, the third scale feature C3, and the fourth scale feature C4:

[0086] B1=Conv1*1(C1)→(56*56*256).

[0087] B2=Conv1*1(C2)→(28*28*256).

[0088] B3=Conv1*1(C3)→(14*14*256).

[0089] B4=Conv1*1(C4)→(7*7*256).

[0090] Then build the feature pyramid from top to bottom:

[0091] A4=B4.

[0092] A3=B3+Upsample(A4) (upsampling using nearest neighbor interpolation).

[0093] A2=B2+Upsample(A3).

[0094] A1=B1+Upsample(A2).

[0095] Use 3*3 convolution smoothing:

[0096] P1=Conv3*3(A1)→(56*56*256).

[0097] P2=Conv3*3(A2)→(28*28*256).

[0098] P3=Conv3*3(A3)→(14*14*256).

[0099] P4=Conv3*3(A4)→(7*7*256).

[0100] Input P1 into the first classification head to obtain the first classification result, input P1 and P2 into the second classification head to obtain the second classification result, input P2, P3 and P4 into the third classification head to obtain the third classification result, input P1, P2, P3 and P4 into the fourth classification head to obtain the fourth classification result, and input P3 and P4 into the fifth classification head to obtain the fifth classification result.

[0101] P1 is input into the first classification head, and color features are gradually extracted through three consecutive convolutional layers. Batch normalization and ReLU activation are performed after each convolution. During the convolution process, the number of channels is gradually reduced, and the features are compressed into a one-dimensional vector using global average pooling. Then, they are mapped to the output space of five categories through a fully connected layer to obtain the first classification result.

[0102] Specifically, the first classification head is for surrounding rock color classification, and its input is: P1 (56*56*256), which represents the high-resolution features of the image. First, P1 passes through three consecutive convolutional layers:

[0103] Conv1: 3*3, 256→128 channels, after BN layer and ReLU activation.

[0104] Conv2: 3*3, 128→64 channels, after BN layer and ReLU activation.

[0105] Conv3: 3*3, 64→32 channels, after BN layer and ReLU activation.

[0106] Through three consecutive convolutional layers, BN and ReLU activation, nonlinearity can be increased, and more advanced features can be gradually extracted while reducing the complexity of the feature map.

[0107] Global average pooling is used: 56*56*32→1*1*32 to retain the average activation of each channel, which is equivalent to a global summary of the entire feature map. After the fully connected layer: 35→5, that is, this 32-dimensional vector is mapped to 5 output units through a fully connected layer, corresponding to 5 surrounding rock color categories. The output of the fully connected layer is converted into a probability distribution through the softmax function.

[0108] P1 and P2 are input into the second classification head, and width-related features are extracted through one 3*3 convolutional layer and two 3*3 hole convolutional layers. The ReLU activation function is applied after each convolutional layer. Then the processed P1 and P2 are upsampled or downsampled to the same size, and then concatenated in the channel dimension. The concatenated features are fused through a 1*1 convolution, and the features are compressed into a one-dimensional vector using global average pooling. Then, they are mapped to the output space of 5 categories through a fully connected layer to obtain the second classification result.

[0109] Specifically, the second classification head is for crack width classification, and its input is: P1 (56*56*256) and P2 (28*28*256), combining information of different scales. The same processing is performed on P1 and P2 respectively:

[0110] 3*3 regular convolution, 256→128 channels, activated by ReLU.

[0111] 3*3 dilated convolution (dilated ratio = 2), 128→64 channels, activated by ReLU.

[0112] 3*3 dilated convolution (dilated ratio = 4), 64→32 channels, activated by ReLU.

[0113] The dilated convolution can expand the receptive field without increasing the number of parameters, which helps capture a wider range of contextual information. This is particularly useful for detecting crack width.

[0114] Feature adjustment: P1 is 56*56*32 after processing, and P2 is 28*28*32 after processing, which is upsampled to 56*56*32.

[0115] Feature splicing: 56*56*64 (32+32 channels), that is, splicing in the channel dimension.

[0116] 1*1 convolution fusion: 64→32 channels to fuse information at two scales.

[0117] Global average pooling: 56*56*32 → 1*1*32.

[0118] Fully connected layer: 32→5, that is, this 32-dimensional vector is mapped to 5 output units through a fully connected layer, corresponding to 5 crack width categories. The output of the fully connected layer is converted into a probability distribution through the softmax function.

[0119] P2, P3, and P4 are input into the third classification head. Each feature undergoes 1*1 convolution and self-attention mechanism independently. The three processed features are upsampled to the same size and spliced ​​in the channel dimension. The spliced ​​features pass through a self-attention layer to fuse multi-scale information. Finally, they are mapped to the output space of 5 categories through a fully connected layer to obtain the third classification result.

[0120] Specifically, the third classification head is for crack length classification, and its input is: P2 (28*28*256), P3 (14*14*256), P4 (7*7*256), covering more scale information. The same processing is performed on the three feature maps:

[0121] 1*1 convolution: 256→128 channels.

[0122] Self-attention mechanism: calculate the Q, K, V matrices, calculate the attention weights, and apply the attention weights to get the output.

[0123] The self-attention mechanism allows the model to focus on important regions in the feature map, which is very useful for detecting crack lengths that may span different scales.

[0124] Feature adjustment: After P2 processing, it is 28*28*128, after P3 processing, it is 14*14*128, which is upsampled to 28*28*128, and after P4 processing, it is 7*7*128, which is upsampled to 28*28*128.

[0125] Feature splicing: 28*28*384 (128*3 channels), that is, splicing in the channel dimension.

[0126] The self-attention mechanism is applied again to further integrate multi-scale information.

[0127] Fully connected layer: 384→5, that is, this 384-dimensional vector is mapped to 5 output units through a fully connected layer, corresponding to 5 crack length categories. The output of the fully connected layer is converted into a probability distribution through the softmax function.

[0128] P1, P2, P3 and P4 are input into the fourth classification head. Each feature is independently processed by 1*1 convolution and channel attention block. The processed 4 features are adjusted to the same size and spliced ​​in the channel dimension. The spliced ​​features are passed through an additional channel attention block to further fuse multi-scale information. The features are compressed into a one-dimensional vector through global average pooling and then mapped to the output space of 5 categories through a fully connected layer to obtain the fourth classification result.

[0129] Specifically, the fourth classification head is for crack group number classification, and its input is: P1 (56*56*256), P2 (28*28*256), P3 (14*14*256), P4 (7*7*256). Each feature map is processed independently:

[0130] 1*1 convolution: 256→128 channels.

[0131] Channel attention block: global average pooling and maximum pooling, shared multi-layer perceptron, fusion of average pooling and maximum pooling results.

[0132] The channel attention block uses global average pooling and max pooling to generate channel descriptors, which are then processed by a shared multi-layer perceptron and finally fused to re-weight the importance of each channel. This helps the model focus on the feature channels that are most relevant for classifying the number of crack groups.

[0133] Feature adjustment: After P1 processing, it is 56*56*128, which is downsampled to 14*14*128; after P2 processing, it is 28*28*128, which is downsampled to 14*14*128; after P3 processing, it is 14*14*128; after P4 processing, it is 7*7*128, which is upsampled to 14*14*128.

[0134] Feature splicing: 14*14*512 (128*4 channels), that is, splicing in the channel dimension.

[0135] The feature representation is further optimized by passing it through a channel attention block again.

[0136] Global average pooling: 14*14*512 → 1*1*512.

[0137] Fully connected layer: 512→5, that is, this 512-dimensional vector is mapped to 5 output units through a fully connected layer, corresponding to 5 crack group number categories. The output of the fully connected layer is converted into a probability distribution through the softmax function.

[0138] P3 and P4 are input into the fifth classification head, and each feature is subjected to two 3*3 convolutional layers to extract water development-related features. Batch normalization and ReLU activation are performed after each convolution. Then, an adaptive maximum pooling layer is used to obtain two feature sequences. These two feature sequences are spliced ​​in the time dimension to form a spliced ​​sequence, and then input into the bidirectional GRU for processing. The output of the GRU is average pooled and then mapped to the output space of 5 categories through the fully connected layer to obtain the fifth classification result.

[0139] Specifically, the fifth classification head is for groundwater development degree classification, and its input is: P3 (14*14*256), P4 (7*7*256), mainly focusing on the medium and low resolution features of the image. Each feature map is processed independently:

[0140] The first 3*3 convolution: 256→128 channels, after BN layer and ReLU activation.

[0141] The second 3*3 convolution: 128→64 channels, after BN layer and ReLU activation.

[0142] Adaptive max pooling: Convert P3 and P4 into sequences of 1*14*64 and 1*7*64 respectively.

[0143] Feature sequence splicing: 1*21*64 (14+7 time dimensions).

[0144] Bidirectional GRU processing: The input is 21 time steps, each step has 64-dimensional features, the GRU hidden state dimension is 128 (64 in each direction), and the output is 21*128. GRU can capture long-term dependencies in the sequence, which is very helpful for understanding the spatial distribution pattern of groundwater development.

[0145] Average pooling: 21*128 → 1*128.

[0146] Fully connected layer: 128→5, that is, this 128-dimensional vector is mapped to 5 output units through a fully connected layer, corresponding to 5 groundwater development degree categories. The output of the fully connected layer is converted into a probability distribution through the softmax function.

[0147] Finally, the five classification heads output the probability distribution of the five categories respectively, and for each task, the category with the highest probability is selected as the prediction result. That is, the final output includes: surrounding rock color category (one of the five categories), crack width category (one of the five categories), crack length category (one of the five categories), crack group number category (one of the five categories), and groundwater development degree category (one of the five categories).

[0148] Specifically, in one embodiment of the present invention, step S3 includes:

[0149] S31 sets a standard value range, divides the standard value range into five sub-ranges, and corresponds five categories of five classification results to the five sub-ranges one by one, wherein the five classification results are respectively recorded as a first classification result, a second classification result, a third classification result, a fourth classification result, and a fifth classification result;

[0150] S32 obtains the predicted probability values ​​of the first classification result, the second classification result, the third classification result, the fourth classification result, and the fifth classification result, and calculates the individual scores of the five classification results according to the conversion relationship between the predicted probability values ​​and the corresponding sub-ranges;

[0151] S33 sets weights for the five classification results and calculates a comprehensive score based on the weights and individual scores.

[0152] Let's take a specific example to illustrate:

[0153] The standard value range is set to 0-100 points, and the standard value range is divided into 5 sub-ranges, namely the first sub-range 90-100 points, the second sub-range 80-90 points, the third sub-range 70-80 points, the fourth sub-range 60-70 points, and the fifth sub-range below 60 points.

[0154] The first sub-range, the second sub-range, the third sub-range, the fourth sub-range, the fifth sub-range and the five categories of the five classification results are matched, as shown in Table 1:

[0155] Table 1

[0156]

[0157]

[0158] According to step S2, the five classification results of the image to be identified are obtained by using the deep learning model, and the sub-ranges corresponding to the five classification results are found according to Table 1. According to the predicted probability values ​​of the five classification results, the individual scores of the five classification results are calculated:

[0159]

[0160] In the formula, Represents the single score of the j1-th classification result, j1 = 1, 2, 3, 4, 5; Indicates the lower limit of the value range of the sub-range j2 corresponding to the j1-th classification result, j2 = 1, 2, 3, 4, 5; Represents the predicted probability value of the j1th classification result; Represents the prediction threshold of the j1th classification result; Represents the adjustment factor for sub-range j2.

[0161] For example, the fourth classification result is of the category of more, that is, it belongs to the fourth sub-range, and its predicted probability value is 0.75, that is, j1=4. The prediction threshold of the classification result is set to 0.6, that is, Set the adjustment factor of the fourth sub-range to twice its range span, that is, the range span of the fourth sub-range is 10. Then the single score of the fourth classification result is: S4=60+(0.75-0.6)*40=66 points.

[0162] Specifically, The value range of falls within the value span of its corresponding subrange.

[0163] According to the formula for single scoring, the single scores of the five classification results are calculated in turn, and finally the comprehensive score is calculated:

[0164]

[0165] In the formula, S total is the comprehensive score, Represents the weight of the j1th classification result.

[0166] The formula for the comprehensive score is a weighted sum, so S total The value range is still 0-100 points.

[0167] In a specific example, the weight of the first classification result is 10%, the weight of the second classification result is 20%, the weight of the third classification result is 25%, the weight of the fourth classification result is 25%, and the weight of the fifth classification result is 20%.

[0168] Specifically, in one embodiment of the present invention, step S4 includes:

[0169] The value range of the comprehensive score is divided into 6 level ranges, and the 6 level ranges are set as surrounding rock level 1, surrounding rock level 2, surrounding rock level 3, surrounding rock level 4, surrounding rock level 5, and surrounding rock level 6 according to the comprehensive scores from high to low. According to the value of the comprehensive score calculated in step S3, it is assigned to the corresponding surrounding rock level to obtain the surrounding rock level of the image to be identified.

[0170] In actual identification, it is very rare for the comprehensive score to be 100 points. This is an ideal situation. It is regarded as surrounding rock level 1, 90-99 points as surrounding rock level 2, 80-89 points as surrounding rock level 3, 70-79 points as surrounding rock level 4, 60-69 points as surrounding rock level 5, and below 60 points as surrounding rock level 6.

[0171] According to the calculated comprehensive score, the surrounding rock level of the image to be identified can be obtained. total =78, then the surrounding rock level of the image to be identified is 4.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligently determining the surrounding rock grade of a tunnel face, characterized in that: include: S1 builds a deep learning model and iteratively trains the deep learning model using the training set until the model converges to obtain a trained deep learning model; S2 obtains the image to be identified, inputs the image to be identified into the trained deep learning model, and outputs five classification results; S3 standardizes the five classification results, and then uses a comprehensive scoring method to comprehensively calculate the five classification results of the standard value range to obtain a comprehensive score; S4 determines the comprehensive score according to the surrounding rock grade rule to obtain the surrounding rock grade of the image to be identified.

2. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 1, characterized in that: The deep learning model includes ResNet50 backbone network, feature pyramid network and classification network, among which: The ResNet50 backbone network includes a one-dimensional convolutional layer, a BN layer, a maximum pooling layer, and 4 residual blocks. The first residual block includes 3 residual units, the second residual block includes 4 residual units, the third residual block includes 6 residual units, and the fourth residual block includes 3 residual units. The feature pyramid network adopts a top-down path, including 4 convolutional layers, and each convolutional layer uses lateral connections; The classification network includes 5 classification heads. The first classification head includes 3 convolutional layers, a global average pooling layer and a fully connected layer. The second classification head includes 3 consecutive convolutional layers, a global average pooling layer and a fully connected layer. The third classification head includes a convolutional layer, a self-attention layer, an average pooling layer and a fully connected layer. The fourth classification head includes a convolutional layer, a channel attention block, a global average pooling layer and a fully connected layer. The fifth classification head includes 2 convolutional layers, an adaptive maximum pooling layer, a bidirectional GRU layer and a fully connected layer.

3. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 2, characterized in that: The first classification head is used for surrounding rock color classification; the second classification head is used for crack width classification; the third classification head is used for crack length classification; the fourth classification head is used for crack group number classification; the fifth classification head is used for groundwater development degree classification.

4. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 3, characterized in that: The process of deep learning model processing the image to be recognized is as follows: The image to be recognized is input into the ResNet50 backbone network, and four residual blocks are used to extract features of four scales in sequence, namely the first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4; The first scale feature C1, the second scale feature C2, the third scale feature C3 and the fourth scale feature C4 are input into the feature pyramid network, and the four scale features C i Both apply 1*1 convolution and adjust the number of channels to the same value. i =Conv1*1(C i ); Following the top-down path, starting from the top layer, the high-level scale features are upsampled and added to the next layer to form a four-layer pyramid feature A i , A i =B i +Upsample(A i+1 ); for each pyramid feature A i Apply 3*3 convolution for smoothing and get 4 smooth features P i , P i =Conv3*3(A i ); where i = 1, 2, 3, 4; Input P1 into the first classification head to obtain the first classification result, input P1 and P2 into the second classification head to obtain the second classification result, input P2, P3 and P4 into the third classification head to obtain the third classification result, input P1, P2, P3 and P4 into the fourth classification head to obtain the fourth classification result, and input P3 and P4 into the fifth classification head to obtain the fifth classification result.

5. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 4, characterized in that: P1 is input into the first classification head, and color features are gradually extracted through three consecutive convolutional layers. Batch normalization and ReLU activation are performed after each convolution. During the convolution process, the number of channels is gradually reduced, and the features are compressed into a one-dimensional vector using global average pooling. Then, they are mapped to the output space of five categories through a fully connected layer to obtain the first classification result. P1 and P2 are input into the second classification head, and width-related features are extracted through one 3*3 convolution layer and two 3*3 hole convolution layers. The ReLU activation function is applied after each convolution layer. After that, the processed P1 and P2 are upsampled or downsampled to the same size, and then spliced ​​in the channel dimension. The spliced ​​features are fused through a 1*1 convolution, and the features are compressed into a one-dimensional vector using global average pooling. Then, they are mapped to the output space of 5 categories through a fully connected layer to obtain the second classification result; P2, P3, and P4 are input into the third classification head. Each feature is independently subjected to 1*1 convolution and self-attention mechanism. The three processed features are upsampled to the same size and spliced ​​in the channel dimension. The spliced ​​features are passed through a self-attention layer to fuse multi-scale information. Finally, they are mapped to the output space of 5 categories through a fully connected layer to obtain the third classification result. P1, P2, P3 and P4 are input into the fourth classification head. Each feature is independently processed by 1*1 convolution and channel attention block. The processed 4 features are adjusted to the same size and spliced ​​in the channel dimension. The spliced ​​features are passed through an additional channel attention block to further fuse multi-scale information. The features are compressed into a one-dimensional vector through global average pooling and then mapped to the output space of 5 categories through a fully connected layer to obtain the fourth classification result. P3 and P4 are input into the fifth classification head, and each feature is subjected to two 3*3 convolutional layers to extract water development-related features. Batch normalization and ReLU activation are performed after each convolution. Then, an adaptive maximum pooling layer is used to obtain two feature sequences. These two feature sequences are spliced ​​in the time dimension to form a spliced ​​sequence, and then input into the bidirectional GRU for processing. The output of the GRU is average pooled and then mapped to the output space of 5 categories through the fully connected layer to obtain the fifth classification result.

6. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 4, characterized in that: The first category results include single, relatively single, relatively turbid, turbid, and extremely turbid; the second category results include extremely narrow, narrow, relatively narrow, relatively wide, and wide; the third category results include less than extremely short, short, relatively short, relatively long, and long; the fourth category results include very little, few, fewer, more, and much; the fifth category results include dry, dot-shaped, drop-shaped, line-shaped, and strand-shaped.

7. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 1, characterized in that: Step S3 includes: S31 sets a standard value range, divides the standard value range into five sub-ranges, and corresponds five categories of five classification results to the five sub-ranges one by one, wherein the five classification results are respectively recorded as a first classification result, a second classification result, a third classification result, a fourth classification result, and a fifth classification result; S32 obtains the predicted probability values ​​of the first classification result, the second classification result, the third classification result, the fourth classification result, and the fifth classification result, and calculates the individual scores of the five classification results according to the conversion relationship between the predicted probability values ​​and the corresponding sub-ranges; S33 sets weights for the five classification results and calculates a comprehensive score based on the weights and individual scores.

8. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 7, characterized in that: The calculation formula for the single score is: In the formula, Represents the single score of the j1th classification result, j1 = 1, 2, 3, 4, 5; Indicates the lower limit of the value range of the sub-range j2 corresponding to the j1-th classification result, j2 = 1, 2, 3, 4, 5; Represents the predicted probability value of the j1th classification result; Represents the prediction threshold of the j1th classification result; Represents the adjustment factor for sub-range j2.

9. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 8, characterized in that: The formula for calculating the comprehensive score is: In the formula, S total is the comprehensive score, Represents the weight of the j1th classification result.

10. The method for intelligently determining the surrounding rock grade of a tunnel face according to claim 1, characterized in that: Step S4 includes: The value range of the comprehensive score is divided into 6 level ranges, and the 6 level ranges are set as surrounding rock level 1, surrounding rock level 2, surrounding rock level 3, surrounding rock level 4, surrounding rock level 5, and surrounding rock level 6 according to the comprehensive scores from high to low. According to the value of the comprehensive score calculated in step S3, it is assigned to the corresponding surrounding rock level to obtain the surrounding rock level of the image to be identified.

Citation Information

Cited By

  • Shooting-based surrounding rock classification method, device, equipment and medium

    CN119785120A