Tunnel smooth blasting design parameter intelligent optimization method based on deep learning and neural network

Abstract

The invention relates to a tunnel smooth blasting design parameter optimization method based on deep learning and a neural network, belongs to the technical field of tunnel blasting construction, solves the problem of optimization design of tunnel blasting design parameters at present, and comprises the following steps: loading processed image data into an SP-Yolov8 model and a Yolov8-seg model; and correspondingly outputting a main structural plane dip angle category and a main structural plane region, extracting main structural plane information to obtain a main structural plane position, inputting blasting design parameters, an integrity coefficient, uniaxial compressive strength, the main structural plane dip angle category and the main structural plane position into a DBN model, and outputting a maximum linear over-excavation predicted value. And comparing the relationship between the maximum linear over-excavation predicted value and a limit value, and outputting a final blasting design parameter. According to the method, the final blasting design parameters can be rapidly and accurately obtained, and the method has the advantages of being accurate in feature information extraction, high in robustness, high in convergence speed and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel blasting construction, and in particular to a method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network. Background Art

[0002] Geological work runs through the entire process of tunnel investigation, design and construction. Due to the complexity and variability of geological conditions, the surrounding rock actually revealed during the construction process is often quite different from the geological survey data, resulting in unreasonable construction parameters, causing economic losses, delays in construction period, safety accidents and other problems. The drilling and blasting method is widely used in tunnel construction due to its low cost and strong geological applicability. However, due to the influence of the "instantaneous" characteristics of the work, it is often difficult to accurately control it in engineering applications, and over-excavation, under-excavation and block size problems often occur. These two problems will directly affect the safety, quality and efficiency of tunnel excavation.

[0003] For a long time, tunnel over-excavation has been one of the major problems that plagues the construction of drilling and blasting in my country. This will not only increase the cost of engineering construction, but also easily cause local stress concentration in the surrounding rock, thereby affecting the overall stability of the surrounding rock. In order to predict and control the over-excavation caused by tunnel drilling and blasting, Ibarra et al. predicted the over-excavation based on the relationship between the surrounding rock classification system and the unit explosive consumption, and found that the increase in unit explosive consumption can reduce the over-excavation, and the increase in Q value can lead to an increase in under-excavation; Singh et al. studied the factors affecting the damage of over-excavation through a series of experiments, and pointed out that all influencing parameters can be divided into three categories, including rock characteristics, explosives and blasting modes; in addition, existing research results also show that there are many factors that affect over-excavation, such as plugging depth, plugging, the last row of charges and total charges, special charges, special charges for each delay, number of drilling rows, rock strength, rock quality index, rock weathering, groundwater, etc. How to minimize the impact of surrounding rock blasting damage during blasting excavation, reduce over-excavation and ensure the stability of surrounding rocks have become basic issues in the process of tunnel blasting construction.

[0004] In recent years, information technology represented by artificial intelligence technology has developed rapidly, and cross-disciplinary research has been carried out in various fields and disciplines, and has been applied to the fields of civil and mining engineering. Xiao Qinghua has introduced deep learning and image recognition technology to automatically extract and segment tunnel over-excavation and under-excavation images, thereby realizing intelligent statistics of over-excavation and under-excavation and intelligent classification of blasting excavation quality levels; Wu Renjie, Nie Jun, Ku-latilake, etc. have established prediction models for blasting block size based on principal component analysis, gene expression programming algorithm, and neural network technologies.

[0005] At present, computer vision detection based on deep learning methods has been widely used in many aspects including tunnel lining disease detection and vehicle identification in tunnels. Theoretical research and engineering facts show that it is a feasible means to apply this technology to tunnel construction. However, tunnel blasting design parameters are a multi-parameter high-dimensional nonlinear problem. In order to finally determine the optimal smooth blasting design parameters under the corresponding surrounding rock conditions and achieve high-quality tunnel blasting, how to use the on-site tunnel face image data to optimize the blasting design parameters is a difficult problem faced in current underground tunnel construction projects. Summary of the invention

[0006] In view of the defects existing in the above-mentioned prior art, the purpose of the present invention is to propose a tunnel smooth surface blasting design parameter optimization method based on deep learning and neural network, to improve the deep learning Yolov8 model, so as to obtain the main structure surface inclination classification and recognition in the tunnel face image and the main structure surface information extraction, and then input the blasting design parameter optimization data set into the DBN deep neural network for training, to obtain the optimal final blasting design parameters, so as to solve the problems existing in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] On the one hand, the present invention discloses a method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network, comprising the following steps:

[0009] S1: Select blasting design parameters, integrity coefficient and uniaxial compressive strength;

[0010] S2: Acquire tunnel face image data and preprocess it;

[0011] S3: Load the preprocessed tunnel face image data into the SP-Yolov8 model and the Yolov8-seg model, and output the main structural surface inclination category and the main structural surface area accordingly;

[0012] S4: extracting the main structural surface information from the main structural surface area to obtain the main structural surface position;

[0013] S5: input blasting design parameters, integrity coefficient, uniaxial compressive strength, main structural surface inclination type, and main structural surface position into the optimized DBN model, and output the maximum linear over-excavation prediction value;

[0014] S6: If the output maximum linear over-break prediction value is higher than the limit value, modify the blasting design parameters and return to step S5 for re-prediction until the output maximum linear over-break prediction value is less than or equal to the limit value, then output the final blasting design parameters.

[0015] As a further preferred solution of the above technical solution: in step S2, the tunnel face image data is subjected to sliding window cutting processing, enhancement processing and normalization processing in sequence, wherein: the formula for normalization processing is:

[0016]

[0017] In the formula: y is the value of the data after normalization; x is the original data before normalization, max is the maximum value of the data before normalization; x min is the minimum value of the data before normalization.

[0018] A further preferred solution is: in step S3, the SP-Yolov8 model is obtained by optimizing the Backbone network structure of the Yolov8 model, and the SP-Yolov8 model is used to classify and identify the main structural surface inclination of the tunnel face. The specific steps are as follows:

[0019] S31: Add an Identity layer in parallel to the three asymmetric convolution blocks to form an SP-DWConv2D convolution layer;

[0020] S32: Bring the SP-DWConv2D convolutional layer into the Bottleneck structure in the Backbone network structure, and replace it upwards in sequence to form a new SP-C2f module network structure;

[0021] S33: Replace the C2f module in the Yolov8 model with the SP-C2f module network structure to obtain the SP-Yolov8 model.

[0022] A further preferred solution is: in step S32, the SP-DWConv2D convolution layer includes a Split layer at the input front end, three deep convolution kernels, an Identity layer, and a concat layer at the back end for feature fusion, wherein the sizes of the three deep convolution kernels are 3×3 convolution kernel, 1×9 band kernel, and 9×1 band kernel, respectively.

[0023] A further preferred solution is: in step S3, when classifying and identifying the inclination angle of the main structural surface, a linear frame is used to mark and classify the main structural surface in the tunnel face image, and the results of the marking and classification are saved in a JSON file. The results of the marking and classification are mobilized together with the tunnel face image through the SP-Yolov8 model to identify the category of the inclination angle of the main structural surface.

[0024] A further preferred solution is: in step S3, based on the Head structure in the Yolov8 model, by adding the Soft-NMS filtering algorithm with the aspect ratio constraint and the distance constraint of the main structural surface, the Yolov8-seg model is obtained after optimization, and the false main structural surface is eliminated by the Yolov8-seg model to obtain the main structural surface area. The specific steps include:

[0025] First, based on the aspect ratio constraint, the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is calculated. Then the aspect ratio constraint is introduced and the aspect ratio of each detection frame is calculated. If the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is greater than or equal to the aspect ratio constraint, the detection frames with too large range or useless are removed to obtain a new set of detection frames.

[0026] Then, based on the distance constraint, the probability scores of the new detection boxes are sorted from high to low, and the center coordinates x are calculated based on the upper left corner coordinates and the lower right corner coordinates of the detection box. o1 、x o2 , with x o1 As the midpoint, the distance constraint R is introduced. When x o2 With x o1 When the distance is less than or equal to R, the detection box with a lower probability score is eliminated.

[0027] A further preferred solution is: in step S4, the step of extracting the main structural surface information includes:

[0028] S41: input the height Y of the tunnel face image and the height h of the tunnel face into the SP-YOLOv8 model, and the coordinates of the lower right corner endpoint of the tunnel face image are taken as the origin;

[0029] S42: The tunnel face image is segmented by the YOLOv8-seg model, and the main structural surface is expanded to form a mask area. The coordinates of the lower right corner of the mask area box are (x1, y1);

[0030] S43: extract the main structural surface area and perform image processing using binarization;

[0031] S44: The Zhang-Suen algorithm is used to extract the skeleton of the binary image. The region to be extracted is iterated multiple times, and a trace is formed by indenting from the edge to the center. At this time, the coordinates of the right end point of the trace are (x2, y2);

[0032] S45: The pixel height y=y1+y2 of the right end point of the main structural surface is obtained by the image pixel height, and the actual height H=y×(h / Y) of the right end point of the main structural surface is obtained according to the ratio of the actual face height, thereby obtaining the position of the main structural surface.

[0033] A further preferred solution is: in step S5, the DBN model is composed of a two-layer RBM and a single-layer BP neural network, and the PSO algorithm is used to assist the DBN model in finding the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the learning rate and the number of training times in the hyperparameters.

[0034] On the other hand, the present invention discloses an electronic device, including a processor and a memory, wherein computer instructions are stored in the memory, and the processor is used to run the computer instructions stored in the memory to implement the steps of a tunnel smooth blasting design parameter optimization method based on deep learning and neural network as described in any of the above items.

[0035] On the other hand, the present invention further discloses a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the steps of any of the above-mentioned methods for optimizing design parameters of tunnel smooth blasting based on deep learning and neural networks.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention studies the intelligent optimization of tunnel smooth surface blasting design parameters, and constructs a blasting design parameter optimization method based on deep learning and deep neural network. First, the deep learning network based on the YOLOv8 model is optimized, and the optimized SP-YOLOv8 model is used to obtain the main structural surface inclination classification, and the optimized YOLOv8-seg model is used to obtain the main structural surface area. The main structural surface position is obtained by cooperating with the YOLOv8 model and the YOLOv8-seg model, so as to construct a tunnel smooth surface blasting design parameter optimization data set, thereby determining the input parameters and output parameters of the DBN neural network model, and obtaining the optimal final blasting design parameters after multiple training of the DBN model.

[0038] 2. The SP-Yolov8 model of the present invention can effectively reduce the image preprocessing steps in the traditional method, and can still complete the main structural surface information extraction when there are many image interference factors. It has the advantages of accurate feature information extraction and fast convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required in the embodiments are briefly introduced below.

[0040] Figure 1 It is a flow chart of a method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network of the present invention;

[0041] Figure 2 The input parameters of the tunnel smooth blasting design parameter optimization model of the present invention;

[0042] Figure 3 The present invention is a process of performing sliding window cutting processing on tunnel face image data;

[0043] Figure 4 It is a network structure diagram of the SP-DWConv2D convolutional layer of the present invention;

[0044] Figure 5 A process diagram of the present invention replacing the C2f module of the YOLOv8 model with the SP-C2f module network to obtain the SP-YOLOv8 model;

[0045] Figure 6 is the network structure of the DBN model of the present invention;

[0046] Figure 7 These are the training and verification fitting curves for the BP model, CNN model, and DBN model in the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the embodiments will be described clearly and completely in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments, not all of the embodiments. For ordinary technicians in this field, other embodiments can be obtained without creative work.

[0048] Reference Figure 1-7 The present invention provides a method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network, comprising the following steps:

[0049] S1: Select blasting design parameters, integrity coefficient and uniaxial compressive strength;

[0050] S2: Acquire tunnel face image data and preprocess it;

[0051] S3: Load the preprocessed tunnel face image data into the SP-Yolov8 model and the Yolov8-seg model, and output the main structural surface inclination category and the main structural surface area accordingly;

[0052] S4: extracting the main structural surface information from the main structural surface area to obtain the main structural surface position;

[0053] S5: input blasting design parameters, integrity coefficient, uniaxial compressive strength, main structural surface inclination type, and main structural surface position into the optimized DBN model, and output the maximum linear over-excavation prediction value;

[0054] S6: If the output maximum linear over-break prediction value is higher than the limit value, modify the blasting design parameters and return to step S5 for re-prediction until the output maximum linear over-break prediction value is less than or equal to the limit value, then output the final blasting design parameters.

[0055] like Figure 2 As shown in FIG. 1 , when considering the input parameters of the tunnel smooth blasting design parameter optimization model, there are mainly two aspects: surrounding rock quality parameters and blasting design parameters. The surrounding rock quality parameters include the integrity coefficient X1 and the uniaxial compressive strength X2 (MPa) in step S1, and the main structural surface inclination X3 and the main structural surface position X4 in step S2, which have a greater impact on the maximum linear over-excavation in tunnel blasting construction. The blasting design parameters include the peripheral eye charge X5 (kg), the peripheral eye hole distance X6 (cm), the outer ring auxiliary eye hole distance X7 (cm), the outer ring auxiliary eye charge X8 (kg), the smooth blasting layer thickness X9 (cm), the peripheral eye charge X10 (kg) at the structural surface, and the peripheral eye charge X2 (kg). 10 (kg), the amount of charge of the outer ring auxiliary eye at the structural surface X 11 (kg), number of blast holes X 12 , total explosive quantity X 13 (kg) and cycle footage X 14 (cm), and the above 14 parameters are used as input parameters of the tunnel blasting design parameter optimization model.

[0056] In step S2, the tunnel face image data is obtained from the construction site, and the collected tunnel face image data is preprocessed. Considering that the tunnel face surrounding rock image collected according to the acquisition standard is large in size, the background is complex, and the main structure surface pixel ratio is small, if it is directly trained on the tunnel blasting design parameter optimization model, there are problems such as excessive time consumption and difficulty in determining the model structure and parameters. Therefore, it is necessary to perform sliding window cutting processing on each tunnel face image. The sliding window cutting processing process is as follows: Figure 3 shown.

[0057] After the tunnel face image is processed by sliding window cutting, the image data enhancement technology is used to increase the quantity and quality of the tunnel face surrounding rock image database. In order to improve the convergence speed and calculation accuracy of the tunnel blasting design parameter optimization model, the enhanced data is normalized to unify the dimension and variation range. The "mapminmax" normalization formula in MATLAB is used for data normalization. The processed data range is [0, 1]. The normalization formula is as follows:

[0058]

[0059] In the formula: y is the value of the data after normalization; x is the original data before normalization, max is the maximum value of the data before normalization; x minis the minimum value of the data before normalization.

[0060] After normalization, a data set suitable for the subsequent prediction of the maximum linear over-break neural network is obtained, namely, the maximum linear over-break data set for tunnel blasting.

[0061] In step S3, the SP-Yolov8 model is obtained by optimizing the Backbone network structure based on the Yolov8 model, and the SP-Yolov8 model is used to classify and identify the inclination angle of the main structural surface of the tunnel face.

[0062] In the YOLOv8 model, the Backbone network structure is mainly used to extract image feature information. When the Backbone network structure is used to train the main structural surface inclination classification and recognition data set, the image feature information is lost.

[0063] The core of the Backbone network is the C2f module, which is used to convert the output of the convolutional layer into the input of the fully connected layer. It is the key module that causes missed detection and false detection after YOLOv8 model training. In the C2f module, it is the Bottleneck structure that can improve the effect. Therefore, this application improves the performance of the Bottleneck structure in the tunnel face main structure surface inclination classification and recognition dataset by adjusting the number and size of the convolution kernels in the Bottleneck structure.

[0064] When optimizing the Bottleneck structure, an Identity layer is added in parallel to the asymmetric convolution block, and a structural plane-DWConv2D (SP-DWConv2D) convolution layer is designed for the tunnel face main structure surface inclination classification and recognition dataset. The principle of the SP-DWConv2D convolution layer is as follows: Figure 4 shown.

[0065] More specifically, the SP-DWConv2D convolution layer includes a Split layer at the input front end, three deep convolution kernels, an Identity layer, and a concat layer for feature fusion at the back end. The sizes of the three deep convolution kernels are 3×3 convolution kernels, 1×9 banded kernels, and 9×1 banded kernels. The SP-DWConv2D convolution layer is brought into the Bottleneck structure and replaced upward in sequence to form a new structural plane-C2f (SP-C2f) module network structure. The C2f module of YOLOv8 is replaced with the SP-C2f module network to obtain the optimized YOLOv8 model, as shown in Figure 1. Figure 5 As shown, this is the SP-YOLOv8 model, which is the tunnel blasting design parameter optimization model mentioned above.

[0066] Then, the SP-YOLOv8 model is used to classify and identify the inclination angle of the main structure surface of the tunnel. Before classification and identification, the main structure surface is marked in the tunnel face image in the form of a linear frame and classified according to the angle.

[0067] Specifically, for the input parameter of the main structural surface inclination X3, engineering geology divides rock mass into 5 categories according to the structural surface inclination, namely, nearly horizontal rock layer (0°~5°), gently inclined rock layer (5°~30°), inclined rock layer (30°~70°), steep rock (70°~85°), and vertical rock layer (85°~90°). This application draws on the rock mass classification of engineering geology and divides the main structural surface of the tunnel into 5 categories, and numbers them 1, 2, 3, 4, and 5 respectively. The main structural surface dataset is generated according to the VOC dataset format, and the main structural surface dataset is input into the SP-YOLOv8 model to output the main structural surface inclination type of the tunnel face.

[0068] For the input parameter X4 of the main structural surface position, a labeling frame is made according to the segmentation and extraction of the main structural surface. Specifically, the main structural surface is labeled in the tunnel face image in the form of a linear frame. The labeling frame should cover the main structural surface. The labeling frame is required to contain the complete main structural surface and the surrounding small area rock background. When there are multiple main structural surfaces in a face image, they are labeled one by one with a single linear labeling frame, and the labeling frame is saved in the form of a JSON file. Then, the labeling frame and the image are called together in deep learning, and the height of the main structural surface is calculated and recorded. The face area is classified according to the calculated height value, and the entire upper step face is divided into three areas according to height. The area within 100 cm from the arch is the arch area, the area 200 cm from the bottom of the step is the middle and lower part of the step, and the rest is the arch shoulder area. The recorded main structural surface heights are classified according to the above design requirements and recorded as 11, 12, and 13 respectively.

[0069] Since the Head structure in the original YOLOv8 model performs poorly in the recognition of the main structural surface area, false detection and re-detection may occur. Therefore, the Head structure of the YOLOv8 model is optimized, and the Soft-NMS filtering algorithm with aspect ratio constraints and distance constraints is added to the Head structure. After optimization, the YOLOv8-seg model is obtained. The aspect ratio constraint condition is used to eliminate some detection frames that are too large or useless. After the aspect ratio constraint condition is eliminated, a new detection frame set will be output. On this basis, the distance constraint condition is introduced to further eliminate the false main structural surfaces of the newly output detection frame to reduce the false detection rate and re-detection rate of the model; the specific method is as follows:

[0070] Based on the constraint algorithm of the aspect ratio constraint, the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is calculated, and then the aspect ratio constraint is introduced to calculate the aspect ratio of each detection frame. If the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is greater than or equal to the aspect ratio constraint, the detection frames with too large range or useless are eliminated to obtain a new set of detection frames;

[0071] Then, based on the constraint algorithm of the distance constraint condition, the probability scores of the new detection boxes are sorted from high to low, and the center coordinates x are calculated based on the coordinates of the upper left corner and the lower right corner of the detection box. o1 、x o2 , with x o1 As the midpoint, the distance constraint R is introduced. When x o2 With x o1 When the distance is less than or equal to R, the detection box with a lower probability score is eliminated.

[0072] Furthermore, the optimized YOLOv8-seg model is used to further segment and extract the main structural surface. After obtaining the main structural surface area, the image is segmented by the YOLOv8-seg model to form a mask image. The formed mask image is binarized, and then the Zhang-Suen skeleton extraction algorithm is used to obtain the endpoint coordinates of the main structural surface, thereby extracting the main structural surface information and obtaining the main structural surface position; specifically:

[0073] First, the height Y of the tunnel face image and the height h of the tunnel face are input into the SP-YOLOv8 model, and the coordinates of the lower right corner endpoint of the tunnel face image are taken as the origin; then the tunnel face image is segmented by the YOLOv8-seg model, the main structural surface is expanded to form a mask area, the lower right corner coordinates of the mask area box are (x1, y1), the main structural surface area is extracted and image processing is performed using binarization, the Zhang-Suen algorithm is used to extract the skeleton of the binarized image, the area to be extracted is iterated multiple times, and a trace is indented from the edge to the center. At this time, the coordinates of the right end point of the trace are (x2, y2), and the pixel height y=y1+y2 of the right end point of the main structural surface is obtained by the image pixel height. According to the ratio of the actual tunnel face height, the actual height H=y×(h / Y) of the right end point of the main structural surface is obtained, thereby obtaining the position of the main structural surface. This method can reduce the image preprocessing steps in the traditional method, and can still complete the main structural surface information extraction when there are many image interference factors.

[0074] In step S5, the maximum linear over-excavation Y1 (cm) is selected as the output parameter of the tunnel smooth blasting design parameter optimization model, and the tunnel contour after blasting is measured by a cross-section scanner. The tunnel blasting maximum linear over-excavation data set obtained above is input into the DBN deep network model for maximum linear over-excavation prediction; the DBN deep neural network is composed of an unsupervised lower layer superimposed by several restricted Boltzmann machines (RBMs) and a supervised upper layer composed of a BP neural network. RBM is a Markov random field proposed by Professor Hinton and is an important component of the deep belief network (DBN). RBM is a neural network structure composed of two layers: a hidden layer (h) and an input layer (v). The basic principle is to adjust the parameter settings through a large amount of training data so that the distribution probability represented by the RBM under the parameters is as consistent as possible with the training data.

[0075] In the present invention, a variety of neural network models (such as BP model, CNN model, DBN model) are compared and selected, and the prediction result fitting curve is obtained through training and verification. The performance of the above three models is comprehensively evaluated, and the DBN model is finally determined. The comparison process of the above three models is referred to Figure 7 As shown, where: Figure 7 Middle (a), Figure 7 (b) represents the prediction value fitting curves of the training results and verification results of the DBN model respectively; Figure 7 Middle (c), Figure 7 (d) represents the predicted value fitting curves of the training results and verification results of the BP model, respectively; (d) BP verification; Figure 7 Middle (e), Figure 7 (f) in the middle represents the predicted value fitting curves of the training results and verification results of the CNN model.

[0076] Depend on Figure 7 It can be seen that during the training and verification process, most of the scattered points of the three models are located on both sides of the standard line, and the scattered point position of the DBN model is closest to the standard line, followed by the BP model, and the CNN model has the farthest scattered point from the standard line. According to the fitting curve, it can also be seen that the fitting curve drawn by the scattered points of the DBN model is closest to the standard line, but there is still underfitting. The fitting curves of the BP model and the CNN model are quite different from the standard line, and the underfitting phenomenon is serious. Combining the above results, it can be concluded that the DBN model has the best fitting performance in the entire training and verification process.

[0077] Usually, the hyperparameters of the DBN model, such as the learning rate, number of hidden layer nodes, and number of training times, are set manually, which leads to the problem that the maximum relative error does not meet the actual requirements of engineering use. Therefore, it is necessary to optimize the setting of hyperparameters.

[0078] In the present invention, the particle swarm optimization algorithm (PSO) is used to optimize the DBN model, mainly to optimize the selection of hyperparameters to avoid the subjectivity of manual parameter determination. The PSO algorithm is used to find the number of nodes in the first hidden layer, the number of nodes in the second hidden layer, the learning rate and the number of training times in the DBN model hyperparameters. Figure 6 As shown, the DBN model in the present invention is composed of two layers of RBM and a single layer of BP neural network, the number of nodes in each RBM hidden layer is 20, the learning rate is 0.01, the number of RBM pre-training times is 500 times, and the number of reverse iterations is 3000 times; the tunnel blasting maximum linear over-excavation data set is input into the optimized DBN model, and the maximum linear over-excavation prediction value is output.

[0079] The maximum linear over-excavation prediction value obtained by the present invention is compared with the limit value. If the maximum linear over-excavation prediction value is higher than the limit value, the blasting design parameters are modified and the optimized DBN model is used again for prediction until the output maximum linear over-excavation prediction value is less than or equal to the limit value, and the final blasting design parameters are output.

[0080] It should be clear that the above detailed description of the embodiments of the present invention provided in the accompanying drawings is only used to further illustrate the present invention and is not intended to limit the scope of the invention claimed for protection, but only represents selected embodiments of the present invention. Based on the embodiments of the present invention, some non-essential improvements and adjustments made by ordinary technicians in this field according to the above content of the present invention belong to the protection scope of the present invention.

Claims

1. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network, characterized in that: The following steps are involved: S1: Select blasting design parameters, integrity coefficient and uniaxial compressive strength; S2: Acquire tunnel face image data and preprocess it; S3: Load the preprocessed tunnel face image data into the SP-Yolov8 model and the Yolov8-seg model, and output the main structural surface inclination category and the main structural surface area accordingly; S4: extracting the main structural surface information from the main structural surface area to obtain the main structural surface position; S5: input blasting design parameters, integrity coefficient, uniaxial compressive strength, main structural surface inclination type, and main structural surface position into the optimized DBN model, and output the maximum linear over-excavation prediction value; S6: If the output maximum linear over-break prediction value is higher than the limit value, modify the blasting design parameters and return to step S5 for re-prediction until the output maximum linear over-break prediction value is less than or equal to the limit value, then output the final blasting design parameters.

2. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network according to claim 1, characterized in that: In step S3, the SP-Yolov8 model is obtained by optimizing the Backbone network structure of the Yolov8 model, and the SP-Yolov8 model is used to classify and identify the main structural surface inclination of the tunnel face. The specific optimization steps are as follows: S31: Add an Identity layer in parallel to the three asymmetric convolution blocks to form an SP-DWConv2D convolution layer; S32: Bring the SP-DWConv2D convolutional layer into the Bottleneck structure and replace it upwards in sequence to form the SP-C2f module network structure; S33: Replace the C2f module in the Yolov8 model with the SP-C2f module network structure to obtain the SP-Yolov8 model.

3. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network according to claim 2, characterized in that: In step S32, the SP-DWConv2D convolution layer includes a Split layer at the input front end, three deep convolution kernels, an Identity layer, and a concat layer at the back end for feature fusion, wherein the sizes of the three deep convolution kernels are 3×3 convolution kernel, 1×9 band kernel, and 9×1 band kernel, respectively.

4. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network according to claim 3, characterized in that: In step S3, when classifying and identifying the inclination angle of the main structural surface, a linear frame is used to mark and classify the main structural surface in the tunnel face image, and the results of the marking and classification are saved in a JSON file. The results of the marking and classification are mobilized together with the tunnel face image through the SP-Yolov8 model to identify the inclination angle category of the main structural surface.

5. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network according to claim 4, characterized in that: In step S3, based on the Head structure in the Yolov8 model, the Yolov8-seg model is obtained after optimization by adding the Soft-NMS filtering algorithm with the aspect ratio constraint and the distance constraint of the main structural surface. The false main structural surfaces are eliminated by the Yolov8-seg model to obtain the main structural surface area. The specific steps include: First, based on the aspect ratio constraint, the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is calculated. Then the aspect ratio constraint is introduced and the aspect ratio of each detection frame is calculated. If the aspect ratio of the upper left corner coordinates and the lower right corner coordinates of the detection frame is greater than or equal to the aspect ratio constraint, the detection frames with too large range or useless are removed to obtain a new set of detection frames. Then, based on the distance constraint, the probability scores of the new detection boxes are sorted from high to low, and the center coordinates x are calculated based on the upper left corner coordinates and the lower right corner coordinates of the detection box. o1 、x o2 , with x o1 As the midpoint, the distance constraint R is introduced. When x o2 With x o1 When the distance is less than or equal to R, the detection box with a lower probability score is eliminated.

6. A method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network according to claim 5, characterized in that: In step S4, the step of extracting the main structural surface information from the main structural surface area includes: S41: input the height Y of the tunnel face image and the height h of the tunnel face into the SP-YOLOv8 model, and the coordinates of the lower right corner endpoint of the tunnel face image are taken as the origin; S42: The tunnel face image is segmented by the YOLOv8-seg model, and the main structural surface is expanded to form a mask area. The coordinates of the lower right corner of the mask area box are (x1, y1); S43: extract the main structural surface area and perform image processing using binarization; S44: The Zhang-Suen algorithm is used to extract the skeleton of the binary image. The region to be extracted is iterated multiple times, and a trace is formed by indenting from the edge to the center. At this time, the coordinates of the right end point of the trace are (x2, y2); S45: The pixel height y=y1+y2 of the right end point of the main structural surface is obtained by the image pixel height, and the actual height H=y×(h / Y) of the right end point of the main structural surface is obtained according to the ratio of the actual face height, thereby obtaining the position of the main structural surface.

7. An electronic device, comprising a processor and a memory, characterized in that: Computer instructions are stored in the memory, and the processor is used to run the computer instructions stored in the memory to implement the steps of a tunnel smooth blasting design parameter optimization method based on deep learning and neural network as described in any one of claims 1-6.

8. A computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable a computer to execute the steps of a method for optimizing design parameters of tunnel smooth blasting based on deep learning and neural network as described in any one of claims 1 to 6.

Images