A vision-based method for detecting the initial shotcrete quality in tunnels

Abstract

This invention relates to the field of tunnel construction quality inspection technology, specifically a method for inspecting the initial shotcrete spraying quality in tunnels based on vision technology. The method includes the following steps: S1: Through multimodal sensor fusion and adaptive exposure denoising algorithms, combined with multiple information joint correction techniques, an automatic cleaning device is installed at the front end of the equipment. Origin-point positioning calibration is used to perform high-precision and robust acquisition of the three-dimensional coordinates and apparent vector data of the shotcrete surface under complex environments. This invention, through multimodal sensor fusion, anti-interference processing, and precise calibration, ensures high-precision and robust acquisition of shotcrete surface data under complex tunnel environments, providing a reliable data foundation for subsequent inspection and control. It effectively improves the accuracy of shotcrete surface smoothness control, reduces material loss rate, and can adaptively adjust strategies according to geological conditions and construction status, balancing construction efficiency and safety.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction quality inspection technology, specifically a method for inspecting the quality of initial shotcrete spraying in tunnels based on visual technology. Background Technology

[0002] Shotcrete application in tunnel construction is a crucial step in the initial support phase. It refers to the first layer of concrete sprayed onto the surface of the surrounding rock after excavation to seal the rock mass, prevent weathering and spalling, stabilize the working face, and provide a foundation for subsequent support structures. Its quality directly impacts the stability of the initial support, the safety of the secondary lining, and the quality of the waterproofing layer, making it a vital link in ensuring tunnel construction safety and structural durability. However, in current technology and actual construction, the complex environment and long-term reliance on manual sampling during initial shotcrete application lead to low inspection efficiency, poor accuracy, and an inability to monitor the construction process in real time. This further results in problems such as difficulty in controlling over- or under-spraying, easy intrusion of the initial support into the secondary lining, repetitive and inefficient construction, and potential quality issues with the waterproofing layer.

[0003] Based on this, the present invention provides a method for detecting the quality of initial shotcrete spraying in tunnels based on vision technology, in order to solve the aforementioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting the quality of initial shotcrete spraying in tunnels based on vision technology. This invention utilizes multimodal sensor fusion, anti-interference processing, and precise calibration to ensure high-precision and robust acquisition of shotcrete surface data in complex tunnel environments, providing a reliable data foundation for subsequent detection and control. Furthermore, it leverages artificial intelligence algorithms to fuse multi-source information to construct a three-dimensional model and dynamically optimize spraying parameters, achieving adaptive adaptation to complex geological conditions and intelligent iteration of construction strategies. Through predictive control and multi-objective optimization, it achieves the linkage adjustment of nozzle pose and parameters, balancing construction quality, efficiency, and safety. This effectively improves the accuracy of shotcrete surface smoothness control, reduces material loss rate, and can adaptively adjust strategies according to geological conditions and construction status, balancing construction efficiency and safety.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a method for detecting the quality of initial shotcrete spraying in tunnels based on vision technology, comprising the following steps:

[0007] S1: By using multimodal sensor fusion and adaptive exposure denoising algorithm, combined with multiple information joint correction technology, an automatic cleaning device is set up at the front end of the equipment. Through origin positioning calibration, high-precision and robust acquisition of three-dimensional coordinates and apparent vector data of the sprayed surface in complex environments is achieved.

[0008] S2: Based on image processing algorithms, it integrates laser scanning, intelligent data processing and real-time monitoring technologies to detect flatness and out-of-limit areas;

[0009] S3: An AI-based self-learning data fusion and parameter optimization model that automatically identifies key features in multi-view visual data, fuses laser and image information, constructs a high-precision three-dimensional sprayed surface model, analyzes the influence of parameters such as spraying distance, angle, pressure, and speed on the smoothness of the sprayed surface, and dynamically optimizes key construction parameters such as spraying distance, angle, and speed through reinforcement learning algorithms.

[0010] S4: Based on the analysis results of S3, dynamic and coordinated adjustment of nozzle pose and injection parameters is carried out based on predictive control algorithm and multi-objective optimization control.

[0011] The specific steps in S1 are as follows:

[0012] S1.1: Capture the original data of the three-dimensional coordinates and apparent vectors of the sprayed surface through multi-sensor fusion of laser and vision.

[0013] S1.2: Adaptive exposure denoising algorithm and automatic cleaning device are used to eliminate environmental interference;

[0014] S1.3: Correction and origin positioning calibration through joint correction of multiple information sources.

[0015] In step S1.2, an adaptive exposure denoising algorithm and an automatic cleaning device are used to eliminate environmental interference. The specific steps are as follows:

[0016] S1.2.1: Real-time collection of dust concentration and light intensity inside the tunnel using dust sensors and light sensors; when dust concentration > 50 mg / m³ 3 Or, when the light intensity is <300 lux, the anti-interference mechanism is triggered;

[0017] S1.2.2: Controlling cleaning frequency based on dust concentration levels:

[0018] ① Concentration 50-100mg / m 3 At that time, the rotating brush is activated every 5 minutes to clean the sensor surface with high-pressure airflow;

[0019] ② Concentration > 100 mg / m³ 3 At that time, continuous trigger cleaning is initiated;

[0020] S1.2.3: Adaptive exposure adjustment is adopted for low-light environments, and a multi-frame superposition averaging denoising algorithm is executed on the acquired images. 3-5 frames of images are acquired continuously, and random noise is eliminated by weighted averaging of pixel gray values.

[0021] S1.2.4: Based on the response delay of the sensor after cleaning, perform time axis calibration on the subsequent acquired data.

[0022] The specific steps in S2 are as follows:

[0023] S2.1: Dynamically acquire data on the flatness and surface morphology details of the sprayed surface using laser scanning technology;

[0024] S2.2: Use image processing algorithms to analyze the collected data, identify areas exceeding limits, and determine the warning level;

[0025] S2.3: The monitoring results are displayed in real time through an intuitive interface, and the construction personnel are given immediate warnings when the limits are exceeded.

[0026] The specific steps in S3 are as follows:

[0027] S3.1: Integrates multi-view visual and laser information to automatically extract key features of the sprayed surface and construct a 3D model;

[0028] S3.2: Analyze the influence of parameters such as spray distance, angle, pressure, and speed on smoothness, and establish a correlation model;

[0029] S3.3: Based on self-learning and reinforcement learning algorithms, dynamically optimize spraying parameters and generate construction adjustment plans.

[0030] In step S3.1, multi-view visual and laser information are integrated to automatically extract key features of the sprayed surface and construct a 3D model. The specific steps are as follows:

[0031] S3.1.1: Perform distortion correction and grayscale normalization on multi-view visual images, and simultaneously perform noise reduction filtering on laser point cloud data;

[0032] S3.1.2: Identify common feature points in visual images and laser point clouds using a feature point matching algorithm, and establish a spatial coordinate mapping relationship between multi-source data;

[0033] S3.1.3: Based on the mapping relationship, the texture information of the visual image is fused with the depth information of the laser point cloud to extract the key features of the concave and convex regions and the abrupt change points of the smoothness of the sprayed surface.

[0034] S3.1.4: The Poisson surface reconstruction algorithm is used to divide the fused 3D point cloud data into meshes to generate a sprayed surface 3D model with texture information. The model resolution is set to contain 500-800 vertices per cubic meter.

[0035] In step S3.3, the spraying parameters are dynamically optimized and a construction adjustment plan is generated based on self-learning and reinforcement learning algorithms. The specific steps are as follows:

[0036] S3.3.1: Construct the state space for reinforcement learning, including parameters such as surface smoothness, spraying distance, angle, velocity, and pressure, as well as geological type labels;

[0037] S3.3.2: Define the action space, including the adjustment range and adjustment step size of the injection distance, angle, and speed;

[0038] S3.3.3: Define a reward function with the following optimization objectives: spraying smoothness compliance rate, material utilization rate, and construction efficiency. The reward function is expressed as follows:

[0039] R = w1·flatness compliance rate + w2·material utilization rate + w3·construction efficiency

[0040] Among them, w1, w2, and w3 are weighting coefficients, satisfying w1 + w2 + w3 = 1;

[0041] S3.3.4: The PPO algorithm is used to train and optimize the model, and the injection parameter strategy is continuously adjusted through interaction with the environment;

[0042] S3.3.5: Outputs the optimal combination of spraying parameters, including spraying distance, angle, and speed, to guide construction adjustments;

[0043] S3.3.6: The model continuously receives new data and updates online, performing self-learning optimization of the spraying strategy.

[0044] The specific steps in S4 are as follows:

[0045] S4.1: Receive the optimization parameters and adjustment instructions output by S3;

[0046] S4.2: Based on predictive control algorithms, the positional parameters of the nozzle's distance from the rock surface and the spray angle are adjusted in real time;

[0047] S4.3: Through multi-objective optimization control, the nozzle position and spray parameters are adjusted in a coordinated manner.

[0048] In step S4.2, the pose parameters of the nozzle distance and spray angle are adjusted in real time based on the predictive control algorithm. The specific steps are as follows:

[0049] S4.2.1: Establish a dynamic prediction model for nozzle position and spray parameters, and input feedback information such as current nozzle position, spray angle, rock surface distance, and spray surface smoothness;

[0050] S4.2.2: Based on the prediction model, the trend of nozzle pose and injection parameters over the next N time steps is calculated in a rolling manner;

[0051] S4.2.3: Define the objective function, with the goal of minimizing injection error and control energy. The objective function is expressed as follows:

[0052]

[0053] Among them, e k Let Δu be the injection error at step k. k To control the changes in the quantity, Q and R are weighted matrices;

[0054] S4.2.4: By solving the optimization problem, the optimal control sequence is obtained, and the optimal control quantity at the current moment is selected and output to the actuator;

[0055] S4.2.5: Update the model state based on the actual feedback error and transmit it back to S3.3.6 in real time via the OPC-UA protocol.

[0056] In S4.3, multi-objective optimization control is used to adjust the nozzle pose and injection parameters in a coordinated manner. The specific operation is as follows:

[0057] S4.3.1: Establish a three-objective optimization model that includes construction quality, efficiency, and cost:

[0058] ① Quality target: Flatness deviation Δh ≤ 3mm;

[0059] ② Efficiency target: injection speed v ≥ 1.8m 2 / min;

[0060] ③ Cost target: Material loss rate η≤5%;

[0061] S4.3.2: The improved NSGA-III algorithm is used to solve for the Pareto optimal solution set, where:

[0062] ① Set the population size to 100 and the number of iterations to ≥50 generations;

[0063] ②For tunnel construction scenarios, add mechanical limit constraints for the nozzles;

[0064] ③ Adopting a dynamic reference point mechanism to adapt to different surrounding rock grades;

[0065] S4.3.3: Select the final parameter combination from the Pareto front using fuzzy decision-making, prioritizing the following:

[0066] ① Class IV surrounding rock: Quality target weight 0.7;

[0067] ② Class V surrounding rock: The safety target weight for initial support thickness compliance rate ≥ 95% is 0.6;

[0068] S4.3.4: When the rock mass displacement rate is detected to be >2mm / h, automatically switch to conservative spraying mode.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] This invention ensures high-precision and robust acquisition of sprayed surface data in complex tunnel environments through multimodal sensor fusion, anti-interference processing, and precise calibration, providing a reliable data foundation for subsequent detection and control. Furthermore, it utilizes artificial intelligence algorithms to fuse multi-source information to construct a three-dimensional model and dynamically optimize spraying parameters, achieving adaptive adaptation to complex geological conditions and intelligent iteration of construction strategies. Through predictive control and multi-objective optimization, it achieves coordinated adjustment of nozzle pose and parameters, balancing construction quality, efficiency, and safety. This effectively improves the accuracy of sprayed surface smoothness control, reduces material loss, and can adaptively adjust strategies according to geological conditions and construction status, balancing construction efficiency and safety. Attached Figure Description

[0071] Figure 1 This is a flowchart of a method for detecting the initial shotcrete spraying quality in tunnels based on vision technology, according to the present invention.

[0072] Figure 2 This is an overall architecture diagram of a method for detecting the initial shotcrete spraying quality in tunnels based on vision technology, according to the present invention.

[0073] Figure 3 This is a flowchart of the multi-objective collaborative control process in a vision-based method for detecting the initial shotcrete spraying quality in tunnels, as described in this invention. Detailed Implementation

[0074] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0075] Example:

[0076] like Figures 1-3As shown, this embodiment provides a method for detecting the initial shotcrete spraying quality in tunnels based on vision technology, including the following steps: S1: By using multimodal sensor fusion and adaptive exposure denoising algorithms, combined with multiple information joint correction techniques, an automatic cleaning device is set at the front end of the equipment. Through origin positioning calibration, high-precision robust acquisition of the three-dimensional coordinates and apparent vector data of the sprayed surface under complex environments is performed; S2: Based on image processing algorithms, laser scanning, intelligent data processing, and real-time monitoring technologies are integrated to detect flatness and out-of-limit areas; S3: Based on an artificial intelligence self-learning data fusion and parameter optimization model, key features in multi-view visual data are automatically identified, laser and image information are fused, a high-precision three-dimensional sprayed surface model is constructed, the influence of spraying distance, angle, pressure, and speed parameters on the flatness of the sprayed surface is analyzed, and key construction parameters such as spraying distance, angle, and speed are dynamically optimized through reinforcement learning algorithms; S4: Based on the analysis results of S3, the nozzle pose and spraying parameters are dynamically and collaboratively adjusted based on predictive control algorithms and multi-objective optimization control.

[0077] In this embodiment, it should also be noted that the specific steps in S1 are as follows: S1.1, capture the original data of the three-dimensional coordinates and apparent vector of the sprayed surface through multi-sensor fusion of laser and vision; S1.2, use an adaptive exposure denoising algorithm and an automatic cleaning device to eliminate environmental interference; the specific steps are as follows: S1.2.1: collect the dust concentration and light intensity in the tunnel in real time through dust sensors and light sensors. When the dust concentration is >50mg / m³ 3 Or, when the light intensity is <300 lux, the anti-interference mechanism is triggered; S1.2.2: Cleaning frequency is controlled based on dust concentration classification: ① Concentration 50-100 mg / m³ 3 At that time, the rotating brush is activated every 5 minutes in conjunction with high-pressure airflow to clean the sensor surface; ② Concentration > 100mg / m³ 3 S1.2.3: In low-light environments, adaptive exposure adjustment is used, and a multi-frame superposition averaging denoising algorithm is executed on the acquired images. 3-5 frames are continuously acquired, and random noise is eliminated by weighted averaging of pixel grayscale values. S1.2.4: Based on the sensor's response delay after cleaning, the subsequent acquired data is calibrated along the time axis. S1.3: Multiple information is used for joint correction and origin positioning calibration.

[0078] Furthermore, it should be noted that the laser wavelength range is 850nm, the visual sensor is a back-illuminated CMOS sensor with ISO ≥ 51200, and the synchronization mechanism of "multi-sensor fusion" uses timestamp synchronization with an error ≤ 1m. "Multi-information joint correction" employs a calibration board-assisted Bundle Adjustment algorithm to simultaneously optimize the intrinsic and extrinsic parameters of the LiDAR and camera, controlling the reprojection error within ±0.5 pixels. The specific information types for "multi-information joint correction" (such as laser point clouds, visual images, and inertial navigation data) are as follows: the origin coordinates for "origin positioning calibration" are based on the center point of the tunnel face.

[0079] In this embodiment, it should also be noted that the specific steps in S2 are as follows: S2.1: Dynamically acquire data on the flatness and surface morphology details of the sprayed surface using laser scanning technology; S2.2: Analyze the acquired data using image processing algorithms, identify areas exceeding limits, and determine the warning level; S2.3: Display the monitoring results in real time through an intuitive interface, and issue immediate warnings to construction personnel for exceeding limits.

[0080] Furthermore, it should be noted that the criteria for determining "over-limit areas" is that a flatness deviation > 5mm is considered over-limit. The classification of "warning levels" is based on the following: Level 1 warning: deviation 5-10mm; Level 2 warning: deviation > 10mm.

[0081] In this embodiment, it should also be noted that the specific steps in S3 are as follows: S3.1: Fuse multi-view visual and laser information, automatically extract key features of the sprayed surface and construct a three-dimensional model; the specific steps are as follows: S3.1.1: Perform distortion correction and grayscale normalization processing on the multi-view visual images, and simultaneously perform noise reduction filtering on the laser point cloud data; S3.1.2: Identify common feature points in the visual images and laser point clouds through feature point matching algorithms, and establish a spatial coordinate mapping relationship between the multi-source data; S3.1.3: Based on the mapping relationship, fuse the texture information of the visual images with the depth information of the laser point clouds, and extract key features of the concave and convex area boundaries and flatness change points of the sprayed surface; S3.1.4: Use the Poisson surface reconstruction algorithm to perform mesh division on the fused three-dimensional point cloud data, generate a three-dimensional model of the sprayed surface with texture information, and set the model resolution to contain 500-800 vertices per cubic meter. S3.2: Analyze the influence of spraying distance, angle, pressure, and speed parameters on smoothness and establish a correlation model; S3.3: Based on self-learning and reinforcement learning algorithms, dynamically optimize spraying parameters and generate construction adjustment plans. Specific steps are as follows: S3.3.1: Construct the state space for reinforcement learning, including parameters for surface smoothness, spraying distance, angle, speed, and pressure, as well as geological type labels; S3.3.2: Define the action space, including the adjustment range and step size for spraying distance, angle, and speed; S3.3.3: Set a reward function, with the smoothness compliance rate, material utilization rate, and construction efficiency as optimization objectives. The reward function is expressed as:

[0082] R = w1·flatness compliance rate + w2·material utilization rate + w3·construction efficiency

[0083] Where w1, w2, and w3 are weighting coefficients, satisfying w1+w2+w3=1; S3.3.4: The PPO algorithm is used to train and optimize the model, and the spraying parameter strategy is continuously adjusted through interaction with the environment; S3.3.5: The optimal combination of spraying parameters, including spraying distance, angle, and speed, is output to guide construction adjustments; S3.3.6: The model continuously receives new data and updates online to perform self-learning optimization of the spraying strategy.

[0084] Furthermore, it should be noted that S3.1.2 uses an improved ORB feature descriptor for matching, and feature point pairs are considered valid matches when the distance is less than 0.8 times the Hamming distance threshold. Geological type labels include coal seam interlayers, karst, and fault fracture zones.

[0085] In this embodiment, it should also be noted that the specific steps in S4 are as follows: S4.1: Receive the optimization parameters and adjustment instructions output by S3; S4.2: Based on the predictive control algorithm, adjust the pose parameters of the nozzle's distance from the rock surface and the spray angle in real time; the specific steps are as follows: S4.2.1: Establish a dynamic prediction model of the nozzle pose and spray parameters, and input the current nozzle position, spray angle, rock surface distance, and spray surface smoothness feedback information; S4.2.2: Based on the prediction model, calculate the changing trend of the nozzle pose and spray parameters over the next N time steps; S4.2.3: Set the objective function to minimize the spray error and control energy, the objective function is expressed as:

[0086]

[0087] Among them, e k Let Δu be the injection error at step k. k For controlling the change in the variable, Q and R are weighted matrices; S4.2.4: By solving the optimization problem, the optimal control sequence is obtained, and the optimal control variable at the current moment is selected and output to the actuator; S4.2.5: The model state is updated according to the actual feedback error and transmitted back to S3.3.6 in real time via the OPC-UA protocol. S4.3: Through multi-objective optimization control, the nozzle pose and injection parameters are adjusted in conjunction. The specific operations are as follows: S4.3.1: Establish a three-objective optimization model that includes construction quality, efficiency, and cost: ① Quality objective: Smoothness deviation Δh ≤ 3mm; ② Efficiency objective: Spraying speed v ≥ 1.8m² / min; ③ Cost objective: Material loss rate η ≤ 5%; S4.3.2: Use the improved NSGA-III algorithm to solve the Pareto optimal solution set, where: ① Population size is set to 100, and the number of iterations is ≥ 50 generations; ② For tunnel construction scenarios, add mechanical limit constraints for the nozzle; ③ Use a dynamic reference point mechanism to adapt to different surrounding rock grades; S4.3.3: Select the final parameter combination from the Pareto front through fuzzy decision-making, prioritizing the following: ① Grade IV surrounding rock: Quality objective weight 0.7; ② Grade V surrounding rock: Safety objective weight 0.6 for initial support thickness compliance rate ≥ 95%; S4.3.4: When the rock mass displacement rate is detected to be > 2mm / h, automatically switch to conservative spraying mode.

[0088] Furthermore, it should be noted that the dynamic prediction model in S4.2.1 is a state-space model: Wherein, the state vector Includes parameters such as position, angle, and velocity, w k v k It is Gaussian noise. The specific rule for "fuzzy decision-making" in S4.3.3 is: when the flatness deviation is <3mm and the material loss rate is <5%, the efficiency parameter is selected first.

[0089] In this embodiment, a method for detecting the initial shotcrete spraying quality in tunnels based on vision technology is described as follows: First, through multi-sensor fusion of laser and back-illuminated CMOS vision sensors, the original data of the three-dimensional coordinates and apparent vectors of the sprayed surface are captured synchronously with timestamps. This is combined with an automatic cleaning device (controlling the cleaning frequency according to dust concentration levels) and an adaptive exposure denoising algorithm to eliminate environmental interference. Then, the data is processed by Bundle... The Adjustment algorithm, combining laser point cloud correction, visual images, and inertial navigation data, completes positioning and calibration with the tunnel face center point as the origin, achieving high-precision and robust data acquisition in complex environments. Next, it dynamically acquires the smoothness data of the sprayed surface through laser scanning, using image processing algorithms to identify areas with smoothness deviations >5mm and classifying them into warning levels based on deviations of 5-10mm (Level 1) and >10mm (Level 2). The results are displayed and warnings are issued in real-time through an intuitive interface. Subsequently, multi-view visual and laser information are fused, and a textured 3D model is constructed through distortion correction and feature point matching. The influence of spraying parameters on smoothness is analyzed, and spraying parameters are optimized based on reinforcement learning and the PPO algorithm, generating a construction adjustment plan that is continuously updated through self-learning. Finally, the optimized parameters are received, and the nozzle pose is adjusted in real-time based on a predictive control algorithm (state-space model). An improved NSGA-III algorithm is used to achieve high quality (Δh≤3mm) and efficiency (v≥1.8m). 2 Multi-objective linkage control with parameters such as per minute and cost (η≤5%) is implemented. Differentiated weight allocation is applied to Class IV / V surrounding rock. When the rock mass displacement is >2mm / h, the system switches to conservative mode, and the control feedback is transmitted back to the optimization model via the OPC-UA protocol.

[0090] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0091] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for detecting the initial shotcrete spraying quality in tunnels based on vision technology, characterized in that, Includes the following steps: S1: By using multimodal sensor fusion and adaptive exposure denoising algorithm, combined with multiple information joint correction technology, an automatic cleaning device is set up at the front end of the equipment. Through origin positioning calibration, high-precision and robust acquisition of three-dimensional coordinates and apparent vector data of the sprayed surface in complex environments is achieved. S2: Based on image processing algorithms, it integrates laser scanning, intelligent data processing and real-time monitoring technologies to detect flatness and out-of-limit areas; S3: An AI-based self-learning data fusion and parameter optimization model that automatically identifies key features in multi-view visual data, fuses laser and image information, constructs a high-precision three-dimensional sprayed surface model, analyzes the influence of parameters such as spraying distance, angle, pressure, and speed on the smoothness of the sprayed surface, and dynamically optimizes key construction parameters such as spraying distance, angle, and speed through reinforcement learning algorithms. S4: Based on the analysis results of S3, dynamic and coordinated adjustment of nozzle pose and injection parameters is carried out based on predictive control algorithm and multi-objective optimization control; The specific steps in S3 are as follows: S3.1: Integrates multi-view visual and laser information to automatically extract key features of the sprayed surface and construct a 3D model; S3.2: Analyze the influence of parameters such as spray distance, angle, pressure, and speed on smoothness, and establish a correlation model; S3.3: Based on self-learning and reinforcement learning algorithms, dynamically optimize spraying parameters and generate construction adjustment plans; S3.3 uses self-learning and reinforcement learning algorithms to dynamically optimize spraying parameters and generate construction adjustment plans. The specific steps are as follows: S3.3.1: Construct the state space for reinforcement learning, including parameters such as surface smoothness, spraying distance, angle, velocity, and pressure, as well as geological type labels; S3.3.2: Define the action space, including the adjustment range and adjustment step size of the injection distance, angle, and speed; S3.3.3: Define a reward function with the following optimization objectives: spraying smoothness compliance rate, material utilization rate, and construction efficiency. The reward function is expressed as follows: in, , , For the weighting coefficients, satisfying ; S3.3.4: The PPO algorithm is used to train and optimize the model, and the injection parameter strategy is continuously adjusted through interaction with the environment; S3.3.5: Outputs the optimal combination of spraying parameters, including spraying distance, angle, and speed, to guide construction adjustments; S3.3.6: The model continuously receives new data and updates online to perform self-learning optimization of the spraying strategy; The specific steps in S4 are as follows: S4.1: Receive the optimization parameters and adjustment instructions output by S3; S4.2: Based on predictive control algorithms, the positional parameters of the nozzle's distance from the rock surface and the spray angle are adjusted in real time; S4.3: Through multi-objective optimization control, the nozzle position and spray parameters are adjusted in a coordinated manner; In S4.2, based on the predictive control algorithm, the pose parameters of the nozzle distance and spray angle are adjusted in real time. The specific steps are as follows: S4.2.1: Establish a dynamic prediction model for nozzle position and spray parameters, and input feedback information such as current nozzle position, spray angle, rock surface distance, and spray surface smoothness; S4.2.2: Based on the prediction model, the trend of nozzle pose and injection parameters over the next N time steps is calculated in a rolling manner; S4.2.3: Define the objective function, with the goal of minimizing injection error and control energy. The objective function is expressed as follows: in, Let the injection error be at step k. To control the changes in the quantity, Q and R are weighted matrices; S4.2.4: By solving the optimization problem, the optimal control sequence is obtained, and the optimal control quantity at the current moment is selected and output to the actuator; S4.2.5: Update the model status based on the actual feedback error and transmit it back to S3.3.6 in real time via the OPC-UA protocol; In S4.3, multi-objective optimization control is used to adjust the nozzle position and injection parameters in a linked manner. The specific operation is as follows: S4.3.1: Establish a three-objective optimization model that includes construction quality, efficiency, and cost: ① Quality target: Flatness deviation Δh ≤ 3mm; ② Efficiency target: Injection speed v ≥ 1.8 m² / min; ③ Cost target: Material loss rate η≤5%; S4.3.2: The improved NSGA-III algorithm is used to solve for the Pareto optimal solution set, where: ① Set the population size to 100 and the number of iterations to ≥50 generations; ②For tunnel construction scenarios, add mechanical limit constraints for the nozzles; ③ Adopting a dynamic reference point mechanism to adapt to different surrounding rock grades; S4.3.3: Select the final parameter combination from the Pareto front using fuzzy decision-making, prioritizing the following: ① Class IV surrounding rock: Quality target weight 0.7; ② Class V surrounding rock: The safety target weight for initial support thickness compliance rate ≥ 95% is 0.6; S4.3.4: When the rock mass displacement rate is detected to be >2mm / h, automatically switch to conservative spraying mode.

2. The method for detecting the initial shotcrete spraying quality in tunnels based on vision technology according to claim 1, characterized in that, The specific steps in S1 are as follows: S1.1: Capture the original data of the three-dimensional coordinates and apparent vectors of the sprayed surface through multi-sensor fusion of laser and vision. S1.2: Adaptive exposure denoising algorithm and automatic cleaning device are used to eliminate environmental interference; S1.3: Correction and origin positioning calibration through joint correction of multiple information sources.

3. The method for detecting the initial shotcrete quality in tunnels based on vision technology according to claim 2, characterized in that, In step S1.2, an adaptive exposure denoising algorithm and an automatic cleaning device are used to eliminate environmental interference. The specific steps are as follows: S1.2.1: The dust concentration and light intensity in the tunnel are collected in real time by dust sensor and light sensor. When the dust concentration is >50mg / m³ or the light intensity is <300lux, the anti-interference mechanism is triggered. S1.2.2: Controlling cleaning frequency based on dust concentration levels: ① When the concentration is 50-100mg / m³, start the rotating brush every 5 minutes to clean the sensor surface with high-pressure airflow; ② When the concentration is >100mg / m³, continuous trigger cleaning will be initiated; S1.2.3: Adaptive exposure adjustment is adopted for low-light environments, and a multi-frame superposition averaging denoising algorithm is executed on the acquired images. 3-5 frames of images are acquired continuously, and random noise is eliminated by weighted averaging of pixel gray values. S1.2.4: Based on the response delay of the sensor after cleaning, perform time axis calibration on the subsequent acquired data.

4. The method for detecting the initial shotcrete quality in tunnels based on vision technology according to claim 1, characterized in that, The specific steps in S2 are as follows: S2.1: Dynamically acquire data on the flatness and surface morphology details of the sprayed surface using laser scanning technology; S2.2: Use image processing algorithms to analyze the collected data, identify areas exceeding limits, and determine the warning level; S2.3: The monitoring results are displayed in real time through an intuitive interface, and the construction personnel are given immediate warnings when the limits are exceeded.

5. The method for detecting the initial shotcrete spraying quality in tunnels based on vision technology according to claim 1, characterized in that, In step S3.1, multi-view visual and laser information are integrated to automatically extract key features of the sprayed surface and construct a 3D model. The specific steps are as follows: S3.1.1: Perform distortion correction and grayscale normalization on multi-view visual images, and simultaneously perform noise reduction filtering on laser point cloud data; S3.1.2: Identify common feature points in visual images and laser point clouds using a feature point matching algorithm, and establish a spatial coordinate mapping relationship between multi-source data; S3.1.3: Based on the mapping relationship, the texture information of the visual image is fused with the depth information of the laser point cloud to extract the key features of the concave and convex regions and the abrupt change points of the smoothness of the sprayed surface. S3.1.4: The Poisson surface reconstruction algorithm is used to divide the fused 3D point cloud data into meshes to generate a sprayed surface 3D model with texture information. The model resolution is set to contain 500-800 vertices per cubic meter.

Images