Self-supervised cross-modal 3D printing concrete pore structure point cloud segmentation method

Abstract

The invention discloses a point cloud segmentation method for a self-supervised cross-modal 3D printing concrete pore structure. The method comprises the following steps of: 1, acquiring and processing point cloud data; 2, a 3D point cloud data set of the 3D printing concrete surface pore structure is manufactured; step 3, obtaining point cloud key points, intra-group feature detail information and local embedding information; step 4, fusing two-dimensional image texture semantic information carried by the ViT model trained by using large-scale two-dimensional image data and depth information of the three-dimensional point cloud data set; step 5, obtaining a pre-training weight through a pre-training stage of the self-supervised cross-modal 3D printing concrete pore structure point cloud segmentation method; step 6, passing through a downstream segmentation model based on Mama and octree sorting; precise segmentation of the surface pore structure of the 3D printing concrete is completed; according to the method, the requirement for intelligent quality detection of the 3D printing concrete pore structure is met, and more accurate and precise segmentation is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual segmentation of 3D printed concrete surface pore structure, in particular to a self-supervised cross-modal 3D printed concrete pore structure point cloud segmentation method. BACKGROUND

[0002] 3D printed concrete technology is one of the important measures to promote green and low-carbon buildings, and has developed rapidly. At present, there are many successful cases of applying 3D printed concrete technology to the construction industry. However, 3D printed concrete is prone to structural defects such as interlayer and surface pores due to the accumulation effect, and is affected by printing materials, environment and other conditions, and existing research shows that pore structure is easy to cause discontinuity of 3D printed concrete in macro and micro aspects, therefore, how to effectively identify and detect pore structure to provide accurate precursor conditions for 3D printed concrete mechanical property research is very important, and has practical research value and engineering application significance.

[0003] An existing method for detecting 3D printed concrete interface pores (publication number: CN114638813A) uses genetic algorithm and threshold segmentation method, which is an intelligent method for quality detection of 3D printed concrete pore structure identification.

[0004] However, this method still belongs to the traditional image processing category, and the genetic algorithm iterative calculation is time-consuming, the application scene is limited, and it is greatly affected by the threshold value, which cannot meet the real demand of intelligent quality detection. SUMMARY

[0005] In order to overcome the defects of the above existing technology, the purpose of the present application is to provide a self-supervised cross-modal 3D printed concrete pore structure point cloud segmentation method, which meets the requirements of intelligent quality detection of 3D printed concrete pore structure and promotes more accurate and precise segmentation on the basis of fusing the complementary advantages of 2D and 3D.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is:

[0007] A self-supervised cross-modal 3D printed concrete pore structure point cloud segmentation method, comprising the following steps:

[0008] Step 1, using a laser scanner to collect point cloud data of 3D printed concrete components, and performing data preprocessing;

[0009] Step 2, for the preprocessed data of step 1, a special 3D printed concrete surface pore structure three-dimensional point cloud data set is made;

[0010] Step 3, build a self-supervised teacher-student model framework by building a point cloud grouping and point cloud embedding module, to obtain key point information, group feature detail information and local embedding information of the point cloud;

[0011] Step 4, based on step 3, build a self-supervised pre-training teacher model in the self-supervised teacher-student model framework, to integrate the advantages of two-dimensional image texture semantic information carried by the ViT model trained using large-scale two-dimensional image data and the depth information advantage of the three-dimensional point cloud dataset created in step 2;

[0012] Step 5, build a self-supervised Mamba-based mask autoencoder student model in the self-supervised teacher-student model framework, cooperate with the pre-training teacher model in step 4, and go through the pre-training phase of the self-supervised cross-modal 3D printed concrete pore structure point cloud segmentation method SCTramba; obtain pre-training weights;

[0013] Step 6, after obtaining the pre-training weights in the pre-training phase, use the downstream segmentation model based on Mamba and octree sorting; optimize the unordered point cloud, and complete the accurate segmentation of the 3D printed concrete surface pore structure;

[0014] Step 7, use the training set of the self-built 3D printed concrete surface pore structure three-dimensional point cloud dataset in step 2 to train the downstream segmentation model based on Mamba and octree sorting in step 6;

[0015] Step 8, use the three-dimensional point cloud data in step 2 as the validation set and test set to verify and predict the trained downstream segmentation model, and use evaluation indicators to evaluate the model performance.

[0016] The step 1 includes: using a 3D scanning device to scan and collect point cloud data of a 3D printed concrete component; after collecting the surface pore structure, remove the noise points and invalid point clouds, and save the file.

[0017] The step 2 includes: importing the collected point cloud data in step 1 using CloudCompare software, performing data labeling, specifically labeling 3D printed concrete pores and non-pore parts, labeling the pore part as 1 and the non-pore part as 0, and dividing the obtained 3D printed concrete surface pore structure three-dimensional point cloud dataset into three categories: training set train, validation set val and test set test according to 1:1:8. The train is mainly used to train the model, the val is used to verify the model, and the test is mainly used to test the model.

[0018] The step 3 comprises a point cloud grouping and point cloud embedding strategy, adopts a farthest point sampling (FPS) algorithm to obtain key points, and uses a nearest neighbor (KNN) algorithm to find K nearest neighbors for each key point, constructs a point cloud embedding module through an encoder (Encoder) and a graph convolution network (DGCNN) and maximum pooling for local features in a group of point clouds, and obtains local block features;

[0019] The Encoder has two layers, and the DGCNN has five layers, and the features extracted by the first four layers are spliced and aggregated together and sent into the fifth layer DGCNN to avoid loss of low-dimensional details.

[0020] Given a point cloud P = {p i |i = 1, 2,..., N} ∈ R N×3 , a farthest point sampling (FPS) algorithm is adopted to obtain G key points, and a nearest neighbor (KNN) algorithm is used to find K nearest neighbors for each key point, which is defined as

[0021]

[0022] The subscript g refers to the gth key point, and g k represents the kth point in the group, and a point cloud grouping module is constructed.

[0023] The local features N g of the point cloud P in the group are constructed through an encoder (Encoder) and a graph convolution network (DGCNN) and maximum pooling to obtain local block features.

[0024] The step 4 comprises using a ViT to train a model of a large-scale two-dimensional traditional image data set as a traditional two-dimensional image texture information input, using a three-dimensional point cloud data as a three-dimensional depth information input, using a Transformer as a pre-training teacher model backbone, and constructing a codebook as a shared semantic space of two-dimensional and three-dimensional information.

[0025] The Transformer encoding process is represented as:

[0026] h l = MLP (LN (MSA (LN (h l-1 ))+ h l-1 )), l = 1,..., L.

[0027] Wherein, MLP represents a multi-layer perception module, LN represents a normalization module, MSA represents a multi-head self-attention mechanism, h represents feature information, l represents the number of model layers, and there are L layers in total.

[0028] The step 5 includes, taking Mamba as the core, proposing to generate a mask matrix by dynamically masking the point cloud token by 80% high proportion random mask, taking Mamba as the encoder structure, taking the pre-training teacher model of step 4 as the target feature, and constantly optimizing the prediction feature result of the decoder in the autoencoder through the teacher model prediction result through the loss function;

[0029] Wherein, the teacher model is represented as F T , the output latent feature Z T = F T (X), the student model is represented as F S , the output prediction feature Z S = F S (X masked ), the cosine similarity is taken as the loss function, and the prediction result of the student model is optimized; wherein, X represents the local feature, X masked represents the mask matrix, and the loss function is represented as:

[0030]

[0031] The step 6 includes proposing a downstream segmentation model based on Mamba and octree sorting, taking Mamba as the backbone network to capture global features, and introducing an octree sorting mechanism to obtain causal relationships in order to solve the unordered nature of point clouds. The sorted point cloud features are converted into one-dimensional sequences Z, which are sent into N*Mamba encoder to extract features; Mamba is composed of a normalization layer, a selective SSM state space model, a deep convolution, and a residual connection.

[0032] Specifically, an octree structure is first constructed, and the same depth octree nodes are sorted by shuffled keys, wherein x i ,y i ,z i represent the coordinates of the i-th point of the octree node, d is the depth of the octree, and the value of the key corresponds to the 3D z-order curve position, which is represented as:

[0033] Key(χ,z,d)=x0y0z0x1y1z1…x d-1 y d-1 z d-1

[0034] The sorted point cloud features are converted into one-dimensional sequences Z, which are sent into N*Mamba encoder to extract features; Mamba is composed of a normalization layer, a selective SSM, a deep convolution, and a residual connection, and global features are obtained through average pooling and maximum pooling after Mamba encoding, and finally a task segmentation head is used to complete the segmentation task of the pores.

[0035] The step 7 comprises training the segmentation network based on Mamba and octree sorting constructed in step 6 using the training set, and obtaining an optimal weight file saving important parameters and weight information of the model when the training is completed.

[0036] The evaluation index related to the step 8 specifically comprises:

[0037] Accuracy (Accuary): the proportion of the number of correct predictions in the total number of positive and negative examples;

[0038]

[0039] Wherein TP (true positive) represents a pore point correctly identified as a pore, TN (true negative) represents a background point correctly classified as non-pore, FP (false positive) refers to a background point incorrectly classified as a pore, and FN (false negative) is a pore point incorrectly labeled as background;

[0040] Mean Intersection over Union (MIoU): the average value of the similarity between the predicted segmentation result and the true segmentation result.

[0041]

[0042] Wherein, C = 2.

[0043] F1 score (F1): the F1 score is the harmonic mean of the prediction rate and the recall rate.

[0044]

[0045] The beneficial effects of the present application are:

[0046] The present application provides a 3D printing concrete pore structure point cloud segmentation method based on self-supervised cross-modal, aiming to make up for the difficulty of point cloud acquisition, and at the same time, to play the advantages of two-dimensional image texture information and three-dimensional point cloud spatial depth information, to solve the problems of difficult pore structure feature recognition, difficult point cloud acquisition and high labor cost of data annotation under the complex background of 3D printing concrete surface by cross-modal idea.

[0047] The application constructs a pre-training-downstream segmentation two-stage based on a teacher-student model, steps 3 and 4 construct a self-supervised pre-training teacher model based on Mmaba and a Transformer, and step 5 constructs a student model of an autoencoder, step 6 constructs a segmentation model based on a Mamba backbone and octree sorting in a downstream segmentation task, jointly constitutes SCTramba, improves the segmentation accuracy, realizes a new idea and method for solving the defect quality detection of 3D printing concrete, and step 8 performs model evaluation, and the results show that the accuracy is improved by 2.45% to 2.98%, the MIoU is improved by 4.74% to 10.79%, and the F1score is improved by 5.81% to 15.7%. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 It is a data acquisition diagram for a laser scanner.

[0049] Figure 2 It is a data set format diagram.

[0050] Figure 3 It is a point cloud embedding module diagram.

[0051] Figure 4 It is a self-supervised pre-training teacher model algorithm flowchart.

[0052] Figure 5 It is a self-supervised mask autoencoder student model algorithm flowchart based on Mamba.

[0053] Figure 6 It is a downstream segmentation model structure diagram based on Mamba and octree sorting.

[0054] Figure 7 It is a Mamba model structure diagram. DETAILED DESCRIPTION

[0055] The application will be further described in detail below with reference to the accompanying drawings.

[0056] The present example discloses a 3D printing concrete surface pore structure point cloud segmentation method based on self-supervised cross-modal, which specifically includes the following steps:

[0057] Step 1, using a laser scanner to collect point cloud data of 3D printing concrete components and performing data preprocessing.

[0058] In step 1, a 3D scanning device Freescan Combo is used to collect point cloud data of 3D printing concrete components at multiple angles and multiple levels. After collecting the surface pore structure, the noise points and invalid points are removed, the file is saved, and the collection process is as shown in Figure 1

[0059] ​Step 2, make a special 3D printing concrete surface pore structure point cloud dataset for the pretreated data of step 1.

[0060] The step 2, use CloudCompare software to annotate the dataset, and specifically classify it as label 0 for non-pore and 1 for pore. The dataset format is (x, y, z, label). x represents the spatial horizontal coordinate, y represents the spatial vertical coordinate, z represents the spatial depth information, and label represents the label.

[0061] The dataset format is as shown in Figure 2 According to 1:1:8, it is divided into three types of training set train, validation set val and test set test. Its main function is the downstream segmentation model in the method, train is mainly used to train the model, val is used to verify the model, and test is mainly used to test the model. Because of the complexity of data acquisition, the point cloud data is horizontally flipped and scaled for data enhancement to expand the point cloud dataset.

[0062] Step 3, design a self-supervised teacher-student model framework, build a point cloud grouping and point cloud embedding module, and obtain point cloud key points and group feature details and local embedding information.

[0063] The step 3, given a point cloud P = {p i |i = 1, 2,..., N} ∈ R N×3 , use the farthest point sampling FPS algorithm to obtain G key points, and then use the nearest neighbor (KNN) algorithm to find K nearest neighbors for each key point, defined as

[0064] N g = {p gk ∈ P | ‖p gk -c g ‖2≤ ‖p-c g ‖2for K nearest}

[0065] The subscript g refers to the gth key point, and g k represents the kth point in the group, and the point cloud grouping module is constructed.

[0066] The local features N g of the point cloud P are obtained by constructing a point cloud embedding module through an encoder Encoder and a graph convolutional network DGCNN and maximum pooling, and the local block features are obtained. |||2 represents the Euclidean distance.

[0067] Specifically, the encoder is composed of a Conv layer, a BN layer and a Relu layer, and the point cloud embedding module contains two Encoder layers, each of which is constructed as Conv1d-BN-Relu-Conv1d, with an input channel of 3 and an output channel of 256, and then maximum pooling is performed to obtain global features. Then, the low-dimensional non-pooled features and the high-dimensional pooled features are spliced, and a second Encoder is used for more fully detailed feature extraction, with an input channel of 512 and an output channel of 1024. The graph convolution network DGCNN is composed of KNN-Conv2d-GroupNorm-MaxPool, with an input channel of 256, and after 5 layers of DGCNN, the output channel is 2304. Finally, the features extracted by the first four layers are spliced and aggregated, and are sent into the fifth layer of DGCNN together to avoid loss of low-dimensional details. The local block features are obtained by multiple fully high and low dimensional extractions of the point cloud embedding module, as shown in Figure 3 .

[0068] Step 4, based on step 3, a self-supervised pre-training teacher model is completely constructed, which combines the advantages of two-dimensional texture semantic information and three-dimensional point cloud depth information.

[0069] In step 4, the training model of a large-scale two-dimensional traditional image dataset is first used as the traditional two-dimensional image information input of the method, the point cloud data is used as the three-dimensional depth information input, and the Transformer is used as the pre-training teacher model backbone.

[0070] The Transformer encoding process is represented as:

[0071] h l = MLP(LN(MSA(LN(h l-1 ))+h l-1 )), l = 1,..., L

[0072] where MLP represents a multi-layer perceptron module, LN represents a normalization module, MSA represents a multi-head self-attention mechanism, h represents feature information, l represents the number of model layers, and there are L layers in total.

[0073] where the codebook is constructed as a shared semantic space for two-dimensional and three-dimensional information, the two-dimensional traditional image information and the three-dimensional point cloud information are used as the core of cross-modal fusion learning using the Transformer, and finally the decoding reconstruction is performed to obtain the weights of the self-supervised pre-training teacher model, and the self-supervised pre-training teacher model is completely built, and the overall algorithm process is shown in Figure 4 .

[0074] Step 5, construct a self-supervised Mamba-based mask autoencoder student model, cooperate with step 4 to pre-train the teacher model, and construct a self-supervised cross-modal 3D printing concrete pore structure point cloud segmentation method SCTramba pre-training stage.

[0075] Said step 5, the proposed self-supervised Mamba-based mask autoencoder student model takes Mamba as the core, uses its full-length modeling capability to capture key features of the data. It is proposed to generate a mask matrix by dynamically masking 80% of the high-proportion random mask points, and fuse it with the globally learnable semantic information label, and use Mamba as the encoder to build an asymmetric autoencoder for extracting point features. At the same time, the pre-trained teacher model in step 4 is used as the target feature output prediction result, and the prediction feature result after decoding in the Mamba autoencoder is constantly optimized through the teacher model prediction result. Among them, the teacher model is represented as F T , the output latent feature Z T = F T (X), and the student model is represented as F S , the output prediction feature Z S = F S (X masked ), and the cosine similarity is used as the loss function to optimize the prediction result of the student model. Among them, X represents the local feature, X masked represents the mask code, and the loss function is represented as:

[0076]

[0077] Among them, L cos represents the loss function, N s represents the total number of masked points, M i represents the mask matrix weight coefficient, that is, whether the point is masked, Z S represents the prediction feature result of the student model, and Z T represents the latent feature of the teacher model.

[0078] A self-supervised Mamba-based mask autoencoder student model is constructed, and the design of the pre-trained teacher-student model is completed, and the student model algorithm process is as shown in Figure 5 .

[0079] Step 6, after obtaining the pre-training weight in the pre-training stage, a downstream segmentation model based on Mamba and octree sorting is proposed to optimize the unordered point cloud and complete the accurate segmentation of the 3D printing concrete surface pore structure.

[0080] Said step 6, a downstream segmentation model based on Mamba and octree sorting is proposed, and the model structure is as shown in Figure 6As shown, the pre-trained teacher-student model weights of steps 5 and 6 are used as downstream segmentation model pre-training weights to capture global features with Mamba as the backbone network. Using the laboratory self-made 3D printed concrete surface pore structure point cloud dataset, first, key point sampling and extraction are performed using the FPS and KNN algorithms, and then local features are obtained through point cloud embedding coding. In order to solve the unordered nature of point clouds and better accurately represent the particularity of pore structure, a special defect structure in spatial information representation, an octree sorting mechanism is introduced to obtain causal relationships. Specifically, first, an octree structure is constructed, and the same depth octree nodes are sorted by a shuffle key, where x i ,y i ,z i represents the i-th point coordinate of the octree node, d is the depth of the octree, and the value of the key corresponds to the 3D z-order curve position, which is represented as:

[0081] Key(χ,z,d)=x0y0z0x1y1z1…x d-1 y d-1 z d-1

[0082] The sorted point cloud features are converted into a one-dimensional sequence Z and input into the N*Mamba encoder to extract features.

[0083] Mamba consists of a normalization layer, a selective SSM, a deep convolution, and a residual connection, and the model structure is as shown in Figure 7 . After Mamba encoding, global features are obtained through average pooling and maximum pooling, and finally the pore structure is segmented through the task segmentation head.

[0084] Step 7, use the training set of the self-built 3D printed concrete surface pore structure point cloud dataset made in step 2 to train the downstream segmentation model based on Mamba and octree sorting in step 6.

[0085] Step 8, use the validation set and test set of step 2 to validate and predict the trained model.

[0086] In step 8, SCTramba is compared with the classic MLP-based benchmark model PointNet++ and the classic Transformer-based network PointMAE, and the model evaluation results are as follows:

[0087] Network Backbone Accuracy MIoU F1score Pointnet++ MLP 81.58 55.93 53.14 PointMAE Transformer 82.11 61.98 62.5 SCTramba Mamba+Trasformer 84.56 66.72 68.31

[0088] In summary, from the evaluation indicators, it can be shown that SCTramba outperforms the other two classic models in terms of accuracy, MIou, and F1score evaluation indicators, significantly better than PointNet++ and PointMAE. Compared with the two classic algorithms, the accuracy of the pores is increased by 2.98% and 2.45% respectively. SCTramba improves the MIoU index by 10.79% and 4.74% respectively, and the F1score is increased by 15.7% and 5.81% respectively. It highlights the accuracy of SCTramba in capturing pore space and background and its better sensitivity to unbalanced and irregular pore-background distribution. It provides strong support for the intelligent quality detection of 3D printed concrete surface defects. At the same time, the novelty and efficiency of the method provide a solid foundation for further research.

Claims

1. A self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method, characterized in that, The method comprises the following steps: Step 1, point cloud data acquisition of 3D printed concrete components is carried out using a laser scanner, and data preprocessing is carried out; Step 2, based on the preprocessed data of step 1, a 3D point cloud data set of the surface pore structure of 3D printed concrete is prepared; Step 3, a point cloud grouping and point cloud embedding module is built to construct a self-supervised teacher-student model framework, and key points and group feature details and local embedding information of the point cloud are obtained; Step 4, a self-supervised pre-training teacher model in the self-supervised teacher-student model framework is completely constructed based on step 3, and the two-dimensional image texture semantic information carried by the ViT model trained using large-scale two-dimensional image data and the depth information of the three-dimensional point cloud data set of step 2 are fused; Step 5, a self-supervised Mamba-based mask autoencoder student model in the self-supervised teacher-student model framework is constructed, and the pre-training teacher model of step 4 is used to obtain pre-training weights through a pre-training phase of the self-supervised cross-modal 3D printed concrete surface pore structure point cloud segmentation method; Step 6, after obtaining the pre-training weights in the pre-training phase, a downstream segmentation model based on Mamba and octree sorting is used; The point cloud disorder is optimized, and the 3D printed concrete surface pore structure is accurately segmented; Step 7, the downstream segmentation model based on Mamba and octree sorting of step 6 is trained using the training set of the self-built 3D printed concrete surface pore structure three-dimensional point cloud data set prepared in step 2; Step 8, the trained downstream segmentation model is verified and predicted using the three-dimensional point cloud data of step 2 as the verification set and the test set, and the model performance is evaluated using evaluation indexes.

2. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 1, characterized in that, The step 1 comprises: using a 3D scanning device to scan and collect point cloud data of the 3D printed concrete component; after collecting the surface pore structure, removing noise points and invalid points, and saving the file.

3. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 2, characterized in that, The step 2 comprises: importing the point cloud data collected in step 1 into CloudCompare software, performing data labeling, specifically labeling the 3D printed concrete pores and non-pore parts, labeling the pore part as 1 and the non-pore part as 0, and dividing the obtained 3D printed concrete surface pore structure three-dimensional point cloud data set into three types of training set train, verification set val and test set test.

4. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 3, characterized in that, The step 3 comprises: using the farthest point sampling FPS algorithm to obtain key points, and using the nearest neighbor KNN algorithm to find K nearest neighbors for each key point; the local features in the group are constructed through an encoder Encoder and a graph convolutional network DGCNN and a maximum pooling to obtain local block features; The Encoder has two layers for more sufficient detail feature extraction, the graph convolutional network DGCNN has five layers, and the features extracted by the first four layers are spliced and aggregated together to be sent into the fifth layer DGCNN to avoid loss of low-dimensional details; Given a point cloud P = {p i |i=1,2,...,N}∈R N×3 The FPS algorithm for farthest point sampling is used to obtain... There are G key points in total. The K nearest neighbor (KNN) algorithm is then used to find K nearest neighbors for each key point, defined as... The subscript g refers to the gth key point, and g k represents the kth point in the group, and the point cloud grouping module is constructed; local features N within the group of point cloud P g The point cloud embedding module is constructed by an encoder Encoder and a graph convolution network DGCNN and maximum pooling to obtain local block features.

5. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 4, characterized in that, The step 4 comprises using a ViT to train a model of a large-scale two-dimensional traditional image data set as two-dimensional image texture information input, using a three-dimensional point cloud data as three-dimensional depth information input, using a Transformer as a pre-training teacher model backbone, and constructing a codebook as a shared semantic space of two-dimensional and three-dimensional information. The Transformer encoding process is represented as: h l = MLP(LN(MSA(LN(h l-1 ))+ h l-1 )), / = 1,..., L; Wherein, MLP represents a multi-layer perception module, LN represents a normalization module, MSA represents a multi-head self-attention mechanism, h represents feature information, and l represents the number of model layers, a total of L layers.

6. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 5, characterized in that, The step 5 comprises using Mamba as an encoder, simultaneously using the pre-training teacher model of step 4 as a target feature, and constantly optimizing the prediction feature result after decoding in the autoencoder through the teacher model prediction result through a loss function. Wherein, the teacher model is represented as F T , the output latent feature Z T T (X), the student model is represented as Fs, and the output predicted feature Z S S (X masked ) is obtained, the cosine similarity is used as the loss function, and the student model prediction result is optimized; wherein, X represents the local feature, X masked represents the mask matrix, and the loss function is represented as:​​ 7. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 6, characterized in that, The step 6 comprises using Mamba as a backbone network to capture global features and introducing an octree ordering mechanism to obtain causal relationships; the ordered point cloud features are converted into a one-dimensional sequence Z and input into an N*Mamba encoder to extract features; the Mamba comprises a normalization layer, a selective SSM state space model, a deep convolution, and a residual connection; Specifically, first construct an octree structure, and sort the same depth octree nodes by a shuffle key, where x i ,y i ,z i represents the i-th point coordinate of the octree node, d is the depth of the octree, and the value of the key corresponds to the z-order curve position in 3D, which is represented as: Key (x, z, d) = x0y0z0x1y1z1... x d-1 y d-1 z d-1 The ordered point cloud features are converted into a one-dimensional sequence Z and input into an N*Mamba encoder to extract features; the Mamba comprises a normalization layer, a selective SSM, a deep convolution, and a residual connection, and the global features are obtained through average pooling and maximum pooling after Mamba encoding, and finally a segmentation task is completed on the pores through a task segmentation head.

8. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 7, characterized in that, The step 7 comprises training the segmentation network based on Mamba and octree ordering constructed in step 6 using a training set, and obtaining a best weight file when the training is completed, which saves important parameters and weight information of the model.

9. The self-supervised cross-modality 3D printed concrete pore structure point cloud segmentation method according to claim 8, characterized in that, The evaluation indicators of the step 8 specifically comprise: Accuracy: the ratio of the number of correct predictions in positive and negative examples to the total number; Wherein, TP represents a pore point correctly identified as a pore, TN represents a background point correctly classified as a non-pore, FP represents a background point incorrectly classified as a pore, and FN represents a pore point incorrectly labeled as a background; Mean Intersection over Union (MIoU): measures the average value of the similarity between the predicted segmentation result and the true segmentation result. C=2; F1 score (F1): the F1 score is the harmonic mean of the prediction rate and the recall rate;

Images