Intelligent analysis method for completeness degree of tunnel face

Abstract

The invention belongs to the technical field of rock mass integrity degree judgment, and particularly discloses a tunnel face integrity degree intelligent analysis method, which comprises the following steps: acquiring a complete tunnel face image, and preprocessing the complete tunnel face image; cutting the preprocessed complete tunnel face image into a local image, and constructing a database by using the local image; constructing a convolutional neural network model, and inputting the image data in the database into the convolutional neural network model for training, testing and verification to obtain a tunnel face integrity intelligent analysis model; and identifying the tunnel face integrity degree by using the tunnel face integrity degree intelligent analysis model, and outputting an identification result. The method has an obvious effect on judging the completeness degree of the tunnel face with uniform rock mass conditions, and has a certain effect on judging the tunnel face with large rock mass condition differences. The problems that an existing rock mass integrity degree judgment mode is high in dependency on personnel and large in risk are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of rock mass integrity determination, and particularly relates to the design of an intelligent analysis method for tunnel face integrity. Background Art

[0002] At present, the basic quality index (BQ) method is widely used in underground engineering construction in my country as an important basis for surrounding rock classification. The integrity of the rock mass at the tunnel face is an important indicator for surrounding rock classification in the BQ method. According to the "Engineering Rock Mass Classification Standard" GB / T 50218-2014, the integrity of the rock mass at the tunnel face can be judged by qualitative and quantitative methods. Among them, qualitative judgment is mainly based on factors such as the number of structural planes, average spacing and the degree of combination of the main structures to determine the integrity of the rock mass; quantitative judgment is based on the rock mass integrity index or the number of rock mass volume joints to determine the integrity of the rock mass.

[0003] For the integrity of the rock mass at the face, both qualitative and quantitative judgments require technicians to observe, measure and even test near the face. This method not only places high demands on the professional quality of technicians, but also the complex environment in underground projects, such as smoke, dust, noise, etc., will also affect the judgment of technicians. In addition, working in front of the face for a long time will also increase the risk of technicians suffering from disasters such as rock bursts and water inrush. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the existing rock mass integrity determination method is highly dependent on personnel and has high risks, and proposes an intelligent analysis method for the integrity of the tunnel face.

[0005] The technical solution of the present invention is: a method for intelligent analysis of the integrity of a tunnel face, comprising the following steps:

[0006] S1. Collecting a complete tunnel face image and preprocessing the complete tunnel face image;

[0007] S2. cropping the preprocessed complete tunnel face image into partial images, and using the partial images to construct a database;

[0008] S3. Construct a convolutional neural network model, and input the image data in the database into the convolutional neural network model for training, testing and verification, to obtain an intelligent analytical model of the integrity of the face;

[0009] S4. Use the intelligent analytical model of the tunnel face integrity to identify the tunnel face integrity and output the identification result.

[0010] The beneficial effects of the present invention are:

[0011] The present invention has an obvious effect on judging the integrity of a tunnel face with uniform rock mass conditions, and also has a good effect on judging a tunnel face with greatly different rock mass conditions.

[0012] Preferably, the preprocessing in step S1 includes image denoising and image enhancement.

[0013] Through image denoising and image enhancement preprocessing, the rock details are made clearer, effectively improving the comprehensive performance of the intelligent analytical model of the tunnel face integrity in the multi-classification problem of rock integrity.

[0014] Preferably, the local images in the database in step S2 are divided into five categories according to the degree of integrity of the rock mass: complete, relatively complete, relatively broken, broken and extremely broken.

[0015] The beneficial effects of the above preferred solution are:

[0016] A database consisting of five categories of local images was established according to the categories of rock mass integrity, which met the training requirements of the neural network and thus established a more accurate correspondence between input and output.

[0017] Preferably, the convolutional neural network model in step S3 adopts a HRNet neural network; the HRNet neural network includes an input module, a first transition module, a first feature fusion module, a second transition module, a second feature fusion module, a third transition module, a third feature fusion module and an output module connected in sequence.

[0018] The beneficial effects of the above preferred solution are:

[0019] By adopting the HRNet neural network, feature interaction is enhanced while maintaining high-resolution representation, which improves classification performance. At the same time, the feature fusion module is used to realize multiple high-resolution and low-resolution information exchange and fusion, so that the network can learn richer high-resolution features, thereby achieving good results in classification tasks.

[0020] Preferably, the step S3 specifically comprises the following steps:

[0021] S31. Construct a convolutional neural network model;

[0022] S32. Divide the database images into a training set, a validation set and a test set;

[0023] S33. Input the training set data into the convolutional neural network model for face integrity recognition training, and adjust the parameters of the convolutional neural network model using the validation set until the training times are reached and the training is completed;

[0024] S34. Input the test set into the trained convolutional neural network model, test the classification performance of the trained convolutional neural network model, and obtain an intelligent parsing model of the integrity of the face with classification performance reaching the standard.

[0025] Preferably, when the convolutional neural network model is trained in step S33, a multi-classification cross entropy loss function is used as the loss function.

[0026] Preferably, the calculation formula of the multi-classification cross entropy loss function is:

[0027]

[0028] Where L represents the multi-classification cross entropy loss function; C represents the number of categories; i represents the category of the integrity of the face; p i Indicates the predicted probability that the observed sample belongs to the category; y i Whether the prediction of the observed sample belonging to a class is correct or not.

[0029] The beneficial effects of the above preferred solution are:

[0030] By using the multi-classification cross entropy loss function, the pixel-level accuracy between the model prediction results and the true labels can be accurately measured, thereby accurately establishing the correspondence between the input and output of the neural network model.

[0031] Preferably, the step S4 specifically comprises the following steps:

[0032] S41. Inputting the local image of the to-be-analyzed tunnel face into the input module of the tunnel face integrity intelligent analysis model, and adjusting the size of the image data by convolution to obtain an adjusted local image;

[0033] S42. Generate a low-resolution subnet with feature maps of different depths and resolutions from the adjusted local image through three transition modules, and perform feature fusion on the low-resolution subnets of feature maps of different depths and resolutions through three feature fusion modules to obtain a feature map with uniform resolution;

[0034] S43. Based on the feature map with uniform resolution, the output module is used to identify the integrity of the tunnel face, and outputs the identification results of five rock integrity levels: complete, relatively complete, relatively broken, broken and extremely broken.

[0035] The beneficial effects of the above preferred solution are:

[0036] The feature fusion module of the HRNet neural network classification model realizes multiple high-resolution and low-resolution information exchange and fusion, which has good effects in classification tasks. At the same time, the HRNet neural network has a good effect on judging the integrity of the face. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Shown is a flow chart of an intelligent analysis method for tunnel face integrity. DETAILED DESCRIPTION

[0038] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0039] like Figure 1 As shown, a method for intelligent analysis of the integrity of a tunnel face comprises the following steps:

[0040] S1. Collecting a complete tunnel face image and preprocessing the complete tunnel face image;

[0041] S2. cropping the preprocessed complete tunnel face image into partial images, and using the partial images to construct a database;

[0042] S3. Construct a convolutional neural network model, and input the image data in the database into the convolutional neural network model for training, testing and verification, to obtain an intelligent analytical model of the integrity of the face;

[0043] S4. Use the intelligent analytical model of the tunnel face integrity to identify the tunnel face integrity and output the identification result.

[0044] In this embodiment, the preprocessing in step S1 includes image denoising and image enhancement.

[0045] In this embodiment, the local images in the database described in step S2 are divided into five categories according to the degree of rock integrity: complete (338 images), relatively complete (419 images), relatively broken (462 images), broken (631 images) and extremely broken (580 images).

[0046] In this embodiment, the convolutional neural network model in step S3 adopts a high-resolution neural network (HRNet); the HRNet neural network includes an input module, a first transition module, a first feature fusion module, a second transition module, a second feature fusion module, a third transition module, a third feature fusion module and an output module connected in sequence.

[0047] In this embodiment, step S3 specifically includes the following steps:

[0048] S31. Construct a convolutional neural network model;

[0049] S32. The database images are divided into training set (1944 images), validation set (243 images) and test set (243 images) in a ratio of 8:1:1. The network is trained using NVIDIA GTX 2080ti GPU and Adam optimizer, with batch size (batchsize) of 8, initial learning rate (lr) of 5e-4, and training epochs (epoches) set to 100;

[0050] S33. Input the training set data into the convolutional neural network model for face integrity recognition training, and adjust the parameters of the convolutional neural network model using the validation set until the training times are reached and the training is completed;

[0051] S34. Input the test set into the trained convolutional neural network model, test the classification performance of the trained convolutional neural network model, and obtain an intelligent parsing model of the integrity of the face with classification performance reaching the standard.

[0052] In this embodiment, when the convolutional neural network model is trained in step S33, the loss function adopts a multi-classification cross entropy loss function.

[0053] In this embodiment, the calculation formula of the multi-classification cross entropy loss function is:

[0054]

[0055] Where L represents the multi-classification cross entropy loss function; C represents the number of categories; i represents the category of the integrity of the face; p i Indicates the predicted probability that the observed sample belongs to the category; y i Indicates whether the prediction of the observed sample belonging to the category is correct or not. If the prediction is correct, the value is 1, otherwise the prediction is wrong and the value is 0.

[0056] In this embodiment, step S4 specifically includes the following steps:

[0057] S41. Inputting the local image of the to-be-analyzed tunnel face into the input module of the tunnel face integrity intelligent analysis model, and adjusting the size of the image data by convolution to obtain an adjusted local image;

[0058] S42. The adjusted local image is used to generate a low-resolution subnet with feature maps of different depths and resolutions by using different numbers of convolution kernels and downsampling times through three transition modules, and the low-resolution subnets of the feature maps of different depths and resolutions are feature-fused through three feature fusion modules to obtain a feature map with uniform resolution;

[0059] S43. Based on the feature map with uniform resolution, the output module is used to identify the integrity of the tunnel face, and the softmax function is used to output the identification results of five types of rock integrity: complete, relatively complete, relatively broken, broken, and extremely broken.

[0060] The effect of the present invention is tested and described below:

[0061] 1. Test the classification performance of the HRNet neural network model for local images

[0062] 243 samples that were not used for training were input into the two trained models for testing, and compared with the actual results. The actual classification performance of the model for rock mass integrity was judged by multiple evaluation indicators. The evaluation indicators selected were true positive (TP), true negative (TN), false positive (FP), false negative (FN), precision (Precision), recall (Recall), MicroF1 Score, and Macro F1 Score. Since it is a multi-classification problem, true positive means that the prediction is category a and the actual category is also category a; true negative means that the prediction is not category a and the actual category is not category a; false positive means that the prediction is category a and the actual category is not category a; false negative means that the prediction is not category a and the actual category is not category a. a refers to any classification category.

[0063] Precision refers to the ratio of the number of samples correctly predicted by the model as category a to the number of samples predicted by the model as category a. The calculation formula for precision is as follows:

[0064]

[0065] Recall rate refers to the ratio of the number of samples correctly predicted by the model as category a to the number of actual category a samples. The calculation formula for recall rate is as follows:

[0066]

[0067] Generally speaking, when precision is higher, recall is lower, and when recall is higher, precision is lower. Micro F1Score combines precision and recall to measure the classification performance of the model on category a. The value of Micro F1 Score is between 0 and 1. The closer to 1, the better the performance of the model in the classification of category a. The calculation formula of Micro F1Score is as follows:

[0068]

[0069] Macro F1 Score averages the Micro F1 Scores of all categories to assess the classification performance of the model on all categories. The calculation formula of Macro F1 Score is as follows:

[0070]

[0071] In the formula, n is the total number of classification categories.

[0072] Among the 243 samples, 34 were complete, 42 were relatively complete, 46 were relatively broken, 63 were broken, and 58 were extremely broken. The test results of the 243 samples were statistically analyzed to form a confusion matrix. The evaluation indicators of each model were calculated based on the confusion matrix. The Macro F1 Score of the HRNet model after image enhancement was 0.938, which was higher than 0.901 of the HRNet model without image enhancement. This shows that image enhancement effectively improves the comprehensive performance of the model in the multi-classification problem of rock mass integrity. The Micro F1 Score of each category of the HRNet model after image enhancement is higher than that of the HRNet model without image enhancement, indicating that image enhancement is effective for rock masses of each integrity category. At the same time, the contrast of the image after image enhancement is stronger, making the details of the rock mass clearer, and the dark areas originally caused by light problems are improved to a certain extent, making the structural surface and individual rock blocks clearer. And because the rock mass integrity classification is not very sensitive to color information (it may even be disturbed by color information), it is more necessary to pay attention to the brightness difference between the rock and the structural surface boundary in the rock mass. After image enhancement, the color information is weakened and the brightness contrast is emphasized, which plays a key role in enhancing the model's ability to capture various rock mass details and features. This confirms the effectiveness of image enhancement technology for rock mass integrity classification problems and provides important support for improving classification accuracy and model practicality.

[0073] In summary, the Macro F1 Score of the HRNet model is 0.938, which shows that the HRNet model can achieve the rock integrity classification task with excellent performance.

[0074] 2. Test the HRNet neural network model’s ability to determine the integrity of the entire face

[0075] By selecting 20 tunnel face images that did not participate in the model training for testing, the process of analyzing the tunnel face integrity of the present invention is explained and the judgment effect of the present invention on the tunnel face integrity is determined.

[0076] First, one of the collected complete face images was processed by image denoising and image enhancement technology. Second, the processed image was cropped to remove useless information for judging the integrity of the face, such as rock debris in front of the face, and then the image size was adjusted and cropped into 8 partial images. Third, the partial images were input into the trained neural network model for integrity judgment. The model predicted 7 of the partial images as broken and 1 as extremely broken, accounting for 7 / 8 of the broken categories. Most of the face was broken rock mass, and it was reasonable to judge the integrity of the face rock mass as broken.

[0077] According to the above method, 20 tunnel face images were subjected to integrity analysis test, and the statistical results are shown in Table 1. It can be seen that the tunnel face image integrity analysis results of the present invention are basically consistent with the integrity of the tunnel face rock mass actually determined on site.

[0078] Table 1 Analysis results of rock mass integrity at the tunnel face

[0079]

[0080] In summary, this method is more effective in judging the integrity of tunnel faces with uniform rock mass conditions, and has a certain effect on tunnel faces with large differences in rock mass conditions. It is worth noting that when there are many integrity categories in the local image of the tunnel face and the proportion of the main integrity category is not large enough, the integrity of the tunnel face should be determined in combination with the actual situation on site.

[0081] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of tunnel face integrity, characterized in that: The following steps are involved: S1. Collecting a complete tunnel face image and preprocessing the complete tunnel face image; S2. cropping the preprocessed complete tunnel face image into partial images, and using the partial images to construct a database; S3. Construct a convolutional neural network model, and input the image data in the database into the convolutional neural network model for training, testing and verification, to obtain an intelligent analytical model of the integrity of the face; S4. Use the intelligent analytical model of the tunnel face integrity to identify the tunnel face integrity and output the identification result.

2. The intelligent analysis method for tunnel face integrity according to claim 1 is characterized by: The preprocessing in step S1 includes image denoising and image enhancement.

3. The intelligent analysis method for tunnel face integrity according to claim 1 is characterized by: The local images in the database in step S2 are divided into five categories according to the degree of integrity of the rock mass: complete, relatively complete, relatively broken, broken and extremely broken.

4. The intelligent analysis method for tunnel face integrity according to claim 1 is characterized by: The convolutional neural network model in step S3 adopts the HRNet neural network; the HRNet neural network includes an input module, a first transition module, a first feature fusion module, a second transition module, a second feature fusion module, a third transition module, a third feature fusion module and an output module connected in sequence.

5. The intelligent analysis method for tunnel face integrity according to claim 1 is characterized in that: The step S3 specifically comprises the following steps: S31. Construct a convolutional neural network model; S32. Divide the database images into a training set, a validation set and a test set; S33. Input the training set data into the convolutional neural network model for face integrity recognition training, and adjust the parameters of the convolutional neural network model using the validation set until the training times are reached and the training is completed; S34. Input the test set into the trained convolutional neural network model, test the classification performance of the trained convolutional neural network model, and obtain an intelligent parsing model of the integrity of the face with classification performance reaching the standard.

6. The intelligent analysis method for tunnel face integrity according to claim 5 is characterized by: When the convolutional neural network model is trained in step S33, the loss function adopts a multi-classification cross entropy loss function.

7. The intelligent analysis method for tunnel face integrity according to claim 6 is characterized in that: The calculation formula of the multi-classification cross entropy loss function is: Where L represents the multi-classification cross entropy loss function; C represents the number of categories; i represents the category of the integrity of the face; p i Indicates the predicted probability that the observed sample belongs to the category; y i Indicates whether the prediction that the observed sample belongs to a category is correct or not.

8. The intelligent analysis method for tunnel face integrity according to claim 1 is characterized in that: The step S4 specifically comprises the following steps: S41. Inputting the local image of the to-be-analyzed tunnel face into the input module of the tunnel face integrity intelligent analysis model, and adjusting the size of the image data by convolution to obtain an adjusted local image; S42. Generate a low-resolution subnet with feature maps of different depths and resolutions from the adjusted local image through three transition modules, and perform feature fusion on the low-resolution subnets of feature maps of different depths and resolutions through three feature fusion modules to obtain a feature map with uniform resolution; S43. Based on the feature map with uniform resolution, the output module is used to identify the integrity of the tunnel face, and outputs the identification results of five rock integrity levels: complete, relatively complete, relatively broken, broken and extremely broken.

Images