Intelligent identification method for roadway surrounding rock loosening circle

Abstract

The invention relates to an intelligent identification method for a roadway surrounding rock loosening circle, and belongs to the technical field of intelligent identification of geotechnical engineering. The method comprises the following steps: S1, acquiring a borehole peeping image which is a picture intercepted from video data measured by a borehole imager; s2, preprocessing and crack identification are carried out on the borehole peeping image; s3, measuring the elastic modulus of the rock through an elastic wave method; s4, calibrating the range of the loose circle through the elastic modulus and the identified fracture, and carrying out category balance improvement through a weighted loss function; s5, training a YoLov5 deep learning model through the calibrated loose circle information, and improving the YoLov5 deep learning model; s6, evaluating and verifying the model; and S7, inputting a borehole peeping image obtained on site and the measured elastic modulus into the model to predict the loose circle. According to the method, the loosening circle can be quickly and accurately identified on an actual mine site, and a scientific basis is provided for roadway support design and mine safety management.

Description

Technical Field

[0001] The present invention relates to an intelligent identification method for the loosening zone of roadway surrounding rock, belonging to the technical field of intelligent identification in geotechnical engineering. Background Art

[0002] The formation of the loosening zone of roadway surrounding rock is usually caused by stress redistribution and deformation of the surrounding rock due to mine exploitation or blasting operations, resulting in the destruction of the rock structure and the formation of areas such as fissures and broken zones. The size and location of the loosening zone have an important impact on the design and implementation of support engineering. Therefore, accurately identifying the distribution, changes and influence range of the loosening zone is of great significance for mine safety management and roadway support design.

[0003] At present, traditional loosening zone identification methods mainly rely on manual exploration, manual measurement and traditional mechanical model calculation. These methods have some limitations, such as subjectivity of measurement results, long time consumption, poor accuracy, and inability to reflect the changes of the loosening zone in real time and dynamically. Therefore, there is an urgent need for an efficient, accurate and real-time loosening zone identification method.

[0004] In recent years, with the continuous development of digital image technology and deep learning technology, the intelligent identification method of loosening zone based on image processing has gradually attracted attention. Digital image technology can provide intuitive information of rock surface, fissures and broken zones by obtaining the peep images of underground drill holes, while deep learning models can efficiently analyze images and perform target recognition. In addition, the elastic modulus of rock can reflect its mechanical properties and deformation ability, and the loosening zone usually shows a significant decrease in elastic modulus. Therefore, obtaining the elastic modulus data of rock by the elastic wave method can further assist in the identification and positioning of the loosening zone.

[0005] Although certain research results have been achieved in the loosening zone identification method based on digital image and deep learning, the current practical application of this technology in mine sites still faces some challenges. First of all, how to effectively combine the image data obtained on site with the elastic modulus data measured by the elastic wave method to construct an efficient and accurate loosening zone identification model is still a difficult point. Secondly, how to use deep learning models to accurately detect fissures and broken zones in complex environments and combine them with rock mechanical parameters to provide real-time and dynamic loosening zone prediction is also a problem to be solved. For this reason, the present invention is proposed. Summary of the Invention

[0006] Aiming at the deficiencies of the prior art, the present invention provides an intelligent identification method for the loosening zone of roadway surrounding rock, which integrates borehole peep images, fissure identification, elastic modulus data measured by the elastic wave method and the YoLov5 deep learning model, and can realize the rapid and accurate identification of the loosening zone in the actual mine site, providing a scientific basis for roadway support design and mine safety management.

[0007] The technical solution of the present invention is as follows: An intelligent identification method for the loosening zone of roadway surrounding rock comprises the following steps: S1: Obtain borehole peep images, which are pictures intercepted from the video data measured by a borehole imager; S2: Preprocess the borehole peep images and identify fractures; S3: Measure the elastic modulus of the rock by the elastic wave method; S4: Calibrate the loosening zone range through the elastic modulus and the identified fractures, and improve the class balance through a weighted loss function; S5: Train the YoLov5 deep learning model with the calibrated loosening zone information and improve the YoLov5 deep learning model; S6: Conduct model evaluation and verification; S7: Input the borehole peep images obtained on-site and the measured elastic modulus into the model for predicting the loosening zone.

[0008] Preferably according to the present invention, in step S1, the specific steps for obtaining the borehole peep images are as follows: S11: Borehole image acquisition: Select a high-resolution borehole imager for real-time image acquisition, ensuring that the image resolution is high enough to capture details such as tiny cracks and structural features of the surrounding rock, especially fractures and possible loosening signs. For boreholes at different depths and positions, different imaging strategies are adopted to ensure that the obtained images cover the entire roadway surrounding rock and can reflect the spatial variations of the rock mass; The specific different imaging strategies are as follows: For borehole imaging at different depths and positions, high-resolution and wide-angle devices are used for shallow boreholes to capture surface information; smaller fractures and subtle structural changes usually exist in deep rock masses. When imaging, high dynamic range technology and multi-resolution sampling can be combined to improve the detail capture ability. In terms of position, the top area captures tensile fracture features through oblique imaging, the side walls record shear fractures by parallel or vertical scanning, and the bottom uses oblique imaging and multi-spectral technology to identify compaction fractures.

[0009] S12: Data sorting and storage: For the video data measured by the borehole imager, play it and intercept pictures as borehole peep images, and save the borehole peep images in a unified standard format, such as saving as PNG format, to ensure the convenience of subsequent processing.

[0010] Preferably according to the present invention, in step S2, the specific steps are as follows: S21: First, perform image preprocessing through operations such as high-pass filtering and smoothing to remove noise in the image and improve the recognition rate of fissures. Then, perform operations such as contrast enhancement and brightness adjustment on the image to highlight the fissures and other structural features in the rock. Subsequently, perform enhancement processing on the image based on the Multi-Scale Retinex algorithm (hereinafter simply referred to as the MSR algorithm). The MSR algorithm separates the reflection component and illumination component of the image through a multi-scale method, estimates the illumination map by retaining the prominent structures in the image, and simultaneously removes redundant texture details to achieve low-light image enhancement estimated through the illumination map. S22: Fissure detection and recognition: Automatically label and segment the fissure area in the image through the U-Net deep learning image segmentation method, extract geometric features such as the length, width, density, and angle of the fissures, provide basic data for subsequent loose circle recognition, label the fissure area, and distinguish different types of fissures (such as cracking fissures, shear fissures, etc.) to provide support for subsequent correlation analysis between fissures and loose circles.

[0011] According to the preference of the present invention, in step S3, the specific steps are as follows: S31: Use a rock elastic modulus measuring instrument to conduct on-site tests on the surrounding rock through the elastic wave method to obtain the elastic modulus value. The elastic modulus value reflects the stiffness and deformation characteristics of the rock and provides important mechanical parameters for the determination of the loose circle. First, determine the measurement positions. Set a measurement position every 1 meter in the borehole. In addition, add measurement positions at locations with larger or denser fissures detected during borehole peeping. Select two directions parallel and perpendicular to the borehole axis for measurement at each measurement position. After that, prepare the rock elastic modulus measuring instrument. Install an ultrasonic sensor at the preset measurement position in the borehole to ensure good contact between the ultrasonic sensor and the rock, avoid signal attenuation, send ultrasonic waves into the rock, and measure the time required for the wave to travel from the transmitting end to the receiving end. By measuring the wave propagation time of the rock at different depths, obtain the wave propagation speed and shear wave propagation speed of the rock at different depths. S32: Process the measured elastic wave data (longitudinal wave propagation speed and shear wave propagation speed) to remove interference factors and obtain the accurate rock elastic modulus. The specific calculation process is as follows: The rock elastic modulus includes Young's modulus and shear modulus. Calculate the Young's modulus E of the rock through the longitudinal wave propagation speed V P , which is an important mechanical parameter reflecting the stiffness of the rock. According to the relationship between wave speed and modulus, use the following formula for calculation:

[0012] Among them, ρ is the density of the rock; v is the Poisson's ratio. Through the shear wave velocity V S and the rock density ρ, calculate the shear modulus G:

[0013] The Poisson's ratio v is a parameter describing the deformation properties of rocks. The relationship between the Poisson's ratio and the wave velocity is as follows:

[0014] Through the above calculations, the elastic modulus of the rock can be deduced from the measured longitudinal and shear wave velocities, rock density and other parameters, providing a mechanical parameter basis for the subsequent determination of the loosening zone.

[0015] According to the preferred embodiment of the present invention, in step S4, the specific steps are as follows: S41: The formation of the loosening zone is closely related to the expansion of cracks. The characteristics such as the number, scale, and distribution of cracks directly affect the elastic modulus of the rock. Generally, the elastic modulus of the rock in the loosening zone is relatively low, and the crack density is relatively high. Based on the relationship between the crack characteristics and the elastic modulus value, a calibration rule is formulated. By a certain threshold range, the surrounding rock is divided into the loosening zone and the unloosened area. The loosening zone usually shows an area with a reduced elastic modulus and a dense crack distribution. Combining the elastic modulus data and the crack distribution, the regression analysis method is used to calibrate the specific range of the loosening zone, determine the boundary of the loosening zone, and use the calibrated loosening zone image to form the data set for model training; S42: During the training process, the imbalance of data will affect the model performance. Especially when the loosening zone area is small or the crack area is difficult to identify, the weighted loss function Focal Loss is used to handle the class imbalance. Especially when the loosening zone or crack occupies a small area, Focal Loss can better handle the recognition problem of small objects.

[0016] The class imbalance refers to the problem that the number distribution of different class samples is uneven during the model training process. In the intelligent identification of the loosening zone of the roadway surrounding rock, the loosening zone area usually accounts for a relatively small proportion of the entire image, while the unloosened area accounts for a relatively large proportion. This imbalance will cause the model to be more inclined to accurately predict the unloosened area while ignoring or misjudging the loosening zone area. To solve this problem, the weighted loss function Focal Loss is used to dynamically adjust the loss weight and improve the attention to difficult-to-predict samples (cracks or boundary areas).

[0017] According to a further preferred embodiment of the present invention, in step S41, the determination basis of the loosening zone calibration rule is Young's modulus and crack density D f , and define the loosening zone characteristic function F(x) as follows:

[0018] Where: E(x) is the Young's modulus of a certain point in the surrounding rock; E ref is the reference value of the Young's modulus of the undamaged surrounding rock; D f (x) is the fracture density of a certain point in the surrounding rock; D f，ref is the reference value of the fracture density of the undamaged surrounding rock; w 1 、w 2 are the weight coefficients of the elastic modulus and the fracture density, satisfying w 1 +w 2 = 1. According to the loosening zone characteristic function F(x), a judgment threshold T is set. When F(x) ≥ T, it is determined that this area is the loosening zone; otherwise, it is the unloosened area.

[0019] Preferably according to the present invention, in step S5, the specific steps are as follows: S51: Use the loosening zone training YoLov5 deep learning model calibrated in step S4 to identify the characteristics of the loosening zone and optimize the detection accuracy of the model. The YoLov5 deep learning model includes an input layer, a backbone network, a feature fusion network, a detection head, a loss function, and an output layer. The input layer preprocesses the image, the backbone network is responsible for extracting basic features, the feature fusion network integrates multi-scale features to enhance the detection ability for targets of different sizes, the detection head outputs the category, position, and confidence of the target, and optimizes the classification, localization, and confidence prediction performance through the loss function. Finally, non-maximum suppression is used in the output layer to generate the detection result. This architecture enables the YoLov5 deep learning model to have high real-time performance and detection accuracy and is suitable for the intelligent identification task of the loosening zone; S52: Make improvements to the YoLov5 deep learning model suitable for the application of the present invention: (1) Use a multi-scale training strategy to strengthen the model's detection ability for objects of different scales. During the training process of the YoLov5 deep learning model, diverse transformations of the image scale are performed to adapt to the detection of loosening zones of different sizes; The multi-scale training strategy enables the YoLov5 deep learning model to learn the characteristics of targets of different sizes and enhance the detection ability for multi-scale objects by dynamically adjusting the resolution of the input image (such as from 320×320 to 640×640) in each training batch. During the adjustment process, the aspect ratio of the image is maintained, and padding is used to avoid deformation affecting feature extraction. At the same time, data augmentation methods such as cropping, rotation, and flipping are combined to increase data diversity. During training, the model extracts multi-level features through the backbone network, uses the multi-scale feature fusion network to fuse global and detailed information, optimizes the feature expression under different scale inputs, and finally ensures through verification that the model can accurately detect various features from small fractures to large-scale loosening zones; (2) Introduce the self-attention mechanism Transformer module to improve the model's capture of context information, especially in areas with complex rock structures, which is particularly important for the identification of loosening zones; (3) Improve the loss function: Introduce the GIoU function to improve the traditional IoU loss function, enhance the model's bounding box prediction accuracy, and improve the model's detection ability for small objects and complex structures; When the shapes of the predicted box and the ground truth box differ significantly, the traditional IoU loss function may not effectively reflect the model's prediction accuracy. The GIoU function introduces additional geometric information on the basis of the IoU loss function, which can consider the distance and relative position relationship between the predicted box and the ground truth box, thus more accurately reflecting the model's prediction situation. The formulas of the IoU loss function and the GIoU function are as follows:

[0020]

[0021] where, A I is the intersection area of the predicted box and the ground truth box; A U is the union area of the predicted box and the ground truth box; A SEB is the area of the smallest bounding rectangle that completely encloses the predicted box and the ground truth box. After introducing the GIoU function, the loss function not only considers the overlapping area between the predicted box and the ground truth box, but also additionally considers the spatial relationship between them, thus improving the bounding box prediction accuracy of small objects and complex targets. In the detection of loosening zones, the GIoU function can better capture details and improve the accuracy of the model.

[0022] According to the preferred embodiment of the present invention, in step S6, the specific steps are as follows: S61: Verify the model: Evaluate the YoLov5 deep learning model using evaluation metrics such as accuracy and recall. At the end of each training cycle, apply the current model to the validation dataset, calculate the validation loss and accuracy, compare the training set loss with the validation set loss, and determine whether the model is overfitting on the training set. If the validation loss does not decrease significantly or starts to rise in multiple consecutive cycles, it indicates that the model may start to overfit. At this time, consider using an early stopping strategy to stop the training process, so as to retain the best model parameters and prevent the model from continuing to optimize on the training set while performing poorly on the validation set; S62: Model optimization and adjustment: According to the verification results, adjust hyperparameters such as the learning rate, batch size, and number of network layers of the YoLov5 deep learning model to improve the model performance. Adjust hyperparameters such as the batch size, learning rate, and number of layers through cross-validation to improve the model performance.

[0023] The above cross-validation selects the K-fold cross-validation method: randomly divide the data set into K subsets, use K-1 subsets to train the model each time, and the remaining one subset is used for validation. This process will be repeated K times, each time selecting a different validation set, and finally calculate the average performance index of K times. In the present invention, K = 4 is selected for four-fold cross-validation, the generalization ability of the model is verified, the best hyperparameters are selected, and then the average accuracy of each cross-validation is calculated, and the best model is selected according to its stability and performance.

[0024] Preferably according to the present invention, in step S7, the borehole peephole images and elastic modulus data collected on site are input into the trained YoLov5 deep learning model. The YoLov5 deep learning model, according to the input images and elastic modulus data, real-time predicts the loosening circle range in the rock mass and marks it on the borehole peephole images.

[0025] The beneficial effects of the present invention are as follows: 1. By combining digital images with the elastic wave method and combining crack identification with elastic modulus measurement, the present invention comprehensively evaluates the loosening situation of rocks, making the obtained loosening circle range more scientific and representative.

[0026] 2. The present invention applies the YoLov5 deep learning model to the identification of the loosening circle, realizes efficient and accurate prediction of the loosening circle. By improving the YoLov5 deep learning model, it is more suitable for the data set of the present invention, effectively excludes misidentified areas, and improves its efficiency and accuracy.

[0027] 3. By training the model with on-site data, the present invention can be directly applied to the site subsequently, provide real-time loosening circle prediction results, assist in mine safety management and support design, provide real-time and accurate loosening circle identification and evaluation for the mine site, and improve the safety and efficiency of mine exploitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the flow chart of the steps of the present invention.

[0029] Figure 2 is the machine learning flow chart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] The present invention will be further described below by way of examples in conjunction with the drawings, but not limited thereto.

[0031] Example 1: As Figure 1 - Figure 2 shown, this example provides an intelligent identification method for the loosening circle of roadway surrounding rock, and the steps are as follows: S1: Obtain borehole peephole images. The borehole peephole images are pictures intercepted from the video data measured by a borehole imager. The specific steps are as follows: S11: Borehole image acquisition: Select a high-resolution borehole imager for real-time image acquisition to ensure that the image resolution is high enough to capture details such as tiny cracks and structural features of the surrounding rock, especially cracks and possible signs of loosening. Different imaging strategies are used for boreholes of different depths and locations to ensure that the images obtained cover the entire tunnel surrounding rock and can reflect the spatial changes of the rock mass; Specifically, for borehole imaging at different depths and locations, high-resolution, wide-viewing-angle equipment is used to capture surface information in shallow boreholes. Deep rock masses usually have smaller cracks and subtle structural changes, and high dynamic range technology and multi-resolution sampling can be combined for imaging to enhance detail capture capabilities. In terms of location, tilted imaging is used to capture tension features in the top area, parallel or vertical scanning is used to record shear cracks on the sidewalls, and tilted imaging and multi-spectral technology are used to identify compacted cracks at the bottom.

[0032] S12: Data collation and storage: The video data measured by the borehole imager is played and captured as a borehole voyeuristic image, and the borehole voyeuristic image is saved in a unified standard format, such as PNG format, to ensure the convenience of subsequent processing.

[0033] S2: Preprocessing and crack identification of the borehole peek image. The specific steps are as follows: S21: First, the image is preprocessed by high-pass filtering, smoothing and other operations to remove noise in the image and improve the recognition of cracks. Then, the image is contrast enhanced and the brightness is adjusted to highlight the cracks and other structural features in the rock. The image is then enhanced based on the Multi-Scale Retinex (MSR) algorithm. The MSR algorithm uses a multi-scale method to separate the reflection component and the illumination component of the image, estimates the illumination mapping by retaining the prominent structure of the image, and removes redundant texture details to achieve low-light image enhancement estimated by the illumination map. Specifically, MSR enhancement is performed during image preprocessing: first, the original image data is read, and then the image is subjected to multi-scale Gaussian blur processing to obtain images of different scales. The reflection component of the image is calculated by subtracting the illumination component. The reflection component R (x, y) is:

[0034] Among them, I (x, y) is the original image, L σi (x,y) is the lighting component, representing the lighting conditions of the scene.

[0035] Afterwards, the reflection components of each scale are combined to form the final enhanced image, and the enhanced image is normalized to ensure that the data range is suitable for input into the neural network.

[0036] S22: Fracture detection and identification: Automatically label and segment the fracture regions in the image through the U-Net deep learning image segmentation method, extract geometric features such as the length, width, density, and angle of the fractures, provide basic data for subsequent loose circle identification, label the fracture regions, and distinguish different types of fractures (such as cracking fractures, shear fractures, etc.), providing support for subsequent correlation analysis between fractures and loose circles. U-Net is a common convolutional neural network used for image segmentation tasks. It achieves precise segmentation of fracture regions through pixel-by-pixel classification of the image.

[0037] S3: Measure the elastic modulus of the rock by the elastic wave method. The specific steps are as follows: S31: Use a rock elastic modulus measuring instrument to conduct on-site tests on the surrounding rock by the elastic wave method to obtain the elastic modulus value. The elastic modulus value reflects the stiffness and deformation characteristics of the rock and provides important mechanical parameters for the determination of the loose circle; First, determine the measurement positions. Set a measurement position every 1 meter in the borehole. In addition, add measurement positions at locations with larger or denser fractures detected during borehole peeping. Select two directions parallel and perpendicular to the borehole axis for measurement at each measurement position; After that, prepare the rock elastic modulus measuring instrument. Install ultrasonic sensors at the preset measurement positions in the borehole to ensure good contact between the ultrasonic sensors and the rock, avoid signal attenuation, send ultrasonic waves into the rock, and measure the time required for the wave to travel from the transmitting end to the receiving end. By measuring the wave propagation times of rocks at different depths, obtain the wave propagation speeds and shear wave propagation speeds of rocks at different depths; S32: Process the measured elastic wave data (longitudinal wave propagation speed and shear wave propagation speed) to remove interference factors and obtain the accurate elastic modulus of the rock. The specific calculation process is as follows: The elastic modulus of the rock includes Young's modulus and shear modulus. Calculate the Young's modulus E of the rock through the longitudinal wave propagation speed V P , which is an important mechanical parameter reflecting the stiffness of the rock. According to the relationship between wave speed and modulus, use the following formula for calculation:

[0038] where ρ is the density of the rock; v is the Poisson's ratio; Calculate the shear modulus G through the shear wave propagation speed V S and the rock density ρ:

[0039] The Poisson's ratio v is a parameter describing the deformation properties of the rock. The relationship between the Poisson's ratio and wave speed is as follows:

[0040] Through the above calculations, the elastic modulus of the rock can be deduced from parameters such as the measured longitudinal wave and transverse wave propagation velocities and the rock density, providing a mechanical parameter basis for the subsequent determination of the loosened zone.

[0041] S4: Calibrate the range of the loosened zone based on the elastic modulus and the identified fissures, and improve the class balance through a weighted loss function. The specific steps are as follows: S41: The formation of the loosened zone is closely related to the expansion of fissures. Characteristics such as the number, scale, and distribution of fissures directly affect the elastic modulus of the rock. Generally, the rock within the loosened zone has a lower elastic modulus and a higher fissure density. Based on the relationship between the fissure characteristics and the elastic modulus value, formulate calibration rules, and divide the surrounding rock into the loosened zone and the un-loosened area through a certain threshold range. The loosened zone usually shows an area with a reduced elastic modulus and a dense fissure distribution. Combining the elastic modulus data and the fissure distribution, use the regression analysis method to calibrate the specific range of the loosened zone, determine the boundary of the loosened zone, and use the calibrated loosened zone image to form the dataset for model training; The judgment basis for the loosened zone calibration rule is the Young's modulus and the fissure density D f , and define the loosened zone characteristic function F(x) as follows:

[0042] where: E(x) is the Young's modulus of a certain point in the surrounding rock; E ref is the reference value of the Young's modulus of the undamaged surrounding rock; D f (x) is the fissure density of a certain point in the surrounding rock; D f，ref is the reference value of the fissure density of the undamaged surrounding rock; w 1 , w 2 are the weight coefficients of the elastic modulus and the fissure density, satisfying w 1 +w 2 = 1. According to the loosened zone characteristic function F(x), set the judgment threshold T. When F(x) ≥ T, determine that this area is the loosened zone; otherwise, it is the un-loosened area.

[0043] S42: During the training process, the imbalance of data will affect the model performance. Especially in the case where the loosened zone area is small or the fissure area is difficult to identify, handle the class imbalance through the weighted loss function Focal Loss. Especially when the loosened zone or fissures occupy a small area, Focal Loss can better handle the recognition problem of small objects.

[0044] The Focal Loss function is a commonly used loss function in tasks dealing with class imbalance, especially in object detection. It can focus on difficult-to-classify samples and reduce the impact of easy-to-classify samples on the loss. It is a weighted extension of the cross-entropy loss, specifically designed to address the class imbalance problem. The formula of the Focal Loss function consists of the following parts:

[0045] Among them, p t is the predicted probability of the correct class, that is, the probability predicted by the model for the correct class. If it is the target class, then p t = p, otherwise p t = 1 - p, α t is a weight factor used to balance the impact of positive and negative samples. Usually, different weights are given to positive and negative samples respectively, and it is set as α t = α (the weight of positive samples). γ is the focal parameter that controls the weighting degree of easy and difficult samples. Usually, the value is γ ∈ [0, 5]. The role of this parameter is to adjust the loss according to the prediction accuracy, reduce the loss impact on easy-to-classify samples, and thus focus more on difficult-to-classify samples.

[0046] By introducing the focal parameter γ and the weight α t , the Focal Loss effectively solves the class imbalance problem. In object detection and classification tasks, by focusing on difficult-to-classify samples and suppressing the impact of easy-to-classify samples on the loss, it can improve the sensitivity of the model to minority classes and difficult-to-classify samples and optimize the model performance.

[0047] S5: Train the YoLov5 deep learning model with the calibrated loose circle information and improve the YoLov5 deep learning model. The specific steps are as follows: S51: Use the loose circle calibrated in step S4 to train the YoLov5 deep learning model to identify the characteristics of the loose circle and optimize the detection accuracy of the model. The YoLov5 deep learning model includes an input layer, a backbone network, a feature fusion network, a detection head, a loss function, and an output layer. The input layer preprocesses the image, the backbone network is responsible for extracting basic features, the feature fusion network integrates multi-scale features to enhance the detection ability for targets of different sizes, the detection head outputs the class, location, and confidence of the target, and optimizes the classification, localization, and confidence prediction performance through the loss function. Finally, non-maximum suppression is used in the output layer to generate detection results. This architecture enables the YoLov5 deep learning model to have high real-time performance and detection accuracy and is suitable for the intelligent identification task of the loose circle; S52: Make improvements to the YoLov5 deep learning model suitable for the application of the present invention: (1)Use a multi-scale training strategy to enhance the model's detection ability for objects of different scales. During the training process of the YoLov5 deep learning model, perform diverse transformations of the image scale to adapt to the detection of loose circles of different sizes. The multi-scale training strategy enables the YoLov5 deep learning model to learn the features of targets of different sizes by dynamically adjusting the resolution of the input image (such as from 320×320 to 640×640) in each training batch, enhancing the detection ability for multi-scale objects. During the adjustment process, the aspect ratio of the image is maintained, and padding is used to avoid deformation affecting feature extraction. At the same time, data augmentation methods such as cropping, rotation, and flipping are combined to increase data diversity. During training, the model extracts multi-level features through the backbone network, uses the multi-scale feature fusion network to fuse global and detailed information, optimizes the feature representation under different scale inputs, and finally ensures through verification that the model can accurately detect various features from small cracks to large-scale loose circles. (2)Introduce the self-attention mechanism Transformer module to improve the model's capture of context information, which is particularly important for the identification of loose circles, especially in areas with complex rock structures. Given the input query vector Q, key vector K, and value vector V, the output calculation formula of the self-attention mechanism is as follows:

[0048] where Q is the query vector; K is the key vector; V is the value vector; d k is the dimension of the key vector, used to scale the inner product; the softmax function is used to calculate the attention weights, emphasizing important information.

[0049] (3)Improve the loss function: Introduce the GIoU function to improve the traditional IoU loss function, enhance the model's bounding box prediction accuracy, and improve the model's detection ability for small objects and complex structures. When the shape difference between the predicted box and the ground truth box is large, the traditional IoU loss function may not effectively reflect the model's prediction accuracy. The GIoU function introduces additional geometric information on the basis of the IoU loss function, which can consider the distance and relative position relationship between the predicted box and the ground truth box, thus more accurately reflecting the model's prediction situation. The formulas of the IoU loss function and the GIoU function are as follows:

[0050]

[0051] where A I is the intersection area of the predicted box and the ground truth box; A U is the union area of the predicted box and the ground truth box; ASEB The area of ​​the minimum circumscribed rectangle that completely surrounds the predicted box and the true box is calculated. After the GIoU function is introduced, the loss function not only considers the overlapping area between the predicted box and the true box, but also additionally considers the spatial relationship between them, thereby improving the bounding box prediction accuracy of small objects and complex targets. In the detection of loose circles, the GIoU function can better capture details and improve the accuracy of the model.

[0052] S6: Evaluate and verify the model. The specific steps are as follows: S61: Validation model: Use accuracy, recall and other evaluation indicators to evaluate the YoLov5 deep learning model. At the end of each training cycle, apply the current model to the validation data set, calculate the validation loss and accuracy, compare the training set loss with the validation set loss, and determine whether the model is overfitting on the training set. If the validation loss does not drop significantly or starts to rise over multiple consecutive cycles, it means that the model may start to overfit. At this time, consider using an early stopping strategy to stop the training process, thereby retaining the best model parameters and preventing the model from continuing to optimize on the training set while performing poorly on the validation set. S62: Model optimization and adjustment: According to the verification results, adjust the learning rate, batch size, number of network layers and other hyperparameters of the YoLov5 deep learning model to improve the model performance. Adjust the batch size, learning rate, number of layers and other hyperparameters through cross-validation to improve the model performance.

[0053] The above cross-validation selects the K-fold cross-validation method: the data set is randomly divided into K subsets, K-1 subsets are used to train the model each time, and the remaining subset is used for verification. This process is repeated K times, and a different verification set is selected each time. Finally, the average performance index of K times is calculated. In the present invention, K=4 is selected, and a four-fold cross-validation is performed. By verifying the generalization ability of the model, the best hyperparameters are selected, and then the average accuracy of each cross-validation is calculated, and the best model is selected according to its stability and performance.

[0054] S7: The borehole peek images and elastic modulus data collected on site are input into the trained YoLov5 deep learning model. The YoLov5 deep learning model predicts the range of the loose zone in the rock mass in real time based on the input images and elastic modulus data and marks it in the borehole peek images.

Claims

1. A method for intelligently identifying loose zones of surrounding rock in tunnels, characterized in that: Here are the steps: S1: Acquire a borehole peek image, which is a picture captured from the video data measured by the borehole imager; S2: preprocessing and crack identification of borehole peep images; S3: rock elastic modulus measured by elastic wave method; S4: The loose zone range is calibrated by elastic modulus and identified cracks, and the category balance is improved by weighted loss function; S5: Train the YoLov5 deep learning model through the calibrated loose circle information and improve the YoLov5 deep learning model; S6: Conduct model evaluation and verification; S7: The borehole peek images obtained on site and the measured elastic modulus are input into the model to predict the loose zone.

2. The method for intelligently identifying loosened surrounding rock zones in tunnels according to claim 1, characterized in that: In step S1, the specific steps of obtaining the borehole viewing image are as follows: S11: Borehole image acquisition: The borehole imager performs real-time image acquisition to ensure that the cracks and structural characteristics of the surrounding rock are captured; S12: Data collation and storage: The video data measured by the borehole imaging device is played and captured as a borehole voyeuristic image, and the borehole voyeuristic image is saved in a unified standard format.

3. The method for intelligently identifying loosened surrounding rock zones in tunnels as claimed in claim 2, characterized in that: In step S2, the specific steps are: S21: First, the image is preprocessed by high-pass filtering and smoothing operations to remove noise in the image. Then, the image is contrast enhanced and the brightness is adjusted to highlight the cracks and structural features in the rock. Then, the image is enhanced based on the multi-scale retinal algorithm to estimate the illumination map by retaining the prominent structure of the image and removing redundant texture details, so as to achieve low-light image enhancement estimated by the illumination map. S22: Crack detection and identification: The crack areas in the image are marked and segmented through the U-Net deep learning image segmentation method, the length, width, density, and angle geometric features of the cracks are extracted, and the crack areas are marked.

4. The method for intelligently identifying loosened surrounding rock zones in tunnels as claimed in claim 3, characterized in that: In step S3, the specific steps are: S31: First, determine the measurement position, set a measurement position every 1 meter in the borehole, and in addition, set additional measurement positions at locations with larger or denser cracks detected during the borehole peeping process, and select two directions parallel and perpendicular to the borehole axis for measurement at each measurement position; After that, prepare the rock elastic modulus measuring instrument, install the ultrasonic sensor at the preset measuring position in the borehole, send ultrasonic waves into the rock, and measure the time required for the wave to travel from the transmitting end to the receiving end. By measuring the wave propagation time of rocks at different depths, the rock propagation velocity and shear wave propagation velocity at different depths are obtained. S32: Rock elastic modulus includes Young's modulus and shear modulus, which is expressed by the longitudinal wave propagation velocity V P , calculate the Young's modulus E of the rock: Among them, ρ is the density of rock; v is Poisson's ratio; The shear wave propagation speed V S And the rock density ρ calculate the shear modulus G: Poisson's ratio v is a parameter that describes the deformation properties of rocks. The relationship between Poisson's ratio and wave velocity is as follows: 。 5. The method for intelligently identifying loosened surrounding rock zones in tunnels as claimed in claim 4, characterized in that: In step S4, the specific steps are: S41: Based on the relationship between crack characteristics and elastic modulus values, a calibration rule is formulated to divide the surrounding rock into loosened areas and non-loosened areas through a certain threshold range, and the calibrated loosened area images are used to form a data set for model training; S42: During the training process, the weighted loss function Focal Loss is used to handle category imbalance.

6. The method for intelligently identifying loosened surrounding rock zones in tunnels as claimed in claim 5, characterized in that: In step S41, the loose circle calibration rule is determined based on Young's modulus and crack density D f , define the loose zone characteristic function F(x) as follows: Where: E(x) is the Young's modulus of a point in the surrounding rock; E ref is the benchmark value of Young's modulus of undamaged surrounding rock; D f (x) is the crack density at a certain point in the surrounding rock; D f，ref is the reference value of crack density of undamaged surrounding rock; w1 and w2 are weight coefficients of elastic modulus and crack density, satisfying w1+w2=1. According to the characteristic function of the loosening zone F(x), the judgment threshold T is set. When F(x)≥T, the area is judged as a loosening zone; otherwise, it is a non-loosening area.

7. The method for intelligently identifying loosened surrounding rock zones in tunnels according to claim 6, characterized in that: In step S41 and step S5, the specific steps are as follows: S51: Use the loose circle calibrated in step S4 to train the YoLov5 deep learning model to identify the features of the loose circle. The YoLov5 deep learning model includes an input layer, a backbone network, a feature fusion network, a detection head, a loss function and an output layer. The input layer preprocesses the image, the backbone network is responsible for extracting basic features, and the feature fusion network integrates multi-scale features to enhance the detection capability of large and small targets. The detection head outputs the category, position and confidence of the target, and optimizes the classification, positioning and confidence prediction performance through the loss function. Finally, the output layer uses non-maximum suppression to generate the detection result; S52: Improvements to the YoLov5 deep learning model: (1) Use a multi-scale training strategy to enhance the model’s ability to detect objects of different scales. During the training process of the YoLov5 deep learning model, perform diversified image scale transformations to adapt to the detection of loose circles of different sizes. The multi-scale training strategy enables the YoLov5 deep learning model to learn the features of targets of different sizes and enhance the detection capability of multi-scale objects by dynamically adjusting the resolution of the input image (e.g., 320×320 to 640×640) in each training batch. The image aspect ratio is maintained during the adjustment process, and deformation is avoided by padding to prevent feature extraction. At the same time, cropping, rotation, and flipping are combined to increase data diversity. During training, the model extracts multi-level features through the backbone network, and uses the multi-scale feature fusion network to fuse global and detail information, optimize the feature expression under different scale inputs, and finally ensures that the model detects a variety of features from small cracks to large-scale loose circles through verification. (2) Introducing the self-attention mechanism Transformer module; (3) Improved loss function: Introducing the GIoU function to improve the traditional IoU loss function; The formulas for the IoU loss function and the GIoU function are as follows: Among them, A I is the intersection area of ​​the predicted box and the real box; A U is the union area of ​​the predicted box and the true box; A SEB It is the area of ​​the minimum bounding rectangle that completely surrounds the predicted box and the true box.

8. The method for intelligently identifying loosened surrounding rock zones in tunnels as claimed in claim 7, characterized in that: In step S41 and step S6, the specific steps are as follows: S61: Validation model: Use accuracy and recall to evaluate the YoLov5 deep learning model. At the end of each training cycle, apply the current model to the validation dataset, calculate the validation loss and accuracy, compare the training set loss with the validation set loss, and determine whether the model is overfitting on the training set. If the validation loss does not drop significantly or starts to rise over multiple consecutive cycles, use the early stopping strategy to stop the training process, thereby retaining the best model parameters. S62: Model optimization and adjustment: Based on the verification results, adjust the hyperparameters of the YoLov5 deep learning model to improve model performance.

9. The method for intelligently identifying loosened surrounding rock zones in tunnels according to claim 8, characterized in that: In step S41 and step S7, the borehole peep images and elastic modulus data collected on site are input into the trained YoLov5 deep learning model. The YoLov5 deep learning model predicts the range of the loose zone in the rock mass in real time based on the input images and elastic modulus data and marks it in the borehole peep image.

Images