An intelligent classification method and device for surrounding rock of tunnel heading face based on deep learning

Abstract

The present invention relates to the field of tunnel exploration, in particular to an intelligent classification method and device for surrounding rock of tunnel heading face based on deep learning. The present invention labels and preprocesses the surrounding rock photo samples of the tunnel heading face to establish a heading face image database for training; then, a surrounding rock classification prediction model based on deep learning is built, and the model is trained according to the heading face image database, so as to construct a more accurate recognition prediction model. Moreover, the present invention uses image cutting to segment the photo of the heading face to be detected, predicts each sub-image separately, and finally makes a comprehensive evaluation prediction according to the prediction results of each sub-image, making the data on which the prediction results are based more comprehensive, and also realizing the efficient and accurate discrimination of the surrounding rock of the heading face.

Description

Technical Field

[0001] The present invention relates to the field of tunnel exploration, and in particular to an intelligent classification method and device for surrounding rock of tunnel face based on deep learning. Background Technique

[0002] Surrounding rock classification is an important basic work in tunnel engineering, which is an important basis for tunnel support structure design and determination of excavation methods, and also the basis for tunnel safety risk assessment. From the development process of surrounding rock classification at home and abroad, the main traditional methods for surrounding rock classification in tunnel engineering include single-factor index classification, multi-factor index classification, multi-factor comprehensive classification method, etc. Its development trend is from single-factor index to multi-factor index comprehensive evaluation, from qualitative analysis and empirical judgment to quantitative analysis, and from manual identification to machine learning intelligent identification. At present, the surrounding rock classification in tunnel engineering mainly focuses on the exploration and design stage, and the rock mass parameter information is obtained through drilling and geophysical exploration means for evaluation and classification. However, the surrounding rock grades divided in the exploration and design stage are relatively macroscopic, and often cannot accurately divide the surrounding rock grades of local tunnel sections during the construction stage. At this time, it is necessary to keep up with the changes of the surrounding rock on site in time and determine the surrounding rock grade, and propose design changes. Therefore, there are certain limitations. To make up for the limitations of surrounding rock classification in the exploration and design stage, it is very necessary to evaluate the surrounding rock grade during the construction stage. At present, the surrounding rock evaluation during the construction stage is mainly carried out through on-site geological reconnaissance and advanced geological prediction methods. This method often requires manual collection of indexes such as surrounding rock strength, rock mass integrity, groundwater, etc. for comprehensive analysis, and often has problems such as large work intensity and low efficiency, and there is a certain lag, which has a greater impact on the tunnel construction progress, resulting in difficulties and challenges in surrounding rock classification by traditional methods. Since the surrounding rock of the tunnel face is basically grade III, IV, and V, the present invention only discusses the surrounding rock classification from grade III to grade V.

[0003] Determining the grade of the surrounding rock ahead through the surrounding rock of the tunnel face during the tunnel construction stage is the most intuitive and effective method. However, this method usually relies on the experience of technicians. With the development of current artificial intelligence technology, it has certain feasibility to carry out the surrounding rock grade evaluation based on the exposed surrounding rock images through deep learning technology instead of the on-site work of technicians. However, due to the large area of the tunnel face and the differences in the lithology and structural features of the surrounding rock in each area within the exposed range, that is, the geological conditions revealed in different parts of the entire tunnel face are different, it is necessary to evaluate the overall stability of the surrounding rock according to the actual situation. However, currently there are few mature technical solutions that can consider the influence of the deviation of lithology and structural features in different regions on the surrounding rock grade evaluation and realize automatic determination based on deep learning technology.

[0004] With the rapid development of information technology, artificial intelligence technology has gradually been introduced into the field of geotechnical engineering. In recent years, a large number of artificial intelligence methods have been applied to surrounding rock classification (Table 1). For example, the entropy weight fuzzy comprehensive evaluation model is used for the surrounding rock classification of extremely soft rock tunnels; the AlexNet convolutional neural network model is used to analyze and evaluate tunnel face images; the structural plane parameters of the tunnel face are extracted based on edge detection technology; and the KNN algorithm is used to achieve dynamic intelligent classification of surrounding rocks. For example, Patent CN202010017540.6 also discloses an intelligent prediction method and device for the surrounding rock level in front of the tunnel face of the drill-blast method. These artificial intelligence methods have brought great convenience to surrounding rock classification.

[0005] Table 1 Related research on artificial intelligence surrounding rock classification

[0006]

[0007] However, the above methods still have the following problems:

[0008] 1. The method for dynamically changing the surrounding rock grade in front of the tunnel face during the tunnel construction process is the advanced geological prediction method, including geological survey, geophysical exploration and drilling methods. However, the current methods generally have problems such as relying on the experience of technical personnel, high work intensity, low efficiency, and the prediction accuracy needs to be further improved.

[0009] 2. The tunnel face images contain a large amount of geological information, but their degree of informatization and automation processing is relatively low, and they cannot be efficiently applied to the classification of surrounding rock grades.

[0010] 3. The conventional image preprocessing of the deep learning model for surrounding rock image classification of the tunnel face, including cropping, position transformation, and color transformation, has certain disadvantages in improving the prediction accuracy, and there is a lack of a set of image preprocessing methods with strong applicability for surrounding rock image classification of the tunnel face.

[0011] 4. Due to the special nature of the surrounding rock of the tunnel face, the lithology and structural structure of the surrounding rock in different areas of the same tunnel face may be different. For example, the surrounding rock of the entire tunnel face is hard, with few fissures and good overall integrity, but there are particularly developed joint fissures in a small local area, or small karst caves and soft interlayers are developed. Since the proportion of the abnormal range is very small, ordinary image recognition models are likely to identify it as surrounding rock of a better type, but the actual surrounding rock quality is poor, that is, the deep learning model is prone to misjudgment.

[0012] That is, the existing methods are inefficient, the evaluation is not comprehensive enough, and the accuracy is not high enough. Therefore, there is a need for an intelligent classification method for tunnel face surrounding rock with higher efficiency and higher accuracy. Summary of the Invention

[0013] The purpose of the present invention is to overcome the problems of low efficiency, incomplete evaluation, and poor accuracy existing in the prior art, and to provide an intelligent classification method for surrounding rock of tunnel heading face based on deep learning.

[0014] In order to achieve the above-mentioned invention purpose, the present invention provides the following technical solutions:

[0015] An intelligent classification method for surrounding rock of tunnel heading face based on deep learning, comprising the following steps:

[0016] Input the photo of the heading face to be predicted, segment the photo of the heading face to be predicted into a group of sub-images, and sequentially input the group of sub-images into a pre-built surrounding rock classification prediction model for prediction, and output the surrounding rock classification prediction result of the photo of the heading face to be predicted according to the statistical information of the surrounding rock classification prediction results of each sub-image;

[0017] Among them, the surrounding rock classification prediction results include Class III, Class IV, and Class V; the segmentation process is to cut the photo of the heading face to be predicted into n*n sub-images of the same size, where n takes values of 3, 4, or 5; the surrounding rock classification prediction model includes the following training steps:

[0018] S1: Input sample data and perform annotation to form annotated sample data; the annotation is the surrounding rock classification prediction result corresponding to the sample data;

[0019] S2: Preprocess the annotated sample data;

[0020] S3: Input the preprocessed annotated sample data into the surrounding rock classification prediction model, and perform model training on the surrounding rock classification prediction model, and output the surrounding rock classification prediction model after model training. The present invention establishes a heading face image database for training by annotating and preprocessing the surrounding rock photo samples of the tunnel heading face; then, by building a surrounding rock classification prediction model based on deep learning and training the model according to the heading face image database, a more accurate prediction model is constructed. Moreover, the present invention uses image cutting to segment the photo of the heading face to be detected, predicts each sub-image separately, and finally comprehensively evaluates and predicts according to the prediction results of each sub-image, making the data on which the prediction results are based more comprehensive, and also realizing the efficient and accurate discrimination of the surrounding rock of the heading face.

[0021] As a preferred solution of the present invention, the surrounding rock classification prediction criteria are as follows:

[0022] (1) When N5≥n, it is determined that the surrounding rock grade of the photo of the heading face to be predicted is Class V;

[0023] (2) When N5 < n and N3≥n*(n - 1), it is determined that the surrounding rock grade of the photo of the heading face to be predicted is Class III;

[0024] (3) When N5 and N3 do not meet the conditions of (1) and (2), determine that the surrounding rock grade of the photo of the heading face to be predicted is grade IV;

[0025] Where N5 and N3 are respectively the numbers of grade V and grade III surrounding rock images in a group of the sub-images.

[0026] As a preferred solution of the present invention, the preprocessing in the step S1 includes image segmentation, image enhancement, and image coding;

[0027] Where the image enhancement includes image cropping, image flipping, image color transformation, image data normalization processing, and image data standardization processing. By performing preprocessing such as image segmentation and enhancement on the image, the present invention eliminates irrelevant information in the image, improves the display effect of the image, highlights the image features, and also expands the data set, facilitating subsequent recognition and analysis, thereby improving the model accuracy and enhancing the model stability.

[0028] As a preferred solution of the present invention, the image cropping includes CenterCrop mode, RandomCrop mode, FiveCrop mode, and / or RandomResizedCrop mode;

[0029] The image flipping includes horizontal flipping and / or vertical flipping;

[0030] The image color transformation includes grayscale transformation, brightness transformation, saturation transformation, contrast transformation, and / or hue transformation;

[0031] The image data normalization processing uses the torchvision.transforms.ToTensor function;

[0032] The image data standardization processing uses the torchvision.transforms.Normalize function.

[0033] As a preferred solution of the present invention, the transformation amplitude of the brightness transformation is ±20%;

[0034] The transformation amplitude of the saturation transformation is ±30%;

[0035] The transformation amplitude of the contrast transformation is ±30%;

[0036] The transformation amplitude of the hue transformation is ±30%.

[0037] As a preferred solution of the present invention, the preprocessing step uses the transforms.Compose() function to integrate the functions in the image enhancement step.

[0038] As a preferred solution of the present invention, the image encoding includes the following steps:

[0039] Encode the labeled sample data after image segmentation and image enhancement, and encode the labeled sample data of level III, level IV, and level V as "2", "1", and "0" in sequence.

[0040] As a preferred solution of the present invention, the surrounding rock classification prediction model adopts the ResNet50 network architecture.

[0041] As a preferred solution of the present invention, the training parameters of the model training in step S3 are set as follows:

[0042] Batch sample number: batch_size = 16;

[0043] Learning rate: lr = 0.001;

[0044] Number of iterations: epoch = 100.

[0045] An intelligent classification device for tunnel face surrounding rock based on deep learning includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described in any one of the above.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. By labeling and preprocessing the tunnel face surrounding rock photo samples, the present invention establishes a face image database for training; then, by building a surrounding rock classification prediction model based on deep learning and training the model according to the face image database, a more accurate prediction model is constructed. Moreover, the present invention uses image cutting to segment the to-be-detected face photo, predicts each sub-image separately, and finally comprehensively evaluates and predicts according to the prediction results of each sub-image, making the data on which the prediction results are based more comprehensive and realizing the efficient and accurate discrimination of the face surrounding rock.

[0048] 2. By preprocessing such as segmenting and enhancing the image, the present invention eliminates the irrelevant information in the image, improves the display effect of the image, highlights the image features, and also expands the data set, facilitating subsequent recognition and analysis, thereby improving the model accuracy and enhancing the model stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flow chart of an intelligent classification method for tunnel face surrounding rock based on deep learning according to Embodiment 1 of the present invention;

[0050] Figure 2 Schematic diagram of the training process of the surrounding rock grading prediction model in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 1 of the present invention;

[0051] Figure 3 Schematic diagram of random cropping of images in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0052] Figure 4 Schematic diagram of image flipping in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0053] Figure 5 Schematic diagram of color transformation in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0054] Figure 6 Schematic diagram of the connection mode between the BP neural network and the convolutional neural network in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0055] Figure 7 Structure diagram of the ResNet50 model in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0056] Figure 8 Structure diagram of the residual block in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0057] Figure 9 Schematic diagram of the train_loss curve in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0058] Figure 10 Schematic diagram of the Val_accuracy curve in a deep learning-based intelligent grading method for tunnel face surrounding rock described in Embodiment 2 of the present invention;

[0059] Figure 11 An intelligent grading device for tunnel face surrounding rock based on deep learning, which utilizes the intelligent grading method for tunnel face surrounding rock based on deep learning described in Embodiment 1 of the present invention. Detailed implementation manner

[0060] The present invention will be further described in detail below in combination with test examples and specific embodiments. However, it should not be understood that the scope of the above-mentioned subject matter of the present invention is limited to the following embodiments. All technologies implemented based on the content of the present invention belong to the scope of the present invention.

[0061] Embodiment 1

[0062] As Figure 1 shown, an intelligent classification method for surrounding rock of tunnel face based on deep learning includes the following steps:

[0063] Input the photo of the tunnel face to be predicted, segment the photo of the tunnel face to be predicted into a set of sub-images, and sequentially input this set of sub-images into a pre-built surrounding rock classification prediction model for prediction, and output the surrounding rock classification prediction result of the photo of the tunnel face to be predicted according to the statistical information of the surrounding rock classification prediction results of each sub-image;

[0064] Among them, the surrounding rock classification prediction results include Class III, Class IV, and Class V; the segmentation process is to cut the photo of the tunnel face to be predicted into n*n sub-images of the same size, where n takes the values of 3, 4, 5 (if the sub-images are too large, it will easily lead to misjudgment of the recognition result, and if the sub-images are too small, it will be difficult to obtain a good recognition effect due to low pixels), and classify the surrounding rock grades according to the following criteria:

[0065] (1) When N5≥n, it is determined that the surrounding rock grade of the photo of the tunnel face to be predicted is Class V;

[0066] (2) When N5 < n and N3≥n*(n - 1), it is determined that the surrounding rock grade of the photo of the tunnel face to be predicted is Class III;

[0067] (3) When N5 and N3 do not meet the conditions of (1) and (2), it is determined that the surrounding rock grade of the photo of the tunnel face to be predicted is Class IV;

[0068] Among them, N5 and N3 are respectively the numbers of Class V and Class III surrounding rock images in a set of the sub-images.

[0069] As Figure 2 shown, the surrounding rock classification prediction model includes the following training steps:

[0070] S1: Input sample data and perform annotation to form annotated sample data.

[0071] Among them, the annotation is the surrounding rock classification prediction result corresponding to the sample data; the preprocessing includes image segmentation, image enhancement, and image coding; the image enhancement includes image cropping, image flipping, image color transformation, image data normalization processing, and image data standardization processing;

[0072] The image cropping includes CenterCrop mode, RandomCrop mode, FiveCrop mode and / or RandomResizedCrop mode;

[0073] The image flipping includes horizontal flipping and / or vertical flipping;

[0074] The image color transformation includes grayscale transformation, brightness transformation, saturation transformation, contrast transformation and / or hue transformation;

[0075] The image data normalization process uses torchvision.transforms.ToTensor function;

[0076] The image data normalization process uses the torchvision.transforms.Normalize function.

[0077] S2: Preprocessing the labeled sample data.

[0078] S3: Input the pre-processed labeled sample data into the surrounding rock classification prediction model, perform model training on the surrounding rock classification prediction model, and output the trained surrounding rock classification prediction model. The surrounding rock classification prediction model adopts the ResNet50 network architecture.

[0079] Example 2

[0080] This embodiment is a specific embodiment of the embodiment 1 when n*n is 3*3.

[0081] Due to the special properties of the surrounding rock of the tunnel face, that is, the surrounding rock of the tunnel face is hard, has few cracks, and has good overall integrity, but joints and cracks are particularly developed in a small local area, or small caves and weak interlayers are developed. Since the abnormal range accounts for a very small proportion, the model prediction result is often Grade III or Grade IV surrounding rock, but the actual surrounding rock grade is Grade V.

[0082] Therefore, in order to make the prediction results closer to the true value, this embodiment first divides the required predicted tunnel face photo into 9 sub-images by 3*3, and then sequentially inputs the sub-images into the surrounding rock classification prediction model for prediction, and outputs the prediction results of each sub-image. Let N5, N4, and N3 be the number of surrounding rocks of grade V, grade IV, and grade III, respectively. The prediction results of each sub-image are processed and output as follows:

[0083] (1) When N5≥3, the surrounding rock grade of the entire tunnel face is determined to be Grade V;

[0084] (2) When N5<3 and N3≥6, the surrounding rock grade of the entire tunnel face is determined to be Grade III;

[0085] (3) When N5 and N3 do not meet the conditions in (1) and (2), it is determined that the surrounding rock grade of the entire heading face is grade IV.

[0086] The following is the construction process of the surrounding rock classification prediction model described in this embodiment

[0087] S1: Input sample data and perform annotation to form annotated sample data.

[0088] Collect the surrounding rock image data of the heading face as sample data:

[0089] The data of this embodiment is sourced from a tunnel along a certain expressway in Guangxi. The image acquisition method mainly relies on high-definition cameras and mobile phone high-definition cameras to take pictures of the heading face and transmit them back to the computer terminal for storage. The collected heading face images cover grade III, grade IV, and grade V surrounding rocks.

[0090] Perform classification of the surrounding rock grade of the heading face for the said sample data:

[0091] The classification of the surrounding rock grade is comprehensively determined by combining the on-site excavation and exposure situation, survey and design data, and the obtained rock physical parameters. The on-site excavation and exposure surrounding rock grade is mainly identified according to the "Code for Design of Highway Tunnels" including lithology, joint fissure development degree, roughness, interlayer development situation, groundwater development situation, dissolution development situation, and integrity, etc. The sample data collected in this embodiment is divided according to the above principles, and an image database of grade III, grade IV, and grade V surrounding rocks is established.

[0092] S2: Preprocess the said annotated sample data.

[0093] Before performing operations such as image segmentation, recognition and classification, and edge detection on the image, the original image needs to be enhanced to eliminate irrelevant information, improve the image effect, highlight the image features, expand the data set, facilitate computer recognition and analysis, so as to improve the model accuracy and enhance the model stability.

[0094] (I) Image cutting (expansion of image data set)

[0095] A large amount of training data is the basis for deep learning models to improve prediction accuracy. The acquisition and accumulation of original data is a rather cumbersome and labor-intensive task. Therefore, in order to obtain a sufficient number of heading face surrounding rock maps, this embodiment uses the image cutting method to cut the collected original image into 9 sub-images of the same size according to 3*3, so as to achieve the purpose of data expansion.

[0096] Since the features contained in some sub-images after the original image is cut will be reduced, it is necessary to re-manually identify and classify the surrounding rock grades of the cut sub-images, and eliminate some sub-images that are greatly affected by the initial support of the tunnel and the bottom surface of the tunnel excavation, so as to quickly and conveniently expand the database to ensure the basic training needs of the model. Through the Image module in the Python image library PIL, batch 3*3 cutting of the original image can be achieved (n*n cutting can also be achieved according to requirements).

[0097] (2) Image Enhancement

[0098] Image enhancement is an important measure to improve the generalization of the model and reduce the phenomenon of overfitting. Conventional image enhancement methods include padding, erasing, scaling, cropping, rotating, flipping, color change, normalization, and standardization, etc. In this embodiment, data enhancement is implemented based on PyTorch. PyTorch is an open-source Python machine learning library that can provide two advanced functions: one is tensor calculation with powerful GPU acceleration (such as NumPy); the other is a deep neural network containing an automatic differentiation system. Compared with other deep learning frameworks, PyTorch has great advantages in terms of flexibility, ease of use, and speed.

[0099] The following are the specific steps of the image enhancement method adopted in this embodiment:[

[0100] (1) Image Cropping:

[0101] CenterCrop mode: Crop the required size from the center of the image;

[0102] RandomCrop mode: Randomly crop the required size at a certain position in the picture;

[0103] FiveCrop mode: Crop 5 pictures of the specified size in the upper left, upper right, lower left, lower right, and center of the image;

[0104] RandomResizedCrop mode: First scale the picture according to a random area ratio, then adjust to a random width-height ratio, and finally randomly crop the picture at a certain position according to the required size. As Figure 3 shown, when RandomResizedCrop = 224, the schematic diagram of the example picture being randomly scaled, then adjusted to a random width-height ratio, and finally randomly cropped at a position according to the size of size=(224,224), where (a), (b), and (c) are the example surrounding rock photos, the random cropping effect 1 of the photo, and the random cropping effect 2 of the photo respectively.

[0105] ((ID=30)) Image Flipping:

[0106] It includes horizontal flipping and vertical flipping. The RandomHorizontalFlip(p=*) class flips the image horizontally with a probability (the default probability is 0.5). As shown in Figure 4 the following figure. Figure 4 (a) Example photo of surrounding rock, Figure 4 (b) is the randomly horizontally flipped example picture; The RandomVerticalFlip class flips the image vertically (i.e., up and down) with a probability. The principle is the same as that of random horizontal flipping.

[0107] (3) Image color transformation:

[0108] It includes grayscale, brightness, saturation, contrast, hue transformation, etc. By calling Grayscale, the image can be converted into a grayscale image (as shown in Figure 5 the following figure. Figure 5 (a) is the example photo of surrounding rock, Figure 5 (b) is the effect picture after being converted into a grayscale image); Calling ColorJitter(brightness, contrast, saturation, hue) can adjust the brightness, contrast, saturation, and hue of the image. In the embodiment, ColorJitter(brightness = 0.2, contrast = 0.3, saturation = 0.3, hue = 0.3) is selected to preprocess the data set, that is, the brightness fluctuates ±20% compared with the original image, the contrast fluctuates ±30% compared with the original image, the saturation fluctuates ±30% compared with the original image, and the hue fluctuates ±30% compared with the original image.

[0109] (4) Image data normalization processing:

[0110] Normalization is for the convenience of subsequent data processing and to ensure the acceleration of convergence during program operation. The specific role of normalization is to summarize and unify the statistical distribution of samples. Normalization between 0 - 1 is the statistical probability distribution, and normalization in a certain interval is the statistical coordinate distribution. Its purpose is to make incomparable data comparable while maintaining the relative relationship between the compared data. In this project, during the normalization process, the torchvision.transforms.ToTensor function is used to convert a PIL.Image with a value range of [0, 255] into a torch.FloadTensor with a shape of [C, H, W] and a value range of [0, 1.0].

[0111] (5) Image data standardization processing:

[0112] In this embodiment, after being normalized to [0, 1] by dividing by 255 through ToTensor, and then through Normalize calculation, the data is normalized to [-1, 1].

[0113] In this embodiment, torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) is used to normalize the images channel by channel (the mean becomes 0 and the standard deviation becomes 1), which can accelerate the convergence of the model. The selected mean and standard deviation values are calculated from millions of images in the Imagenet database.

[0114] Finally, through the transforms.Compose() function, multiple image preprocessing steps can be integrated together, making data preprocessing more flexible and convenient.

[0115] (III) Database Packing and Encoding of Labels

[0116] To facilitate model training and improve training efficiency, in this embodiment, Dataset in PyTorch is used to package the source files, scale, and other non-essential functions of the heading face image dataset, and then DataLoader is used to load the data for training. At the same time, the photos of grade III, IV, and V heading face surrounding rocks in Dataset are encoded, and they are encoded as "2", "1", and "0" in sequence.

[0117] S3: Input the preprocessed labeled sample data into the surrounding rock classification prediction model, and perform model training on the surrounding rock classification prediction model, and output the surrounding rock classification prediction model after model training.

[0118] The image recognition technology used in this embodiment is a deep learning technology based on a convolutional neural network. The convolutional neural network learns by directly extracting image features, and can obtain higher-level and more abstract feature expressions of the image, so that more essential features in the image can be obtained, making the learned features more accurate and generalizable. A convolutional neural network generally consists of an input layer, a hidden layer, and an output layer, where the hidden layer is mainly composed of a convolutional layer, a pooling layer, and a fully connected layer. The convolutional neural network uses a local connection method, which can effectively reduce the number of parameters in the neural network compared with the BP neural network, improve the calculation efficiency, and can also read the mutual connections between influencing factors, enhancing the reliability of the model. The connection methods of the BP neural network and the convolutional neural network are shown in Figure 6 as follows.

[0119] To give full play to the advantages of deep learning, for the classification of the surrounding rock grades of tunnel faces, through comparative analysis, in this embodiment, a ResNet50 model is selected to train the database and classify the surrounding rock grades. The advantage of ResNet50 lies in the proposed residual structure, and a very deep network structure (breaking through 1000 layers) is built. By preprocessing the data and using the BN (Batch Normalization) layer in the network, the problem of gradient disappearance or gradient explosion can be solved. Through research and comparison, it is found that the ResNet model is superior to the Alexnet model in terms of the number of network layers, number of iterations, accuracy, and loss rate, and can better handle the feature extraction of tunnel face pictures.

[0120] As Figure 7 shown, it is the structure diagram of the ResNet50 model. ResNet is divided into 5 stages. The structure of Stage 1 is relatively simple and can be regarded as the preprocessing of the INPUT. The latter 4 stages are all composed of Bottlenecks and have a relatively similar structure. Stage 2 has 3 Bottlenecks, including 1 Conv Block and 2 ID Blocks. The remaining 3 stages include 4, 6, and 3 Bottlenecks respectively. Among them, the Conv Block represents the residual block for adding dimensions, and each residual block contains three convolutional layers. ID Block×2 represents 2 residual blocks without changing dimensions. In this embodiment, the ResNet50 model will be built according to the following structure. After the model is built, the three types of tunnel face photos will be divided into a training set and a test set according to a ratio of 4:1. Through the operations of Dataset and DataLoader in pytorch, and then shuffle and mini-batch operations, the training set and validation set data can be divided into small data sets, and then the data will be passed into the ResNet50 convolutional neural network model for training.

[0121] The ResNet50 model incorporates a residual block structure. When building a deep network structure, the residual structure can reduce the network parameters and computational amount. As Figure 8 shown, it is the structure diagram of the residual block. ResNet contains two types of mappings: one is the identity mapping, which refers to the identity pointing line in the figure; the other is the residual mapping, which refers to the part except the "identity pointing line". So the final output is y = F(x) + x.

[0122] To enable the training model to give full play to its maximum advantages, reasonable parameter settings are crucial. In this embodiment, the parameters and hyperparameters are selected as follows:

[0123] (1) The batch size of the batch samples is batch_size = 16. If the batch_size is too small, the model won't have enough time to converge; if it's too large, the video memory will explode and the model won't converge to the bottom of the valley.

[0124] (2) The learning rate is lr = 0.001. If the learning rate is too large, the model may not converge and the loss will oscillate up and down continuously; if the learning rate is too small, the model will converge slowly, require more training time, and may be overfitted.

[0125] (3) The number of epochs is epoch = 100. If the value of epoch is too small, the model is not easy to converge; if it's too large, it's easy to cause overfitting.

[0126] (4) The selection of the convolution kernel size (kernel_size), number (kernels), stride, and padding parameters of the ResNet50 model is determined by the model itself and will not be modified in this embodiment.

[0127] After building the model according to the above scheme, the image dataset is trained and good training results are obtained. Finally, the ResNet50 model converges to train_loss = 0.18 (as Figure 9 shown), and the highest accuracy of the model is acc = 88.4% (as Figure 10 shown).

[0128] Verification of the surrounding rock classification prediction model:

[0129] In this embodiment, photos of the tunnel face of two expressways in Guangxi are selected for model training, and the trained model is used for the prediction of the actual surrounding rock grade of the tunnel face. The prediction method of cutting the original image into 9 sub-images of the same size according to 3 * 3 is adopted. The specific results are as follows:

[0130] (1) For a whole tunnel face photo of a certain tunnel section of a certain expressway, the predicted surrounding rock grade is grade IV. Through the sub-image prediction method, the obtained surrounding rock grade is grade V (N5 = 4, N4 = 5, N3 = 0), and the actual surrounding rock grade is grade V.

[0131] (2) For a whole tunnel face photo of a certain tunnel section of a certain expressway, the predicted surrounding rock grade is grade IV. Through the sub-image prediction method, the obtained surrounding rock grade is grade IV (N5 = 0, N4 = 9, N3 = 0), and the actual surrounding rock grade is grade IV.

[0132] (3) For a whole tunnel face photo of a certain tunnel section of a certain expressway, the predicted surrounding rock grade is grade IV. Through the sub-image prediction method, the obtained surrounding rock grade is grade V (N5 = 0, N4 = 8, N3 = 1), and the actual surrounding rock grade is grade IV.

[0133] Through the above prediction verification, the prediction result of the tunnel face surrounding rock grade prediction method adopted in this embodiment has a high consistency with the actual surrounding rock situation, further verifying the rationality and availability of an intelligent classification method for tunnel face surrounding rock based on deep learning involved in the present invention.

[0134] Embodiment 3

[0135] As Figure 11 shown, an intelligent classification device for tunnel face surrounding rock based on deep learning includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute an intelligent classification method for tunnel face surrounding rock based on deep learning as described in the foregoing embodiment. The input / output interface may include a display, a keyboard, a mouse, and a USB interface for inputting and outputting data; the power supply is used to provide electrical energy for the intelligent classification device for tunnel face surrounding rock based on deep learning.

[0136] Those skilled in the art can understand that all or part of the steps to implement the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM), magnetic disks, or optical disks and other various media that can store program codes.

[0137] When the above integrated unit of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, magnetic disks, or optical disks and other various media that can store program codes.

[0138] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligent classification of surrounding rock at tunnel face based on deep learning, characterized in that: It includes the following steps: Input the photo of the heading face to be predicted, segment the photo of the heading face to be predicted into a set of sub-images, and sequentially input this set of sub-images into a pre-built surrounding rock classification prediction model for prediction, and output the surrounding rock classification prediction result of the photo of the heading face to be predicted according to the statistical information of the surrounding rock classification prediction results of each sub-image; Among them, the surrounding rock classification prediction results include grade III, grade IV, and grade V; the segmentation process is to cut the photo of the heading face to be predicted into n*n sub-images of the same size, and n takes values of 3, 4, or 5; the surrounding rock classification prediction model includes the following training steps: S1: Input sample data and perform annotation to form annotated sample data; the annotation is the surrounding rock classification prediction result corresponding to the sample data; S2: Preprocess the annotated sample data; S3: Input the preprocessed annotated sample data into the surrounding rock classification prediction model, and perform model training on the surrounding rock classification prediction model, and output the surrounding rock classification prediction model after model training; The preprocessing in step S2 includes image segmentation, image enhancement, and image encoding; Among them, the image enhancement includes image cropping, image flipping, image color transformation, image data normalization processing, and image data standardization processing; The image cropping includes CenterCrop mode, RandomCrop mode, FiveCrop mode, and / or RandomResizedCrop mode; The image flipping includes horizontal flipping and / or vertical flipping; The image color transformation includes grayscale transformation, brightness transformation, saturation transformation, contrast transformation, and / or hue transformation; The image data normalization processing uses the torchvision.transforms.ToTensor function; The image data standardization processing uses the torchvision.transforms.Normalize function.

2. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 1, characterized in that: The surrounding rock classification prediction criteria are: (1) When N5≥n, it is determined that the surrounding rock grade of the photo of the heading face to be predicted is grade V; (2) When N5 < n and N3≥n*(n - 1), it is determined that the surrounding rock grade of the photo of the heading face to be predicted is grade III; (3) When N5 and N3 do not meet the conditions of (1) and (2), it is determined that the surrounding rock grade of the photo of the heading face to be predicted is grade IV; Among them, N5 and N3 are respectively the numbers of grade V and grade III surrounding rock images in a set of the sub-images.

3. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 1, characterized in that: The transformation range of the brightness transformation is ±20%; The transformation range of the saturation transformation is ±30%; The transformation range of the contrast transformation is ±30%; The transformation range of the hue transformation is ±30%.

4. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 1, characterized in that: The preprocessing step uses the transforms.Compose() function to integrate the functions in the image enhancement step.

5. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 1, characterized in that: The image encoding includes the following steps: Encode the annotated sample data after image segmentation and image enhancement, and sequentially encode the annotated sample data of grade III, grade IV, and grade V as "2", "1", and "0".

6. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 1, characterized in that: The surrounding rock classification prediction model adopts the ResNet50 network architecture.

7. The method for intelligent classification of surrounding rock at a tunnel face based on deep learning according to claim 6, characterized in that: The training parameters of the model training in step S3 are set as follows: Batch number: batch_size = 16; Learning rate: lr = 0.001; Number of iterations: epoch = 100.

8. A tunnel face surrounding rock intelligent grading device based on deep learning, characterized in that: The invention comprises at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method according to any one of claims 1 to 7.

Images