A neural network-based tunnel face surrounding rock intelligent classification method

Abstract

A neural network-based intelligent grading method for tunnel face surrounding rock includes obtaining surrounding rock strength, obtaining rock integrity, and grading the surrounding rock. The steps for obtaining surrounding rock strength are as follows: selecting the rebound value and geological development degree of the tunnel face surrounding rock from field tests as input indicators; constructing an RBF learning model based on the RBF neural network intelligent algorithm, inputting learning sample data into the RBF learning model to optimize parameters; establishing an RBF learning model using the optimized parameters, inputting test sample data into the RBF prediction model, and obtaining the surrounding rock strength output result. The steps for obtaining rock integrity are as follows: taking images of the surrounding rock at the tunnel face; obtaining complete images of the tunnel face using image recognition processing methods; constructing a Mask RCNN learning model based on the Mask RCNN convolutional neural network algorithm, inputting learning sample data into the Mask RCNN learning model to optimize parameters, etc. This invention intelligently grades the surrounding rock at the tunnel face based on a neural network algorithm, and can accurately predict the surrounding rock level of the tunnel face.

Description

Technical Field

[0001] This invention relates to the technical field of real-time grading of the surrounding rock quality at the tunnel face. Background Technology

[0002] To assess the stability of tunnel surrounding rock, guide safe construction, and improve the relevance, accuracy, and scientific rigor of tunnel support design and excavation, real-time dynamic classification of surrounding rock has become a crucial and paramount task in tunnel engineering. According to the BQ method, rock strength and integrity are important bases and indicators for surrounding rock classification.

[0003] Rock strength can be obtained through laboratory testing and field testing. Laboratory testing is not suitable for obtaining real-time dynamic rock strength data due to its complex, costly, and time-consuming processes, including sampling, sample preparation, testing, and analysis. Field testing can employ point load methods, acoustic wave testing, and rebound testing, but the conversion relationship between point load index and rock strength varies depending on engineering geological conditions and lacks universality. Furthermore, point load testing instruments are inconvenient to carry at construction sites, making them unsuitable for on-site testing. Acoustic wave testing cannot reflect the stress state and surface characteristics of the rock, while rebound testing has limitations in reflecting the internal structure of the rock.

[0004] Regarding methods for obtaining rock mass integrity, the acoustic method cannot comprehensively characterize the integrity of the surrounding rock in various parts of the tunnel working face due to the limited layout of measuring points; the relationship between the Z-RBI value of the comprehensive quantitative index of rock mass structure and the integrity of rock mass in the rock mass block size method is not yet mature; and the adaptability of various algorithms and the identification effect of joints in the tunnel working face of the planar bar count method based on digital image + image algorithm still need to be improved.

[0005] In summary, the following problems exist in using standard methods for real-time classification of surrounding rock at tunnel faces: (1) Indoor testing has limited sampling and is time-consuming, resulting in large errors in the obtained rock strength results, and it is difficult to meet the needs of real-time data analysis; (2) The results of field testing methods for single rock mass strength are not representative, and point load tests are inconvenient to carry instruments, making it difficult to meet the requirements of on-site testing at any time during construction; (3) The methods for obtaining rock mass integrity cannot accurately characterize the integrity of surrounding rock at various parts of the tunnel face.

[0006] Therefore, it is difficult to achieve real-time grading of the surrounding rock at the tunnel face using existing standard methods. The present invention provides a method for extracting tunnel surrounding rock features and automatically grading the surrounding rock based on neural networks and image recognition, which is of great significance and can largely ensure the accuracy and scientific nature of the surrounding rock grading. Summary of the Invention

[0007] To overcome the technical deficiencies of existing technologies, this invention discloses an intelligent grading method for surrounding rock at tunnel faces based on neural networks. The method uses on-site testing and image recognition technology to identify geological information of the surrounding rock in front of the tunnel face in real time. Artificial intelligence algorithms are then used to analyze and process the surrounding rock parameter data, establishing a real-time quality grading model for the surrounding rock at tunnel faces based on RBP neural networks and Mask R-CNN convolutional neural network analysis. This model has been validated in the Luanlu Expressway Nianpan Tunnel and the Zhejiang Hangwen Railway Tunnel project. The grading results show good agreement with the actual exposed surrounding rock, demonstrating good real-time performance and applicability.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0009] A method for intelligent classification of surrounding rock at tunnel faces based on neural networks, comprising the following steps:

[0010] Step 1: Obtaining the strength of the surrounding rock includes:

[0011] S1. Select the rebound value and geological development degree of the surrounding rock at the working face from the field test as input indicators;

[0012] S2. Construct an RBF learning model based on the RBF neural network intelligent algorithm, and input the learning sample data obtained from S1 into the RBF learning model to optimize the parameters;

[0013] S3. Establish an RBF learning model using optimized parameters, input the test sample data obtained in S1 into the RBF prediction model, and obtain the surrounding rock strength output results.

[0014] Step Two: Assessing the integrity of the rock includes:

[0015] S4. Acquire image data of the surrounding rock at the tunnel face;

[0016] S5. Obtain a complete image of the working face using image recognition processing methods;

[0017] S6. Construct a Mask RCNN learning model based on the Mask RCNN convolutional neural network algorithm, and input the learning sample data into the Mask RCNN learning model to optimize the parameters;

[0018] S7. Build a Mask RCNN learning model using optimized parameters, input the test sample data into the Mask RCNN prediction model, and obtain a binary image of the joints on the working face;

[0019] S8. Calculate the volumetric joint number of the surrounding rock mass at the working face, and then obtain the rock mass integrity coefficient.

[0020] Step 3: Surrounding rock classification includes:

[0021] S9. Using the rock strength and rock integrity coefficients obtained in steps 1 and 2, the BQ method is used to obtain the surrounding rock classification results.

[0022] The method for selecting the rebound value of the surrounding rock at the working face as the input index in S1 is as follows: the working face is divided into zones. When dividing the zone, the asymmetry of the properties of the surrounding rock on the left and right sides needs to be considered in the horizontal direction, and the lithological differences of the surrounding rock in the arch, middle and bottom need to be considered in the vertical direction; a rebound hammer is used to perform 16 rebound tests on the rock mass of each zone, and the distance between every two rebound points is not less than the diameter of the impact rod to ensure that the rebound hammer is perpendicular to the surface of the rock mass; the three maximum values ​​and three minimum values ​​are removed from the obtained data, and the average of the remaining 10 data is taken as the rebound value.

[0023] The method for selecting the geological development degree of the tunnel face as the input index in S1 is as follows: 1 horizontal survey line is arranged at the arch, middle and bottom of the tunnel working face to detect the rock and soil of the tunnel face; the geological radar map at the three survey lines is measured using a geological radar instrument; the parameter - geological development degree - is calculated based on the typical geological radar map of the tunnel face. The geological development degree is calculated by formula (1).

[0024] The learning sample data in S2 is a selection of learning samples (x) from a portion of the tunnel face data. i ,y i ), forming the sample training set D={(x i ,y i |i=1,2,…,n},x i ∈R n ,y i ∈+, where n is the number of samples, x i A two-dimensional vector representing factors influencing rock mass strength, including the rebound value of the surrounding rock at the working face and the degree of geological development, y i It is the output vector and its value is the corresponding uniaxial compressive strength of the rock.

[0025] The method for constructing the RBF learning model in S2 is as follows: based on the RBF three-layer neural network, x in the normalized learning samples... i The input layer vector is nonlinearly transformed to the hidden layer, and then the hidden layer is linearly transformed to the output layer. Finally, the output layer result y′ is processed. i Inverse normalization will yield the predicted uniaxial compressive strength y of the rock. i .

[0026] x in the sample i The normalization process is as follows:

[0027]

[0028] Where: x i For x′ i Normalized data; xmin For x′ i The minimum value of x; max For x′ i The maximum value.

[0029] Output layer result y′ i The denormalization process is as follows.

[0030] y′ i =y i ×(y max -y min )+y min (3)

[0031] Where: y max The maximum value in the original data; y min This is the minimum value in the original data.

[0032] The method for nonlinearly transforming the input layer vector to the hidden layer is as follows: The hidden layer space is constructed using Radial Basis Functions (RBFs) as the "basis" of the hidden units. The hidden layer transforms the input vector, converting the low-dimensional pattern input data to a high-dimensional space, thus enabling linearly separable problems in the low-dimensional space to become linearly separable in the high-dimensional space. The radial basis functions chosen for the nonlinear transformation are:

[0033]

[0034] Where X is the input vector; C j C is the center vector of the j-th neuron in the hidden layer. j =[c j1 ,c j2 ,…,c jn ] T σ is the variance of the radial basis functions;

[0035] Through the above transformation, the output value z of the j-th neuron in the hidden layer is... j for:

[0036]

[0037] Among them, D j D is the width vector of j neurons in the hidden layer. j =[d j1 ,d j2 ,…,d jn ] T p represents the total number of hidden layer neurons.

[0038] Center vector parameter c ji Initial value:

[0039]

[0040] Where mini is the minimum value of all input information for the i-th metric in the training set, and maxi is the maximum value of all input information for the i-th metric in the training set.

[0041] Width vector parameter d ji Initial value:

[0042]

[0043] Where, d f It is the width adjustment coefficient, and its value is less than 1.

[0044] The network output is a linearly weighted sum of the hidden layer outputs, and the output of the output layer neurons is:

[0045] Y = [y1, y2, ..., y q ] T (8)

[0046]

[0047] Among them, w kj q is the weight between the k-th neuron in the output layer and the j-th neuron in the hidden layer; q is the total number of neurons in the output layer.

[0048] The parameters that need to be optimized in S2 include the center parameter c of the radial basis function. ji Variance (width) parameter d ji and the weights w from the hidden layer to the output layer kj The gradient descent method is used to update and optimize to the optimal value. The update calculation is as follows:

[0049]

[0050] Among them, c ji (t) represents the center vector parameter of the j-th hidden layer neuron in the t-th update calculation of the i-th input neuron; d ji (t) is related to the central parameter c ji (t) corresponds to the width vector parameter; w kj (t) represents the weight of the k-th output neuron relative to the j-th hidden neuron during the t-th update calculation; η is the learning factor; E is the RBF neural network evaluation function:

[0051]

[0052] Among them, O lk y is the expected output value of the k-th output neuron with the l-th input sample; lk It is the network output value of the k-th output neuron when the l-th input sample is received.

[0053] The method for acquiring the image data of the tunnel face surrounding rock in S4 is as follows: based on the on-site shooting conditions and the size of the tunnel face area (span and height), the tunnel face is divided into N equal parts (the area of ​​each shooting section is controlled between 5 and 20 m²). 2 (For best results) Use a high-definition digital camera to capture images of the surrounding rock at the working face of each section. The optimal time to acquire images is after the muck removal process is completed.

[0054] The step S5, acquiring a complete image of the tunnel face, includes stitching and merging the segmented images of the tunnel face. Specifically, the N segmented images of the tunnel face taken on-site are roughly arranged in Photoshop according to their actual positions. The positions of the segmented images are then fine-tuned based on the overlapping image details at the stitching points to ensure they are aligned as closely as possible. Subsequently, the segmented images are rasterized, and the process is completed by clicking Edit → Auto Blend Layers → Panorama → OK to stitch and merge the segmented images.

[0055] The Mask RCNN convolutional neural network structure in S6 includes five main parts: the backbone architecture and the Feature Pyramid Network (FPN), the Region Recommendation Network (RPN), the Region of Interest Alignment (ROIAlign), the classifier, and the segmentation mask generation network.

[0056] The backbone architecture employs a deep residual network, ResNet101, to extract feature maps from the image. The Feature Pyramid Network (FPN) fuses feature maps from the lower to higher levels, thereby fully utilizing the features extracted at each stage.

[0057] The anchor labeling rules in the Region Candidate Network (RPN) differ from simply dividing anchors into positive and negative samples based on IoU values. Instead, the Non-Maximum Suppression (NMS) method should be used, employing iterative calculations based on the principles illustrated in equation (*.**). First, the highest-scoring bounding box is obtained. IoU values ​​are then calculated between this box and other bounding boxes. If the IoU value is greater than 0.6, this box is labeled as a negative sample. Next, the second-highest-scoring bounding box is selected, and the next iteration continues until all bounding boxes are labeled.

[0058]

[0059] In the formula: s i Let M be the score of the i-th box; M is the target box, b i The suggestion box with the highest confidence; N t This is the threshold for IoU, which is set to 0.6 here.

[0060] The pooling method is the ROIAlign pooling method, which uses bilinear interpolation to retain the floating-point number of pixels, in order to obtain more accurate joint and crack recognition results.

[0061] The classifier in the Mask RCNN convolutional neural network structure uses a fully connected network of 1024 neurons to classify and regress each ROI, obtaining the final category and bounding box for crack detection.

[0062] The segmentation mask generation network in the Mask RCNN convolutional neural network structure achieves the effect of segmenting the target by masking the target to be detected (joints and fissures) through mask regions.

[0063] The Mask RCNN convolutional neural network in S6 defines a multi-task loss, and its loss function is:

[0064] L = L cls +L box +L mask (13)

[0065] Where L cls L box and L mask These refer to classification loss, detection box loss, and mask loss, respectively.

[0066] The input image annotation method of the Mask RCNN convolutional neural network in S6 is as follows: based on the image enhancement such as improving contrast and expanding the image set such as flipping, the obtained sample image is used to calibrate the joints on the face of the tunnel using the image annotation tool provided by Matlab, that is, to calibrate the fissure (joint) target region, and the calibrated image is exported to the work area for subsequent processing.

[0067] The number of volumetric joints in the surrounding rock mass at the working face in S8 is calculated using the planar joint count method.

[0068]

[0069] Among them, J V The number of volumetric joints in the rock mass; m i M is the number of joints within the unit measurement area of ​​the i-th image-taking zone; M is the number of image-taking zones at the tunnel face.

[0070] The rock mass integrity coefficient of the working face in S8 is based on the rock mass integrity coefficient K in the "Engineering Rock Mass Classification Standard". v With the number of joints in the rock mass J v The relationship table is converted.

[0071] Table 1 Rock mass integrity coefficient K v With the number of joints in the rock mass J v Comparison Table

[0072]

[0073] Generally, based on Table 1, the specific calculation formula obtained by interpolation is shown below:

[0074]

[0075] The beneficial effects of this invention are as follows: Geological parameters are acquired in real time using field testing and digital image processing methods. A real-time rock grading learning model is constructed by modeling the acquired geological parameters using RBP neural networks and Mask R-CNN convolutional neural networks, and the model parameters are optimized. Based on the optimized model parameters, a real-time online grading prediction model for the surrounding rock at the tunnel face is constructed. Addressing the current limitations of indoor and field testing methods in accurately and in real-time characterizing the geological parameters of the surrounding rock at the tunnel face, this invention proposes a real-time online grading method for surrounding rock quality under mining construction conditions, based on intelligent algorithms and image recognition. This aims to accurately predict the surrounding rock quality of the current excavated strata, enabling timely adjustment and optimization of excavation plans and construction parameters, and providing timely advance support when adverse geological conditions occur, ensuring the stability of the surrounding rock and safe and efficient construction during tunnel excavation. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the process of the present invention;

[0077] Figure 2 This is a schematic diagram of the rebound zone of a typical highway and railway tunnel working face according to the present invention;

[0078] Figure 3 This is the layout diagram of the geological radar survey line described in this invention;

[0079] Figure 4 This is a schematic diagram illustrating the calculation of geological development degree in this invention;

[0080] Figure 5 This is a comparison chart of predicted values ​​and measured values ​​in the training set of the RBF neural network of this invention;

[0081] Figure 6 This is a schematic diagram of the on-site image capture scheme for the surrounding rock at the working face according to the present invention;

[0082] Figure 7 This is a complete diagram of the surrounding rock at the working face of this invention;

[0083] Figure 8 This is a labeled image of the joints in this invention;

[0084] Figure 9 This is a graph showing the change in the loss value of this invention;

[0085] Figure 10 This is a binarized image of the joints on the face of the machine tool according to the present invention. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0087] like Figure 1 As shown, a method for intelligent classification of surrounding rock at tunnel faces based on neural networks comprises the following steps:

[0088] Step 1: Obtaining Rock Strength

[0089] S1: Select the rebound value and geological development degree of the surrounding rock at the working face from the field test as the input indicators for the strength of the surrounding rock.

[0090] Based on the principle of speed and effectiveness, real-time obtainable geological parameters of the surrounding rock are selected as input indicators for the real-time classification model of the surrounding rock, including:

[0091] 1) Select the rebound value of the surrounding rock at the tunnel face as the input index: Divide the tunnel face into zones. When dividing, the asymmetry of the surrounding rock properties on the left and right sides needs to be considered horizontally, and the lithological differences between the arch, middle, and bottom sections need to be considered vertically. Typical rebound zoning of highway and railway tunnel working faces is shown below. Figure 2 As shown. A rebound hammer was used to conduct 16 rebound tests on each section of the rock mass. The distance between every two rebound points was not less than the diameter of the impact rod, and the rebound hammer was kept perpendicular to the rock mass surface. The three maximum values ​​and three minimum values ​​were removed from the obtained data, and the average of the remaining 10 data was taken as the rebound value.

[0092] 2) Select the geological development degree of the surrounding rock at the tunnel face as an input indicator: Arrange one horizontal survey line at the top, middle, and bottom of the tunnel face to detect the rock and soil mass, as shown below. Figure 3 As shown; the ground-penetrating radar (GPR) maps were measured at three survey lines using a ground-penetrating radar instrument; the geological development degree was calculated based on the typical GPR map of the tunnel face. A schematic diagram for calculating the geological development degree is shown below. Figure 4 The calculation formula is shown in equation (1).

[0093]

[0094] Where: D—geological development degree; S n — Total area of ​​energy clusters within the region; S — Total area of ​​the region.

[0095] S2. Construct an RBF learning model based on the RBF neural network intelligent algorithm, and input the learning sample data obtained from S1 into the RBF learning model to optimize the parameters.

[0096] (1) Select a portion of the tunnel face data as a training sample (x i ,y i ), forming the sample training set y i ∈+, where n is the number of samples, x i A two-dimensional vector representing factors influencing rock mass strength, including the normalized rebound value of the surrounding rock at the working face and the normalized geological development degree, y i It is the output vector and its value is the normalized value of the corresponding uniaxial compressive strength of the rock.

[0097] In this embodiment, the learning sample set consists of 15 sets of surrounding rock parameters from the tunnel faces of the Luanlu Expressway Tunnel and the Hangzhou-Wenzhou Railway Tunnel. The input indicators, surrounding rock rebound value and geological development degree, and the output indicator, rock strength, are normalized to obtain the input and output of the RBF neural network.

[0098]

[0099] Where: x i For x′ i Normalized data; x min For x′ i The minimum value of x; max For x′ i The maximum value.

[0100] Table 2 shows the rebound values ​​and geological development of the surrounding rock at the training face obtained from the field test, and Table 3 shows the input and output of the normalized RBF neural network.

[0101] Table 2. Rock strength parameters at the working face from field training.

[0102]

[0103] Table 3 Input and output parameters of the RBF neural network in the training set

[0104]

[0105]

[0106] (2) According to the RBF algorithm, the input variable X is first subjected to a nonlinear transformation to transform the low-dimensional pattern input data into a high-dimensional space, making the linearly inseparable problem in the low-dimensional space linearly separable in the high-dimensional space. The radial basis function selected for the nonlinear transformation is:

[0107]

[0108] Where X is the input vector; C j C is the center vector of the j-th neuron in the hidden layer. j =[c j1 ,c j2 ,…,c jn ] T σ is the variance of the radial basis functions;

[0109] Through the above transformation, the output value z of the j-th neuron in the hidden layer is... j for:

[0110]

[0111] Among them, D j D is the width vector of j neurons in the hidden layer. j =[d j1 ,d j2 ,…,d jn ] T p represents the total number of hidden layer neurons.

[0112] Center vector parameter c ji Initial value:

[0113]

[0114] Where mini is the minimum value of all input information for the i-th metric in the training set, and maxi is the maximum value of all input information for the i-th metric in the training set.

[0115] Width vector parameter d ji Initial value:

[0116]

[0117] Where, d f It is the width adjustment coefficient, and its value is less than 1.

[0118] (3) Next, the output of the output layer neuron is obtained by the linear weighted sum of the hidden layer outputs:

[0119] Y = [y1, y2, ..., y q ] T

[0120]

[0121] Among them, w kj q is the weight between the k-th neuron in the output layer and the j-th neuron in the hidden layer; q is the total number of neurons in the output layer.

[0122] (4) Finally, the errors are compared and the optimization parameters are updated to their optimal values ​​using gradient descent. The optimization parameters include the center parameter c of the radial basis function. ji Variance (width) parameter d ji and the weights w from the hidden layer to the output layer kj The parameter update calculation is as follows:

[0123]

[0124]

[0125]

[0126] Among them, c ji (t) represents the center vector parameter of the j-th hidden layer neuron in the t-th update calculation of the i-th input neuron; d ji (t) is related to the central parameter c ji (t) corresponds to the width vector parameter; w kj (t) represents the weight of the k-th output neuron relative to the j-th hidden neuron during the t-th update calculation; η is the learning factor; E is the RBF neural network evaluation function:

[0127]

[0128] Among them, O lk y is the expected output value of the k-th output neuron with the l-th input sample; lk It is the network output value of the k-th output neuron when the l-th input sample is received.

[0129] Verification of the calculation results of the surrounding rock strength at the tunnel face: First, examine the calculation results of the rock mass training sample to check whether the calculation accuracy of the training set meets the test requirements (see...). Figure 5 The training accuracy reaches 90% to meet the testing requirements. If the requirements are met, the test sample data is put into the pre-RBF prediction model for calculation and compared with the indoor test calculation results to see if its accuracy meets the requirements.

[0130] S3. Using the obtained optimized parameters, establish an RBF learning model. Input the test sample dataset obtained in S1 into the RBF prediction model to obtain the surrounding rock strength output results. The comparison results between the predicted values ​​and measured values ​​of the test samples are shown in Table 4.

[0131] Table 4 Comparison of Rock Strength Prediction Results of RBF Neural Network

[0132]

[0133] As shown in Table 4, the average deviation between the rock strength prediction results and the measured results of the RBF neural network for the test group sample data is 1.9 MPa, which is a relatively ideal result.

[0134] Step Two: Assessing the integrity of the rock includes:

[0135] S4. Acquire image data of the surrounding rock at the tunnel face.

[0136] Based on the on-site shooting conditions and the size of the tunnel face (span and height), the tunnel face will be divided into N equal parts (each shooting section will have an area of ​​5-20m²). 2 (For best results) Use a high-definition digital camera to photograph the surrounding rock at the working face of each section. The shooting plan is as follows: Figure 6 As shown, a digital camera was placed 10 meters directly in front of the tunnel working face on the tunnel centerline for taking pictures, with the shooting direction perpendicular to the tunnel working face. This article relies on the fact that the working face area of ​​the tunnels in the engineering projects is generally less than 100 square meters. 2 The tunnel working face was divided into six detailed zones based on the tunnel height (three equal parts) and span (two equal parts): the left abutment zone, the crown zone, the right abutment zone, the left side wall zone, the floor zone, and the right side wall zone. Images were taken after muck removal from the tunnel working face and before the steel arch frame was erected.

[0137] S5. Obtain a complete image of the working face using image recognition processing methods.

[0138] The images of the tunnel face are stitched together to obtain a complete image of the tunnel face (e.g., Figure 7 The specific steps are as follows: The six images of the tunnel face taken on-site are roughly arranged in Photoshop according to their actual positions. The positions of the images are then fine-tuned based on the overlapping details at the stitching points to ensure they are aligned as closely as possible. The images are then rasterized, and the process is completed by clicking Edit → Auto Blend Layers → Panorama → OK to stitch and blend the images together.

[0139] S6. Construct a Mask RCNN learning model based on the Mask RCNN convolutional neural network algorithm, and input the learning sample data into the Mask RCNN learning model to optimize the parameters.

[0140] (1) First, the sample images are divided into training and validation sets at a ratio of 80% and 20%, respectively. Joint annotation is performed on the complete images in the training set. This is done by enhancing the image set by increasing contrast and expanding the image set by flipping, and then using Matlab's built-in Image Labeler tool to annotate the resulting sample images, identifying the fissure (joint) target regions. The annotated images are then exported to the workspace (see...). Figure 8 (This is for use in subsequent processing.)

[0141] (2) The preprocessed image dataset is input into the Mask RCNN convolutional neural network for training. The initial learning rate is set to 0.001, and the maximum number of iterations (max epochs) is set to 200. The classification loss (loss_classifier), localization loss (loss_box_reg), segmentation loss (loss_mask), and total loss of this network are calculated respectively, and their relationship is shown in Equation (13). The changes of each loss function are as follows: Figure 9 As shown, the model tends to converge when the loss function reaches its minimum and stabilizes. The loss value of the training set sample image data decreases continuously during the training process and finally stabilizes at epoch=35, where the localization loss < classification loss < segmentation loss.

[0142] S7. Utilize optimized parameters to establish a Mask R-CNN learning model. Input 20% of the test sample data into the Mask R-CNN prediction model to obtain a binarized image of the joints on the tunnel face (e.g., ...). Figure 10 );

[0143] S8. Calculate the volumetric joint number of the surrounding rock mass at the working face, and then obtain the rock mass integrity coefficient.

[0144] The number of volumetric joints in the surrounding rock mass at the tunnel face is calculated using the planar joint count method, as shown in formula (14).

[0145] The rock mass integrity coefficient of the tunnel face surrounding rock is based on the rock mass integrity coefficient K in the "Engineering Rock Mass Classification Standard". v With the number of joints in the rock mass J v The conversion is performed using Table 1. Generally, the result is obtained by interpolation based on Table 1, and the specific calculation is shown in Equation (15).

[0146] Step 3: Surrounding rock classification includes:

[0147] S9. Using the rock strength and rock integrity coefficients obtained in steps one and two, the BQ method is used to obtain the surrounding rock classification results.

[0148] Engineering testing and verification: Thirteen tunnels, including Gutangyuan No. 2 Tunnel, Jinzhuping Tunnel, and Mufeiling Tunnel, on the Hangzhou-Wenzhou Railway were the first to apply this classification system for intelligent classification of surrounding rock. The comparison between the system's predicted classification, the classification during the exploration stage, and the actual excavation results is shown in the table below.

[0149] Table 5 Comparison of Classification Results of Tunnels on the Hangzhou-Wenzhou Railway

[0150]

[0151] As shown in the table above, the accuracy rate of the classification results obtained using the intelligent classification system for mountain tunnel surrounding rock is 92.3%, which is a significant improvement compared to the 53.8% accuracy rate of the classification during the survey stage. However, the DK74+240 excavation face at the entrance of the Shijian Tunnel is located at the junction of the cross passage and the entrance, where poor ventilation inside the tunnel leads to low visibility, which may interfere with the system's image recognition of the tunnel working face, resulting in a more conservative prediction result.

[0152] This invention utilizes a neural network algorithm for intelligent classification of surrounding rock at tunnel faces. It accurately predicts the rock mass level at the tunnel face, enabling timely adjustment and optimization of excavation plans and construction parameters, thus ensuring safe and efficient tunnel construction. In engineering applications, this invention applies RBF neural network technology to rock strength calculation, while simultaneously using image recognition and Mask RNNN convolutional neural networks to calculate the rock mass integrity coefficient. Finally, it obtains the real-time classification result of the surrounding rock at the tunnel face based on the BQ method. Verification has shown that this method meets the requirements for engineering classification applications, providing a novel approach for real-time rock mass classification during tunnel excavation.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent classification of surrounding rock at tunnel faces based on neural networks, characterized in that: This includes the sequential steps of obtaining surrounding rock strength, obtaining rock integrity, and classifying surrounding rock, among which: Step 1: Obtaining the strength of the surrounding rock includes: S1. Select the rebound value and geological development degree of the surrounding rock at the working face from the field test as input indicators; S2. Construct an RBF learning model based on the RBF neural network intelligent algorithm, and input the learning sample data obtained from S1 into the RBF learning model to optimize the parameters; S3. Build an RBF learning model using optimized parameters, input the test sample data obtained in S1 into the RBF prediction model, and obtain the surrounding rock strength output result; Step Two: Assessing the integrity of the rock includes: S4. Acquire image data of the surrounding rock at the tunnel face; S5. Obtain a complete image of the working face using image recognition processing methods; S6. Construct a Mask RCNN learning model based on the Mask RCNN convolutional neural network algorithm, and input the learning sample data into the Mask RCNN learning model to optimize the parameters; S7. Build a Mask RCNN learning model using optimized parameters, input the test sample data into the Mask RCNN prediction model, and obtain a binary image of the joints on the working face; S8. Calculate the volumetric joint number of the surrounding rock mass at the tunnel face, and then obtain the rock mass integrity coefficient; Step 3: Surrounding rock classification includes: S9. Using the rock strength and rock integrity coefficients obtained in steps one and two, the BQ method is used to obtain the surrounding rock classification results; Geological development degree, which is calculated by the following formula: （1） In the formula: —Geological development level; —Total area of ​​energy clusters within the region; —Total area of ​​the region.

2. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, The method for selecting the rebound value of the surrounding rock at the working face as the input index in S1 is as follows: the working face is divided into zones. When dividing the zone, the asymmetry of the properties of the surrounding rock on the left and right sides is considered in the horizontal direction, and the lithological differences of the surrounding rock in the arch, middle and bottom are considered in the vertical direction. A rebound hammer is used to conduct 16 rebound tests on the rock mass of each zone. The distance between every two rebound points is not less than the diameter of the impact rod, and the rebound hammer is perpendicular to the surface of the rock mass. The three maximum values ​​and three minimum values ​​are removed from the obtained data, and the average of the remaining 10 data is taken as the rebound value. The method for selecting the geological development degree of the tunnel face as the input index in S1 is as follows: 1 horizontal survey line is arranged at the arch, middle and bottom of the tunnel working face to detect the rock and soil of the tunnel face; the geological radar map at the three survey lines is measured using a ground-penetrating radar instrument; the parameter—geological development degree—is calculated based on the typical geological radar map of the tunnel face.

3. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, The learning sample data in S2 is a selection of data from the tunnel face. , forming the sample training set , For the number of samples, A two-dimensional vector representing factors influencing rock mass strength, including the rebound value of the surrounding rock at the working face and the degree of geological development. It is an output vector and its value is the corresponding uniaxial compressive strength of the rock.

4. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, The method for constructing the RBF learning model in S2 is as follows: based on the RBF three-layer neural network, the normalized learning samples are... The input layer vector is nonlinearly transformed to the hidden layer, and then the hidden layer is linearly transformed to the output layer. Finally, the output layer result is analyzed. Inverse normalization was performed to obtain the predicted results of the uniaxial compressive strength of the rock. ; In the sample The normalization process is shown in equation (2): （2） in: ; ; The method for nonlinearly transforming the input layer vector to the hidden layer is as follows: Radial basis functions (RBFs) are used as the "basis" of the hidden units to construct the hidden layer space. The hidden layer transforms the input vector, converting the low-dimensional pattern input data to a high-dimensional space, making the linearly inseparable problem in the low-dimensional space linearly separable in the high-dimensional space. The radial basis function (RBF) chosen for the nonlinear transformation is: （4） in, It is the input vector; It is the hidden layer. The center vector of each neuron ; It is the variance of the radial basis functions; Output layer results The inverse normalization process is shown in equation (3): （3） in:

5. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, In S4: The method for obtaining image data of the surrounding rock at the tunnel face is as follows: Based on the on-site shooting conditions and the size of the tunnel face, the tunnel face was divided into N equal parts, with each shooting section having an area controlled between 5 and 20 square meters. 2 High-definition digital cameras were used to capture images of the surrounding rock at the working face of each section; the optimal time for image acquisition was after the muck removal process was completed.

6. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, In S5: Obtaining a complete image of the tunnel face includes stitching and merging the segmented images of the tunnel face. The specific steps are as follows: Place the N segmented images of the tunnel face taken on-site in Photoshop according to their actual positions, and fine-tune the positions of the segmented images based on the overlapping image details at the stitching points to ensure alignment; then rasterize the segmented images and click Edit → Auto Blend Layers → Panorama → OK to achieve the stitching and merging of the segmented images.

7. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, In S6: The Mask RCNN convolutional neural network structure consists of five main parts: the backbone architecture and feature pyramid network FPN, region recommendation network RPN, region of interest alignment ROIAlign, classifier, and segmentation mask generation network. The backbone architecture uses a deep residual network ResNet101 to extract feature maps from the image. The anchor labeling rule in the Region Recommendation Network (RPN) adopts the Non-Maximum Suppression (NMS) method, and iterative calculation is performed according to the principle shown in Equation (12). First, the highest-scoring box is obtained, and the IoU value of other boxes is calculated with this box. When the IoU value is greater than 0.6, this box is labeled as a negative sample. Then, the second highest-scoring box is selected, and the next iteration is performed until all are completed. （12） In the formula: s i Let M be the score of the i-th box; M is the target box, b i The suggestion box with the highest confidence; N t This is the threshold for IoU, which is set to 0.6 here; The pooling method in the Region of Interest Align (ROIAlign) calibration is a bilinear interpolation method that retains the floating-point number of pixels in the ROIAlign pooling method, so as to obtain more accurate joint and crack recognition results. The classifier uses a fully connected network of 1024 neurons to classify and regress each ROI, obtaining the final category and bounding box for crack detection. The segmentation mask generation network achieves the effect of segmenting the target by masking the joints and fissures of the target to be detected through mask regions.

8. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, In S6: Mask RCNN convolutional neural network defines a multi-task loss, and its loss function is: L=L cls +L box +L mask (13) Where L cls L box and L mask These refer to classification loss, detection box loss, and mask loss, respectively.

9. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, The input image annotation method of the Mask RCNN convolutional neural network in S6 is as follows: based on the image enhancement and image set expansion of the sample images, the obtained sample images are annotated using the Image Labeler image annotation tool provided by Matlab to mark the fissure target region, and the labeled image is exported to the workspace for subsequent processing.

10. The intelligent classification method for surrounding rock at a tunnel face based on a neural network according to claim 1, characterized in that, In S8: The number of volumetric joints in the surrounding rock mass at the tunnel face is calculated using the planar joint count method. （14） in, The number of joints in the rock mass; It is the number of joints within the measured area of ​​the i-th image capture unit; It refers to the number of camera zones at the tunnel face; The rock mass integrity coefficient of the tunnel face is based on the rock mass integrity coefficient in the "Engineering Rock Mass Classification Standard". Number of joints in rock mass The relational table is converted to Table 1; Table 1 Rock mass integrity coefficient Number of joints in rock mass Comparison Table Based on Table 1, the specific calculation formula obtained by interpolation is shown below: （15）。 。

Images