Blasting design parameter optimization method and system based on deep learning and deep neural network

Abstract

The invention discloses a blasting design parameter optimization method and system based on deep learning and a deep neural network, and belongs to the technical field of engineering blasting safety to solve the problem of how to improve the precision and efficiency of blasting design parameter optimization, and the method comprises the specific steps: 1, loading a tunnel face image into an SP-YOLOv8 model and an SPM-YOLOv8-seg model, and carrying out the calculation of a blasting design parameter; respectively obtaining a main structural plane inclination angle category and a main structural plane area; 2, locally inputting the main structural plane into a main structural plane information extraction method to obtain the position of the main structural plane; 3, blasting design parameters, surrounding rock quality parameters, main structural plane dip angle categories and main structural plane positions are input into the CIPSO-DBN model, and a maximum linear over-excavation predicted value is output; 4, if the maximum linear over-excavation predicted value is higher than the limit value, the blasting design parameters are modified, the step 3 is executed again for prediction till the maximum linear over-excavation predicted value meets the limit value requirement, and the blasting design parameters are output.

Description

Technical Field

[0001] The present invention belongs to the technical field of engineering blasting safety, and in particular relates to a blasting design parameter optimization method and system based on deep learning and deep neural network. Background Art

[0002] In recent years, the state has proposed that transportation construction should shift from a scale and speed-oriented development to a quality and efficiency-oriented development period, which requires that tunnel excavation construction needs to solve problems such as over-excavation and under-excavation of the face and landslides. At present, among the tunnel excavation construction methods, the drilling and blasting method occupies an important position in the construction of mountain tunnel projects due to its advantages of low cost and high efficiency. However, during the construction process, due to the complex and changeable nature of the surrounding rock, the problem of over-excavation still exists, which directly affects the cost, safety and efficiency of subsequent construction. Therefore, how to minimize the damage to the surrounding rock while ensuring small over-excavation and stability of the face during blasting excavation is a basic problem of tunnel blasting construction. Realizing the control of blasting over-excavation will directly reduce construction costs and improve construction efficiency.

[0003] Due to the existence of structural surfaces in natural rock mass, the mechanical properties of rock mass and intact rock are very different. This difference is reflected in the blasting process as the explosion stress wave "spreads" along the structural surface. If the structural surface intersects with the tunnel contour, over-excavation will occur at the intersection after blasting. When a structural surface (main structural surface) that penetrates the rock mass of the tunnel face appears, the over-excavation phenomenon becomes more serious. At this stage, the position and angle of the main structural surface of the face are mainly obtained by on-site personnel. With the development of computer technology, it is particularly necessary to carry out research on accurate and efficient acquisition models of the main structural surface of the tunnel surrounding rock.

[0004] At present, there is no rapid optimization method for smooth blasting design parameters that takes into account both geological conditions and blasthole parameters. The traditional blasting design parameter optimization method based on theoretical analysis and field tests requires a lot of manpower, material resources and time, and requires multiple blasting cycle verification. The process is complicated and the cost is high. Combining the face feature extraction based on deep learning with the intelligent optimization method of blasting design parameters based on deep neural networks can achieve rapid adjustment of blasting design parameters and improve the quality of tunnel contour after blasting. Summary of the invention

[0005] In order to improve the accuracy and efficiency of blasting design parameter optimization, the present invention develops a blasting design parameter optimization method, and then combines the SP-YOLOv8 tunnel main structure surface inclination classification and recognition model, the SPM-YOLOv8-seg main structure surface information extraction method and the CIPSO-DBN tunnel smooth surface blasting design parameter optimization model to construct a blasting design parameter optimization method based on deep learning and deep neural network.

[0006] An object of the present invention is to provide a method for optimizing blasting design parameters based on deep learning and deep neural networks, the optimization method comprising the following steps:

[0007] Step 1: Load the tunnel face image into the SP-YOLOv8 model and the SPM-YOLOv8-seg model to obtain the main structural surface inclination category and the main structural surface area respectively;

[0008] Step 2: Input the main structural surface locally into the main structural surface information extraction method to obtain the main structural surface position;

[0009] Step 3: Input blasting design parameters, surrounding rock quality parameters, main structural surface inclination type and main structural surface position into the CIPSO-DBN model, and output the maximum linear overbreak prediction value;

[0010] Step 4: If the maximum linear over-excavation prediction value is higher than the limit, modify the blasting design parameters and return to step 3 to re-predict until the maximum linear over-excavation prediction value meets the limit requirements and output the blasting design parameters.

[0011] Furthermore, in step 1, the SP-YOLOv8 model is based on the YOLOv8 model structure, integrating feature extraction optimization based on asymmetric convolution blocks and a main structural surface filtering algorithm based on aspect ratio and distance constraints.

[0012] Furthermore, the feature extraction optimization based on asymmetric convolution blocks includes the optimization of the Backbone network of the YOLOv8 model. The specific operations are as follows:

[0013] An Identity layer is added in parallel to the asymmetric convolution block, and an SP-DWConv2D convolution layer is designed for the classification and identification of the inclination angle of the main structure surface of the tunnel face. The SP-DWConv2D convolution layer is substituted into the Bottleneck network of the YOLOv8 model, and then replaced upward to form the SP-C2f module network, and the SP-C2f module network replaces the C2f module of YOLOv8.

[0014] Furthermore, the SP-DWConv2D convolution layer includes a Split layer, three depth convolution kernels, an Identity layer and a concat layer, wherein the three depth convolution kernels include a 3×3 convolution kernel, a 1×9 banded kernel and a 9×1 banded kernel, respectively.

[0015] Furthermore, the main structural surface filtering algorithm based on aspect ratio and distance constraints is optimized for the Head structure of the YOLOv8 model. A Soft-NMS filtering algorithm with aspect ratio constraints and distance constraints is proposed to remove false main structural surfaces through aspect ratio constraints and distance constraints. The specific steps are as follows:

[0016] (1) The detection frames obtained from the tunnel face image are classified and recorded as set A = {A 1 ,A 2 …A 5}, in each category, according to the score and number of detection boxes, the detection box set m = {m 1 ,m 2 …m 5} and the corresponding probability score set n = {n 1 ,n 2 …n 5};

[0017] (2) Calculate the aspect ratio of each detection box based on the coordinates of the upper left corner of the detection box m (x 1 ,y 1 ) and the lower right corner coordinate (x 2 ,y 2 ), calculate its aspect ratio

[0018] (3) Aspect ratio filtering: introduce the aspect ratio constraint L and remove the detection boxes with l ≤ L;

[0019] (4) Introducing m, n, and A after aspect ratio constraint processing;

[0020] (5) Take the m of set m corresponding to the largest probability score in set n max =m 1 ;

[0021] (6) Calculate the center coordinates based on the detection box m i The upper left corner coordinate (x i1 ,y i1 ) and the lower right corner coordinate (x i2 ,y i2 ), calculate its center coordinates (x oi ,y oi ), x oi =(x i1 -x i2 ) / 2,y oi =(y i1 -y i2 ) / 2;

[0022] (7) Calculate m 1 With m i The center coordinate distance r,

[0023] (8) Detection box elimination based on distance constraint: introduce the distance constraint R and delete the detection boxes with r≤R;

[0024] (9) Output m, n, and A after eliminating the aspect ratio condition and distance constraint condition.

[0025] Furthermore, in step 2, the specific steps of extracting the main structural surface position by the main structural surface information extraction method are as follows:

[0026] Step 201: The tunnel face image is segmented using the SPM-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 (x 1 ,y 1 );

[0027] Step 202: After the mask area is binarized, the Zhang-Suen skeleton extraction algorithm is used to iterate the mask area multiple times, and a trace is formed by indenting from the edge to the center. The coordinates of the right end point of the trace are (x 2 ,y 2 );

[0028] Step 203: Obtain the pixel height y of the right end point of the main structural surface by using the image pixel height. 2 +y 1 ), and then the actual height of the right end point of the main structural surface is calculated according to the ratio of the actual face height H = y × (h / Y).

[0029] Another object of the present invention is to provide a blasting design parameter optimization system based on deep learning and deep neural network, the optimization system comprising:

[0030] An image acquisition module, used to acquire tunnel face images;

[0031] The image processing module calls the SP-YOLOv8 model and SPM-YOLOv8-seg model deployed in the background to process the input tunnel face image respectively to obtain the main structural surface inclination category and main structural surface area;

[0032] The main structural surface processing module calls the main structural surface information extraction method to process the main structural surface area and obtain the main structural surface position;

[0033] The prediction module calls the CIPSO-DBN model to process the input blasting design parameters, surrounding rock quality parameters, main structural surface inclination type and main structural surface position to obtain the maximum linear overbreak prediction value;

[0034] A judgment module is used to judge whether the maximum linear over-excavation prediction value is higher than the preset value. If the maximum linear over-excavation prediction value is higher than the preset value, the blasting design parameters are modified and input into the prediction module for re-prediction until the maximum linear over-excavation prediction value meets the limit requirement;

[0035] Output module, outputs the modified blasting design parameters.

[0036] Another object of the present invention is to provide an electronic device, comprising a processor and a memory, wherein the memory stores computer instructions, and the processor is used to run the computer instructions stored in the memory to implement the steps of a blasting design parameter optimization method based on deep learning and deep neural networks.

[0037] Another object of the present invention is to provide a computer-readable storage medium storing computer instructions, which are used to execute the steps of a blasting design parameter optimization method based on deep learning and deep neural networks.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] 1. The present invention designs an improved Structural Plane-YOLOv8 model (SP-YOLOv8) suitable for classification and identification of the inclination of the main structural surface of the tunnel, which can avoid the loss of image feature information and add a Soft-NMS filtering algorithm with aspect ratio and distance constraints. The false main structural surfaces are eliminated through the aspect ratio and distance constraints to reduce the false detection rate and re-detection rate of the model.

[0040] 2. The SPM-YOLOv8-seg model designed in the present invention is suitable for main structural surface segmentation and extraction, which can avoid interference feature extraction. At the same time, the main structural surface area extracted by the SPM-YOLOv8-seg model is used to extract the main structural surface position using the main structural surface information extraction method, which 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.

[0041] 3. The present invention can realize rapid adjustment of blasting design parameters and improve the forming quality of the tunnel profile after blasting.

[0042] 4. The traditional blasting design parameter optimization method requires a lot of manpower, material resources and time, and requires multiple blasting cycle verification, which is complicated and costly. However, this application can improve the efficiency and accuracy of blasting design parameter optimization while reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flowchart of parameter optimization method for blasting design based on deep learning and deep neural networks.

[0044] Figure 2 This is the design principle diagram of the SP-DWConv2D convolution layer.

[0045] Figure 3 This is the structural diagram of the SP-C2f module network.

[0046] Figure 4 The schematic diagram of the Soft-NMS main structure surface filtering algorithm based on aspect ratio constraints and distance constraints.

[0047] Figure 5 Schematic diagram of the main structural surface information extraction method for extracting the main structural surface position; (a) principle; (b) calculation of the structural surface endpoint coordinates.

[0048] Figure 6 This is the feature extraction flow chart of the AKConv module.

[0049] Figure 7 Schematic diagram of the AKConv module.

[0050] Figure 8 This is the prediction flow chart of the CIPSO-DBN model.

[0051] Fig. 9 Classify and identify the inclination angle of the main structure surface of the tunnel and quickly extract the coordinates of the endpoints;

[0052] (a) IDK57+820-IDK57+818.8 palm face image; (b) SP-YOLOv8 recognition image

[0053] (c) SPM-YOLOv8-seg main structure surface information extraction results.

[0054] Fig.10 To optimize the blasting scheme and blasting effect; (a) blasting effect at the spandrel; (b) blasting effect at the crown.

[0055] Fig.11 The blasting effects of three on-site blasting tests; (a) IDK57+820-IDK57+818.8;

[0056] (b) IDK57+818.8-IDK57+817.6; (c) IDK57+817.6-IDK57+816.4. DETAILED DESCRIPTION

[0057] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0058] This embodiment combines the SP-YOLOv8 tunnel main structure surface inclination classification and recognition model, the SPM-YOLOv8-seg main structure surface information extraction method, and the CIPSO-DBN tunnel smooth surface blasting design parameter optimization model to construct a blasting design parameter optimization method based on deep learning and deep neural network. The specific process is as follows: Figure 1shown.

[0059] Combination Figure 1 The blasting design parameter optimization process is explained in detail. The blasting design parameter optimization method based on deep learning and deep neural network is mainly divided into the following steps, as shown in Table 1.

[0060] Table 1 Flowchart of blasting design parameter optimization method based on deep learning and deep neural network

[0061]

[0062] The following introduces each of the above models and processing flows.

[0063] 1. SP-YOLOv8 tunnel main structure surface inclination classification and recognition model

[0064] Based on the YOLOv8 model structure, this paper designs an improved Structural plane-YOLOv8 model (SP-YOLOv8) suitable for classification and identification of the inclination angle of the main structural surface of the tunnel, integrating the feature extraction optimization based on asymmetric convolution blocks and the main structural surface filtering algorithm based on aspect ratio and distance constraints. The two optimization strategies are:

[0065] (1) Feature extraction optimization based on asymmetric convolution blocks

[0066] The Backbone network of the YOLOv8 model is mainly used to extract image feature information. When training the main structural surface inclination classification and recognition data set, the image feature information is lost. The core of the Backbone network is the C2f module, which is used to convert the output of the convolution layer into the input of the fully connected layer. It is the key module that causes missed detection and false detection after model training. The Bottleneck structure can improve the effect in the C2f module. This paper improves its performance in the tunnel face main structural surface inclination classification and recognition data set by adjusting the number and size of convolution kernels in Bottleneck. Ding et al. proposed that asymmetric convolution blocks can effectively improve the accuracy and performance of CNN. Asymmetric convolution blocks are composed of a square convolution kernel, a horizontal convolution kernel, and a vertical convolution kernel. This paper draws on the structure of asymmetric convolution blocks and considers the feature loss phenomenon caused by the increase in the number of convolution kernels. An Identity (identity mapping) layer is added to the asymmetric convolution block in parallel, and a structural plane-DWConv2D (SP-DWConv2D) convolution layer optimized for the classification and recognition dataset of the main structural surface inclination of the tunnel face is designed. The principle of the SP-DWConv2D convolution layer designed in this section is as follows: Figure 2 shown.

[0067] The operation process of the SP-DWConv2D convolution layer optimized for the tunnel face main structural surface inclination classification and recognition dataset includes the Split layer at the input front end, 3 deep convolution kernels and 1 Identity layer, and the concat layer at the back end for feature fusion. The sizes of the 3 convolution kernels are 3×3 convolution kernel, 1×9 strip kernel and 9×1 strip kernel. The SP-DWConv2D convolution layer is brought into the Bottleneck structure and replaced upward in turn to form a new structural plane-C2f (SP-C2f) module network structure. The SP-C2f module network replaces the C2f module of YOLOv8. Figure 3 shown.

[0068] (2) Main structural surface filtering algorithm based on aspect ratio and distance constraints

[0069] In order to solve the problems of false detection and repeated detection, the Head structure of the YOLOv8 model was optimized, and a Soft-NMS filtering algorithm with added aspect ratio and distance constraints was proposed. The false main structure surfaces were eliminated through the aspect ratio and distance constraints to reduce the false detection rate and repeated detection rate of the model.

[0070] The constraint algorithm flow based on aspect ratio is as follows Figure 4 As shown, first, based on the coordinates of the upper left corner of the detection box (x 1 ,y 1 ) and the lower right corner coordinate (x 2 ,y 2 ) calculates its aspect ratio, then introduces the aspect ratio constraint L, and finally calculates the aspect ratio l of each detection box. If l≤L, the detection box is removed.

[0071] The constraint algorithm flow based on distance constraints is as follows Figure 4 As shown, first sort the probability scores of all detection boxes from high to low, namely m 1 、m 2 , based on the coordinates of the upper left corner of the detection box (x 11 ,y 11 )、(x 21 ,y 21 ) and the lower right corner coordinate (x 12 ,y 12 )、(x 22 ,y 22 ), calculate its center coordinate x o1 、x o2 , with x o1 As the midpoint, the distance constraint R is introduced. When x o2 With x o1 If the distance r≤R, then remove m 2 Detection box.

[0072] The Soft-NMS algorithm process based on aspect ratio and distance constraints is as follows:

[0073] Aspect ratio constraint algorithm principle process

[0074] (1) Input: The detection boxes obtained by applying the object detection technology are classified into categories and recorded as a set A = {A 1 ,A 2 …A 5}, in each category, according to the score and number of detection boxes, the detection box set m = {m 1 ,m 2 ...m 5} and the corresponding probability score set n = {n 1 ,n 2 ...n 5}, aspect ratio L.

[0075] (2) Calculate the aspect ratio based on the coordinates of the upper left corner of the detection box m (x 1 ,y 1 ) and the lower right corner coordinate (x 2 ,y 2 ), calculate its aspect ratio

[0076] (3) Aspect ratio filtering: remove the detection frames with l ≤ L.

[0077] (4) Output: m, n, A.

[0078] Distance constraint algorithm principle process

[0079] (1) Input: Introduce m, n, A, and distance constraint R output by the aspect ratio constraint algorithm process.

[0080] (2) Initialization: Take the m of set m corresponding to the largest probability score in set n max =m 1 .

[0081] (3) Calculate the center coordinates: based on the detection box m i The upper left corner coordinate (x i1 ,y i1 ) and the lower right corner coordinate (x i2 ,y i2 ), calculate its center coordinates (x oi ,y oi ), x oi =(x i1 -x i2 ) / 2,y oi =(y i1 -y i2 ) / 2.

[0082] (4) Calculate m 1 With m i The center coordinate distance r:

[0083] (5) Detection box elimination based on distance constraint: Delete the detection boxes with r ≤ R.

[0084] (6) Output: m, n, A.

[0085] (7) Input into Soft-NMS algorithm for post-processing.

[0086] 2. Main structural surface information extraction method based on SPM-YOLOv8-seg model

[0087] Based on the selected SPM-YOLOv8-seg model, a method for extracting the main structural surface information of the face was established. The core principle is that when the image is segmented by the SPM-YOLOv8-seg model, it will diffuse to both ends with the main structural surface as the center to form a mask image. After binarization, the Zhang-Suen skeleton extraction algorithm is used. Its function is to indent the detection zone from the edge to the center through multiple iterations. The obtained endpoint coordinates can be approximately considered as the endpoint coordinates of the main structural surface, thereby extracting the main structural surface information. 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, such as Figure 5 shown.

[0088] Combination Figure 5 The information extraction process is described in detail as follows:

[0089] (1) Input: image height Y, face height h, and the coordinates of the lower right corner of the image as the origin (0,0).

[0090] (2) The tunnel face image is segmented using the SPM-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 (x 1 ,y 1 ).

[0091] (3) Extract the main structural surface area and use binarization for image processing.

[0092] (4) The skeleton extraction method uses the Zhang-Suen algorithm. The core of this algorithm is to iteratively indent the area to be extracted from the edge to the center to form a trace. At this time, the coordinates of the right end point of the trace are (x 2 ,y 2 ).

[0093] (5) Then, the pixel height of the right end point of the main structural surface can be obtained by the image pixel height y = (y 2 +y 1 ), and then the actual height of the right end point of the main structural surface is calculated according to the ratio of the actual face height H = y × (h / Y).

[0094] The AKConv module replaces the Convmodule module in the YOLOv8-seg model. The SPM-YOLOv8-seg model suitable for main structural surface segmentation and extraction is designed. The AKConv module changes the traditional convolution kernel that can only sample features of images with fixed size and fixed window. AKConv can sample features in any shape. The principle process is as follows: Figure 6 shown.

[0095] Depend on Figure 6 It can be seen that AKConv assumes that the initial convolution kernel shape is square. When learning the image data detection frame, it automatically adjusts the size and shape. In this way, when extracting features, the interference features are not extracted. Therefore, this section uses AKConv to optimize the Convmodule module of YOLOv8-seg and designs the AKConv module group structure as follows: Figure 7 shown.

[0096] 3. CIPSO-DBN model

[0097] Particle swarm optimization (PSO) is a swarm intelligence algorithm that simulates bird flight and foraging. This method finds the optimal area in a complex search space through the interaction between particles. However, the PSO algorithm is prone to fall into local optimality. Therefore, in order to further improve the optimization accuracy of the PSO algorithm and improve the algorithm performance, the CIPSO algorithm optimized by the improved second-order oscillation method based on the curve increasing strategy is adopted. The CIPSO algorithm is shown in the figure below. Figure 8 shown.

[0098] Combination Figure 8 The principle of CIPSO optimization algorithm is explained in detail. In the process of foraging in nature, the individual behavior of birds is affected by individual cognition and group behavior. This process is converted into a mathematical model. The calculation principle is as follows:

[0099] (1) Suppose there are f particles in a D-dimensional space, where the position of the i-th particle is x i =(x i1 ,x i2 ,...x iD ), with a velocity of v i =(v i1 ,v i2 ,...v iD ), the optimal position of the i-th particle is pi =(p i1 ,p i2 ,...p iD ), the optimal position of the entire particle swarm is p zbest =(p zbest1 ,p zbest2 ,...p zbestD ).

[0100] (2) The i-th particle at t * The speed of the iteration calculation is as follows:

[0101] v ij (t+1)=wv ij (t)+c 1 r 1 (p ij -x ij (t))+c 2 r 2 (P zbest -x ij (t))(1)

[0102] Where: i = 1, 2...f; j = 1, 2...D; r 1 , r 2 ∈rand(0,1); t is the current iteration number; w is the inertia parameter; c 1 、c 2 is the learning factor.

[0103] (3) The position update is shown as follows:

[0104] x ij (t+1)=v ij (t)+x ij (t)(2)

[0105] The above formula is the standard calculation principle of the PSO algorithm. The improved CIPSO algorithm is as follows:

[0106] (1) The curve increasing strategy for the inertia parameter w is as follows:

[0107]

[0108] Where: w max is the initial inertia parameter; t max is the maximum number of iterations.

[0109] (2) The improved second-order oscillation optimization method based on the curve increasing strategy introduces dual population calculation, and the particle velocity improvement formula is shown as follows:

[0110]

[0111] Where: x ij * is the particle in the new population; u i is the second-order vibration factor.

[0112] Simulation experiment and result analysis

[0113] 1. Obtaining the parameters of the main structure surface of the tunnel

[0114] According to the advanced geological drilling and geological survey data of Guangming Tunnel, the inclination of the main structural surface of the test section IDK57+820-IDK57+805 has not changed, so only the tunnel face image of IDK57+820 is needed to correct the surrounding rock quality parameters. First, through the tunnel face image of the Guangming Tunnel blasting test section, the inclination classification and main structural surface trace of the tunnel face are obtained using the developed software, such as Fig. 9 shown.

[0115] Depend on Fig. 9 It can be seen that the main structural surface inclination of the IDK57+820-IDK57+818.8 face of Guangming Tunnel is classified as nearly horizontal, and it intersects with the tunnel contour line at a large angle. To calculate the coordinates of the main structural surface, first call the relevant data of the recognition image, and calculate that the intersection with the contour line is 450cm high from the bottom of the step, located in the spandrel area. The structural surface inclination obtained by the tunnel main structural surface inclination classification recognition model and the information obtained by calculation, together with the integrity coefficient and uniaxial compressive strength of the geological survey data, are used as the surrounding rock parameters of the software.

[0116] 2. Optimization of smooth blasting design parameters

[0117] This section mainly focuses on the peripheral eye loading amount X 5 (kg), peripheral eyelet distance X 6 (cm), the amount of medicine in the outer ring auxiliary eye X 8 (kg), the amount of charge in the peripheral eye at the structural surface X 10 (kg), the amount of charge of the outer ring auxiliary eye at the structural surface X 11 (kg), light explosion layer thickness X 9 Firstly, the range of relevant parameters for tunnel smooth blasting was determined according to the Blasting Manual, as shown in Table 1.

[0118] Table 1 Range of parameters related to smooth blasting

[0119]

[0120] Guangming Tunnel IDK57+820-IDK57+818.8 is a Class V soft rock tunnel, mainly composed of sandstone. Combined with Table 1, the relevant parameters of smooth blasting are preliminarily determined to form an optimization plan, as shown in Table 2.

[0121] Table 2 Smooth surface blasting optimization data

[0122]

[0123] Among them, surrounding rock parameters: integrity coefficient X 1 , Uniaxial compressive strength X 2 (MPa), main structural surface inclination angle X 3 , Main structural surface position X 4 ; The adjusted parameters are input into the CIPSO-DBN model. The prediction results show that after parameter optimization, the maximum linear over-excavation of the optimized scheme is 15.04 cm. The optimized scheme is used as the blasting optimization scheme for field blasting tests. The original blasting scheme and the optimized blasting scheme are shown in Table 3.

[0124] Table 3 Blasting plan

[0125]

[0126] 3. On-site verification

[0127] The maximum linear over-excavation of the optimized scheme predicted by the blasting design parameter optimization software is 15.04 cm, which has met the 25 cm limit specified in the railway tunnel smooth surface blasting quality assessment standard. Therefore, this section selects the Guangming Tunnel smooth surface blasting test section for on-site blasting tests, and then analyzes the blasting effect of the optimized blasting scheme.

[0128] 3.1 On-site blasting test results

[0129] During the on-site verification phase, the main factor that easily affects the blasting effect is the arrangement of drilling holes. To eliminate the influence of the arrangement of holes, the spacing between the peripheral holes and the spacing between the outer ring auxiliary holes were checked after drilling holes in the test section to ensure that the hole arrangement parameters were consistent with the optimized blasting design parameters. Using the optimized blasting design parameters, blasting excavation was carried out three times in the test section IDK57+820-IDK57+805. The blasting effect of the IDK57+818.8 section is as follows: Fig.10 In front of the tunnel face where ballast is accumulated after blasting, according to the exposed tunnel face, the excavation contour line of the upper step after blasting is relatively smooth, without obvious jagged distribution, and the blast hole traces are clearly visible, but there is still a small over-excavation of the surrounding rock.

[0130] 3.2 Maximum linear overcut of tunnel surrounding rock

[0131] After each blast, the blasting debris in the tunnel is cleaned up, and then the maximum linear over-excavation of the tunnel is measured by a 3D section scanner, and the maximum linear over-excavation result of the upper step is extracted. Fig.11 shown.

[0132] Depend on Fig.11It can be seen that the maximum linear over-excavation of the three on-site blasting tests are 15.7 cm, 15.4 cm, and 16.1 cm respectively. The maximum linear over-excavation data of the blasting test and the blasting design parameter optimization software are statistically analyzed, as shown in Table 4.

[0133] Table 4 Statistics of maximum linear overbreak of tunnels (unit: cm)

[0134]

[0135] According to Table 4, after the blasting scheme is optimized, its maximum linear over-excavation has met the 25cm limit specified in the railway tunnel smooth blasting quality assessment standard, and the errors between the field measurement results and the blasting design parameter optimization software are 4.38%, 2.39%, and 7.04%, respectively. Based on the above analysis, it can be concluded that the blasting design parameter optimization software is feasible and can assist on-site personnel in optimizing the blasting scheme, thereby improving the blasting effect.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A method for optimizing blasting design parameters based on deep learning and deep neural network, characterized in that: The optimization method includes the following steps: Step 1: Load the tunnel face image into the SP-YOLOv8 model and the SPM-YOLOv8-seg model to obtain the main structural surface inclination category and the main structural surface area respectively; Step 2: Input the main structural surface locally into the main structural surface information extraction method to obtain the main structural surface position; Step 3: Input the blasting design parameters, surrounding rock quality parameters, main structural surface inclination category and main structural surface position into the CIPSO-DBN model, and output the maximum linear over-excavation prediction value. If the maximum linear over-excavation prediction value is higher than the limit, modify the blasting design parameters and repeat step 3 until the maximum linear over-excavation prediction value meets the limit requirements, and then output the blasting design parameters.

2. The method for optimizing blasting design parameters based on deep learning and deep neural network according to claim 1, characterized in that: In step 1, the SP-YOLOv8 model is based on the YOLOv8 model structure, integrating feature extraction optimization based on asymmetric convolution blocks and main structure surface filtering algorithm based on aspect ratio and distance constraints.

3. The method for optimizing blasting design parameters based on deep learning and deep neural network according to claim 2, characterized in that: The feature extraction optimization based on asymmetric convolution blocks includes the optimization of the Backbone network of the YOLOv8 model. The specific operations are as follows: An Identity layer is added in parallel to the asymmetric convolution block, and an SP-DWConv2D convolution layer is designed for the classification and identification of the inclination angle of the main structure surface of the tunnel face. The SP-DWConv2D convolution layer is substituted into the Bottleneck network of the YOLOv8 model, and then replaced upward to form the SP-C2f module network, and the SP-C2f module network replaces the C2f module of YOLOv8.

4. The method for optimizing blasting design parameters based on deep learning and deep neural network according to claim 3, characterized in that: The SP-DWConv2D convolution layer includes a Split layer, three depth convolution kernels, an Identity layer, and a concat layer. The three depth convolution kernels include a 3×3 convolution kernel, a 1×9 banded kernel, and a 9×1 banded kernel, respectively.

5. The method for optimizing blasting design parameters based on deep learning and deep neural network according to claim 3, characterized in that: The main structural surface filtering algorithm based on aspect ratio and distance constraints is optimized for the Head structure of the YOLOv8 model. The Soft-NMS filtering algorithm with aspect ratio constraints and distance constraints is proposed to remove false main structural surfaces through aspect ratio constraints and distance constraints. The specific steps are as follows: (1) The detection frames obtained from the tunnel face image are classified into a set A = {A1, A2…A5}. In each category, the score and number of detection frames are recorded as a detection frame set m = {m1, m2…m5} and a corresponding probability score set n = {n1, n2…n5}. (2) Calculate the aspect ratio of each detection box based on the coordinates of the upper left corner (x1, y1) and the lower right corner (x2, y2) of the detection box m. (3) Aspect ratio filtering: introduce the aspect ratio constraint L and remove the detection boxes with l ≤ L; (4) Introducing m, n, and A after aspect ratio constraint processing; (5) Take the m of set m corresponding to the largest probability score in set n max =m1; (6) Calculate the center coordinates based on the detection box m i The upper left corner coordinate (x i1 ,y i1 ) and the lower right corner coordinate (x i2 ,y i2 ), calculate its center coordinates (x oi ,y oi ), x oi =(x i1 -x i2 ) / 2,y oi =(y i1 -y i2 ) / 2; (7) Calculate m1 and m i The center coordinate distance r, (8) Detection box elimination based on distance constraint: introduce the distance constraint R and delete the detection boxes with r≤R; (9) Output m, n, and A after eliminating the aspect ratio condition and distance constraint condition.

6. The method for optimizing blasting design parameters based on deep learning and deep neural network according to claim 5, characterized in that: In step 2, the specific steps of extracting the main structural surface position by the main structural surface information extraction method are as follows: Step 201: The tunnel face image is segmented using the SPM-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); Step 202: After the mask area is binarized, the Zhang-Suen skeleton extraction algorithm is used to iterate the mask area multiple times, and a trace is indented from the edge to the center, and the coordinates of the right end point of the trace are (x2, y2); Step 203: The pixel height y=(y2+y1) of the right end point of the main structural surface is obtained by the image pixel height, and then 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 tunnel face height.

7. A blasting design parameter optimization system based on deep learning and deep neural network, characterized in that: The optimization system includes: An image acquisition module, used to acquire tunnel face images; The image processing module calls the SP-YOLOv8 model and the SPM-YOLOv8-seg model obtained by any one of claims 1 to 5 deployed in the background to process the input tunnel face image respectively, and obtains the main structural face inclination category and the main structural face area; A main structural surface processing module, calling the main structural surface information extraction method according to claim 6 to process the main structural surface area to obtain the main structural surface position; The prediction module calls the CIPSO-DBN model to process the input blasting design parameters, surrounding rock quality parameters, main structural surface inclination type and main structural surface position to obtain the maximum linear overbreak prediction value; A judgment module is used to judge whether the maximum linear over-excavation prediction value is higher than the preset value. If the maximum linear over-excavation prediction value is higher than the preset value, the blasting design parameters are modified and input into the prediction module for re-prediction until the maximum linear over-excavation prediction value meets the limit requirement; Output module, outputs the modified blasting design parameters.

8. An electronic device comprising a processor and a memory, characterized in that: The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory to implement the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to execute the steps of the method as claimed in any one of claims 1 to 6.

Images