Tunnel guniting intelligent detection method and system based on neural network and point cloud processing
By combining 3D laser scanning and deep neural networks, the problems of low efficiency and poor accuracy in traditional tunnel shotcrete detection have been solved, achieving efficient and accurate tunnel shotcrete defect detection and generating a comprehensive report that includes defect type and geometric parameters.
Patent Information
- Application Number
- CN202511115532.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-11
- Publication Date
- 2025-11-28
AI Technical Summary
Traditional tunnel shotcrete detection methods are inefficient and inaccurate, making them unsuitable for large-scale testing needs and highly susceptible to the experience and subjective factors of the testing personnel.
Point cloud data is acquired using 3D laser scanning technology, and data preprocessing and annotation are performed using a deep neural network model. A ResNet-50 neural network is then constructed for defect detection, generating a comprehensive inspection report.
It achieves efficient, accurate, and automated tunnel shotcrete inspection, can identify multiple defect types and obtain defect geometric parameters, and generate comprehensive inspection reports.
Smart Images
Figure CN121027107A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tunnel engineering detection technology, specifically relating to an intelligent detection method and system for tunnel shotcrete based on neural networks and point cloud processing. Background Technology
[0002] In tunnel construction, shotcreting is a crucial step in ensuring the stability and safety of the tunnel structure. The quality of shotcreting directly affects the tunnel's load-bearing capacity, waterproofing performance, and durability; therefore, defect detection on the shotcrete surface is paramount. Traditional methods for inspecting tunnel shotcrete primarily rely on manual inspections, where inspectors visually observe and tap the surface to determine the presence of defects such as voids and cracks. However, this method has several drawbacks: firstly, manual inspection is inefficient and cannot meet the large-scale inspection needs of tunnel projects; secondly, the inspection results are significantly influenced by the inspectors' experience and subjective factors, making it difficult to guarantee accuracy and reliability.
[0003] The development of 3D laser scanning and artificial intelligence technologies has provided new insights for tunnel shotcrete inspection. 3D laser scanning can quickly and accurately acquire point cloud data of the shotcrete surface in tunnels, containing rich spatial information; neural network technology has powerful feature learning and pattern recognition capabilities, enabling it to learn defect features from large amounts of data. Combining these two technologies for tunnel shotcrete inspection holds promise for achieving intelligent and automated inspection processes, improving efficiency and accuracy. Currently, there is still room for improvement in data preprocessing accuracy, model training effectiveness, and the comprehensiveness of defect detection. Therefore, a more sophisticated intelligent inspection method and system for tunnel shotcrete based on neural networks and point cloud processing is needed. Summary of the Invention
[0004] This invention aims to solve the problems of low efficiency and poor accuracy of traditional tunnel shotcrete detection methods, and provides an intelligent detection method and system for tunnel shotcrete based on neural networks and point cloud processing, so as to achieve efficient and accurate detection of surface defects in tunnel shotcrete.
[0005] To address the aforementioned technical problems, this invention provides an intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing, comprising the following steps:
[0006] Point cloud data of the shotcrete surface in the tunnel was acquired using a 3D laser scanner. The point cloud data includes spatial coordinates and reflection intensity information.
[0007] The point cloud data of the tunnel shotcrete surface is preprocessed, and then the preprocessed point cloud data of the tunnel shotcrete surface is labeled.
[0008] A deep neural network model was constructed, and the neural network was trained using labeled point cloud sample data of the tunnel shotcrete surface.
[0009] Point cloud data of the tunnel shotcrete surface is collected in real time, and the point cloud data of the tunnel shotcrete surface is input into a trained neural network model for defect detection.
[0010] A comprehensive inspection report is generated based on the defect detection results.
[0011] Furthermore, the preprocessing includes noise reduction, deduplication, and registration.
[0012] Furthermore, the annotation of the point cloud data of the pre-processed tunnel shotcrete surface includes:
[0013] The point cloud data of the pre-processed tunnel shotcrete surface is manually classified, and the defect types and geometric parameters of the shotcrete surface are labeled to generate a training sample dataset. The defect types include voids, cracks, insufficient thickness, or uneven surfaces. The geometric parameters of the defects include location coordinates, size, depth, and area.
[0014] Furthermore, the deep neural network model is ResNet-50, specifically including:
[0015] Input layer: Receives feature vectors of point cloud regions. The feature vectors are generated by extracting features from spatial coordinates and reflection intensity. Each point cloud sample contains 1024 points, and each point is represented by a 4-dimensional feature vector. The point cloud is converted into a regular grid using a voxelization method.
[0016] Feature extraction layer: Contains residual blocks in 4 stages, each residual block consists of 1×1, 3×3 and 1×1 convolutional layers, and alleviates gradient vanishing through skip connections. Combined with feature pyramid network, it fuses multi-scale features to capture defects of different sizes.
[0017] Classification and Regression Layers: The classification layer outputs the probability of defect type through a fully connected layer and a Softmax function, including voids, cracks, insufficient thickness, and surface irregularities; the regression layer outputs the geometric parameters of the defect through a fully connected layer, including location coordinates, size, depth, and area.
[0018] Output layer: Outputs the defect type and its confidence level, as well as the defect geometry parameters represented in bounding box form.
[0019] Furthermore, the neural network is trained using labeled point cloud sample data of the tunnel shotcrete surface, specifically as follows:
[0020] The training samples were divided into training and validation sets in an 8:2 ratio. The cross-entropy loss function was used to calculate the defect classification loss, and the mean squared error loss function was used to calculate the geometric parameter regression loss. The total loss was the weighted sum of the two class losses. The Adam optimizer was used with an initial learning rate of 0.001. A learning rate decay strategy was adopted, where the learning rate was halved when the validation set loss did not decrease for 5 consecutive epochs. The training was iterated until the loss converged or the maximum number of iterations was reached. The model with the best performance on the validation set was saved.
[0021] Furthermore, it also includes calculating the sprayed grout thickness distribution by comparing the deviation between the sprayed grout surface point cloud and the tunnel design model, and then fusing the thickness distribution results with the defect detection results to generate a comprehensive inspection report containing thickness distribution and defect information.
[0022] The present invention also provides a tunnel shotcrete intelligent detection system based on neural networks and point cloud processing, the system comprising:
[0023] A three-dimensional laser scanning module is used to acquire point cloud data of the shotcrete surface of the tunnel. The point cloud data includes spatial coordinates and reflection intensity information.
[0024] The data preprocessing and annotation module is used to preprocess the point cloud data of the tunnel shotcrete surface and then annotate the preprocessed point cloud data of the tunnel shotcrete surface.
[0025] The neural network training module is used to construct a deep neural network model and train the neural network using labeled point cloud sample data of the tunnel shotcrete surface.
[0026] The defect detection module is used to input real-time point cloud data into a trained neural network model for defect detection.
[0027] The report generation module is used to generate a comprehensive inspection report based on the defect detection results.
[0028] The data storage module is used to store the collected point cloud data, preprocessed data, training sample dataset, trained neural network model, and detection report.
[0029] Compared with the prior art, the present invention has the following advantages:
[0030] High detection efficiency: It uses three-dimensional laser scanning technology to quickly collect data and combines it with a neural network model for automatic detection, which greatly shortens the detection time and improves the detection efficiency.
[0031] High detection accuracy: By improving data quality through data preprocessing and utilizing the powerful feature learning capabilities of neural network models, various defects on the sprayed surface can be accurately identified, reducing subjective errors in manual inspection.
[0032] Excellent detection comprehensiveness: It can not only detect defect types, but also obtain the geometric parameters of defects. At the same time, combined with the information on the distribution of sprayed grout thickness, the generated detection report is more comprehensive and provides sufficient basis for engineering decision-making. Attached Figure Description
[0033] The accompanying drawings illustrate various embodiments generally by way of example rather than limitation, and are used, together with the specification and claims, to explain embodiments of the invention. Where appropriate, the same reference numerals are used in all drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the apparatus or method.
[0034] Figure 1 A schematic diagram of the method flow of the present invention is shown. Detailed Implementation
[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0036] like Figure 1 As shown, this embodiment of the invention provides a smart detection method for tunnel shotcrete based on neural networks and point cloud processing, including the following steps:
[0037] Point cloud data of the shotcrete surface in the tunnel was acquired using a 3D laser scanner. The point cloud data includes spatial coordinates and reflection intensity information.
[0038] The point cloud data of the tunnel shotcrete surface is preprocessed, and then the preprocessed point cloud data of the tunnel shotcrete surface is labeled.
[0039] A deep neural network model was constructed, and the neural network was trained using labeled point cloud sample data of the tunnel shotcrete surface.
[0040] Point cloud data of the tunnel shotcrete surface is collected in real time, and the point cloud data of the tunnel shotcrete surface is input into a trained neural network model for defect detection.
[0041] A comprehensive inspection report is generated based on the defect detection results.
[0042] Example 1
[0043] 1. Point cloud data acquisition
[0044] At a railway tunnel construction site, a high-precision 3D laser scanner (model: Faro Focus S350) was used to scan the shotcrete surface at a point spacing of 0.01m to obtain point cloud data. The data volume was approximately 10 million points, including spatial coordinates (X,Y,Z) and reflection intensity information.
[0045] 2. Point cloud preprocessing and annotation
[0046] Preprocessing: Point cloud processing software (such as CloudCompare) is used to perform noise reduction, deduplication, and registration on the raw point cloud data. Noise reduction uses a statistical outlier removal algorithm with a neighborhood radius of 0.05m and a standard deviation threshold of 1.0. Deduplication is based on merging duplicate points using a Euclidean distance threshold (0.005m). Registration uses an iterative nearest point (ICP) algorithm to align the point cloud data with the tunnel design model, with the registration error controlled within 0.01m.
[0047] Labeling: The preprocessed point cloud data is classified manually, and the defect types (voids, cracks, insufficient thickness, uneven surface) and geometric parameters (location coordinates, size, depth, area) are labeled. A training sample dataset containing 1200 samples is generated (800 cracks, 300 voids, and 100 insufficient thickness).
[0048] 3. Deep Neural Network Model Construction and Training
[0049] Model construction: The ResNet-50 architecture is adopted, specifically including:
[0050] Input layer: Receives feature vectors from local regions of the point cloud. Each point cloud sample contains 1024 points, each represented by a 4-dimensional feature vector (X, Y, Z, intensity). A voxelization method is used to convert the point cloud into a regular mesh (voxel size 0.01m). 3 ).
[0051] Feature extraction layer: Contains residual blocks in four stages (3, 4, 6 and 3 residual blocks respectively). Each residual block consists of 1×1, 3×3 and 1×1 convolutional layers. Skip connections are used to alleviate gradient vanishing. Multi-scale features are fused in combination with Feature Pyramid Network (FPN) to capture defects of different sizes.
[0052] Classification and Regression Layers: The classification layer outputs the probability of defect type (void, crack, insufficient thickness, uneven surface) through a fully connected layer and a Softmax function; the regression layer outputs the geometric parameters of the defect (location coordinates, size, depth, area) through a fully connected layer and is optimized using a mean squared error loss function.
[0053] Output layer: Outputs the defect type and its confidence level, as well as the defect geometry parameters represented in bounding box form.
[0054] Model training:
[0055] The training samples were divided into a training set and a validation set in an 8:2 ratio.
[0056] The cross-entropy loss function is used to calculate the defect classification loss, and the mean squared error loss function is used to calculate the geometric parameter regression loss. The total loss is the weighted sum of the two types of losses (with weights of 0.6 and 0.4, respectively).
[0057] Using the Adam optimizer, the initial learning rate is set to 0.001, and a learning rate decay strategy is adopted, where the learning rate is halved when the validation set loss does not decrease for 5 consecutive epochs.
[0058] After 100 epochs of iterative training, the model with the best performance on the validation set was saved, achieving a defect classification accuracy of 96.5% and a geometric parameter regression error (RMSE) of 0.02m.
[0059] 4. Real-time defect detection
[0060] Point cloud data of the shotcrete surface in the tunnel was acquired in real time using a 3D laser scanner and input into a trained ResNet-50 model for defect detection. The model's inference time for processing 1 million point cloud data points was 0.4 seconds. The detection results included defect type (e.g., crack), location coordinates (X, Y, Z), size (length 10cm, width 0.5cm), depth (0.3cm), and area (50cm²). 2 ).
[0061] By comparing the deviation between the point cloud of the shotcrete surface and the tunnel design model (based on the BIM model), the shotcrete thickness distribution is calculated. The thickness distribution results are then integrated with the defect detection results to generate a comprehensive inspection report, which includes information such as the location, type, parameters, and thickness distribution of defects.
[0062] Example 2
[0063] This embodiment provides an intelligent tunnel shotcrete detection system based on neural networks and point cloud processing. The system includes: a 3D laser scanning module, a data preprocessing and annotation module, a neural network training module, a defect detection module, and a report generation module. The 3D laser scanning module collects point cloud data; the data preprocessing and annotation module preprocesses and annotates the data; the neural network training module constructs and trains a model; the defect detection module performs real-time defect detection; and the report generation module generates a detection report. All modules work collaboratively to realize the entire process of intelligent tunnel shotcrete detection.
[0064] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for intelligent detection of tunnel shotcrete based on neural networks and point cloud processing, characterized in that, Includes the following steps: Point cloud data of the shotcrete surface in the tunnel was acquired using a 3D laser scanner. The point cloud data includes spatial coordinates and reflection intensity information. The point cloud data of the tunnel shotcrete surface is preprocessed, and then the preprocessed point cloud data of the tunnel shotcrete surface is labeled. A deep neural network model was constructed, and the neural network was trained using labeled point cloud sample data of the tunnel shotcrete surface. Point cloud data of the shotcrete surface of the tunnel is collected in real time, and the point cloud data of the shotcrete surface of the tunnel is input into a trained neural network model for defect detection. A comprehensive inspection report is generated based on the defect detection results.
2. The intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing according to claim 1, characterized in that, The preprocessing includes noise reduction, deduplication, and registration.
3. The intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing according to claim 1, characterized in that, The annotation of the point cloud data of the pre-processed tunnel shotcrete surface includes: The point cloud data of the pre-processed tunnel shotcrete surface is manually classified, and the defect types and geometric parameters of the shotcrete surface are labeled to generate a training sample dataset. The defect types include voids, cracks, insufficient thickness, or uneven surfaces. The geometric parameters of the defects include location coordinates, size, depth, and area.
4. The intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing according to claim 1, characterized in that, The deep neural network model is ResNet-50, specifically including: Input layer: Receives feature vectors of point cloud regions. The feature vectors are generated by extracting features from spatial coordinates and reflection intensity. Each point cloud sample contains 1024 points, and each point is represented by a 4-dimensional feature vector. The point cloud is converted into a regular grid using a voxelization method. Feature extraction layer: Contains residual blocks in 4 stages, each residual block consists of 1×1, 3×3 and 1×1 convolutional layers, and alleviates gradient vanishing through skip connections. Combined with feature pyramid network, it fuses multi-scale features to capture defects of different sizes. Classification and Regression Layers: The classification layer outputs the probability of defect type through a fully connected layer and a Softmax function, including voids, cracks, insufficient thickness, and surface irregularities; the regression layer outputs the geometric parameters of the defect through a fully connected layer, including location coordinates, size, depth, and area. Output layer: Outputs the defect type and its confidence level, as well as the defect geometry parameters represented in bounding box form.
5. The intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing according to claim 1, characterized in that, The neural network is trained using labeled point cloud sample data of the tunnel shotcrete surface, specifically as follows: The training samples were divided into training and validation sets in an 8:2 ratio. The cross-entropy loss function was used to calculate the defect classification loss, and the mean squared error loss function was used to calculate the geometric parameter regression loss. The total loss was the weighted sum of the two class losses. The Adam optimizer was used with an initial learning rate of 0.
001. A learning rate decay strategy was adopted, and the learning rate was halved when the validation set loss did not decrease for 5 consecutive epochs. Iterate through the training until the loss converges or the maximum number of iterations is reached, and save the model with the best performance on the validation set.
6. The intelligent detection method for tunnel shotcrete based on neural networks and point cloud processing according to claim 1, characterized in that, It also includes calculating the sprayed grout thickness distribution by comparing the deviation between the sprayed grout surface point cloud and the tunnel design model, and then integrating the thickness distribution results with the defect detection results to generate a comprehensive inspection report containing thickness distribution and defect information.
7. A tunnel shotcrete intelligent detection system based on neural networks and point cloud processing, characterized in that, The system includes: A three-dimensional laser scanning module is used to acquire point cloud data of the shotcrete surface of the tunnel. The point cloud data includes spatial coordinates and reflection intensity information. The data preprocessing and annotation module is used to preprocess the point cloud data of the tunnel shotcrete surface and then annotate the preprocessed point cloud data of the tunnel shotcrete surface. The neural network training module is used to construct a deep neural network model and train the neural network using labeled point cloud sample data of the tunnel shotcrete surface. The defect detection module is used to input real-time point cloud data into a trained neural network model for defect detection. The report generation module is used to generate a comprehensive inspection report based on the defect detection results.
8. The intelligent tunnel shotcrete detection system based on neural networks and point cloud processing according to claim 7, characterized in that, The system also includes a data storage module for storing the collected point cloud data, preprocessed data, training sample datasets, trained neural network models, and detection reports.