Tunnel blasting scheme generation optimization system and method based on image recognition

Abstract

The invention belongs to the technical field of tunnel engineering, and particularly relates to a tunnel blasting scheme generation optimization system and method based on image recognition, according to the scheme, the influence of confusion variables can be effectively stripped by introducing a deep causal perception model, and a real causal mechanism between blasting parameters and effects can be accurately recognized; the output causal contribution degree matrix enables constructors to know which parameter should be adjusted and also to know what the parameter should be adjusted, and through combination with the adaptive optimizer, the problem that only correlation among variables is considered in a traditional mode, and whether the correlation is direct causal or not is not considered is effectively solved. And thus, optimization suggestions obtained based on an existing method may cause a poorer blasting effect.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering technology, specifically relating to a tunnel blasting scheme generation and optimization system and method based on image recognition. Background Technology

[0002] In the drill-and-blast method of tunnel construction, the effectiveness of blasting depends on the rationality of blasting parameters (such as hole spacing, row spacing, charge amount, and detonation time difference), specifically in terms of half-hole retention rate, over- and under-excavation control, rock debris size, and cycle advance.

[0003] The inventors discovered that existing methods for generating and optimizing tunnel blasting parameters have the following problems: During on-site construction, blasting parameters that are suitable for the current geological conditions are usually determined based on the construction experience of technical personnel and internal technical data. This method relies heavily on the personal experience and judgment of technical personnel and has a certain degree of subjectivity. At the same time, it is difficult for technical personnel to be proficient in blasting construction under various geological conditions. This leads to a high probability of initial blasting failure in some new projects or under new geological conditions due to the reliance on the construction experience of technical personnel and the blindness in parameter design. During construction, blasting parameters are usually manually adjusted based on the observations of on-site technicians regarding the blasting effect. In recent years, with the widespread application of machine learning in many industries, some machine learning-based blasting parameter optimization schemes have emerged. However, these methods are mainly based on correlation analysis and do not consider the inherent causal mechanism between blasting parameters and effects. For example, existing methods use machine learning models to consider the high correlation between "charge density" and "over-excavation," but do not consider whether this correlation is a direct causal relationship or caused by the confounding variable of "rock mass quality." Consequently, optimization suggestions based on existing methods may lead to worse blasting effects. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a tunnel blasting scheme generation and optimization system and method based on image recognition, effectively solving the problems existing in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A tunnel blasting scheme generation and optimization system based on image recognition, comprising: A portable data acquisition terminal is configured to acquire images of the tunnel contour surface after blasting, and generate half-hole detection results based on a built-in lightweight half-hole detection model based on deep learning; and acquire cross-sectional point cloud data, and obtain over- and under-excavation feature data based on a built-in point cloud processing algorithm. The generated half-hole detection results are displayed to on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and textual evaluations are transmitted to an edge server. An edge server is configured to receive multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting operation through a pre-built adaptive optimizer.

[0006] Furthermore, the deep causal perception model specifically performs the following processing steps: Using the geometric features of the half-hole, the over-excavation and under-excavation features, the current geological parameters, and the text feature vector representation of the text evaluation from the half-hole detection results as input, the fused features are obtained based on the cross-attention mechanism; Using the original numerical features and fused features as input, the PC algorithm combined with a neural network model is used to obtain the causal adjacency matrix; Based on the obtained causal adjacency matrix and fusion features, combined with predefined target variables, a structured causal model is constructed, and the optimized causal adjacency matrix is ​​obtained through iterative optimization. Intervention calculations are performed based on the optimized causal adjacency matrix to estimate the causal effect of each feature on the target variable; By calculating the causal effects of all features on multiple target variables, a causal effect matrix is ​​formed, and based on the causal effect matrix, a causal contribution matrix is ​​obtained.

[0007] Furthermore, the adaptive optimizer adopts a multilayer perceptron network structure. The adaptive optimizer takes the contribution matrix, the current blasting parameters, and the current geological parameters as a fusion vector as input, and outputs the blasting parameter adjustment amount. Based on the obtained blasting parameter adjustment amount, the current blasting parameters are corrected to obtain the optimized blasting parameters for the next time step.

[0008] Furthermore, the causal contribution matrix is ​​obtained as follows: Based on the obtained causal effect matrix, and combined with predefined feature importance weights and target variable weights, the initial causal contribution is calculated; Based on the initial contribution values, the final causal contribution matrix is ​​obtained through normalization.

[0009] Furthermore, the causal contribution matrix is ​​calculated using the following formula: in, Let be the initial causal contribution of the j-th variable to the k-th target variable. For the causal contribution matrix, Let the feature importance weights of the j-th variable be denoted as . The weight of the k-th target variable, This represents the normalized contribution of variable feature j to the target variable k.

[0010] Furthermore, the system also includes a historical tunnel engineering database, which stores pre-collected tunnel blasting scheme data in different regions. In the initial stage of tunnel blasting, the initial blasting parameters are determined through the historical tunnel engineering database.

[0011] Furthermore, the determination of initial blasting parameters through the historical tunnel engineering database specifically involves: when initiating a new tunnel blasting project, based on the current environmental characteristics of the tunnel, selecting a preset number of blasting cases whose similarity meets preset requirements from the historical tunnel engineering database through similarity calculation; selecting the optimal blasting case based on the effect indicators of the blasting cases; and determining the initial blasting parameters based on the blasting parameters of the optimal blasting case combined with expert evaluation.

[0012] A method for generating and optimizing tunnel blasting schemes based on image recognition, which is based on the aforementioned image recognition-based tunnel blasting scheme generation and optimization system, the method comprising: The tunnel contour image and cross-sectional point cloud data after blasting are acquired by a portable data acquisition terminal. Based on the acquired tunnel contour image and cross-sectional point cloud data after blasting, the half-hole detection results and over- and under-excavation feature data are generated by the built-in algorithm of the portable data acquisition terminal. The generated half-hole detection results are displayed to the on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and text evaluation are transmitted to the edge server; The edge server receives multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model, and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting through a pre-built adaptive optimizer.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: (1) The solution described in this application can effectively remove the influence of confounding variables by introducing a deep causal perception model, and can accurately identify the real causal mechanism between blasting parameters and effects. Its output causal contribution matrix can enable construction personnel to understand which parameter should be adjusted and why it should be adjusted. By combining with an adaptive optimizer, it effectively solves the problem that the traditional method only considers the correlation between variables without considering whether the correlation is a direct cause, which leads to the optimization suggestions obtained based on the existing method potentially resulting in a worse blasting effect.

[0014] (2) The solution described in this application uses a portable data acquisition terminal to collect images and point cloud data after blasting. Compared with traditional large data acquisition equipment, it is more flexible in operation and more in line with real needs. At the same time, the solution does not rely entirely on machine learning models. Instead, by introducing on-site personnel’s textual evaluation of the half-hole detection results, the detection effect of the machine learning model can be effectively supervised, further ensuring the accuracy of the optimization and adjustment of blasting parameters.

[0015] (3) The solution described in this application can provide an initial blasting scheme based on extensive historical experience and with a high success rate when the project starts without its own historical data, by constructing and querying a historical tunnel engineering database. This solves the risk of relying on the subjective experience of individual experts for blasting parameter design. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below: Figure 1 An optimized system structure diagram is generated for the tunnel blasting scheme based on image recognition described in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the data interaction process of the tunnel blasting scheme generation and optimization system based on image recognition as described in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the deep causal perception model processing procedure described in Embodiment 1 of the present invention; Figure 4 This is a flowchart of the tunnel blasting scheme generation and optimization method based on image recognition as described in Embodiment 2 of the present invention. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0018] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0019] Example 1: The following is a detailed description of the tunnel blasting scheme generation and optimization system based on image recognition described in Embodiment 1, with reference to the accompanying drawings.

[0020] like Figure 1 and Figure 2 As shown, the tunnel blasting scheme generation and optimization system based on image recognition includes: A portable data acquisition terminal is configured to acquire images of the tunnel contour surface after blasting, and generate half-hole detection results based on a built-in lightweight half-hole detection model based on deep learning; and acquire cross-sectional point cloud data, and obtain over- and under-excavation feature data based on a built-in point cloud processing algorithm. The generated half-hole detection results are displayed to on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and textual evaluations are transmitted to an edge server. An edge server is configured to receive multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting operation through a pre-built adaptive optimizer.

[0021] In specific implementation, the portable data acquisition terminal can be a mobile phone with a high-definition camera and laser scanning function, or a standalone mobile phone with a high-definition camera and a handheld laser scanner. Taking a mobile phone with a high-definition camera and laser scanning function as an example, after blasting, on-site technicians use a matching application installed on the phone to obtain images of the tunnel outline and scanned cross-section point cloud data. The application uses a lightweight half-hole detection model and point cloud processing algorithm based on deep learning to process the image data and point cloud data respectively, generating half-hole detection results and over- and under-excavation feature data. At the same time, the generated half-hole detection results are displayed on the interface, and on-site technicians provide textual evaluation of the half-hole detection results based on the actual situation on site. They also input the current geological parameters and transmit the above data to the edge server through the application.

[0022] The application needs to be developed in advance, and its main functions include: (1) Use the phone's high-definition camera and laser scanning function to perform image capture and point cloud data acquisition; (2) Built-in lightweight half-hole detection model and point cloud processing algorithm based on deep learning are used to process image data and point cloud data; (3) Provide a display interface for half-hole detection results, and provide input boxes for text evaluation and address parameters in the display interface; (4) Provides the function of transmitting half-hole detection results, over-excavation and under-excavation feature data, current geological parameters and text evaluation to the edge server.

[0023] As is understandable, software function development is based on the above-mentioned functions, so the details of how to conduct software function development will not be elaborated here.

[0024] Furthermore, the following processing steps are performed on the acquired tunnel outline image after blasting: (1) Data preprocessing: including but not limited to geometric correction and perspective transformation, contrast enhancement and noise filtering, image segmentation and ROI (Region of Interest) extraction; (2) Half-hole detection: The scheme described in this embodiment uses the MobileNet-Faster R-CNN lightweight model for half-hole detection. This model takes the pre-processed tunnel contour image after blasting as input and outputs the bounding box and segmentation mask for each half-hole.

[0025] (3) Geometric feature calculation: Based on the half-hole detection results, the following parameters are calculated (i.e., based on the contour and camera calibration parameters, they can be obtained through standard calculation formulas) to obtain the half-hole geometric feature vector. The vector includes the following parameters: : Average depth of half-hole (unit: mm); Standard deviation of half-hole depth; : Deviation rate between actual and design values ​​of half-hole spacing; : Half-hole integrity index (0-1); : The maximum difference in depth between adjacent half-holes; Half-hole axis deflection angle (unit: degrees); Percentage of the area of ​​the broken zone at the orifice; : Half-hole outline clarity score.

[0026] Preferably, after obtaining the half-hole detection results, they are displayed through an application interface (the half-hole detection results can be visually overlaid on the original image, and the identified half-holes can be marked with bounding boxes while displaying the confidence level). Simultaneously, on-site technicians can evaluate the half-hole detection results through an interactive evaluation interface, which can be designed as follows: Method 1: Structured Evaluation: Provides checkboxes and sliders, allowing users to quickly annotate by selecting options, for example: Detection accuracy: Very poor, Poor, Average, Accurate, Very accurate; Problems identified: missed detections, false detections, inaccurate boundaries, and others; Special considerations: fractured rock mass, significant water seepage, well-developed joints, and other factors. Method 2: Unstructured evaluation: Users input free text.

[0027] By combining the two methods described above, a textual evaluation of the half-hole detection results is generated.

[0028] Furthermore, the processing of the cross-sectional point cloud data mainly adopts the following process: (1) Point cloud registration and denoising; (2) Compare the designed cross-section with the actual cross-section; (3) Based on the comparison results, perform over-mining and under-mining feature calculation to obtain the following parameters and obtain the over-mining and under-mining feature vector. ,in: Average linear overcut (unit: mm); Cross-sectional flatness (i.e., the standard deviation of point cloud elevation values).

[0029] In practical implementation, the textual evaluation of the half-hole detection results needs to be converted into a machine-readable vector representation. A suitable language coding model can be selected based on actual needs. In this embodiment, the BERT model is used for text encoding to obtain the text feature vector of the text evaluation. .

[0030] Furthermore, the user needs to input the current geological parameters. ,in: : Comprehensive quality evaluation index of rock mass; : Uniaxial compressive strength of rock (unit: MPa); Joint density (unit: joints / m); Groundwater condition level (the degree of influence of groundwater on the stability of surrounding rock, which is set from 1 to 5).

[0031] In specific implementation, such as Figure 3 As shown, the deep causal perception model learns the causal graph structure between variables and estimates the causal effect of each input feature on the explosion effect through counterfactual reasoning to obtain a causal contribution matrix. Specifically, the deep causal perception model performs the following processing steps: (1) Multi-feature fusion processing Geometric features of the half-hole in the half-hole detection results Over-excavation and under-excavation characteristics Current geological parameters and text evaluation (text feature vectors) ) as input; The geometric features, over-excavation and under-excavation features, and current geological parameters are encoded by a preset numerical feature encoder to obtain a numerical feature vector. The text feature vector is projected onto the text feature vector to obtain the projected text feature vector. The purpose of the text feature projection is to ensure that the numerical features and text features are in the same dimension. Based on the obtained numerical feature vectors and the projected text feature vectors, a fused feature is obtained through a cross-attention mechanism. .

[0032] (2) Causal structure learning In the scheme described in this embodiment, the causal structure learning specifically includes the following processing steps: Based on original numerical features and fusion features As input; The PC algorithm (i.e., the Peter-Clark algorithm) is used for initial causal discovery to obtain the initial causal adjacency matrix. The basic idea is to construct a completely undirected graph by taking each feature variable in the original numerical features (there are 14 variables in the original numerical features) as a node; and infer the potential causal graph structure based on this completely undirected graph (usually represented by a directed acyclic graph, i.e., a Bayesian network). Based on the initial causal adjacency matrix and fusion features By continuously learning the causal adjacency matrix through a neural network model, a more accurate causal adjacency matrix can be obtained. ; Based on the obtained causal adjacency matrix Combined with prior knowledge mask By incorporating DAG constraints (i.e., ensuring the graph is acyclic), the final causal adjacency matrix is ​​obtained. Its specific representation is as follows: Constraints: Satisfy DAG constraints.

[0033] Among them, prior knowledge mask The determination is based on industry knowledge, with the aim of prohibiting causal edges that are impossible between variables.

[0034] (3) Estimation of causal effects Based on the obtained causal adjacency matrix and fusion features Combining predefined target variables (such as over-excavation and half-hole ratio), a structured causal model (SCM) is constructed. The construction process of the structured causal model is as follows: Based on the causal adjacency matrix, a causal mechanism function is constructed for each variable (the scheme described in this embodiment uses a neural network model for construction, for example, it can be implemented by using a fully connected layer + activation function). This function uses the parent node variable of the current variable as input to predict the value of the current variable. By using observational data to co-train all causal mechanism functions, the parameters of each neural network (i.e., causal mechanism function) are optimized by minimizing the difference between the predicted value and the actual observed value of each variable. Through iterative optimization, the structured causal model can accurately reflect the causal dependencies between variables, and the optimized causal adjacency matrix is ​​obtained. Intervention calculations are performed based on the optimized causal adjacency matrix to estimate the causal effect of each feature on the target variable; A causal effect matrix is ​​formed by calculating the causal effects of all features on multiple target variables. In this matrix, each element represents the causal effect value of each variable feature on the target variable, and its dimension corresponds to the number of variables in the original numerical features (i.e., 14 variables) and the number of predefined target variables (e.g., considering only over-excavation and half-hole ratio). =2).

[0035] (4) Calculation of causal contribution Based on the obtained causal effect matrix Combined with predefined feature importance weights and target variable weights Calculate the initial causal contribution: Among them, feature importance weight Based on predefined feature variance or random forest importance, the target variable weights are... The weighting is set based on the importance of the project. For example, the weighting of over-excavation is set to 0.6, and the weighting of half-hole quality is set to 0.4.

[0036] Based on the initial contribution values, the final causal contribution matrix is ​​obtained through normalization: in, This represents the normalized contribution of variable feature j to the target variable k.

[0037] In practice, the adaptive optimizer uses the aforementioned contribution matrix to adjust the blasting parameters, specifically: The adaptive optimizer employs a classic 3-layer MLP (Multilayer Perceptron) network structure, and the adaptive optimizer uses a contribution matrix... Current blasting parameters and current geological parameters The constructed fusion vector is used as input, and the output is the adjustment amount of the explosion parameters. Adjustment based on obtained blasting parameters The current blasting parameters are corrected to obtain the optimized blasting parameters for the next moment. This is used for the next blast, thereby generating a new blasting result. By repeating the aforementioned process, an adaptive loop is formed.

[0038] In specific implementation, the system also includes a historical tunnel engineering database, which stores pre-collected tunnel blasting scheme data in different regions. In the initial stage of the scheme described in this embodiment, the initial blasting parameters are determined through this historical tunnel engineering database.

[0039] Specifically, each data entry in the historical tunnel engineering database includes not only specific blasting parameters, but also environmental characteristics, blasting effect index data, and tunnel-related information corresponding to the current blasting parameters, among which: Environmental characteristics include, but are not limited to: basic rock mass quality indicators, uniaxial compressive strength, joint density, burial depth, and groundwater grade, etc. Performance indicators include, but are not limited to: half-hole ratio, average linear over-excavation, cycle advance, and large block ratio; Information related to the tunnel includes, but is not limited to: tunnel name, construction date, construction team, and team contact information.

[0040] Furthermore, the determination of initial blasting parameters through the historical tunnel engineering database specifically involves: When initiating a new tunnel blasting project, based on the current environmental characteristics of the tunnel, a preset number of blasting cases with similarity that meet preset requirements are selected from the historical tunnel project database through similarity calculation; the optimal blasting case is selected based on the effect indicators of the blasting cases; and the initial blasting parameters are determined based on the blasting parameters of the optimal blasting case combined with expert evaluation.

[0041] Furthermore, the historical tunnel engineering database will periodically update the optimized blasting parameters based on the blasting effect.

[0042] In specific implementation, the edge server is set up at or near the tunnel entrance and command center. Considering that the signal inside the tunnel will be severely affected as the tunnel is excavated, in order to ensure stable data transmission, the solution described in this embodiment uses portable 5G / 4G base stations to build a local area network to ensure stable data transmission between the portable data acquisition terminal and the edge server.

[0043] Example 2: like Figure 4 As shown, this embodiment provides a method for generating and optimizing tunnel blasting schemes based on image recognition. It is based on the aforementioned system for generating and optimizing tunnel blasting schemes based on image recognition. The method includes: The tunnel contour image and cross-sectional point cloud data after blasting are acquired by a portable data acquisition terminal. Based on the acquired tunnel contour image and cross-sectional point cloud data after blasting, the half-hole detection results and over- and under-excavation feature data are generated by the built-in algorithm of the portable data acquisition terminal. The generated half-hole detection results are displayed to the on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and text evaluation are transmitted to the edge server; The edge server receives multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model, and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting through a pre-built adaptive optimizer.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A tunnel blasting scheme generation and optimization system based on image recognition, characterized in that, include: A portable data acquisition terminal is configured to acquire images of the tunnel contour surface after blasting, and generate half-hole detection results based on a built-in lightweight half-hole detection model based on deep learning; and acquire cross-sectional point cloud data, and obtain over- and under-excavation feature data based on a built-in point cloud processing algorithm. The generated half-hole detection results are displayed to on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and textual evaluations are transmitted to an edge server. An edge server is configured to receive multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting operation through a pre-built adaptive optimizer.

2. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 1, characterized in that, The deep causal perception model specifically performs the following processing steps: Using the geometric features of the half-hole, the over-excavation and under-excavation features, the current geological parameters, and the text feature vector representation of the text evaluation from the half-hole detection results as input, the fused features are obtained based on the cross-attention mechanism; Using the original numerical features and fused features as input, the PC algorithm combined with a neural network model is used to obtain the causal adjacency matrix; Based on the obtained causal adjacency matrix and fusion features, combined with predefined target variables, a structured causal model is constructed, and the optimized causal adjacency matrix is ​​obtained through iterative optimization. Intervention calculations are performed based on the optimized causal adjacency matrix to estimate the causal effect of each feature on the target variable; By calculating the causal effects of all features on multiple target variables, a causal effect matrix is ​​formed, and based on the causal effect matrix, a causal contribution matrix is ​​obtained.

3. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 1, characterized in that, The adaptive optimizer adopts a multilayer perceptron network structure. The adaptive optimizer takes the fusion vector constructed from the contribution matrix, the current blasting parameters, and the current geological parameters as input and outputs the blasting parameter adjustment amount. Based on the obtained blasting parameter adjustment amount, the current blasting parameters are corrected to obtain the optimized blasting parameters for the next moment.

4. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 2, characterized in that, The causal contribution matrix is ​​obtained in the following way: Based on the obtained causal effect matrix, and combined with predefined feature importance weights and target variable weights, the initial causal contribution is calculated; Based on the initial contribution values, the final causal contribution matrix is ​​obtained through normalization.

5. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 4, characterized in that, The causal contribution matrix is ​​calculated using the following formula: in, Let be the initial causal contribution of the j-th variable to the k-th target variable. For the causal contribution matrix, Let the feature importance weights of the j-th variable be denoted as . The weight of the k-th target variable, This represents the normalized contribution of variable feature j to the target variable k.

6. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 1, characterized in that, The system also includes a historical tunnel engineering database, which stores pre-collected data on tunnel blasting schemes in different regions. In the initial stage of tunnel blasting, the initial blasting parameters are determined through the historical tunnel engineering database.

7. The tunnel blasting scheme generation and optimization system based on image recognition as described in claim 1, characterized in that, The process of determining initial blasting parameters through a historical tunnel engineering database involves the following steps: When initiating a new tunnel blasting project, based on the current environmental characteristics of the tunnel, a preset number of blasting cases with similarity scores meeting preset requirements are selected from the historical tunnel engineering database through similarity calculation; the optimal blasting case is selected based on the effect indicators of the blasting cases; and the initial blasting parameters are determined based on the blasting parameters of the optimal blasting case combined with expert evaluation.

8. A method for generating and optimizing tunnel blasting schemes based on image recognition, which is based on the tunnel blasting scheme generation and optimization system based on image recognition as described in any one of claims 1-7, characterized in that, The method includes: The tunnel contour image and cross-sectional point cloud data after blasting are acquired by a portable data acquisition terminal. Based on the acquired tunnel contour image and cross-sectional point cloud data after blasting, the half-hole detection results and over- and under-excavation feature data are generated by the built-in algorithm of the portable data acquisition terminal. The generated half-hole detection results are displayed to the on-site personnel, who then provide textual evaluations. The half-hole detection results, over-excavation and under-excavation feature data, current geological parameters, and text evaluation are transmitted to the edge server; The edge server receives multi-source data from a portable data acquisition terminal. Based on the multi-source data, it learns the causal graph structure between variables through a pre-built deep causal perception model, and estimates the causal effect of each input feature on the blasting effect through counterfactual reasoning to obtain a causal contribution matrix. Based on the obtained causal contribution matrix, the current blasting parameters, and the current geological parameters, it obtains the optimized parameters for the next blasting through a pre-built adaptive optimizer.

Images