A two-stage denoising method suitable for point cloud data of a tunneling machine working environment
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-04-23
- Publication Date
- 2026-07-03
Smart Images
Figure CN122335599A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional point cloud data processing, specifically a two-stage noise reduction method for point cloud data applicable to the working environment of tunneling machines. Background Technology
[0002] Point clouds of underground coal mine tunnels are crucial for constructing accurate 3D models of tunnels and ensuring safe production and intelligent, efficient, and green mining in coal mines. Currently, acquiring point clouds of underground coal mine tunnels mainly relies on 3D laser scanning technology. While this technology can quickly collect large amounts of point cloud data with high accuracy, density, and rich semantic information, its application is significantly affected by the complex working environment of underground coal mines. Due to the enclosed space and poor satellite positioning signals underground, point cloud data acquisition and stitching are difficult, and the acquired point cloud data is often subject to significant noise due to interference from dust, humidity, and mechanical vibration. Traditional noise reduction methods are ineffective in handling such complex noise and cannot meet the high-precision point cloud data requirements of coal mine tunneling. Therefore, a more effective noise reduction method is urgently needed to improve the quality of point cloud data. Summary of the Invention
[0003] The purpose of this invention is to provide a two-stage noise reduction method for point cloud data in the working environment of tunneling machines, so as to solve the problems mentioned in the background art.
[0004] To address the aforementioned technical problems, according to one aspect of this application, a two-stage noise reduction method for point cloud data in a tunneling machine operating environment is provided, the method comprising:
[0005] Step 1: Use normalization to scale and translate the point cloud to the vicinity of the origin, and use k-nearest neighbors to complete the neighborhood construction;
[0006] Step 2: Update the learning parameters of the shared weights through backpropagation and gradient descent. After iterative training, obtain the shared weights of the multilayer perceptron. Input the neighborhood constructed by the k-nearest neighbor algorithm into the shared weights to extract initial features, and then obtain point-by-point features. Then, combine the graph attention mechanism to enhance the expression of important points in the point-by-point features, and complete the neighborhood generation of point-by-point feature encoding.
[0007] Step 3: Input the neighborhood encoded by point-by-point features into the SA hierarchical aggregation model using PointNet++, use the farthest point sampling algorithm and grouping algorithm to complete the construction of multi-level local sets, and then apply a multilayer perceptron to each set to generate multiple neighborhoods composed of multiple local geometric features;
[0008] Step 4: Use graph convolution to aggregate multiple local geometric features in a neighborhood calculated by the SA model to obtain the aggregated features of the neighborhood;
[0009] Step 5: In the outlier detection module, the neighborhood aggregation features output by graph convolution are input into the activation function. The activation function calculates the outlier probability of each point in the tunneling machine's operating environment point cloud. If the probability is greater than a preset threshold, the point is judged as an outlier; if the probability is less than the preset threshold, the point is judged as a non-outlier. This completes the outlier detection and removal of the tunneling machine's operating environment point cloud.
[0010] Step 6: The point cloud of the tunneling machine's operating environment after outlier removal is entered into the noise reduction module;
[0011] The noise reduction module differs from the outlier detection module only in the last step. The neighborhood aggregation information output by the graph convolution calculation is input into the single-point convolution to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The prediction result cancels out the noise components, thus completing the point cloud noise reduction.
[0012] As a further preferred approach, in step one, normalization is used to scale and translate the underground tunnel point cloud to the vicinity of the origin, and the k-nearest neighbor algorithm is used to complete the neighborhood construction, including:
[0013] ; (1)
[0014] ; (2)
[0015] In the formula, n represents the number of points contained in the point cloud, and x i y i z i This represents the coordinate components of the i-th point. , , This represents the coordinates after a translation transformation of the original coordinates x, y, z. , , Represents the coordinates after translation transformation , , The coordinates after scaling transformation, where D represents the distance between the two farthest points in the point cloud.
[0016] Using k-nearest neighbors to build the point cloud neighborhood of the tunneling machine's operating environment facilitates the aggregation of neighborhood information in graph convolution.
[0017] As a further preferred approach, in step two, a multilayer perceptron is used to generate pointwise features, and an attention mechanism is used to enhance feature representation, including:
[0018] Graph attention mechanisms are used to enhance the feature representation of important points. The working principle of graph attention mechanisms is as follows:
[0019] ① Learn weight features during the model training phase ;
[0020] ② Use the learned weight features W to perform feature transformation on the input feature matrix H to obtain ;
[0021] ③ Use attention function a to compute the feature transformation Wh of node i. i Wh feature transformation of node j j Attention score;
[0022] ; (3)
[0023] ④ Perform softmax normalization on the attention score.
[0024] ; (4)
[0025] Among them, I i Let i represent the set of neighbors of node i;
[0026] ⑤ Finally, the output feature of node i is:
[0027] ; (5)
[0028] Where σ is the activation function.
[0029] As a further preferred approach, step three uses PointNet++'s SA hierarchical aggregation module to generate local geometric features, including:
[0030] A multi-layered local region is constructed using the farthest point sampling (FPS) algorithm and grouping algorithm. In each region, a multilayer perceptron with max pooling is applied to generate local geometric features f. i .
[0031] As a further preferred approach, in step four, graph convolution is used to aggregate neighborhood information. The working principle of graph convolution is as follows:
[0032] ① Calculate the KNN neighborhood using Euclidean distance to form the initial graph structure;
[0033] ② Use the EdgeConv layer to perform feature difference aggregation. The calculation formula is as follows:
[0034] ; (6)
[0035] In the formula, The characteristics of the center node i in the l-th layer are represented. The feature difference between neighboring node j and center node i can capture local geometric relationships. It is a nonlinear transformation implemented by a multilayer perceptron. This transformation can map the feature difference to a high-dimensional space. max represents max pooling, which is used to aggregate neighborhood features and ensure permutation invariance.
[0036] As a further preferred solution, in step five, an outlier detection module is used to remove outliers from the point cloud of the tunneling machine's operating environment, including:
[0037] The graph convolution module calculates the neighborhood aggregation information of the output, and then concatenates the neighborhood aggregation information and local geometric features. The concatenated result is then passed through a fully connected layer and then enters the activation function.
[0038] The activation function is selected as the sigmoid function, which transforms the outlier detection task of the point cloud of the tunneling machine's working environment into a classification task.
[0039] The calculation formula for outlier detection in the point cloud of the tunneling machine's operating environment, using the neighborhood aggregation information output by the graph convolution module, is as follows:
[0040] ; (7)
[0041] In the formula, p i The probability f calculated for the sigmoid function. global The neighborhood aggregation feature is calculated by formula (6), f i [;] represents local geometric features, [;] represents feature connections, W c represents the weights in the fully connected layer, and b represents the biases in the fully connected layer; the sigmoid function can output the probability that each point in the point cloud of the tunneling machine's operating environment belongs to an outlier, and the output probability p i Compare with a pre-set outlier threshold t, if p i If p is greater than t, then the point is an outlier. i If the value is less than t, then the point is considered a non-outlier point; the outlier detection task of the point cloud of the tunneling machine's operating environment is completed using the above classification approach.
[0042] As a further preferred solution, in step six, a noise reduction module is used to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The predicted coordinates will cancel out the noise components, thereby achieving point cloud noise reduction, including:
[0043] After outlier detection and removal, the point cloud of the tunneling machine's working environment is sequentially translated, scaled, k-nearest neighbor construction, multilayer perceptron, graph attention mechanism, SA layer, and graph convolution to obtain neighborhood aggregation information.
[0044] The neighborhood aggregation information is fed into a single-point convolution to reduce the dimensionality of the high-dimensional features. Finally, through prediction, the three-dimensional coordinates of the point cloud of the tunneling machine's operating environment are generated.
[0045] The predicted coordinates can cancel out the noise components, transforming the point cloud denoising task into a regression task. By regressing the coordinates of the noise points, the denoising of the point cloud in the tunneling machine's operating environment is completed.
[0046] Compared with traditional noise reduction methods, the two-stage noise reduction method proposed in this patent, which is adapted to point cloud data of tunneling machine operating environment, can more effectively learn the point cloud features of tunneling machine operating environment. Based on the learned features, this method can better distinguish between cluster points, noise points and effective points, and effectively overcome the problems of under-noise reduction and over-noise reduction in traditional noise reduction during the noise reduction process.
[0047] Specifically, it has the following advantages:
[0048] 1. By extracting key features point-by-point through a "multilayer perceptron + graph attention mechanism," and combining PointNet++'s SA module to construct multi-level local sets and graph convolution modules to aggregate neighborhood information, this method can accurately identify outliers, noise points, and valid points in the point cloud of the tunneling machine's operating environment. Compared with the fuzzy distinction between the three types of points by traditional methods, this method significantly reduces the false positive rate of valid points, making the noise reduction results more consistent with the true structural features of the point cloud.
[0049] 2. To address the problems of "under-denoising (residual noise affecting subsequent processing)" and "over-denoising (mistaken deletion of valid points leading to structural loss)" that are prone to occur in traditional methods, this method adopts a two-stage design of "outlier prediction module to specifically remove outliers + denoising module to specifically adjust the coordinates of noise points" to achieve "on-demand denoising". This not only completely removes invalid noise, but also completely preserves the valid point cloud information of walls, pipelines, supporting structures, etc., ensuring the accuracy of subsequent modeling and spatial analysis.
[0050] 3. A two-stage noise reduction method for point cloud data of tunneling machine operating environment is adopted, which uses a partial training and partial freezing strategy. First, the outlier detection module is trained separately and the noise reduction module is frozen. Then, the noise reduction module is trained separately and the outlier detection module is frozen. Finally, the whole network is finely adjusted, which can adapt to dust, water mist outliers and vibration noise generated by dynamic underground operations and improve training efficiency. Attached Figure Description
[0051] Figure 1 A structural block diagram of a two-stage noise reduction method for point cloud data in the tunneling machine operating environment provided in the embodiments of this application;
[0052] Figure 2 A schematic diagram of the attention mechanism in a two-stage noise reduction model for point cloud data in a tunneling machine operating environment provided for embodiments of this application;
[0053] Figure 3A network structure diagram of a two-stage noise reduction model for point cloud data in a tunneling machine operating environment, provided for embodiments of this application;
[0054] Figure 4 A flowchart for training a two-stage noise reduction model for point cloud data in the tunneling machine operating environment, provided for embodiments of this application. Detailed Implementation
[0055] 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.
[0056] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0057] It should be noted that the terms "first," "the," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] As described in the background section, due to the enclosed tunneling environment of tunneling machines, poor satellite positioning signals, and difficulties in collecting and stitching point cloud data, the collected point cloud data often contains a large amount of noise due to interference from factors such as dust, humidity, and mechanical vibration. Traditional noise reduction methods are not very effective in dealing with such complex noise and cannot meet the needs of tunneling machines for high-precision tunnel point cloud data.
[0060] Therefore, in order to solve the quality problems of point clouds in the existing tunneling machine operating environment and to solve the problems of under-noise reduction and over-noise reduction in traditional noise reduction, the embodiments of this application provide a two-stage noise reduction method suitable for point cloud data in the tunneling machine operating environment.
[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
[0062] This embodiment provides a two-stage noise reduction method for point cloud data in the tunneling machine operating environment. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.
[0063] Figure 1 This is a schematic flowchart of a two-stage noise reduction method for point cloud data in a tunneling machine operating environment, provided according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:
[0064] Step S101: Use normalization to scale and translate the point cloud to the vicinity of the origin, and then use the k-nearest neighbor algorithm to complete the neighborhood construction;
[0065] ; (1)
[0066] ; (2)
[0067] In the formula, n represents the number of points contained in the point cloud, and x i y i z i This represents the coordinate components of the i-th point. , , This represents the coordinates after a translation transformation of the original coordinates x, y, z. , , Represents the coordinates after translation transformation , , The coordinates after scaling transformation, where D represents the distance between the two farthest points in the point cloud;
[0068] Specifically, formulas (1) and (2) are used to complete the translation and scaling of the tunneling machine's working environment point cloud. Finally, all the tunneling machine's working environment point clouds will be moved to the vicinity of the origin, which simplifies the complexity of model learning and avoids the model learning too much 3D spatial position information that is unrelated to outlier detection and noise reduction.
[0069] Using k-nearest neighbors to build the point cloud neighborhood of the tunneling machine's operating environment facilitates the aggregation of neighborhood information in graph convolution.
[0070] In step S101, by performing translation and scaling normalization on the point cloud, different input data can have uniform pose characteristics, thereby reducing the data distribution differences caused by the dynamic working environment of the tunneling machine. This also alleviates the problem of uneven point cloud density caused by different manual data collection habits, reducing the probability of the network model misclassifying outliers and noise points. Furthermore, pre-constructing k-nearest neighbor topology relationships before inputting point cloud data into the network can help the network effectively distinguish between valid feature points and outliers in low-density point cloud regions during the feature discrimination stage.
[0071] Step S102: The learning parameters of the shared weights are updated through backpropagation and gradient descent. After iterative training, the shared weights of the multilayer perceptron are obtained. The neighborhood constructed by the k-nearest neighbor algorithm is input into the shared weights to extract initial features, thereby obtaining point-by-point features. Then, the important points in the point-by-point features are enhanced by graph attention mechanism to complete the neighborhood generation of point-by-point feature encoding.
[0072] Specifically, assuming the input point cloud of the tunneling machine's operating environment is [1024, 3, 2], where 1024 represents the number of points in the point cloud, 3 represents the point cloud's attributes (3D coordinates), and 2 indicates that the model processes two point cloud datasets simultaneously (batch processing). Abstract features are extracted using a multilayer perceptron with shared weights, for example, 3→32→64→99, outputting a high-dimensional abstract feature set [1024, 99, 2], thus generating point-by-point features. , where N represents the number of points in the point cloud (1024) and d is the feature dimension (99).
[0073] Figure 2 This is a schematic diagram of the graph attention mechanism. The graph attention mechanism is used to enhance the feature representation of important points. The working principle of the graph attention mechanism is as follows:
[0074] ① Learn weight features during the model training phase ;
[0075] ② Use the learned weight features W to perform feature transformation on the input feature matrix H to obtain ;
[0076] ③ Use attention function a to compute the feature transformation Wh of node i. i Wh feature transformation of node j j Attention score;
[0077] ; (3)
[0078] ④ Perform softmax normalization on the attention score.
[0079] ; (4)
[0080] Among them, I i Let i represent the set of neighbors of node i;
[0081] ⑤ Finally, the output feature of node i is:
[0082] ; (5)
[0083] Where σ is the activation function.
[0084] In step S102, by introducing a graph attention mechanism, a single point can retain its own feature information while adaptively fusing the spatial and feature information of its neighboring points. For point clouds with uneven density acquired under the dynamic operating environment of a tunneling machine, this method can effectively characterize the feature differences between effective feature points and outliers in low-density areas. It also highlights the distinguishability between key points with rapidly changing normal vectors (such as anchor bolts and cables), linear targets, and noise points, preventing the network model from over-smoothing important geometric features. Furthermore, since the point cloud neighborhood has already been constructed in step S101, this step does not require repeated neighborhood search operations for the graph attention mechanism, significantly saving time on neighborhood construction and ensuring the timeliness of point cloud processing under tunneling machine operating conditions.
[0085] Step S103: Input the neighborhood encoded by point-by-point features into the SA hierarchical aggregation model using PointNet++, and use the farthest point sampling algorithm and grouping algorithm to complete the construction of multi-level local sets. Then apply a multilayer perceptron to each set to generate multiple neighborhoods composed of multiple local geometric features.
[0086] The local geometric features are generated using PointNet++'s SA hierarchical aggregation module, including:
[0087] A multi-layered local region is constructed using the farthest point sampling (FPS) algorithm and grouping algorithm. In each region, a multilayer perceptron with max pooling is applied to generate local geometric features f. i .
[0088] Then, graph convolution is used to aggregate neighborhood information, where graph convolution works as follows:
[0089] ① Calculate the KNN neighborhood using Euclidean distance to form the initial graph structure;
[0090] ② Use the EdgeConv layer to perform feature difference aggregation. The calculation formula is as follows:
[0091] ; (6)
[0092] In the formula, xli represents the feature of the center node i in the l-th layer, xlj-xli represents the feature difference between the neighboring node j and the center node i, which can capture local geometric relationships, and ϕ l It is a nonlinear transformation implemented by a multilayer perceptron. This transformation can map the feature difference to a high-dimensional space. max represents max pooling, which is used to aggregate neighborhood features and ensure permutation invariance.
[0093] In step S103, for the differentiated features extracted by the graph attention mechanism, global feature fusion is completed through the feature aggregation module. The features that determine whether a point is an outlier are highlighted in the outlier detection module, and the features that distinguish noise points are enhanced in the noise reduction module. This approach highlights key features while retaining non-core features, achieving selective feature expression and adaptive allocation. It avoids redundant calculations of the graph attention mechanism and neighborhood feature aggregation in the noise reduction module, and can further refine and optimize based on existing features without repeating feature extraction.
[0094] Step S104: In the outlier detection module, the neighborhood aggregation features output by graph convolution calculation are input into the activation function. The activation function calculates the outlier probability of each point in the tunneling machine's working environment point cloud. If the probability is greater than a preset threshold, the point is judged as an outlier; if the probability is less than the preset threshold, the point is judged as a non-outlier. This completes the outlier detection and removal of the tunneling machine's working environment point cloud.
[0095] Specifically, the point cloud of the tunneling machine's operating environment, after outlier removal, enters the noise reduction module;
[0096] An outlier detection module is used to remove outliers from the point cloud of the tunneling machine's operating environment, including:
[0097] The graph convolution module calculates the neighborhood aggregation information of the output, and then concatenates the neighborhood aggregation information and local geometric features. The concatenated result is then passed through a fully connected layer and then enters the activation function.
[0098] The activation function is selected as the sigmoid function, which transforms the outlier detection task of the point cloud of the tunneling machine's working environment into a classification task.
[0099] The calculation formula for outlier detection in the point cloud of the tunneling machine's operating environment, using the neighborhood aggregation information output by the graph convolution module, is as follows:
[0100] ; (7)
[0101] In the formula, p i The probability f calculated for the sigmoid function. global The neighborhood aggregation feature is calculated by formula (6), f i[;] represents local geometric features, [;] represents feature connections, W c represents the weights in the fully connected layer, and b represents the biases in the fully connected layer; the sigmoid function can output the probability that each point in the point cloud of the tunneling machine's operating environment belongs to an outlier, and the output probability p i Compare with a pre-set outlier threshold t, if p i If p is greater than t, then the point is an outlier. i If the value is less than t, then the point is considered a non-outlier point; the outlier detection task of the point cloud of the tunneling machine's operating environment is completed using the above classification approach.
[0102] The noise reduction module differs from the outlier detection module only in the last step. The neighborhood aggregation information output by the graph convolution calculation is input into the single-point convolution to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The prediction result cancels out the noise components, thus completing the point cloud noise reduction.
[0103] Specifically, a noise reduction module is used to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The predicted coordinates cancel out the noise components, thus achieving point cloud noise reduction, including:
[0104] After outlier detection and removal, the point cloud of the tunneling machine's working environment is sequentially translated, scaled, k-nearest neighbor construction, multilayer perceptron, graph attention mechanism, SA layer, and graph convolution to obtain neighborhood aggregation information.
[0105] The neighborhood aggregation information is fed into a single-point convolution to reduce the dimensionality of the high-dimensional features. Finally, through prediction, the three-dimensional coordinates of the point cloud of the tunneling machine's operating environment are generated.
[0106] The predicted coordinates can cancel out the noise components, transforming the point cloud denoising task into a regression task. By regressing the coordinates of the noise points, the denoising of the point cloud in the tunneling machine's operating environment is completed.
[0107] In step S104, the point cloud of the tunneling machine's operating environment is processed step-by-step through the outlier detection module and the noise reduction module. First, outlier removal is completed, and then noise point filtering is performed. This step-by-step processing strategy has a clear division of labor and logical structure, effectively avoiding problems such as incomplete noise reduction and insufficient outlier removal that are prone to occur in one-step processing. At the same time, it can prevent the loss of important feature information during the processing.
[0108] Figure 3This diagram illustrates a two-stage denoising network model suitable for point cloud data from a tunneling machine's operating environment. The original point cloud data from the tunneling machine's operating environment sequentially passes through an outlier detection module and a denoising module to remove outliers and reduce noise. In the outlier detection module, the 3D features of the point cloud are extracted, and the probability of each point being an outlier is calculated using an activation function. Outlier identification is then performed based on a pre-set threshold. In the denoising module, the 3D features of the point cloud are extracted, and the offset distance of noisy points is predicted using single-point convolution to complete the point cloud denoising.
[0109] Figure 4 This invention presents a training process for a two-stage denoising model suitable for point cloud data in tunneling machine operating environments. To better leverage the model's two-stage detection capabilities, the invention employs a partial training and partial freezing strategy during the model training phase. Specifically, it trains the outlier detection module separately while freezing the denoising module, trains the denoising module separately while freezing the outlier detection module, and finally fine-tunes the overall network structure.
[0110] The above description is only 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 scope of the technology 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 two-stage noise reduction method for point cloud data applicable to tunneling machine operating environments, comprising: Step 1: Use normalization to scale and translate the point cloud to the vicinity of the origin, and then use the k-nearest neighbor algorithm to complete the neighborhood construction; Step 2: Update the learning parameters of the shared weights through backpropagation and gradient descent. After iterative training, obtain the shared weights of the multilayer perceptron. Input the neighborhood constructed by the k-nearest neighbor algorithm into the shared weights to extract initial features, and then obtain point-by-point features. Then, the graph attention mechanism is used to enhance the representation of important points in the point-by-point features, thereby completing the neighborhood generation of the point-by-point feature encoding; Step 3: Input the neighborhood encoded by point-by-point features into the SA hierarchical aggregation model using PointNet++, use the farthest point sampling algorithm and grouping algorithm to complete the construction of multi-level local sets, and then apply a multilayer perceptron to each set to generate multiple neighborhoods composed of multiple local geometric features; Step 4: Use graph convolution to aggregate multiple local geometric features in a neighborhood calculated by the SA model to obtain the aggregated features of the neighborhood; Step 5: In the outlier detection module, the neighborhood aggregation features output by graph convolution are input into the activation function. The activation function calculates the outlier probability of each point in the tunneling machine's operating environment point cloud. If the probability is greater than a preset threshold, the point is judged as an outlier; if the probability is less than the preset threshold, the point is judged as a non-outlier. This completes the outlier detection and removal of the tunneling machine's operating environment point cloud. Step 6: The point cloud of the tunneling machine's operating environment after outlier removal is entered into the noise reduction module; The noise reduction module differs from the outlier detection module only in the last step. The neighborhood aggregation information output by the graph convolution calculation is input into the single-point convolution to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The prediction result cancels out the noise components, thus completing the point cloud noise reduction.
2. The two-stage noise reduction method for point cloud data in the tunneling machine operating environment according to claim 1, characterized in that: In step one, normalization is used to scale and translate the underground tunnel point cloud to the vicinity of the origin, and the k-nearest neighbor algorithm is used to complete the neighborhood construction, including: Use formulas (1) and (2) to perform translation and scaling of the point cloud of the tunneling machine's working environment; ； （1） ； （2） In the formula, n represents the number of points contained in the point cloud, and x i y i z i This represents the coordinate components of the i-th point. , , This represents the coordinates after a translation transformation of the original coordinates x, y, z. , , Represents the coordinates after translation transformation , , The coordinates after scaling transformation, where D represents the distance between the two farthest points in the point cloud; Using k-nearest neighbors to build the point cloud neighborhood of the tunneling machine's operating environment facilitates the aggregation of neighborhood information in graph convolution.
3. The two-stage noise reduction method for point cloud data in the tunneling machine operating environment according to claim 1, characterized in that: In step two, a multilayer perceptron is used to generate pointwise features, and an attention mechanism is used to enhance feature representation, including: Graph attention mechanisms are used to enhance the feature representation of important points. The working principle of graph attention mechanisms is as follows: ① Learn weight features during the model training phase ; ② Use the learned weight features W to perform feature transformation on the input feature matrix H to obtain ; iii. using an attention function a to compute a feature transformation Whof node i i a feature transformation Whof node j j an attention score; ； （3） ④ Perform softmax normalization on the attention score. ； （4） Among them, I i Let i represent the set of neighbors of node i; ⑤ Finally, the output feature of node i is: ； （5） Where σ is the activation function.
4. The two-stage noise reduction method for point cloud data in the tunneling machine operating environment according to claim 1, characterized in that: In step three, the SA hierarchical aggregation module of PointNet++ is used to generate local geometric features, including: A multi-layered local region is constructed using the farthest point sampling (FPS) algorithm and grouping algorithm. In each region, a multilayer perceptron with max pooling is applied to generate local geometric features f. i .
5. A two-stage noise reduction method for point cloud data adapted to the working environment of a tunneling machine, as described in claim 1, is characterized in that: In step four, graph convolution is used to aggregate neighborhood information. The working principle of graph convolution is as follows: ① Calculate the KNN neighborhood using Euclidean distance to form the initial graph structure; ② Use the EdgeConv layer to perform feature difference aggregation. The calculation formula is as follows: ； （6） In the formula, The characteristics of the center node i in the l-th layer are represented. The feature difference between neighboring node j and center node i can capture local geometric relationships. It is a nonlinear transformation implemented by a multilayer perceptron. This transformation can map the feature difference to a high-dimensional space. max represents max pooling, which is used to aggregate neighborhood features and ensure permutation invariance.
6. A two-stage noise reduction method for point cloud data in a tunneling machine operating environment according to claim 1, characterized in that: In step five, the outlier detection module is used to remove outliers from the point cloud of the tunneling machine's operating environment, including: The graph convolution module calculates the neighborhood aggregation information of the output, and then concatenates the neighborhood aggregation information and local geometric features. The concatenated result is then passed through a fully connected layer and then enters the activation function. The activation function is selected as the sigmoid function, which transforms the outlier detection task of the point cloud of the tunneling machine's working environment into a classification task. The calculation formula for outlier detection in the point cloud of the tunneling machine's operating environment, using the neighborhood aggregation information output by the graph convolution module, is as follows: ； （7） In the formula, p i The probability f calculated for the sigmoid function. global The neighborhood aggregation feature is calculated by formula (6), f i [;] represents local geometric features, [;] represents feature connections, W c represents the weights in the fully connected layer, and b represents the biases in the fully connected layer; the sigmoid function can output the probability that each point in the point cloud of the tunneling machine's operating environment belongs to an outlier, and the output probability p i Compare with a pre-set outlier threshold t, if p i If p is greater than t, then the point is an outlier. i If the value is less than t, then the point is considered a non-outlier point; the outlier detection task of the point cloud of the tunneling machine's operating environment is completed using the above classification approach.
7. A two-stage noise reduction method for point cloud data in a tunneling machine operating environment according to claim 1, characterized in that: In step six, a noise reduction module is used to predict the coordinates of noise points in the point cloud of the tunneling machine's operating environment after outlier removal. The predicted coordinates cancel out the noise components, thus achieving point cloud noise reduction. This includes: After outlier detection and removal, the point cloud of the tunneling machine's working environment is sequentially translated, scaled, k-nearest neighbor construction, multilayer perceptron, graph attention mechanism, SA layer, and graph convolution to obtain neighborhood aggregation information. The neighborhood aggregation information is fed into a single-point convolution to reduce the dimensionality of the high-dimensional features. Finally, through prediction, the three-dimensional coordinates of the point cloud of the tunneling machine's operating environment are generated. The predicted coordinates can cancel out the noise components, transforming the point cloud denoising task into a regression task. By regressing the coordinates of the noise points, the denoising of the point cloud in the tunneling machine's operating environment is completed.