Tunnel boring machine surrounding rock integrity grading method based on image classification

Abstract

The invention relates to the technical field of surrounding rock integrity grading, and discloses a tunnel boring machine surrounding rock integrity grading method based on image classification, which comprises the following steps: acquiring a surrounding rock image of a tunnel boring machine construction tunnel; preprocessing the acquired surrounding rock image of the tunnel boring machine construction tunnel; constructing a surrounding rock integrity classification data set according to the preprocessed surrounding rock image; constructing a hierarchical classification model based on a ResNet34 feature extraction network, and carrying out training optimization by adopting a multi-hierarchical classifier structure; and carrying out tunnel boring machine surrounding rock integrity grading by utilizing the trained and optimized hierarchical classification model. According to the method, features are extracted based on ResNet34, high-definition digital images are combined, the completeness of the surrounding rock in the TBM tunneling process can be detected in real time, and then intelligent grading of the breaking condition of the surrounding rock is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of surrounding rock integrity classification, and in particular to a tunnel boring machine surrounding rock integrity classification method based on image classification. Background Art

[0002] TBMs (tunnel boring machines) are increasingly used in complex geological conditions. TBM excavation efficiency and tool wear are closely related to the hardness, wear resistance, and integrity of the surrounding rock mass. However, traditional drill-and-blast methods for grading surrounding rock are no longer suitable for the demands of modern tunneling.

[0003] Advances in deep learning technology have made image analysis-based rock mass integrity assessment possible. In particular, machine vision and deep learning methods can analyze and intelligently classify cracks on the rock surface. However, existing methods still face challenges, such as differences between rock samples and actual rock mass characteristics during construction, and the extraction of key features from complex images.

[0004] With the development of deep learning technology, using images of rock debris or rock edges captured during TBM excavation to assess surrounding rock integrity has become a new research trend. By combining image processing methods such as Canny edge detection and MobileNet-v2, researchers have proposed using parameters such as the crack ratio to assess surrounding rock integrity. However, image-based rock integrity assessment still suffers from certain errors, especially when rock textures are complex and pixel distribution is uneven. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a tunnel boring machine surrounding rock integrity classification method based on image classification.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A tunnel boring machine surrounding rock integrity classification method based on image classification includes the following steps: Collect images of the surrounding rock of a tunnel being constructed by a tunnel boring machine; Preprocessing of the surrounding rock images collected from tunnel boring machine-constructed tunnels; Construct a surrounding rock integrity classification dataset based on the preprocessed surrounding rock images; Construct a hierarchical classification model based on the ResNet34 feature extraction network, and use a multi-level classifier structure for training optimization; The trained and optimized hierarchical classification model is used to classify the surrounding rock integrity of the tunnel boring machine.

[0007] Furthermore, when collecting images of the surrounding rock of a tunnel constructed by a tunnel boring machine, a terminal-switch-camera topology is used for equipment layout, and an industrial camera is installed on the tunnel boring machine.

[0008] Furthermore, when pre-processing the collected surrounding rock images of the tunnel constructed by the tunnel boring machine, the original images are cropped to an adapted pixel size.

[0009] Furthermore, when pre-processing the surrounding rock images of the tunnel constructed by the tunnel boring machine, the brightness of the cropped images is adjusted using a brightness adjustment formula; wherein the brightness adjustment formula is:

[0010] in, is the brightness value of the image after adjustment, is the brightness value of the image before adjustment, is the median of all image brightness values, is the brightness adjustment factor (the value is 1 or -1, depending on whether to increase or decrease the brightness).

[0011] Furthermore, when pre-processing the collected surrounding rock images of the tunnel constructed by the tunnel boring machine, the contrast between the cracks and the background is improved in the brightness-adjusted images.

[0012] Furthermore, when constructing a surrounding rock integrity classification dataset based on the preprocessed surrounding rock images, the apparent integrity of the local surrounding rock is divided into complete, relatively complete, relatively broken, and broken, and non-surrounding rock information pixels are included in the dataset in the surrounding rock integrity classification.

[0013] Furthermore, the ResNet34-HDC network is defined, and the hierarchical classification model based on the ResNet34 feature extraction network is constructed, including: A ResNet34 feature extraction network, an adaptive pooling layer, a fully connected layer, and multiple modular independent weighted small classifiers are set in parallel; the fully connected layer extracts high-level features of the input image to perform large-category classification, and the modular independent weighted small classifier selects the corresponding small classifier for classification based on the large-category classification result.

[0014] Furthermore, the ResNet34 feature extraction network determines the weight by calculating the ratio of the total number of samples in each category to the number of samples in that category, specifically:

[0015] Where, Representation category i The weight of represents the total number of samples in the dataset, C represents the number of categories, Representation category i The number of samples.

[0016] The present invention has the following beneficial effects: To address the potential risk of machine jams during TBM construction and improve surrounding rock classification methods, this paper proposes a new assessment method. Based on ResNet34 feature extraction and combined with acquired high-definition digital images, this method can detect surrounding rock integrity in real time during TBM excavation, enabling intelligent classification of surrounding rock fragmentation, thereby supporting refined tunnel design and mechanized construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of a tunnel boring machine surrounding rock integrity classification method based on image classification; Figure 2 This is a data flow diagram of a tunnel boring machine surrounding rock integrity classification method based on image classification; Figure 3 Schematic diagram of equipment installation; Figure 4 This is a schematic diagram of the network model structure; Figure 5 Schematic diagram comparing manual classification and model classification results (the left picture is manual classification, and the right picture is model classification). DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0019] like Figure 1 As shown, an embodiment of the present invention provides a method for grading the integrity of surrounding rock of a tunnel boring machine based on image classification, comprising the following steps S1 to S5: S1, collecting images of the surrounding rock of the tunnel under construction by the tunnel boring machine; In an optional embodiment of the present invention, when collecting the surrounding rock image of the tunnel under construction by the tunnel boring machine in step S1, the equipment is arranged using a PC terminal-switch-camera topology, such as Figure 2 As shown in the figure, an industrial camera is installed at the front end of the tunnel boring machine to facilitate surrounding rock image acquisition. A PC terminal is placed in the TBM cockpit for image acquisition and storage. The industrial camera has a pixel resolution of 2448 × 2048 pixels.

[0020] When installing industrial cameras in an open TBM to collect images of surrounding rocks, due to the complex environment and short shooting distance, it is recommended to use a wide-angle lens to ensure the shooting range and take into account the clarity of the surrounding rocks. The industrial cameras are arranged along the tunnel ring on the TBM, such as Figure 3 shown.

[0021] During the TBM excavation process, the camera takes pictures at regular intervals, taking one picture every 15-20 minutes, ensuring that the overlap between adjacent images taken continuously is no less than 20%.

[0022] S2, preprocessing the collected surrounding rock images; In an optional embodiment of the present invention, step S2 pre-processes the acquired surrounding rock images.

[0023] In this embodiment, in order to meet the input size requirements of the mathematical model for surrounding rock integrity classification, and considering that when the noise occupies a relatively small area in the image, it has little impact on the surrounding rock integrity classification, the image noise is not filtered. Only the basic physical dimensions of the rock mass integrity and the noise ratio are considered for analysis. The original image is cropped to a physical visible area size of approximately 3×3 steel grid units (each grid unit is 20×20 cm in size), with an actual pixel size of 456×456 pixels.

[0024] When pre-processing the collected surrounding rock image in step S2, the brightness of the cropped image is adjusted. The brightness adjustment formula is:

[0025] in, is the brightness value of the image after adjustment, is the brightness value of the image before adjustment, is the median of all image brightness values, is the brightness adjustment factor (the value is 1 or -1, depending on whether to increase or decrease the brightness).

[0026] In a grayscale image, brightness refers to the average intensity or lightness or darkness of the pixel values in the image. It is obtained by calculating the grayscale average of the sample image set.

[0027] When pre-processing the collected surrounding rock image in step S2, the contrast of the image after brightness adjustment is enhanced to make the crack features more prominent, thereby improving the accuracy of image classification.

[0028] S3, constructing a surrounding rock integrity classification dataset based on the preprocessed cropped surrounding rock images; In an optional embodiment of the present invention, when constructing a surrounding rock integrity classification dataset based on the pre-processed surrounding rock image in step S3, the apparent integrity of the local surrounding rock is divided into complete, relatively complete, relatively broken, and broken, and non-surrounding rock information pixels are included in the dataset in the surrounding rock integrity classification, such as Figure 4 (left picture).

[0029] This example refers to traditional surrounding rock classification, combines the degree of joint development and the analysis of TBM excavability, and divides the apparent integrity of local surrounding rock into four categories: complete, relatively complete, relatively broken, and broken.

[0030] In the classification of surrounding rock integrity, non-surrounding rock information pixels, such as steel arches and steel mesh, were also included in the dataset. While removing these steel arches and steel mesh is complex and labor-intensive, testing has shown that the occluded area accounts for less than 15% of the image, minimizing the impact on image classification. Including non-surrounding rock information in the dataset improves the model's generalization capabilities, enabling it to focus more on the characteristics of the rock mass itself, rather than the supporting structure, even in the presence of occlusion.

[0031] S4. Build a hierarchical classification model based on the ResNet34 feature extraction network and use a multi-level classifier structure for training and optimization.

[0032] In an optional embodiment of the present invention, step S4 constructs a ResNet34-HDC network, that is, a hierarchical classification model based on the ResNet34 feature extraction network, and adopts a multi-level (Hierarchical) dynamic classifier structure (DynamicClassification) for training optimization, such as Figure 4 As shown; In a standard neural network, the front convolutional layer is mainly used to extract low-level and high-level features of the image, such as edges, textures, shapes, etc., rather than targeting a specific category. These features are shared parameters, and all categories use the same feature extractor. In the subsequent layers of the network, the features are gradually abstracted and passed to the classification layer. The classification layer makes decisions based on these features, obtains the original prediction value, and finally outputs the class probability through the activation function. The weight of each category does not change individually, but the shared features are obtained through network learning, and the decision is made in the final classification layer. However, it is difficult to deal with hierarchical categories, unbalanced data sets and high-dimensional feature spaces using only ordinary classification layers. In order to better realize the identification of the integrity of the surrounding rock of TBM tunnels, this embodiment improves the classification accuracy by modifying the classifier.

[0033] In an optional embodiment of the present invention, the hierarchical classification model based on the ResNet34 feature extraction network constructed in step S4 includes: ResNet34 is a widely proven convolutional neural network architecture that performs well in many visual tasks. By leveraging ResNet34's powerful feature extraction capabilities, we can build a new network by modifying only the classification layer, avoiding the need to retrain a convolutional network from scratch. This saves computing resources, training time, and achieves high classification accuracy. Therefore, ResNet34 is used as the convolutional layer to extract image features. The Flatten layer flattens the multidimensional feature map into a one-dimensional vector for input into the fully connected layer for classification or regression.

[0034] The ResNet34-HDC model first uses ResNet34 as a shared backbone network (removing the original fully connected layers) to extract global features and further compress the feature maps through adaptive pooling. The model then flattens the features through a Flatten layer and passes them to a fully connected module (FC module), which consists of fully connected layers, ReLU activation functions, and Dropout regularization to enhance the model's expressiveness and prevent overfitting.

[0035] In the classification part, the model first uses the large-category classifier to predict the large category of the input image, and then selects the corresponding small-category classifier for more refined classification based on the prediction results of the large category. Each small-category classifier consists of a fully connected layer, ReLU activation, and Dropout to improve the robustness of small-category classification. Finally, the prediction results of the small-category are only filled into the corresponding index position, and other positions are filled with extremely small values to avoid misleading optimization. This structure effectively utilizes the large-category information, making the small-category classification more accurate while reducing computational overhead. In the classification stage, the model adopts a two-level classification strategy: ① The large-category classifier predicts large categories (complete or fragmented) and uses ReLU activation + Dropout to prevent overfitting; ② Based on the results of the large-category classification, the small-category classifier selects the corresponding large-category classifier (such as "complete" to "relatively complete" or "relatively fragmented" to "fragmented") for fine-grained classification. The small-category classifier corresponding to each large-category has an independent weight to avoid interference between different large-category classes.

[0036] In loss calculation, only the cross-entropy loss is calculated for the selected small class labels, while the small class losses corresponding to the unselected large classes are ignored. Finally, the weighted sum is added to the large class loss to perform gradient optimization. This ensures the accuracy of large class discrimination while improving the ability to classify small classes, effectively reducing misclassification.

[0037] To avoid the negative impact of data imbalance between categories on model training, the weight is determined by calculating the ratio of the total number of samples in each category to the number of samples in that category, so that the model pays more attention to categories with fewer samples during training. The calculation formula is as follows:

[0038] Where, Representation category i The weight of represents the total number of samples in the dataset, C represents the number of categories, Representation category i The number of samples.

[0039] S5. Use the training model to classify the surrounding rock integrity.

[0040] The present invention uses imaging equipment to collect continuous surrounding rock images during the TBM construction phase for subsequent identification; considering factors such as the on-site environment and image brightness, the collected surrounding rock images are preprocessed, and a surrounding rock integrity classification sample data set is established based on the preprocessing results; referring to existing standards and combining with the actual project situation, the surrounding rock is divided into four categories, and the samples are classified accordingly; classification is obtained based on the standard ResNet34 network with multi-level classification labels and independent classifier modules, providing an intelligent identification method for surrounding rock integrity classification, which can provide a reference basis for TBM excavability classification, such as Figure 5 shown.

[0041] The present invention is described in terms of flowcharts and / or block diagrams of devices (systems), implementation methods, and computer program products. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0042] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0043] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0044] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0045] 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 it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A tunnel boring machine surrounding rock integrity classification method based on image classification, characterized in that: The following steps are involved: Collect images of the surrounding rock of a tunnel being constructed by a tunnel boring machine; Preprocessing of the surrounding rock images collected from tunnel boring machine-constructed tunnels; Construct a surrounding rock integrity classification dataset based on the preprocessed surrounding rock images; Construct a hierarchical classification model based on the ResNet34 feature extraction network, and use a multi-level classifier structure for training optimization; The trained and optimized hierarchical classification model is used to classify the surrounding rock integrity of the tunnel boring machine.

2. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 1, characterized in that: When collecting images of the surrounding rock of a tunnel being constructed by a tunnel boring machine (TBM), a terminal-switch-camera topology was used for equipment layout. Industrial cameras were installed on the TBM, ensuring a 360-degree circumferential viewing angle and a set degree of image overlap.

3. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 1, characterized in that: When preprocessing the surrounding rock images of the tunnel constructed by the tunnel boring machine, the original images are cropped to a pixel size suitable for network training.

4. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 3, characterized in that: When pre-processing the surrounding rock images of the tunnel constructed by the tunnel boring machine, the brightness adjustment formula is used to adjust the brightness of the cropped images; The brightness adjustment formula is: in, is the brightness value of the image after adjustment, is the brightness value of the image before adjustment, is the median of all image brightness values, is the brightness adjustment factor (the value is 1 or -1, depending on whether to increase or decrease the brightness).

5. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 4, characterized in that: When preprocessing the surrounding rock images collected from tunnel boring machine-constructed tunnels, the contrast between cracks and background is improved by adjusting the brightness of the images; The contrast adjustment formula is: in, is the adjusted contrast, is the original contrast, are the coordinates of the image pixels, is the scaling factor.

6. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 1, characterized in that: When constructing the surrounding rock integrity classification dataset based on the preprocessed surrounding rock images, the apparent integrity of the local surrounding rock is divided into complete, relatively complete, relatively broken, and broken, and non-surrounding rock noise information pixels are included in the dataset in the surrounding rock integrity classification.

7. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 1, characterized in that: The hierarchical classification model constructed based on the ResNet34 feature extraction network includes: A ResNet34 feature extraction network, an adaptive pooling layer, a fully connected layer, and multiple modular independent weighted small classifiers are arranged in parallel; the fully connected layer extracts high-level features of the input image to perform large-category classification, and the modular independent weighted small classifier selects the corresponding small classifier for classification based on the large-category classification result.

8. The method for grading the surrounding rock integrity of a tunnel boring machine based on image classification according to claim 7, characterized in that: The ResNet34 feature extraction network determines the weight by calculating the ratio of the total number of samples in each category to the number of samples in that category, specifically: Where, Representation category i The weight of represents the total number of samples in the dataset, C represents the number of categories, Representation category i The number of samples.

Images