Shield muck three-dimensional point cloud segmentation and volume calculation method based on deep learning

Abstract

The invention provides a shield muck three-dimensional point cloud segmentation and volume calculation method based on deep learning, and belongs to the technical field of deep learning and shield, and the method comprises the steps: S1, point cloud data collection, preprocessing and sample labeling; s2, constructing and training a deep learning segmentation model; s3, carrying out muck point cloud segmentation; and S4, calculating the volume of the muck. According to the method, efficient interaction of point cloud attention is realized by adopting four serialization modes, and the influence of point cloud disorder on calculation is avoided.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and shield technology, and in particular to a method for three-dimensional point cloud segmentation and volume calculation of shield slag based on deep learning. Background Art

[0002] Shield tunneling has become a key construction technology for modern urban rail transit due to its advantages such as high operational safety, high construction efficiency, and minimal environmental impact. However, when traversing complex geological environments such as water-rich sand layers and unevenly hard and soft strata, shield over-excavation is very likely to occur, resulting in voids behind the lining wall and ultimately causing ground collapse. Ground collapse directly impacts urban traffic and may even damage existing building structures, resulting in extremely adverse effects on urban construction. Existing shield slag volume measurement methods mainly rely on manual visual inspection of the slag bucket loading level to estimate the slag volume. This method is highly dependent on manual operation and judgment, and has a high degree of subjectivity and uncertainty.

[0003] 3D point cloud data often contains numerous noise and background points, which can introduce errors into volume calculations. High-precision volume calculations require obtaining a pure surface point cloud of the soil, removing both background and noise points. However, soil surfaces are diverse in morphology and lack universal distribution patterns or spatial characteristics. Existing clustering and regression algorithms struggle to accurately segment soil surface and background point clouds.

[0004] Early approaches to point cloud computing involved voxelizing point clouds into a dense grid and applying 3D convolutions to this voxel grid. However, because the number of voxels increases cubically with resolution, this strategy incurs significant computational and memory overhead. Drawing on the success of convolutional neural networks in visual tasks, some point cloud segmentation methods have adopted this approach, projecting 3D point clouds onto different image planes and then extracting features using 2D convolutional neural networks. However, point clouds are disordered and unstructured, making standard convolutional neural networks ineffective. With the significant success of the Transformer in natural language processing, research has discovered that attention mechanisms are uniquely suited to the structure of point cloud data. Self-attention is calculated based on capturing associations between elements in a set, and point clouds are collections of elements with associated positional attributes. However, due to the sheer volume of point cloud data, directly performing global attention computation on these datasets is beyond the processing power of general-purpose computers. Therefore, some methods have successfully improved the performance of point cloud models by converting global attention into local attention. The muck point cloud has both a wide range of scales and a dense point distribution, which requires the content of the attention window to be flexible enough and the interaction between windows to be efficient, so as to expand the receptive field of local attention and maintain the global correlation of point cloud features. However, there are still two challenges in applying existing deep learning point cloud segmentation methods to shield construction. First, construction sites usually do not have large computing servers, and the GPU performance of computers is low, which limits the size of local attention windows. Second, the muck point cloud has both a wide spatial scale and a diverse spatial distribution. This requires the content of the attention window to be rich enough and the attention windows to have efficient interaction capabilities to expand the receptive field of local attention and maintain the global correlation of point cloud features. Summary of the Invention

[0005] The present invention provides a three-dimensional point cloud segmentation and volume calculation method for shield slag based on deep learning, which adopts four serialization methods to achieve efficient interaction of point cloud attention and avoid the influence of point cloud disorder on calculation.

[0006] To achieve the above object, the present invention adopts the following technical solutions: The deep learning-based shield slag 3D point cloud segmentation and volume calculation method includes: S1. Point cloud data acquisition, preprocessing, and sample annotation: Scan a passing muck transport railcar using a line lidar to collect 3D point cloud data, including the muck bucket and the muck surface. Automatically determine and store this 3D point cloud data to obtain initial muck point cloud data. Based on this initial muck point cloud data, manually annotate the muck surface points, background points, and noise points to obtain an annotated muck point cloud sample dataset. S2. Construction and training of a deep learning segmentation model: Construct a deep learning-based soil segmentation model with an encoder-decoder structure. The encoder includes a downsampling module and a Transformer block. The Transformer block includes a normal vector relative position encoding module and a serialized point attention interaction module. The decoder includes an upsampling module and an Upsampleblock. Use the labeled soil point cloud sample dataset to train the deep learning-based soil segmentation model to obtain a trained deep learning-based soil segmentation model. S3. Muck point cloud segmentation: The initial muck point cloud data is input into the trained deep learning-based muck segmentation model. After data initialization, feature extraction, downsampling, and upsampling, the muck surface point cloud data is obtained after removing background points and noise points. S4. Calculate the volume of slag. Using a single slag bucket as the calculation unit, divide the unit cube into units and calculate the volume based on the slag surface point cloud data after removing background points and noise points. Combined with the actual size of the slag bucket, construct a virtual plane to correct the volume result and obtain the actual loading volume of the slag.

[0007] In this specification, the automatic determination and storage processing in S1 is specifically as follows: by comparing the z-axis coordinates of each point in the point cloud with the z-axis coordinates of the top of the lining, it is determined whether it is a rail car loaded with slag, and only the three-dimensional point cloud data corresponding to the rail car loaded with slag is stored.

[0008] In this specification, the manual labeling of slag surface points, background points and noise points in S1 is specifically as follows: based on the point cloud coordinates and spatial distribution characteristics, the point cloud is divided into slag surface points, background points and noise points, and category labels are added to different categories of points.

[0009] In this specification, the encoder-decoder structure in S2 is specifically as follows: the encoder includes 4 downsampling stages, the first two downsampling stages use 2 Transformer blocks each, and the last two downsampling stages use 3 Transformer blocks each, with feature channels of 16, 32, 64, 128, and 256 respectively; the decoder includes 4 upsampling stages, each upsampling stage uses 2 Upsample blocks, and feature channels of 256, 128, 64, and 32 respectively, and finally outputs the segmentation result through the SubMConv3d block module and the fully connected layer.

[0010] In this manual, the specific processing process of the normal vector relative position encoding module in S2 is: receiving the point cloud feature map and the point cloud normal vector, mapping the point cloud normal vector to a feature vector through the MLP layer, fusing it with the point cloud feature map through the cross attention module, and then passing through the 3D sparse convolution, Linear layer, and LayerNorm layer in sequence to form a relative position encoding vector, fusing the normal vector feature and the relative position encoding vector through jump connection, and outputting the processing result.

[0011] In this manual, the specific processing process of the serialized point attention interaction module in S2 is: after the input feature map is adjusted by the Linear layer and the LayerNorm layer, it is serialized through four space filling curves: Z-Index, Hilbert, Spiral, and Cantor. The serialized feature channels are compressed to the original number of channels, and the attention is calculated by the SMHA layer. Then, it passes through the Linear layer, LayerNorm layer, and MLP layer in sequence to output the processing results.

[0012] In this specification, the downsampling module of the encoder adopts a grid downsampling method, specifically: the point cloud space is divided into a three-dimensional grid, and multiple points in the grid are aggregated into a single representative point by averaging, taking the maximum value or randomly selecting, and then processed by the Linear layer, BatchNorm layer, and GELU layer in sequence to achieve downsampling.

[0013] In this specification, the upsampling module of the decoder adopts a grid upsampling method, specifically: after adjusting the channel of the input feature map through the Linear layer, the grid scale is reduced and the number of grids is increased, the feature copies saved in the downsampling stage are introduced, the upsampled features and the feature copies are fused through element-wise addition, and then processed through the BatchNorm layer and the GELU layer to achieve upsampling.

[0014] In this specification, S2 uses the labeled slag point cloud sample dataset to train the slag segmentation model based on deep learning, specifically including: data enhancement of the sample data, data enhancement including random rotation, scaling and adding Gaussian noise; dividing the enhanced sample data into a training set and a validation set, using the cross entropy loss function as the optimization target, and iteratively training until the model converges.

[0015] In this specification, the surface point cloud data of the slag that has been excluding background points and noise points in S4 is divided into unit cubes and the volume is calculated as follows: the area of ​​the unit cube is determined according to the number of points on the slag surface, the height of the unit cube is determined according to the coordinates of the points, the volumes of all unit cubes are counted and summed up, and the initial volume from the surface point cloud of the slag to the constructed bottom surface is obtained; the volume correction result is constructed by constructing a virtual plane in combination with the actual size of the slag bucket, and as follows: a virtual plane lower than the actual bottom surface is constructed according to the actual length, width and volume of the slag bucket, the initial volume and the volume difference between the virtual plane and the actual bottom surface are calculated, and the volume difference is subtracted from the initial volume to obtain the actual loading volume of the slag.

[0016] In summary, the present invention has at least the following beneficial effects: The existing technology does not yet have a point cloud scanning and data processing method for slag transport railcars. The present invention fills this technical gap and greatly improves the accuracy and timeliness of shield slag volume measurement. Compared with manual data analysis, the proposed algorithm increases the data processing speed by 66 times.

[0017] Compared with the general point cloud segmentation algorithm, the segmentation accuracy of the muck point cloud data of the present invention is improved by 13%.

[0018] Compared with commonly used point cloud volume calculation methods, such as triangulation and slicing methods, the proposed special calculation method for slag volume has higher adaptability to scanning data. The average calculation error is reduced by 3.12% compared with the existing methods, and the calculation speed is increased by 167%.

[0019] The NCPE calculation module of the present invention introduces the normal vector of the point cloud through cross attention on the basis of the original relative position encoding calculation method, improves the effect of relative position encoding, enhances the spatial information expression of the point cloud, reduces the spatial information loss of point cloud features in each downsampling stage, and improves the model accuracy.

[0020] The SPAI calculation module of the present invention proposes a point cloud attention calculation method through the fusion of four point cloud serializations. Point cloud serialization avoids the impact of the disorder of point clouds on deep learning. The fusion of four serialization orders solves the problem that a single serialization order cannot fully adapt to the changing soil surface; in addition, the fusion of different serialization orders can also enrich the receptive field of the local attention window, so that the Transformer module can obtain better computing results. Based on this idea, when constructing the attention window, two different construction paths are adopted for the four serialization orders, which further improves the interaction efficiency of the attention window and ultimately achieves significant accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 This is a schematic diagram of the deep learning-based shield slag three-dimensional point cloud segmentation and volume calculation method involved in the present invention.

[0023] Figure 2 It is a schematic diagram of the shield slag volume three-dimensional point cloud scanning method involved in the present invention.

[0024] Figure 3 This is a schematic diagram of the network structure of the soil point cloud segmentation model based on deep learning involved in the present invention.

[0025] Figure 4 Schematic diagram of the NCPE structure of the relative position encoding module combined with the normal vector involved in the present invention.

[0026] Figure 5 Schematic diagram of the structure of the point cloud mesh downsampling calculation module involved in the present invention.

[0027] Figure 6 Schematic diagram of the point cloud mesh downsampling calculation principle involved in the present invention.

[0028] Figure 7 Schematic diagram of the point cloud mesh upsampling calculation module structure involved in the present invention.

[0029] Figure 8 Schematic diagram of the sparse convolution calculation module structure for point cloud involved in the present invention.

[0030] Figure 9 This is a schematic diagram of the serialized point attention interaction calculation module structure involved in the present invention.

[0031] Figure 10 Schematic diagram of the four space-filling curve paths used by the serialized point attention interaction module involved in the present invention.

[0032] Figure 11 This is a schematic diagram of the point cloud data structure change process during calculation of the serialized point attention interaction module involved in the present invention.

[0033] Figure 12 This is a schematic diagram of the visualization results of the SMPN involved in the present invention for the soil surface point cloud segmentation.

[0034] Figure 13 Schematic diagram of the method for calculating the point cloud volume of the slag surface involved in the present invention.

[0035] Figure numerals: 1. Line scanning laser radar; 2. Laser radar mounting bracket; 3. Shield transport railcar; 4. Muck bucket; 5. Muck surface. DETAILED DESCRIPTION

[0036] Hereinafter, only certain exemplary embodiments are briefly described. As will be appreciated by those skilled in the art, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and description are to be regarded as illustrative in nature and not restrictive.

[0037] The disclosure below provides many different embodiments or examples for implementing different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the components and configurations of specific examples are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. In addition, the embodiments of the present invention may repeat reference numerals and / or reference letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or configurations discussed.

[0038] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, this embodiment provides a method for shield muck 3D point cloud segmentation and volume calculation based on deep learning, including: S1. Point cloud data acquisition, preprocessing, and sample annotation: Scan a passing muck transport railcar using a line laser radar to collect three-dimensional point cloud data including the muck bucket 4 and the muck surface 5; automatically determine and store the three-dimensional point cloud data to obtain initial muck point cloud data; based on the initial muck point cloud data, manually annotate the muck surface points, background points, and noise points to obtain an annotated muck point cloud sample dataset; S2. Construction and training of a deep learning segmentation model: Construct a deep learning-based soil segmentation model with an encoder-decoder structure. The encoder includes a downsampling module and a Transformer block. The Transformer block includes a normal vector relative position encoding module and a serialized point attention interaction module. The decoder includes an upsampling module and an Upsampleblock. Use the labeled soil point cloud sample dataset to train the deep learning-based soil segmentation model to obtain a trained deep learning-based soil segmentation model. S3. Muck point cloud segmentation: The initial muck point cloud data is input into the trained deep learning-based muck segmentation model. After data initialization, feature extraction, downsampling, and upsampling, the muck surface point cloud data is obtained after removing background points and noise points. S4. Calculate the volume of the slag; using a single slag bucket 4 as the calculation unit, divide the unit cube into units and calculate the volume based on the slag surface point cloud data after removing background points and noise points. Combined with the actual size of the slag bucket, a virtual plane is constructed to correct the volume result to obtain the actual loading volume of the slag.

[0040] In some embodiments, the automatic determination and storage processing in S1 is specifically as follows: by comparing the z-axis coordinates of each point in the point cloud with the z-axis coordinates of the top of the lining, it is determined whether it is a rail car loaded with slag, and only the three-dimensional point cloud data corresponding to the rail car loaded with slag is stored.

[0041] In some embodiments, the manual labeling of slag surface points, background points and noise points in S1 is specifically as follows: based on the point cloud coordinates and spatial distribution characteristics, the point cloud is divided into slag surface points, background points and noise points, and category labels are added to different categories of points.

[0042] In some embodiments, the encoder-decoder structure in S2 is specifically as follows: the encoder includes 4 downsampling stages, the first two downsampling stages use 2 Transformer blocks each, and the last two downsampling stages use 3 Transformer blocks each, and the feature channels are 16, 32, 64, 128, and 256 respectively; the decoder includes 4 upsampling stages, each upsampling stage uses 2 Upsample blocks, and the feature channels are 256, 128, 64, and 32 respectively, and finally the segmentation result is output through the SubMConv3d block module and the fully connected layer.

[0043] In some embodiments, the specific processing process of the normal vector relative position encoding module in S2 is: receiving the point cloud feature map and the point cloud normal vector, mapping the point cloud normal vector into a feature vector through the MLP layer, fusing it with the point cloud feature map through the cross-attention module, and then passing through the 3D sparse convolution, Linear layer, and LayerNorm layer in sequence to form a relative position encoding vector, fusing the normal vector feature and the relative position encoding vector through jump connection, and outputting the processing result.

[0044] In some embodiments, the specific processing process of the serialized point attention interaction module in S2 is: after the input feature map is adjusted by the Linear layer and the LayerNorm layer, it is serialized through four space filling curves: Z-Index, Hilbert, Spiral, and Cantor, and the serialized feature channels are compressed to the original number of channels. The attention is calculated by the SMHA layer, and then the processing results are output in sequence through the Linear layer, LayerNorm layer, and MLP layer.

[0045] In some embodiments, the downsampling module of the encoder adopts a grid downsampling method, specifically: dividing the point cloud space into a three-dimensional grid, aggregating multiple points in the grid into a single representative point by averaging, taking the maximum value or randomly selecting, and then processing them in sequence through the Linear layer, BatchNorm layer, and GELU layer to achieve downsampling.

[0046] In some embodiments, the upsampling module of the decoder adopts a grid upsampling method, specifically: after adjusting the channel of the input feature map through the Linear layer, the grid scale is reduced and the number of grids is increased, the feature copies saved in the downsampling stage are introduced, the upsampled features and the feature copies are fused through element addition, and then processed through the BatchNorm layer and the GELU layer to achieve upsampling.

[0047] In some embodiments, S2 uses the labeled slag point cloud sample data set to train the deep learning-based slag segmentation model, specifically including: performing data enhancement on the sample data, the data enhancement includes random rotation, scaling and adding Gaussian noise; dividing the enhanced sample data into a training set and a validation set, using the cross entropy loss function as the optimization target, and iteratively training until the model converges.

[0048] In some embodiments, the surface point cloud data of the slag that has background points and noise points removed in S4 is divided into unit cubes and the volume is calculated as follows: the area of ​​the unit cube is determined according to the number of points on the slag surface, the height of the unit cube is determined according to the coordinates of the points, the volumes of all unit cubes are counted and summed up to obtain the initial volume from the surface point cloud of the slag to the constructed bottom surface; the volume result is corrected by constructing a virtual plane in combination with the actual size of the slag bucket as follows: a virtual plane lower than the actual bottom surface is constructed according to the actual length, width and volume of the slag bucket, the initial volume and the volume difference between the virtual plane and the actual bottom surface are calculated, the volume difference is subtracted from the initial volume to obtain the actual loading volume of the slag.

[0049] The technical concept of the present invention is as follows: A deep learning-based method for shield tunneling muck 3D point cloud segmentation and volume calculation enables high-precision, real-time measurement of shield tunneling muck volume. A proposed point cloud scanning method is used to obtain a 3D point cloud of the muck bucket loading on a muck transport railcar. A proposed muck surface point cloud segmentation model is used to remove background and noise points, and a proposed point cloud volume calculation method is used to calculate pure muck surface points. An attention interaction method based on point serialization addresses the limitations of point clouds due to their disordered and sparse local attention receptive fields. To segment the dynamic boundary between the muck surface 5 and the muck bucket 4, a normal vector relative position encoding module is designed to enhance the spatial feature representation of the point cloud. This invention enables automated data collection and processing without disrupting the construction process, ultimately achieving highly accurate muck volume calculation results. This invention improves the accuracy and reliability of shield tunneling control, reduces excavation disturbances, minimizes stratum losses, eliminates the risk of ground collapse, and ensures the safety of shield tunneling and urban construction.

[0050] By scanning a moving train using a line laser radar, a complete 3D point cloud of a shield tunneling slag transport vehicle was acquired. A developed slag point cloud segmentation model was used to remove background points and noise, such as the slag bucket 4 and the invert lining. Volumes were calculated based on the extracted pure slag surface point cloud. A deep learning-based slag segmentation model, called the Shield Muck Point Network (SMPN), was developed. The SMPN effectively combines sparse convolution with the Transformer, improving both computational efficiency and accuracy. Furthermore, the SMPN utilizes four space-filling curves to serialize the point cloud and enhances the representation of spatial position in the point cloud through normalized conditional position encoding (NCPE). This invention incorporates a sequenced point attention interaction (SPAI) technique, which adaptively adjusts the perception range and point distribution by partitioning point clouds with different serialization orders into windows. The interaction range of the attention window is expanded by combining different area paths. Finally, the invention includes a volume calculation method specifically based on slag surface point clouds.

[0051] The model's network structure improves upon the encoder-decoder framework. The encoder primarily consists of four downsampling stages, each consisting of a downsampling module and a Transformer block. The Transformer block consists of an NCPE module and a SPAI. Before downsampling, the model first passes through a Transformer block to generate point cloud features. The first two downsampling stages each use two Transformer blocks to refine features, while the last two downsampling stages each use three Transformer blocks. The encoder's five stages have feature channels of 16, 32, 64, 128, and 256, respectively. The decoder is the reverse of the encoder, but each upsampling stage uses two upsample blocks. Compared to the Transformer block, the upsample block adds an SCB before the NCPE to smooth the increase in spatial information. The feature channels corresponding to each upsampling stage are 256, 128, 64, and 32, respectively. After upsampling, the segmentation result is output through a single SCB and a fully connected layer.

[0052] NCPE is a normal vector relative position encoding module that fuses normal vectors and feature maps to emphasize the representation of the original point cloud spatial information in the point cloud feature map. It consists of an MLP, a cross-attention module, a 3D sparse convolution, a Linear layer, and a LayerNorm layer.

[0053] The downsampling module applies the grid downsampling method, which consists of Grid Pooling, Linear, BatchNorm, and GELU in sequence.

[0054] The upsampling module applies the grid upsampling method, which consists of Linear, Grid Pooling, Linear, element addition, BatchNorm, and GELU in sequence.

[0055] The SPAI module is used to form a serialized point cloud and calculate the point cloud attention of the four serialized order fusions. This module is the core computational module of the point cloud segmentation method of this invention. This module is composed of Linear, LayerNorm, Serialization, Linear, SMHA, Linear, Element-wise Addition, LayerNorm, and MLP. SMHA is the serialized multi-head attention layer. This module uses two different attention windows to divide the path based on the characteristics of the serialized point cloud, enriching the interaction range of the local attention window.

[0056] The soil point cloud volume calculation method uses a single soil bucket 4 as the calculation unit, dividing the space between the soil surface 5 and the bottom into a large number of unit cubes. The number and area of ​​the unit cubes are determined by the number of scan points on the soil surface 5, and the height of the unit cubes is determined by the coordinates of the scan points. The volume enclosed by the soil surface 5 and the soil bucket 4 is obtained by summing all the unit cubes. Considering that the actual bottom surface of the soil bucket is an irregular plane, a virtual plane is constructed based on the actual volume of the soil bucket to facilitate program execution. The specific calculation method is shown in the following formula.

[0057] ; ; ; in the formula Indicates the volume of the muck calculated from the point cloud of the muck surface; and They represent the Z coordinate values ​​of the soil surface point and the construction plane respectively; represents the area of ​​a unit cube; Represents the volume difference between the construction plane and the true plane; and Respectively represent the actual length and width of the hopper; Indicates the total amount of soil surface points; Indicates the actual loading volume of the hopper; Represents the resulting construction plane.

[0058] The specific plan is as follows: 1. Muck point cloud collection: like Figure 2 As shown, a laser radar mounting bracket 2 is installed at the top of the shield tunnel, and a line-scanning laser radar 1 is mounted below the bracket, scanning vertically downwards any objects passing through the radar. During shield tunneling, a shield transport railcar 3 travels back and forth between the shield machine and the tunnel entrance. When the shield transport railcar 3 passes the line-scanning laser radar 1, it acquires a point cloud of the railcar's upper surface, including the slag bucket 4 and the slag surface 5. The laser radar is connected to the host computer in the construction monitoring room via optical fiber, and the scanned data is transmitted to the host computer in real time. The host computer automatically determines whether to store the returned slag point cloud data. Specifically, the radar immediately begins storing data when slag transport passes by, while no radar data is stored when a railcar without slag passes by. The collected slag point cloud data is automatically analyzed using the invented point cloud segmentation and volume calculation algorithm, and the final output volume result is uploaded to the subway company's monitoring platform. The criteria for determining whether a railcar has passed by the laser radar are as follows: ; In the formula Represents the z-axis coordinate of each point in the point cloud; Indicates the z-axis coordinate of the top of the lining in the point cloud; Indicates the removal point operation; and Respectively represent the total number of contour points scanned by the radar each time and the number of points eliminated; It is a preset threshold value, which is determined by the height distance from the front of the railcar to the radar.

[0059] 2. Muck point cloud segmentation: See Figure 3 The input model's point cloud data first passes through the data initialization module, which generates six cached data based on the input data: point cloud coordinates, point cloud mesh, point cloud normal vector, four point cloud serialization paths, point cloud features, and pooled cache results. The point cloud serialization path is formed by four space-filling curves, which can be, but are not limited to, Z-Index, Hilbert, Spiral, and Cantor. Figure 3 Point initializing means point cloud initialization operation; SCB means sparse convolution calculation module. The calculation process is shown in Figure 8 NCPE represents the relative position encoding module of the invention. The detailed calculation process is shown in Figure 4 ; SPAI represents the invented serialized point attention interaction module, and its calculation process is shown in Figure 9 ; Downsampling represents the grid downsampling module, and its calculation process is shown in Figure 5 Linear represents a fully connected layer, a common computational method in deep learning. Parent feat represents the original feature parameters cached during the downsampling process. It is fused with the feature map during the model's upsampling process to improve the original spatial information expression of the upsampled point cloud feature map.

[0060] The initialized point cloud passes through the SCB (SubMConv3d block) module to initially form point cloud features. SCB consists of 3D sparse convolution, BatchNorm (data regularization), and GELU (activation function).

[0061] The point cloud features are then fed into the first Transformer block, which includes the invented NCPE and SPAI modules. The results then pass through the downsampling module, which reduces the grid size of the point cloud features, achieving downsampling. Before downsampling, the feature maps generated by the Transformer block are stored in a pooling buffer vector.

[0062] After the first downsampling, the point cloud features are gradually condensed through two consecutive Transformer blocks. A second downsampling is then performed, and the pre-downsampling feature results are stored in a pooled cache vector.

[0063] After the second downsampling, the point cloud features are again passed through two consecutive Transformer blocks, gradually condensing the point cloud features. Then, a third downsampling is performed, and the pre-downsampling feature results are retained in the pooled cache vector.

[0064] After the third downsampling, the point cloud features pass through three consecutive Transformer blocks. As the number of downsampling stages increases, the number of feature channels in the point cloud increases, requiring more Transformer blocks to extract and refine the features. Finally, the fourth downsampling occurs, and the pre-downsampling feature results are retained in the pooled cache vector. After the fourth downsampling, the model passes through three consecutive Transformer blocks again. At this point, the feature encoding stage of the model is complete, forming rich semantic features of the point cloud.

[0065] The downsampled feature extraction results are fed into the upsampling decoding network. The point cloud features are first fed into the first upsampling module and then pass through two consecutive upsample blocks. The upsample block consists of SCB, NCPE, and SPAI, in that order. Compared to the downsampling stage, the upsampling process adds an additional SCB before each Transformer module to better restore the spatial information of the point cloud. After the two consecutive upsample blocks, the feature map enters the next upsampling stage. Before the next upsampling stage, the feature map is fused with the pooled cache result of the same spatial resolution.

[0066] The process of the four upsampling stages is the same. After four upsamplings in sequence, the spatial resolution of the feature map will be consistent with the input point cloud. Finally, the point cloud features are converted into point cloud segmentation results through one SCB and one Linear.

[0067] See Figure 4NCPE accepts two inputs: a point cloud feature map and a point cloud normal vector. The point cloud normal vector first passes through an MLP layer and is then fused with the feature map through a cross-attention module. The MLP layer is used to map the normal vector to a feature vector and unify the feature channel of the normal vector with Feat. After adding the normal vector, the feature map passes through the SubMConv3d layer, the Linear layer, and the LayerNorm layer in sequence. The three layers form a relative position encoding vector based on the feature information. Finally, the normal vector feature fusion result is element-wise added to the relative position encoding vector through a skip connection, and the NCPE calculation result is output. It should be noted that skip connections should be used after the fusion of the normal vector and the feature, that is, after Cross-Attention, because skip connections allow skipping the calculation layers within the link. However, the operation of fusing the normal vector in the feature map is necessary to supplement the original coordinate information for the subsequent SPAI calculation. Figure 4 In it, Normal represents the normal vector input; Feat represents the feature map input; Cross-Attention represents the cross-attention calculation module, which is an improved module of the general Multi-Head Attention and is used to calculate the attention results of different input sources; SubMconv3D represents the 3D sparse convolution calculation; Linear represents the fully connected layer, which is a commonly used calculation method in deep learning; LayerNorm represents layer normalization, which is a technology used in deep learning to reduce internal covariate shift, and improve the model convergence speed and stability by stabilizing the neural network training process; ⊕ represents element addition.

[0068] The downsampling in SMPN adopts grid downsampling, and the calculation module refers to Figure 5 The downsampled feature maps are sequentially passed through the Linear, BatchNorm, and GELU layers. Linear transformations are first used to transform and combine feature dimensions. Batch normalization then adjusts the transformed feature distribution to a stable state (avoiding it from falling into the saturation region of the activation function, such as at either end of a Sigmoid function, which would cause gradient vanishing) while also accelerating training. Finally, the activation function introduces nonlinearity to enhance the model's expressiveness. Figure 5 In [1], Feat represents feature map input; Grid Pooling represents grid pooling, which divides the point cloud into a regular grid and generates high-order features of the local area by aggregating the point cloud features within the grid; BatchNorm represents batch normalization, which is a key technology for accelerating training and improving stability in deep learning, and mainly solves the "internal covariate shift" problem in deep network training; GELU is an activation function with smooth characteristics.

[0069] Grid pooling method see Figure 6, grid pooling divides the point cloud space into several three-dimensional grids. When there are multiple points in the grid, a single point is used to represent all the points in the grid to achieve downsampling. Multiple points in the grid can be aggregated by averaging, taking the maximum value, random selection, etc. By increasing the grid size, downsampling can aggregate points in a larger range. Grid pooling makes the number of spatial points no longer reduced by a fixed multiple or pattern, but is determined by the distribution of the points, thereby achieving more efficient data dimensionality reduction and feature fusion, while retaining key geometric structure information and enhancing the model's adaptability to complex spatial distributions. In addition, the grid size is a hyperparameter that can be set according to the actual distribution of the points to reasonably control the downsampling ratio of the samples. Figure 6 In , Grid size indicates the grid size; Down indicates grid downsampling, and Up indicates grid upsampling.

[0070] For the calculation process of the upsampling calculation module, refer to Figure 7 Before upsampling the grid, the feature map first passes through a Linear layer to adjust the output feature channels. The grid scale is then reduced and the number of grid cells is increased. A Linear layer then passes through the layer to make the added points more consistent with the feature representation. At this point, feature copies saved during the downsampling process are introduced and fused with the upsampled features through element-wise addition. The feature copies pass through a Linear layer before fusion, allowing the model to fine-tune the feature copies during training to achieve better feature fusion. Figure 7 In the figure, Linear represents the fully connected layer, which is a commonly used calculation method in deep learning; Grid Pooling represents grid pooling, which divides the point cloud into a regular grid and generates high-order features of the local area by aggregating the point cloud features within the grid; BatchNorm represents batch normalization; GELU is an activation function with smoothing characteristics; ⊕ represents element-wise addition.

[0071] The output of the upsampling module is input into the spatial convolution block (SCB), see Figure 8 This module consists of three sequentially connected submodules: SubMConv3d, BatchNorm, and GELU. The input data first passes through the SubMConv3d layer, which performs a convolution operation on the original point cloud in three dimensions to extract local features. Because it is a submanifold convolution, it does not change the spatial distribution of the point cloud (i.e., the position and number of points remain unchanged). BatchNorm and GELU are then used to achieve normalization and nonlinear transformations. Figure 8 In it, SubMconv3D represents 3D sparse convolution calculation; BatchNorm represents batch normalization; GELU is an activation function with smooth characteristics.

[0072] The core calculation module of the point cloud segmentation model SMPN proposed in this invention is SPAI. The calculation process is shown in Figure 9 The input feature map first passes through the Linear layer and LayerNorm, the purpose is to allow the feature map before serialization to be adaptively adjusted. Then, the point cloud is given different arrangement orders through four space filling curves: Z-Index, Hilbert, Spiral, and Cantor. The connection order of the four paths can be found in Figure 10 . After four different sorting methods, the point cloud feature channels are expanded to four times the original number. At this time, a linear layer is used to restore the number of channels to the input channel to avoid excessive expansion of parameters during attention calculation. Subsequently, the feature map participates in the SMHA calculation. SMHA is the proposed serialized multi-head attention module, which is characterized by using different paths to form attention windows. After the feature map is input into SMHA, the number of channels is expanded to three times the original number through the linear layer to form the q, k, v matrix. The q, k, v matrix is ​​further divided into q1, q2, k1, k2, v1, v2 matrices, and the number of channels of each matrix is ​​one-sixth of the input channel. q1, q2, k1, k2, v1, v2 are divided into two groups, and the local attention windows are divided by different paths. Then q1 and q2, k1 and k2, v1 and v2 are spliced ​​on the channel to form q, k, v matrices, thereby realizing attention calculation. After the attention calculation is completed, the feature map is input into a Linear layer again to strengthen the expression of the attention result, and then fused with the original input features on the jump connection through element addition. Finally, the result is input into the LayerNorm and MLP layers to achieve feature standardization and enhance nonlinear expression capabilities. Figure 9 In this example, Linear represents a fully connected layer, a common computational method in deep learning. LayerNorm represents layer normalization. Linear represents a fully connected layer, a common computational method in deep learning. Z-Index, Hilbert, Spiral, and Cantor are four space-filling curves used. Serialization Multi-Head Attention (SMHA) represents multi-head serialized attention computation for point clouds. q1, q2, k1, k2, v1, and v2 are the two sets of attention matrices calculated by SMHA. Patch path represents different attention window construction paths. Attention represents the attention computation module, the standard attention computation unit in the Transformer.

[0073] The morphological changes of the feature map tensor in SPAI refer to Figure 11, the point cloud feature map is mapped into 4 different serialization orders and stacked on the feature channel. The feature map is evenly divided into two parts according to the format of the 4 serialization orders, and the attention window is divided into different paths. The first method is to traverse the first part of the feature map horizontally. At this time, the attention windows contain point clouds in different orders. The second method is to traverse the second part of the feature map vertically. At this time, each attention window will contain point clouds in 4 different orders. Finally, attention calculation is performed on all windows to enrich the interaction results of the attention windows.

[0074] The result of soil point cloud segmentation of the present invention is as follows Figure 12 As shown. From the overall sample, the model has fully learned the basis for dividing the slag surface 5 and the slag bucket 4, and the segmentation effect is as expected. The model has a clear definition of the boundaries of the slag surface 5 and the slag bucket 4, and the side walls and bottom of the box are also clearly divided. Although the contours of the contact between the slag and the side wall are relatively smooth (the two surfaces are not directly tangent, but gradually transition), the model can still clearly distinguish between the two classifications. For the case where the speed fluctuation of the slag transport railcar causes uneven distribution of samples along the X-axis, SMPN has sufficient generalization ability to cope with changes in sample density or spatial distribution. The results show that SMPN is close to the effect of manual segmentation in slag point segmentation and has significant engineering value.

[0075] The calculation method of point cloud volume is shown in Figure 13 Because the bucket partition blocks the laser scanning, there is a discontinuity in the middle of the soil surface point. Before performing volume calculation, the discontinuity needs to be filled. According to the X coordinate distribution of the point cloud, it is judged that the X coordinate discontinuity is greater than L x The coordinates of the points on both sides are used to further detect the discontinuity length L d , extract the X coordinates on both sides of the discontinuity as L d When calculating the volume, first count the number of input points and calculate the area of ​​the rectangular unit represented by each point based on the actual size of the bucket truck. Then, construct the virtual bottom surface of the slag bucket 4, that is, the actual volume of the slag bucket 4 Add some volume on the basis of , and then divided by the actual , obtain a plane lower than the real bottom surface, record the plane Axis coordinates. Let the Z-axis coordinate of each soil surface point minus Get the distance from each point to the bottom of the structure, sum the distances from all points to the bottom of the structure and multiply it by the area of ​​the rectangular unit , get the volume from the soil point cloud to the bottom of the structure. Finally, subtract the result Get the actual slag volume of each bucket.

[0076] The comparison results of the proposed SMPN and other point cloud models are shown in Table 1. Compared with models with fewer parameters such as PointNet++ and PTv2, the segmentation accuracy of SMPN has been greatly improved, with its mAcc, mIoU, and oAcc increased by 13.6%, 6.5%, and 6.6%, respectively. Compared with large-scale models such as advanced Swin3D and Octformer, the proposed SMPN still has an accuracy advantage under the condition of shortened inference time. The mAcc, mIoU, and oAcc of SMPN are improved by 1.9%, 0.7%, and 0.4% compared with Swin3D, respectively. In Table 1, Methods represents the model or calculation method used for comparison; Para (M) represents the number of model parameters, and the unit is 10 6 mAcc, mIoU, and oAcc represent the three evaluation methods of the model; Muck (IoU), Box (IoU), and Rail (IoU) represent the IoU results of the three categories of muck surface, muck bucket, and track, respectively; Training time (h) represents the training time of the model, in hours; Inference time (s) represents the inference time of the model, in seconds.

[0077] Table 1. Comparison results of SMPN and other point cloud models Methods Para (M) mAcc mIoU oAcc Muck (IoU) Box (IoU) Rail (IoU) Training time (h) Inference time (s) PointNet++ 3.5 0.778 0.791 0.820 0.756 0.813 0.804 8.2 8.8 MinkUNet 37.9 0.851 0.841 0.860 0.834 0.822 0.867 28.5 36.4 PTv2 12.8 0.913 0.906 0.917 0.898 0.921 0.899 20.7 15.6 OctFormer 44.0 0.952 0.947 0.962 0.941 0.944 0.956 33.4 38.6 Swin3D 71.1 0.968 0.964 0.979 0.946 0.965 0.981 34.8 54.7 SMPN 34.46 0.987 0.971 0.983 0.974 0.968 0.975 26.8 32.5 Table 2. Test results of four sequential fusion attention methods

[0078] Table 2 shows the test results of the four proposed serialization-fusion attention methods. The experiments first tested the effectiveness of four space-filling curves based on the multi-head self-attention mechanism (MHA). The results show that serializing point clouds improves the Transformer's computational efficiency. This is reflected in the improvements of mAcc, mIoU, and oAcc by 4.7%, 6.4%, and 5.9%, respectively, compared to the model without the serialization path. However, a fixed serialization order leads to a decrease in the stability of segmentation accuracy, as reflected by an increase in the interquartile range (IQR). Due to the diversity of soil surface boundaries, any given serialization order is unlikely to be fully applicable to all samples. To address these issues, our SPAI implements attention calculations for point clouds with different serialization orders. The results show that allowing the model to selectively use different spatial curves effectively improves both accuracy and stability. For example, using both Z-Index and Hilbert ordering improved mAcc, mIoU, and oAcc by 6.1%, 5.0%, and 4.8%, respectively, compared to using only Z-Index ordering. However, no optimal curve combination was found after testing different curves, and the four test results for the two-curve combination were very close, indicating that the fusion of the two curves is not the upper limit of the model. These experimental results show that order diversity enhances the adaptability and interactivity of the attention receptive field and improves model performance. Based on these findings, a diverse patch path was introduced to enhance the variation of point cloud ordering, achieving the best experimental performance in benchmark evaluations. The final implementation of SPAI further improved mAcc, mIoU, and oAcc by 3.6%, 2.5%, and 2.7%, compared to using Z-Index and Hilbert ordering alone. It should be noted that if the number of sequential orders is further increased, the feature representation of a single order may be incomplete due to channel limitations. Therefore, under the current SMPN parameters, the number of sequential orders was selected to be four. The implementation of the patch path is similar to adding attention heads, requiring consideration of both the number of feature channels and the patch size. Both paths have reached the upper limit of existing devices. In Table 2, "Attention method" indicates the attention calculation method used by the test model; "Serialization" indicates the point cloud spatial serialization path used by the test model; "Four-curve merge (only path-1)" indicates the experimental results of using four paths for fusion but only one serialization path; "Four-curve merge (only path-1)" indicates the experimental results of using four paths for fusion and two serialization paths; "mAcc", "mIoU", "oAcc", and "IQR" indicate the evaluation metrics of the test model.

[0079] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Therefore, changes in illustrative values ​​or substitutions of equivalent components should still fall within the scope of the present invention.

[0080] From the above detailed description, it will be clear to those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0081] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as covering the preferred embodiments and all changes and modifications that fall within the scope of the invention. The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

[0082] It should be noted that the above description of the relevant processes is for illustration and purpose only and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the processes under the guidance of this specification. However, such modifications and changes are still within the scope of this specification.

[0083] The basic concepts have been described above. It will be apparent to those skilled in the art after reading this application that the above disclosures are merely illustrative and do not constitute limitations on this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and amendments to this application. Such modifications, improvements, and amendments are suggested in this application and remain within the spirit and scope of the exemplary embodiments of this application.

[0084] At the same time, this application uses specific terms to describe the embodiments of this application. For example, "one embodiment," "an embodiment," and / or "some embodiments" refer to a certain feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "one embodiment," "an embodiment," or "an alternative embodiment" mentioned twice or more in different places in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application may be appropriately combined.

[0085] Furthermore, those skilled in the art will appreciate that various aspects of the present application may be illustrated and described in terms of a number of patentable categories or situations, including any new and useful process, machine, product, or combination of substances, or any new and useful improvement thereof. Thus, various aspects of the present application may be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software. Each of the above hardware and software may be referred to as a "unit," "module," or "system." Furthermore, various aspects of the present application may take the form of a computer program product embodied in one or more computer-readable media, with computer-readable program code embodied therein.

[0086] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages ​​such as C programming language, Visual Basic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages ​​such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or as a standalone software package on the user's computer, or partly on the user's computer and partly on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0087] In addition, unless expressly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some embodiments of the invention that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0088] Similarly, it should be noted that in order to simplify the presentation of this disclosure and thereby facilitate understanding of one or more of the invention's embodiments, the foregoing descriptions of the embodiments of this disclosure sometimes combine multiple features into a single embodiment, figure, or description thereof. However, this approach should not be interpreted as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject matter of the invention may possess fewer features than the single embodiment described above.

Claims

1. A method for 3D point cloud segmentation and volume calculation of shield slag based on deep learning, characterized by: include: S1. Point cloud data collection, preprocessing and sample annotation; Using a line laser radar to scan a passing muck transport railcar, three-dimensional point cloud data, including the muck bucket and the surface of the muck, is collected. The three-dimensional point cloud data is automatically judged, stored, and processed to obtain initial muck point cloud data. Based on the initial muck point cloud data, the muck surface points, background points, and noise points are manually annotated to obtain an annotated muck point cloud sample dataset. S2. Construction and training of a deep learning segmentation model: Construct a deep learning-based soil segmentation model with an encoder-decoder structure. The encoder includes a downsampling module and a Transformer block. The Transformer block includes a normal vector relative position encoding module and a serialized point attention interaction module. The decoder includes an upsampling module and an Upsampleblock. Use the labeled soil point cloud sample dataset to train the deep learning-based soil segmentation model to obtain a trained deep learning-based soil segmentation model. S3. Muck point cloud segmentation: The initial muck point cloud data is input into the trained deep learning-based muck segmentation model. After data initialization, feature extraction, downsampling, and upsampling, the muck surface point cloud data is obtained after removing background points and noise points. S4. Calculate the volume of slag. Using a single slag bucket as the calculation unit, divide the unit cube into units and calculate the volume based on the slag surface point cloud data after removing background points and noise points. Combined with the actual size of the slag bucket, construct a virtual plane to correct the volume result and obtain the actual loading volume of the slag.

2. The method for shield slag 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: The automatic determination and storage processing in S1 is specifically as follows: by comparing the z-axis coordinates of each point in the point cloud with the z-axis coordinates of the lining top, it is determined whether it is a rail car loaded with slag, and only the three-dimensional point cloud data corresponding to the rail car loaded with slag is stored.

3. The method for shield slag 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: The manual labeling of slag surface points, background points and noise points in S1 is as follows: based on the point cloud coordinates and spatial distribution characteristics, the point cloud is divided into slag surface points, background points and noise points, and category labels are added to different categories of points.

4. The method for shield slag 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: The encoder-decoder structure in S2 is as follows: the encoder contains 4 downsampling stages, the first two downsampling stages use 2 Transformer blocks each, and the last two downsampling stages use 3 Transformer blocks each, with feature channels of 16, 32, 64, 128, and 256, respectively; the decoder contains 4 upsampling stages, each upsampling stage uses 2 Upsample blocks, with feature channels of 256, 128, 64, and 32, respectively. Finally, the segmentation result is output through the SubMConv3d block module and the fully connected layer.

5. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: The specific processing process of the normal vector relative position encoding module in S2 is: receiving the point cloud feature map and the point cloud normal vector, mapping the point cloud normal vector to a feature vector through the MLP layer, fusing it with the point cloud feature map through the cross attention module, and then passing through the 3D sparse convolution, Linear layer, and LayerNorm layer in sequence to form a relative position encoding vector, fusing the normal vector feature and the relative position encoding vector through jump connection, and outputting the processing result.

6. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: The specific processing process of the serialized point attention interaction module in S2 is as follows: after the input feature map is adjusted by the Linear layer and the LayerNorm layer, it is serialized through four space filling curves: Z-Index, Hilbert, Spiral, and Cantor. The serialized feature channels are compressed to the original number of channels, and the attention is calculated by the SMHA layer. Then, it passes through the Linear layer, the LayerNorm layer, and the MLP layer in sequence to output the processing results.

7. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 4 is characterized in that: The downsampling module of the encoder adopts a grid downsampling method. Specifically, the point cloud space is divided into a three-dimensional grid, and multiple points in the grid are aggregated into a single representative point by averaging, taking the maximum value, or randomly selecting. The point is then processed through the Linear layer, BatchNorm layer, and GELU layer in sequence to achieve downsampling.

8. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 4 is characterized in that: The upsampling module of the decoder adopts the grid upsampling method. Specifically, after adjusting the channel of the input feature map through the Linear layer, the grid scale is reduced and the number of grids is increased. The feature copies saved in the downsampling stage are introduced, and the upsampled features and feature copies are fused through element-wise addition. The upsampling is then achieved through the BatchNorm layer and the GELU layer.

9. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 1 is characterized in that: In S2, the labeled slag point cloud sample dataset is used to train the slag segmentation model based on deep learning, specifically including: data enhancement of the sample data, which includes random rotation, scaling and adding Gaussian noise; dividing the enhanced sample data into training set and validation set, using the cross entropy loss function as the optimization target, and iterative training until the model converges.

10. The method for shield tunneling muck 3D point cloud segmentation and volume calculation based on deep learning according to claim 1, characterized in that: In S4, the surface point cloud data of the slag soil after removing background points and noise points are divided into unit cubes and the volume is calculated as follows: the area of ​​the unit cube is determined according to the number of surface points of the slag soil, the height of the unit cube is determined according to the coordinates of the points, the volumes of all unit cubes are counted and summed up, and the initial volume from the surface point cloud of the slag soil to the constructed bottom surface is obtained; the volume correction result is constructed by constructing a virtual plane based on the actual size of the slag bucket, and the specific results are as follows: a virtual plane lower than the actual bottom surface is constructed according to the actual length, width and volume of the slag bucket, the initial volume and the volume difference between the virtual plane and the actual bottom surface are calculated, and the volume difference is subtracted from the initial volume to obtain the actual loading volume of the slag soil.

Images