3D printing concrete wall quality monitoring method and system based on point cloud segmentation

Abstract

The present invention is a 3D printing concrete wall quality monitoring method and system based on point cloud segmentation, and the monitoring method includes the following contents: obtaining point cloud data in the construction process of 3D printing concrete wall and normalizing the point cloud coordinates; constructing a neural network model PSC-Net to obtain the boundary point cloud of the printed concrete wall construction layer; performing layer height monitoring based on the boundary point cloud of the printed concrete wall construction layer, directly skipping the layer height calculation step, and performing layer height detection by detecting the number of deviation points, which greatly improves the detection efficiency. It can realize real-time monitoring and analysis of construction data, timely discover problems and take measures, optimize construction plans, and ensure construction safety and quality.

Description

Technical Field

[0001] The present invention relates to the field of intelligent construction technology, and in particular to a 3D printed concrete wall quality monitoring method and system based on point cloud segmentation. Through the layer height monitoring of the 3D printed concrete wall based on point cloud segmentation deep learning, real-time monitoring and control of the printing quality can be performed to achieve stable printing. Background Art

[0002] In the traditional construction field, the construction of concrete walls requires multiple steps, such as mold making, pouring, curing, etc., which takes a long time to complete; it also requires a lot of manpower, including material handling, construction and cleaning, which means that a large work team needs to be hired and managed, increasing labor costs and management difficulties, resulting in a significant increase in costs. Therefore, it is crucial to manufacture concrete walls that meet the requirements in a short time and at a low cost.

[0003] 3D printing, as an advanced manufacturing technology, is also known as additive manufacturing. It mainly creates objects by stacking materials layer by layer. The advantages of this technology are high design freedom, high manufacturing efficiency, high resource utilization, and the ability to produce complex structures. However, the printing quality of 3D printed concrete walls is affected by parameters such as the speed of the extrusion nozzle, the distance between the nozzle and the deposition surface, the pressure of the dosing pump, and the rheological properties of the extruded concrete. When these parameters cannot reach a coordinated state, the single-layer height of the printed strips during the printing process will be inconsistent, thereby increasing the interlayer interface cracks of the printed structure, which will seriously affect the quality and structural stability of the final printed structure, causing structural safety and risks to personnel life. Therefore, the single-layer height of the 3D printed concrete wall must be accurately monitored to achieve quality stability control.

[0004] The current quality monitoring of the 3D printing concrete printing process mainly relies on manual visual inspection, which consumes a lot of manpower and economic costs. In addition, manual inspection cannot monitor places above the height of the human field of vision, and cannot quantitatively measure the single-layer height of the printed strips during the construction process of the 3D printing concrete wall. As a result, it is impossible to evaluate the real-time status of the stability of the printing quality of the 3D printing concrete wall in real time. At present, computer vision technology can realize automatic real-time monitoring, but computer vision technology based on 2D images does not have the coordinate information of the target and cannot accurately measure the spatial information of the target. In addition, the traditional 2D image segmentation method has the problems of information loss and increased computational complexity when processing complex point cloud data. It cannot effectively process high-dimensional features and extract key information, and the precise control and stability of the concrete wall 3D printing process cannot be guaranteed. Summary of the invention

[0005] In order to overcome the shortcomings of the above technologies, the present invention provides a 3D printed concrete wall quality monitoring method and system based on point cloud segmentation. Compared with traditional technologies, the present invention utilizes data enhancement, contrast learning, and point cloud segmentation technologies in deep learning and independently constructs a neural network model PSC-Net (where "P" represents the data enhancement part in the neural network model, based on the point cloud enhancement network PF-Net, and "S" and "C" represent the sub-scene boundary mining module and the contrast boundary learning module in the neural network model PSC-Net, respectively) to realize the quality monitoring of 3D printed concrete walls, and can accurately and efficiently monitor whether the concrete wall layer height during the printing process meets the standard, and can also realize real-time monitoring, early warning and control, effectively ensuring the printing quality.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention provides a 3D printed concrete wall quality monitoring method based on point cloud segmentation, the monitoring method comprising the following contents:

[0008] Obtain point cloud data during the construction of 3D printed concrete walls and normalize the point cloud coordinates;

[0009] Constructing a neural network model PSC-Net, wherein the neural network model PSC-Net includes an encoder, a decoder, a sub-scene boundary mining module, and a contrast boundary learning module;

[0010] The point cloud data in the concrete wall construction process is used to train the neural network model PSC-Net to obtain the trained neural network model PSC-Net, and the trained neural network model PSC-Net is used to perform point cloud boundary segmentation to obtain the boundary point cloud of the printed concrete wall construction layer;

[0011] Floor height detection:

[0012] Step 5.1 Voxel downsampling and dimensionality reduction

[0013] The boundary point cloud of the printed concrete wall construction layer is downsampled to obtain a rectangular dot matrix; the x, y and z coordinate values ​​of each point in the rectangular dot matrix are mapped to the corresponding axes in the three-dimensional space to form a spatial point cloud; then the y-axis coordinates of all points are deleted, the three-dimensional space is converted into two-dimensional space, the sum of the x and z coordinates of all points is calculated, the minimum sum is found, and the point corresponding to the minimum sum is set as the coordinate origin. At the same time, the rectangular dot matrix formed by the entire boundary point cloud is translated so that the origin coordinate becomes (0,0), completing the coordinate transformation;

[0014] Step 5.2 Dot grouping

[0015] According to the z coordinates of the points after coordinate transformation, the points with a z coordinate difference within 5% are grouped together to obtain several point sets, and each point set is marked with a layer number in ascending order of the z coordinates; there are m groups of point sets in the printed concrete wall construction layer, and m layers have been printed, where m is a positive integer;

[0016] Step 5.3 Count the number of deviation points

[0017] Preset the layer height threshold, and compare the z coordinates of the point set of the lower layer after translating the layer height threshold with the point set of the adjacent upper layer in turn, record the coordinate points whose z coordinates in the point set of the adjacent upper layer are greater than the coordinates of the point set of the lower layer after translating the layer height threshold as deviation points, and count the number of deviation points in each layer;

[0018] Step 5.4 deletes the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer, and determines whether the number of remaining points in the boundary point cloud of the printed concrete wall construction layer is not greater than 50%. If it is greater than 50%, repeat steps 5.1 to 5.3, re-perform voxel downsampling, dimension reduction, dot matrix grouping, and statistically count the number of deviation points on the remaining points, and obtain the number of deviation points and layer numbers of all printed concrete wall construction layers in the next round; if it is not greater than 50%, stop searching for deviation points;

[0019] Add up the number of deviation points of each layer in all rounds to obtain the final number of deviation points of each layer of the printed concrete wall construction layer;

[0020] Step 5.5 Stability Control

[0021] Set a stable ratio threshold, and use the final number of deviation points in each layer and the corresponding layer number to draw a bar chart. The horizontal axis of the bar chart is the layer number, and the vertical axis is the ratio of the number of deviation points in a layer to all the points in the current layer. If the ratio of the number of deviation points in the top layer to all the points in the layer exceeds the stable ratio threshold by 15%, it means that the height of the top layer exceeds the layer height threshold, the current printing process is unstable, and early warning adjustment is required; if it does not exceed the stable ratio threshold, the layer height meets the standard and the printing process is stable.

[0022] In the encoder, the original point cloud with an input dimension of N×3 is iteratively sampled at the farthest point to obtain a point cloud with a dimension of The downsampled point cloud is N, N is the number of points in the point cloud, and 3 is the spatial three-dimensional coordinate of the point; the original point cloud and the downsampled point cloud are respectively input into two combined multi-layer perceptron modules to obtain two potential feature vectors of the same dimension, and the two potential feature vectors are concatenated to obtain a potential feature map. The potential feature map is then passed through a multi-layer perceptron and an L2 norm normalization layer to obtain a normalized one-dimensional feature vector V', which is recorded as the output of the encoder;

[0023] The decoder is used to remap V' back to the original point cloud data space. In the decoder, V' first passes through the first linear layer to obtain a one-dimensional vector V1, and V1 passes through the second linear layer to obtain a one-dimensional vector V2; the length of V2 is half the length of V1, and V1 and V2 are connected to the first fully connected layer FC1 and the second fully connected layer FC2 respectively; the outputs of the first fully connected layer FC1 and the second fully connected layer FC2 are added to output the enhanced point cloud Y detail , Y detail Denote as the output of the decoder;

[0024] The output of the decoder is connected to a sub-scene boundary mining module and a contrast boundary learning module, the sub-scene boundary mining module is used to determine the label of the point, and the contrast boundary learning module is used to perform boundary segmentation, the sub-scene boundary mining module includes n times of downsampling, the contrast boundary learning module includes n times of upsampling, the result of the first downsampling of the sub-scene boundary mining module is jump-connected with the result of the nth upsampling of the contrast boundary learning module, and the result of the nth downsampling of the sub-scene boundary mining module is jump-connected with the result of the first upsampling of the contrast boundary learning module;

[0025] The contrastive boundary learning module outputs a boundary point cloud.

[0026] The combined multi-layer perceptron module includes four combined perceptron layers, each of which is composed of 1, 2, 3, and 4 perceptrons connected in series. The input point cloud data is processed by the four combined perceptron layers to generate four feature vectors. Then, the four feature vectors are subjected to average pooling operations, and the results of the four average pooling operations are spliced ​​to obtain potential feature vectors.

[0027] The loss function L of the neural network model PSC-Net is composed of the cross entropy function L of the sub-scene boundary mining module cross entropy And the loss function L of the contrast boundary learning module CBL Composition, expressed by the following formula:

[0028]

[0029]

[0030] L cross entropy =-∑(c i *log(p i ))

[0031] Among them B l represents the boundary point set in the real data, x i represents the selected center point, x j Represents x i A point in the neighborhood; l iand l j Represents point x i and x j The true label of i 、f j Represents point x i 、x j The features of ; d() is the distance function; τ represents the temperature hyperparameter in contrastive learning; c i The category to which the point belongs, p i A vector representing the probability distribution of the model prediction, n represents the nth sampling stage, λ is the loss weight; N i represents the center point set.

[0032] The stable ratio threshold is 15%.

[0033] In a second aspect, the present invention provides a 3D printed concrete wall quality monitoring system based on point cloud segmentation, the system comprising:

[0034] Point cloud acquisition module, used to obtain point cloud data during the construction process of 3D printed concrete walls;

[0035] Point cloud preprocessing module, used to preprocess point cloud data to obtain standardized and high-quality point cloud data;

[0036] Layer height detection module, used to monitor in real time whether the layer height during the construction process meets the printing requirements;

[0037] The early warning adjustment module is used to warn of abnormal stability detected by the layer height detection module, and adjust the control unit of the 3D printing equipment to adjust the 3D printing parameters, including speeding up or slowing down the movement speed of the printer nozzle, lowering or raising the nozzle height, speeding up or slowing down the material discharge speed; and at the same time, reminding the staff to check the material status;

[0038] A display module is used to display the monitoring results output by the layer height detection module during real-time printing;

[0039] A storage module, used to store the collected data and the data in the monitoring process;

[0040] The point cloud boundary segmentation model is used to perform point cloud boundary segmentation to obtain the boundary point cloud of the printed concrete wall construction layer; the specific implementation process of the layer height detection module is:

[0041] Step 5.1 Voxel downsampling and dimensionality reduction

[0042] The boundary point cloud of the printed concrete wall construction layer is downsampled to obtain a rectangular dot matrix; the x, y and z coordinate values ​​of each point in the rectangular dot matrix are mapped to the corresponding axes in the three-dimensional space to form a spatial point cloud; then the y-axis coordinates of all points are deleted, the three-dimensional space is converted into two-dimensional space, the sum of the x and z coordinates of all points is calculated, the minimum sum is found, and the point corresponding to the minimum sum is set as the coordinate origin. At the same time, the rectangular dot matrix formed by the entire boundary point cloud is translated so that the origin coordinate becomes (0,0), completing the coordinate transformation;

[0043] Step 5.2 Dot grouping

[0044] According to the z coordinates of the points after coordinate transformation, the points with a z coordinate difference within 5% are grouped together to obtain several point sets, and each point set is marked with a layer number in ascending order of the z coordinates; there are m groups of point sets in the printed concrete wall construction layer, and m layers have been printed, where m is a positive integer;

[0045] Step 5.3 Count the number of deviation points

[0046] Preset the layer height threshold, and compare the z coordinates of the point set of the lower layer after translating the layer height threshold with the point set of the adjacent upper layer in turn, record the coordinate points whose z coordinates in the point set of the adjacent upper layer are greater than the coordinates of the point set of the lower layer after translating the layer height threshold as deviation points, and count the number of deviation points in each layer;

[0047] Step 5.4 deletes the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer, and determines whether the number of remaining points in the boundary point cloud of the printed concrete wall construction layer is not greater than 50%. If it is greater than 50%, repeat steps 5.1 to 5.3, re-perform voxel downsampling, dimension reduction, dot matrix grouping, and statistically count the number of deviation points on the remaining points, and obtain the number of deviation points and layer numbers of all printed concrete wall construction layers in the next round; if it is not greater than 50%, stop searching for deviation points;

[0048] Add up the number of deviation points of each layer in all rounds to obtain the final number of deviation points of each layer of the printed concrete wall construction layer;

[0049] Step 5.5 Stability Control

[0050] Set a stable ratio threshold, and use the final number of deviation points in each layer and the corresponding layer number to draw a bar chart. The horizontal axis of the bar chart is the layer number, and the vertical axis is the ratio of the number of deviation points in a layer to all the points in the current layer. If the ratio of the number of deviation points in the top layer to all the points in the layer exceeds the stable ratio threshold by 15%, it means that the height of the top layer exceeds the layer height threshold, the current printing process is unstable, and early warning adjustment is required; if it does not exceed the stable ratio threshold, the layer height meets the standard and the printing process is stable.

[0051] The point cloud boundary segmentation model is a neural network model PSC-Net, and the neural network model PSC-Net includes an encoder, a decoder, a sub-scene boundary mining module and a comparative boundary learning module connected in sequence.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention realizes the detection of the printing quality of 3D printed concrete walls based on point cloud segmentation. First, the high-dimensional characteristics of point cloud data are utilized to realize the accurate evaluation of the printing quality of 3D printed concrete walls. Traditional methods often rely on one-dimensional or two-dimensional data representation and are difficult to capture complex three-dimensional structural information. Secondly, the present invention utilizes the neural network model PSC-Net, which can enhance and segment the input point cloud data. Compared with traditional point cloud segmentation models, such as RandLA-Net and CloserLook3D, PSC-Net first performs data enhancement before segmenting the point cloud data, which can increase the diversity of data and thus improve the generalization ability of the model. Using the enhanced data for training can also help reduce the risk of model overfitting and improve the reliability and robustness of the model. At the same time, data enhancement can effectively increase the size of the data set, thereby reducing the risk of model overfitting. The segmentation accuracy of the present invention on the indoor data set S3DIS can reach 90.6%, which is better than 87.5% of RandLA-Net and 90.0% of CloserLook3D.

[0054] 2. The traditional floor height detection method usually needs to calculate the height of each floor and then compare it with the preset height threshold. However, this method involves a large number of calculation operations, which is time-consuming and prone to errors. The present invention adopts a more efficient method, directly skipping the floor height calculation step, and performing floor height detection by detecting the number of deviation points, which greatly improves the detection efficiency.

[0055] 3. The present invention can monitor and analyze construction data in real time, find problems in time and take measures, optimize construction plans, and ensure construction safety and quality. The present invention provides the possibility of real-time monitoring of 3D printing quality, further promotes the development of this field, has great significance and broad application prospects, and also provides strong support and guarantee for the development of concrete 3D printing technology.

[0056] 4. The present invention uses point cloud segmentation technology to obtain various information such as the spatial coordinates of the object, and can accurately identify the interfaces and boundaries between different parts of the printed model. By efficiently and accurately identifying the interfaces and boundaries, automatic detection can be achieved, which can optimize data representation and feature extraction, improve printing quality, and bring more reliable solutions to the field of retaining wall construction. Point cloud segmentation technology will bring significant advantages to the quality control of 3D printed concrete walls. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a structural diagram of the neural network model PSC-Net in the present invention.

[0058] Figure 2 It is a specific structural diagram of the combined multi-layer perceptron module in the neural network model PSC-Net in the present invention.

[0059] Figure 3 It is a schematic diagram of the structure of the decoder of the neural network model PSC-Net in the present invention.

[0060] Figure 4 It is a single training flow chart of the neural network model PSC-Net in the present invention. DETAILED DESCRIPTION

[0061] In order to more clearly describe the technical problems, technical solutions and advantages of the present invention, the following will be described in detail with reference to the drawings and embodiments. It should be noted that these embodiments are only used to illustrate the principles and application scope of the present invention and should not be regarded as limiting the present invention.

[0062] Embodiment 1: The 3D printing concrete wall quality monitoring method based on point cloud segmentation in this embodiment mainly includes the following steps:

[0063] The first step is to obtain the point cloud data during the construction of the 3D printed concrete wall and normalize the point cloud coordinates:

[0064] Step 1.1 Point cloud data acquisition

[0065] A number of 3D printed concrete walls were randomly scanned using a depth camera scanning device, and a total of 12,000 3D printed concrete wall point cloud data were obtained. Each point cloud data contains a series of three-dimensional points, and each point is represented by (x, y, z) coordinates. The depth camera can adjust the shooting range according to the printing height, and can obtain images of retaining walls that are higher than the height of the human body.

[0066] Step 1.2 Point cloud dataset preparation

[0067] The 12,000 point cloud data sets are divided into training set, validation set and test set in a ratio of 4:1:1 and stored in a computer; the training set contains 8,000 point cloud data sets, and the validation set and test set each contain 2,000 point cloud data sets.

[0068] Step 1.3 Point cloud data preprocessing

[0069] Step 1.3.1 Use the point cloud processing software "CloudCompare" to manually annotate the acquired point cloud data and mark the boundary and non-boundary parts in the point cloud data.

[0070] Step 1.3.2 normalizes the coordinates of all point cloud data to the interval [-1,1] with the origin as the center, that is, the coordinates of all points are proportionally reduced so that the coordinate values ​​of the points are within the range of [-1,1]. The purpose is to make the coordinate range of the point cloud data consistent and facilitate the training of the neural network model. Assuming the coordinates of a point are (x, y, z), use the following formula for normalization:

[0071]

[0072]

[0073]

[0074] Among them, x min 、x max Represents the minimum and maximum coordinate values ​​of all points in a set of point clouds on the x-axis, min ,y max 、z min 、z max Similarly, represents the minimum and maximum coordinate values ​​of all points in a set of point clouds on the y-axis and z-axis, and x', y', and z' represent the normalized values ​​of the corresponding coordinate axes.

[0075] Step 1.4 Preliminary sampling of point cloud data: 2048 points are uniformly sampled from each point cloud data as the original point cloud to ensure the uniformity of the input data.

[0076] Step 1.5: Secondary sampling of point cloud data. Perform iterative farthest point sampling on the original point cloud data obtained in step 1.4 to obtain a downsampled point cloud. Iterative farthest point sampling is a sampling method for point cloud data that aims to effectively select a set of representative sampling points from a large-scale point cloud. The principle is to gradually select the farthest points in the point cloud data as sampling points to construct a sparse sampling point set with a smaller number, and continuously iterate this process until the sampling point set reaches a predetermined number or meets other termination conditions. The downsampled point cloud is obtained by performing iterative farthest point sampling on the original point cloud, and is input into the neural network model PSC-Net together with the original point cloud.

[0077] Through the above steps, standardized and high-quality point cloud data can be obtained, providing high-quality input data for subsequent point cloud data enhancement.

[0078] The second step is to build the neural network model PSC-Net

[0079] like Figure 1 As shown, the neural network model PSC-Net includes four parts, namely encoder, decoder, sub-scene boundary mining module (SBM) and contrast boundary learning module (CBL). The four parts are connected in series, and the original point cloud passes through the encoder, decoder, sub-scene boundary mining module, and contrast boundary learning module in turn, and finally outputs the segmented boundary point cloud. The "encoder-decoder" structure is mainly used to interpolate, complement and enhance the input point cloud. The process of the "encoder-decoder" structure is end-to-end, which can reduce the number of data transmission and calculation, improve the efficiency of data enhancement, and thus improve the robustness and generalization ability of the model.

[0080] Step 2.1 Encoder: Use the encoder to extract data features

[0081] In the encoder, the original point cloud with a dimension of 2048×3 (where 2048 is the number of input points and 3 is the xyz coordinate data of the point) is iteratively sampled at the farthest point to obtain a dimension of (in is the number of points in the point cloud after sampling, and 3 is the xyz coordinate data of the point). Next, the original point cloud and the downsampled point cloud are respectively input into a combined multi-layer perceptron module to obtain two potential feature vectors, each of which represents the features extracted from point clouds of different resolutions. Subsequently, the two potential feature vectors are concatenated to form a potential feature map M with a dimension of 1920×3. Finally, a multi-layer perceptron and an L2 norm normalization layer are used to map and normalize the potential feature map into a normalized one-dimensional feature vector V'. V' is recorded as the output of the encoder, and its length is 1920. For the specific process, refer to Figure 1 .

[0082] The specific structure of the combined multi-layer perceptron module is as follows Figure 2 As shown in the figure, the module contains four combined perceptron layers, each of which is composed of 1, 2, 3, and 4 perceptrons. The input point cloud is processed by four combined perceptron layers, and each combined perceptron layer extracts a feature vector of a specific dimension, which can generate a total of four feature vectors. Then, the four feature vectors are averaged and pooled. Average pooling can average features of different dimensions to better capture the overall information of the features. After the input point cloud passes through four combined perceptron layers, four feature vectors with dimensions of 64×3, 128×3, 256×3, and 512×3 can be obtained. Then, these four feature vectors are spliced ​​into a potential feature vector with a dimension of 960×3 (960=512+256+128+64). The original point cloud and the downsampled point cloud can generate two potential feature vectors with dimensions of 960×3 after combining the multi-layer perceptron module. These two potential feature vectors are concatenated to obtain a potential feature map with a dimension of 1920×3 (1920=960+960). The potential feature map consists of a series of feature map pixels, and each pixel represents the importance of the feature at that position in the input data. After combining the multi-layer perceptron module, a multi-layer perceptron layer and an L2 norm normalization layer are connected to map and normalize the potential feature map into a normalized one-dimensional feature vector V' with a length of 1920.

[0083] L2 norm normalization is also called Euclidean norm normalization, which is to divide each element of the vector by the L2 norm of the vector (the length of the vector), so that the modulus of the vector is 1. Normalizing the input vector by L2 norm can simplify the data processing and calculation process and improve the efficiency of the algorithm. Assume that the vector V = [v1,v2,...,v 1920 ], then the L2 norm of V is

[0084] The normalized vector V' is

[0085]

[0086] Step 2.2 Decoder, use the decoder to remap the features back to the original point cloud data space. In the decoder, the normalized one-dimensional feature vector V' first passes through the first linear layer to obtain a one-dimensional vector V1 with a length of 1024. Then the vector V1 passes through the first fully connected layer FC1. After passing through the linear layer in the first fully connected layer FC1, the vector V1 is reshaped into a tensor of dimension M×M1. This tensor is convolved to generate a tensor of dimension , and then reshape it into a tensor of dimension At the same time, let the vector V1 pass through the second linear layer to obtain a one-dimensional vector V2 with a length of 512. V2 is reshaped into a tensor with a dimension of M1×3 after passing through the linear layer in the second fully connected layer FC2. It is convolved and expanded to generate a tensor V2' with a dimension of M1×1×3. Finally, the obtained tensors V1' and V2' are spliced ​​and reshaped to obtain the enhanced point cloud Y detail , whose dimension is M×3, the specific process can be referred to Figure 3 .

[0087] The enhanced point cloud is obtained through the two steps of encoding and decoding, thereby providing high-quality point cloud data for the subsequent point cloud boundary segmentation module.

[0088] Step 2.3 Sub-scene boundary mining module

[0089] The sub-scene boundary mining module is used to determine the labels of the points. In order to better explore the scene boundaries, it is necessary to check the boundaries in the sub-sampled point cloud at multiple scales, so that boundary points can be identified at different sub-sampling stages of the backbone model. Since it is difficult to get a correct definition of the boundary point set after sub-sampling, a sub-scene boundary mining module is introduced before boundary point identification to determine the labels of the points in each sub-sampling stage. Since each sub-sampled point is aggregated from a set of points in its previous point cloud; therefore, the enhanced point cloud is first down-sampled n times to iteratively determine the label of the point. For the nth sampling stage, the sub-sampled point cloud is X n , where the i-th point x i n Tags i n for l i n =AVG({l j n-1 |x j n-1 ∈N n-1 (x i n )})

[0090] Where N n-1 (x i n ) represents the n-1th stage point x i n The local neighborhood of x j n-1 Represents the n-1 stage point x i n A point in the neighborhood, in the present invention, the neighborhood is set to point x i nis a spherical area with a radius of 1 cm and AVG represents the average pooling operation. By comparing this formula with the real data labels, the distribution of labels in the sub-sampled point cloud can be obtained.

[0091] Step 2.4 Comparison of boundary learning modules

[0092] Use the contrast boundary learning module to perform point cloud boundary segmentation. The label of the point is determined by the sub-scene boundary mining module, and the boundary segmentation can be performed by the contrast boundary learning module. By comparing the data sample with other related or unrelated samples, similar samples are closer in the feature space, and dissimilar samples are farther away from each other, which can improve the accuracy of segmentation. In the contrast boundary learning module, the sub-sampled point cloud in the previous step is upsampled n times, and the sub-sampled point cloud is gradually restored to the original density. At the same time, a neighborhood search is performed on each point during the upsampling process to find the neighboring points around it. For each point, several positive pairs and negative pairs are sampled according to its neighboring points. Positive pairs refer to neighboring point pairs that belong to the same category as the current point, while negative pairs refer to neighboring point pairs that belong to different categories from the current point. Then, the model is trained by contrast learning. For each point, the loss function value between its positive and negative point pairs is calculated. The model parameters are optimized by the loss function value so that the boundary points can be better distinguished. Finally, the input point cloud data is classified by the boundary information obtained in the contrast learning process to obtain the boundary point cloud.

[0093] The connection between the sub-scene boundary mining module and the contrast boundary learning module adopts a skip connection structure, which can establish a direct connection between different layers of the network, so that information can skip some levels and be directly transmitted to subsequent levels, thus forming a "shortcut" in the network. This structure can improve information flow and gradient propagation, thereby improving the performance and training effect of the neural network. It can realize boundary segmentation of the input point cloud data and obtain the boundary point cloud for subsequent detection and early warning.

[0094] The third step is to train the neural network model. For details, see Figure 4

[0095] Step 3.1 randomly shuffles the 3D printed concrete wall point cloud data in step 1.2 to ensure that the order of the data has no effect on the training and verification results of the neural network model; randomly generates initial training parameters, such as learning rate, etc.

[0096] Step 3.2 uses the training set to train the neural network model PSC-Net for 300 rounds. At the same time, during the training process, after every 20 rounds of training, the validation set is used for validation to understand the generalization performance of the neural network model PSC-Net and avoid overfitting problems.

[0097] Step 3.3 Use the loss function to adjust the neural network parameters and determine whether the loss function is less than the model threshold. If not, adjust the parameters through error back propagation. If less than the model threshold, use the validation set to validate the model. Each round of training of the neural network model PSC-Net will generate a loss function value. The loss function of the neural network model PSC-Net is represented by L, and the calculation formula is as follows:

[0098]

[0099] L cross entropy =-∑(c i *log(p i )),

[0100]

[0101] Among them B l represents the boundary point set in the real data, x i represents the selected center point, x j Represents x i A point in the neighborhood; l i and l j Represents point x i and x j The true label of i 、f j Represents point x i 、x j The feature of ; d() is the distance function used to measure the similarity between two features; τ represents the temperature hyperparameter in contrastive learning, which is used to adjust the sensitivity of the similarity metric; c i The category to which the point belongs, p i A vector representing the probability distribution of the model prediction, n represents the nth sampling stage, λ is the loss weight; N i represents the center point set.

[0102] The loss function selected by the sub-scene boundary mining module is the cross entropy function (L cross entropy ), this loss function is used to measure the difference between a model's prediction result and the true label. When the true label is closer to the prediction result, the value of the cross entropy function is smaller; otherwise, the value is larger. The loss function of the comparison boundary learning module selects the mutual information maximization loss function (L CBL), the mutual information maximization loss function is mainly used to measure the similarity between different samples of the model. In the present invention, it is used to measure the similarity between the real boundary and the predicted boundary. The final loss function L is a combination of the cross entropy function and the mutual information maximization loss function. The smaller the value of L, the better the effect of boundary segmentation. At the same time, error backpropagation is continuously performed during the training process, and the model parameters are dynamically adjusted using the adaptive moment estimation gradient descent strategy. The neural network model PSC-Net is then trained using the updated model parameters.

[0103] Step 3.4 Repeat the above steps 10 times, and train the neural network model PSC-Net for a total of 3000 rounds. At this time, the value of the loss function is less than the model threshold (generally set to within 0.1%), indicating that the model has reached convergence and can be verified using the validation set. It should also be noted that during the training process, after every 20 rounds of training, the validation set is used for verification to understand the performance of the model and avoid overfitting. Overfitting refers to the phenomenon that the model performs very well on the training data, but performs poorly on unseen test data.

[0104] Step 3.5 Use the test set to evaluate the segmentation effect of the neural network model PSC-Net, and use the boundary intersection over union ratio B-IoU to evaluate the point cloud boundary segmentation effect

[0105]

[0106] Among them, B l is the real point cloud boundary point set, B p is the point cloud boundary point set in the prediction result. The closer the boundary intersection ratio is to 1, the better the segmentation effect of the neural network model PSC-Net is.

[0107] Before performing the segmentation operation, the neural network model PSC-Net first performs a data enhancement on the input point cloud data, which can increase the diversity of point cloud samples and make the data more complete, thereby improving the robustness of the neural network model. In addition, the data enhancement part of the neural network model PSC-Net reduces the original two samplings to one, and adds a normalization layer after the multi-layer perceptron; at the same time, the three fully connected layers in the decoder are reduced to two, which speeds up the operation of the model while ensuring the integrity of the point cloud data, ensuring real-time performance. Finally, the contrastive learning method is used in the boundary segmentation part, and its ability to make similar sample points closer and different types of samples farther apart improves the accuracy of the entire model. Such a point cloud segmentation model has robustness, real-time performance and accuracy, thereby improving the stability of the printing quality of the 3D concrete retaining wall and ensuring that the printing results meet the design requirements.

[0108] Step 4: Boundary segmentation

[0109] Step 4.1 sets the operating parameters of the 3D printing equipment, such as the nozzle movement speed, nozzle height and discharge speed. The parameter values ​​are determined based on previous actual operations. After the settings are completed, start loading and printing.

[0110] Step 4.2 uses a depth camera scanning device to perform real-time scanning on the 3D printed concrete wall being printed to obtain point cloud data of the 3D printed concrete wall. The depth camera scanning device can capture point cloud information on the surface of the 3D printed concrete wall in real time and convert it into a three-dimensional coordinate representation.

[0111] Step 4.3 inputs the acquired 3D printed concrete wall point cloud data into the trained neural network model PSC-Net.

[0112] Step 4.4 After being processed by the neural network model PSC-Net, the boundary point clouds of the 3D printed concrete wall can be accurately segmented from the point cloud data. These boundary point clouds represent the boundary lines between different layers of the 3D printed concrete wall. Then, the segmented 3D printed concrete retaining wall boundary point clouds are stored in the device for subsequent quality stability control and analysis.

[0113] Step 5: 3D printed concrete wall height detection

[0114] Step 5.1 Voxel downsampling and dimensionality reduction

[0115] A fixed voxel size is selected to perform voxel downsampling on the boundary point cloud of the 3D printed concrete wall output by the neural network model PSC-Net. Voxel downsampling is a method in point cloud data processing that reduces the quantity and complexity of point cloud data by dividing the point cloud space into several cubic voxel grids and selecting a representative point in each cubic voxel grid. The boundary point cloud is converted into several evenly distributed cubic voxel grids to obtain a rectangular point matrix, so that the points in the boundary point cloud are arranged in space like chess pieces on a chessboard.

[0116] Map the x, y, and z coordinate values ​​of each point in the rectangular dot matrix to the corresponding axes in the three-dimensional space to form a spatial point cloud. Then delete the y-axis coordinates of all points, convert the three-dimensional space into two-dimensional space, reduce the dimension of the data and speed up the calculation. Calculate the sum of the x and z coordinates of all points, find the smallest sum, and set the corresponding point as the coordinate origin. At the same time, translate the entire dot matrix so that the origin coordinate becomes (0,0).

[0117] Step 5.2 Dot grouping

[0118] The points are grouped according to the z coordinates after the coordinate transformation, and the points with z coordinate differences within 5% are grouped together to obtain several point sets, and each point set is marked with a layer number in ascending order of the z coordinates; there are a total of m groups of point sets in the printed concrete wall construction layer, and m layers have been printed, where m is an integer greater than 0.

[0119] Step 5.3 Count the number of deviation points

[0120] Preset the layer height threshold, and compare the z coordinates of the point set of the lower layer after translating the layer height threshold with the point set of the adjacent upper layer in turn, record the coordinate points whose z coordinates in the point set of the adjacent upper layer are greater than the coordinates of the point set of the lower layer after translating the layer height threshold as deviation points, and count the number of deviation points in each layer;

[0121] That is, the i-th point set is translated along the positive direction of the z-axis by the preset layer height threshold, and the number of points in the i+1-th point set whose z coordinate values ​​are greater than the z coordinate values ​​of the corresponding points in the i-th point set after translation is counted. These points are called the deviation points of the i+1-th point set. The number of deviation points in the i+1-th point set is matched with the layer number of the i+1-th layer, and the number of deviation points and layer numbers of all printed concrete wall construction layers are obtained.

[0122] Step 5.4 deletes the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer, and determines whether the number of remaining points in the boundary point cloud of the printed concrete wall construction layer is not greater than 50%. If greater than 50%, repeat steps 5.1 to 5.3, re-perform voxel downsampling, dimension reduction, dot matrix grouping, and deviation point number statistics on the remaining points to obtain the number of deviation points and layer numbers of all printed concrete wall construction layers in the next round; if less than 50%, stop searching for deviation points; add up the number of deviation points of each layer in all rounds to obtain the final number of deviation points of each layer of the printed concrete wall construction layer;

[0123] The computer stores the number of deviation points obtained in each round of statistics and the corresponding layer number in the storage device, and deletes the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer to avoid repeated sampling. Subsequently, the voxel downsampling, dimension reduction and grouping operations are performed again for the remaining data, and the new data is counted. This process is repeated until 50% of the points in the point cloud are deleted.

[0124] Step 5.5 Stability Control

[0125] A stable ratio threshold is set, which is set to 15% in this embodiment. A bar graph is drawn using the final number of deviation points in each layer and the corresponding layer number. The horizontal axis of the bar graph is the layer number, and the vertical axis is the ratio of the number of deviation points in a layer to all the points in the current layer. If the ratio of the number of deviation points in the top layer to all the points in the current layer exceeds the stable ratio threshold, it means that the height of the layer exceeds the layer height threshold, the current printing process is unstable and needs to be adjusted, and a warning signal is sent. If it does not exceed the stable ratio threshold, the layer height meets the standard and the printing process is stable.

[0126] The bar chart reflects the proportion of deviation points in each layer of the 3D printed concrete wall, which is more intuitive.

[0127] Step 6: Control and Exception Recording

[0128] Step 6.1 When receiving the early warning signal, adjust the parameters of the printing equipment, including speeding up or slowing down the movement speed of the printer nozzle, lowering or raising the nozzle height, speeding up or slowing down the material discharge speed, stabilize the layer height within the layer height threshold range through data adjustment, and remind relevant personnel to check the printing status.

[0129] Step 6.2 records the data of the abnormal situation to the storage device while processing the abnormal situation. The recorded data includes the time when the abnormal situation occurred, the specific parameter deviation of the abnormal situation, the adjustment results during the processing process and other related information.

[0130] The hardware devices used in the present invention include:

[0131] Processor: As the core component of the present invention, the processor bears key control and management responsibilities in the entire system, covering multiple key functions such as data acquisition, data processing, image recognition, and alarm. Its powerful computing power and parallel processing capabilities ensure the efficient operation of the system and the satisfaction of real-time requirements. According to the requirements of different application scenarios, the processor can be implemented in a variety of different forms such as single-chip microcomputers, microprocessors, and computers. These superior performance characteristics ensure that the present invention achieves outstanding results in practical applications and provides reliable technical support for 3D printed concrete wall printing quality monitoring.

[0132] Sensors: used to monitor parameters and status during the printing process in real time. For example, laser rangefinders and visual sensors are used to measure the height, shape and position of the retaining wall, as well as to detect deviations and errors during the printing process.

[0133] 3D printing equipment: This is the core equipment for realizing 3D concrete wall printing, which is used to print the concrete material layer by layer according to the predetermined path and layer height to form the structure of the concrete wall. These 3D printing equipment are usually large-scale architectural 3D printers that can realize continuous printing of concrete according to complex design requirements.

[0134] Memory: used to store collected data and historical data for subsequent processing and analysis. It has excellent characteristics such as high speed, high reliability and scalability to meet the needs of long-term stable operation of the system.

[0135] Control unit: used to control the movement and operation of the 3D printing device, including a computer or embedded controller, responsible for planning and controlling the printing path.

[0136] Data processing and analysis equipment: used for processing, analysis and real-time feedback of point cloud data. These devices include computers, GPUs, etc., which are used to run point cloud segmentation algorithms and process point cloud data in real time.

[0137] Communication equipment: used to transmit data and instructions, and ensure the communication connection between the system and sensors and printing equipment.

[0138] Network interface: used for data exchange and communication.

[0139] Other auxiliary equipment: such as power supply equipment, cooling devices, etc., used to support the normal operation of 3D printing equipment.

[0140] This embodiment is mainly used for quality monitoring during the printing process of 3D printed concrete walls. It uses a depth camera installed on a 3D printing device and a neural network to automatically identify the layer boundaries of the 3D printed concrete wall, monitor the layer height of the 3D printed concrete wall, and automatically alarm when it exceeds the threshold, thereby automatically adjusting the printing device parameters and recording the abnormality.

[0141] Embodiment 2: This embodiment is a 3D printing layer quality monitoring system based on point cloud boundary segmentation, the system comprising: a point cloud acquisition module for acquiring point cloud data during the construction process of 3D printed concrete wall;

[0142] Point cloud preprocessing module, used to preprocess point cloud data to obtain standardized and high-quality point cloud data;

[0143] The point cloud acquisition module and the point cloud preprocessing module are responsible for acquiring the point cloud data of 3D printed concrete walls in the actual scene, and preprocessing and optimizing these data for subsequent point cloud enhancement and segmentation tasks.

[0144] Layer height detection module, used to monitor in real time whether the layer height during the construction process meets the printing requirements;

[0145] The early warning adjustment module is used to warn of abnormal stability detected by the layer height detection module, and adjust the control unit of the 3D printing equipment to adjust the 3D printing parameters, including speeding up or slowing down the movement speed of the printer nozzle, lowering or raising the nozzle height, speeding up or slowing down the material discharge speed; and at the same time, reminding the staff to check the material status;

[0146] A display module is used to display the monitoring results output by the layer height detection module during real-time printing;

[0147] A storage module, used to store the collected data and the data in the monitoring process;

[0148] Point cloud boundary segmentation model, used to perform point cloud boundary segmentation and obtain the boundary point cloud of the printed 3D printed concrete wall construction layer;

[0149] The specific implementation process of the floor height detection module is:

[0150] Step 5.1 Voxel downsampling and dimensionality reduction

[0151] The boundary point cloud of the printed concrete wall construction layer is downsampled to obtain a rectangular dot matrix; the x, y and z coordinate values ​​of each point in the rectangular dot matrix are mapped to the corresponding axes in the three-dimensional space to form a spatial point cloud; then the y-axis coordinates of all points are deleted, the three-dimensional space is converted into two-dimensional space, the sum of the x and z coordinates of all points is calculated, the minimum sum is found, and the point corresponding to the minimum sum is set as the coordinate origin. At the same time, the rectangular dot matrix formed by the entire boundary point cloud is translated so that the origin coordinate becomes (0,0), completing the coordinate transformation;

[0152] Step 5.2 Dot grouping

[0153] According to the z coordinates of the points after the coordinate transformation, the points with a z coordinate difference within 5% are grouped into a group to obtain several point sets, and the layer number is marked for each point set in ascending order of the z coordinates; there are m groups of point sets in the printed concrete wall construction layer, and m layers have been printed, where m is an integer greater than 0;

[0154] Step 5.3 Count the number of deviation points

[0155] Preset the layer height threshold, and compare the z coordinates of the point set of the lower layer after translating the layer height threshold with the point set of the adjacent upper layer in turn, record the coordinate points whose z coordinates in the point set of the adjacent upper layer are greater than the coordinates of the point set of the lower layer after translating the layer height threshold as deviation points, and count the number of deviation points in each layer;

[0156] Step 5.4 deletes the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer, and determines whether the number of remaining points in the boundary point cloud of the printed concrete wall construction layer is not greater than 50%. If it is greater than 50%, repeat steps 1.1 to 1.3, re-perform voxel downsampling, dimension reduction, dot matrix grouping, and deviation point number statistics on the remaining points, and obtain the deviation point number and layer number of all printed concrete wall construction layers in the next round; if it is not greater than 50%, stop searching for deviation points;

[0157] Add up the number of deviation points of each layer in all rounds to obtain the final number of deviation points of each layer of the printed concrete wall construction layer;

[0158] Step 5.5 Stability Control

[0159] A stable ratio threshold is set, which is set to 15% in this embodiment. A bar graph is drawn using the final number of deviation points in each layer and the corresponding layer number. The horizontal axis of the bar graph is the layer number, and the vertical axis is the ratio of the number of deviation points in a layer to all the points in the current layer. If the ratio of the number of deviation points in the top layer to all the points in the current layer exceeds the stable ratio threshold, it means that the height of the layer exceeds the layer height threshold, the current printing process is unstable, and early warning adjustment is required. If it does not exceed the stable ratio threshold, the layer height meets the standard and the printing process is stable.

[0160] In this specification, the specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples.

[0161] The present invention aims to solve the problems caused by the current reliance on manual monitoring in the 3D printing process. Traditional manual monitoring methods consume a lot of manpower and economic costs, and are prone to omissions or misjudgments. Manual monitoring is also difficult to quickly and accurately measure the real-time layer height stability during the construction of 3D printed concrete walls, and it is impossible to make a quick and accurate judgment. The present invention can automatically adjust the printing equipment parameters and perform material status inspections. Compared with the prior art, this technical solution has the following advantages and application prospects: ensuring the safety of construction personnel, reducing labor costs, improving printing efficiency, ensuring the quality of printed products, and providing the possibility of pursuing larger-scale architectural-level 3D printing. This is of great significance to the 3D printing industry and has broad application prospects.

[0162] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system.

[0163] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0164] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A 3D printed concrete wall quality monitoring method based on point cloud segmentation, characterized in that: The monitoring method includes the following contents: Obtain point cloud data during the construction of 3D printed concrete walls and normalize the point cloud coordinates; Constructing a neural network model PSC-Net, wherein the neural network model PSC-Net includes an encoder, a decoder, a sub-scene boundary mining module, and a contrast boundary learning module; In the encoder, the original point cloud with an input dimension of N×3 is iteratively sampled at the farthest point to obtain a point cloud with a dimension of The downsampled point cloud is N, N is the number of points in the point cloud, and 3 is the dimension of the spatial three-dimensional coordinates of the point; the original point cloud and the downsampled point cloud are respectively input into two combined multi-layer perceptron modules to obtain two potential feature vectors with the same dimensions, and the two potential feature vectors are concatenated to obtain a potential feature map. The potential feature map is then passed through a multi-layer perceptron and an L2 norm normalization layer to obtain a normalized one-dimensional feature vector V', which is recorded as the output of the encoder; The decoder is used to remap V' back to the original point cloud data space. In the decoder, V' first passes through the first linear layer to obtain a one-dimensional vector V1, and V1 passes through the second linear layer to obtain a one-dimensional vector V2; the length of V2 is half the length of V1, and V1 and V2 are connected to the first fully connected layer FC1 and the second fully connected layer FC2 respectively; the outputs of the first fully connected layer FC1 and the second fully connected layer FC2 are added to output the enhanced point cloud Y detail , Y detail Denote as the output of the decoder; The output of the decoder is connected to a sub-scene boundary mining module and a contrast boundary learning module, the sub-scene boundary mining module is used to determine the label of the point, and the contrast boundary learning module is used to perform boundary segmentation, the sub-scene boundary mining module includes n times of downsampling, the contrast boundary learning module includes n times of upsampling, the result of the first downsampling of the sub-scene boundary mining module is jump-connected with the result of the nth upsampling of the contrast boundary learning module, and the result of the nth downsampling of the sub-scene boundary mining module is jump-connected with the result of the first upsampling of the contrast boundary learning module; The contrast boundary learning module outputs the boundary point cloud; The point cloud data in the concrete wall construction process is used to train the neural network model PSC-Net to obtain the trained neural network model PSC-Net, and the trained neural network model PSC-Net is used to perform point cloud boundary segmentation to obtain the boundary point cloud of the printed concrete wall construction layer; Floor height detection: Step 5.1 Voxel downsampling and dimensionality reduction The boundary point cloud of the printed concrete wall construction layer is downsampled to obtain a rectangular dot matrix; the x, y and z coordinate values ​​of each point in the rectangular dot matrix are mapped to the corresponding axes in the three-dimensional space to form a spatial point cloud; then the y-axis coordinates of all points are deleted, the three-dimensional space is converted into two-dimensional space, the sum of the x and z coordinates of all points is calculated, the minimum sum is found, and the point corresponding to the minimum sum is set as the coordinate origin. At the same time, the rectangular dot matrix formed by the entire boundary point cloud is translated so that the origin coordinate becomes (0,0), completing the coordinate transformation; Step 5.2 Dot grouping According to the z coordinates of the points after coordinate transformation, the points with a z coordinate difference within 5% are grouped together to obtain several point sets, and each point set is marked with a layer number in ascending order of the z coordinates; there are m groups of point sets in the printed concrete wall construction layer, and m layers have been printed, where m is a positive integer; Step 5.3 Count the number of deviation points Preset the layer height threshold, and compare the z coordinates of the point set of the lower layer after translating the layer height threshold with the point set of the adjacent upper layer in turn, record the coordinate points whose z coordinates in the point set of the adjacent upper layer are greater than the coordinates of the point set of the lower layer after translating the layer height threshold as deviation points, and count the number of deviation points in each layer; Step 5.4: Delete the voxel downsampling points from the boundary point cloud of the printed concrete wall construction layer, and determine whether the number of remaining points in the boundary point cloud of the printed concrete wall construction layer is not greater than 50%. If it is greater than 50%, repeat steps 5.1 to 5.3, re-perform voxel downsampling, dimension reduction, dot matrix grouping, and count the number of deviation points on the remaining points to obtain the number of deviation points and layer numbers of all printed concrete wall construction layers in the next round; if it is not greater than 50%, stop searching for deviation points; Add up the number of deviation points of each layer in all rounds to obtain the final number of deviation points of each layer of the printed concrete wall construction layer; Step 5.5 Stability Control Set the stable ratio threshold, and use the final number of deviation points in each layer and the corresponding layer number to draw a bar chart. The horizontal axis of the bar chart is the layer number, and the vertical axis is the ratio of the number of deviation points in a layer to the total number of points in the current layer. If the ratio of the number of deviation points in the top layer to the total number of points in the layer exceeds the stable ratio threshold by 15%, it means that the height of the top layer exceeds the layer height threshold, the current printing process is unstable, and early warning adjustment is required. If it does not exceed the stable ratio threshold, the layer height meets the standard and the printing process is stable.

2. The monitoring method according to claim 1, characterized in that: The combined multi-layer perceptron module includes four combined perceptron layers, each of which is composed of 1, 2, 3, and 4 perceptrons connected in series. The input point cloud data is processed by the four combined perceptron layers to generate four feature vectors. Then, the four feature vectors are subjected to average pooling operations, and the results of the four average pooling operations are spliced ​​to obtain potential feature vectors.

3. The monitoring method according to claim 1, characterized in that: The loss function L of the neural network model PSC-Net is composed of the cross entropy function L of the sub-scene boundary mining module cross entropy And the loss function L of the contrast boundary learning module CBL Composition, expressed by the following formula: Among them B l represents the boundary point set in the real data, x i represents the selected center point, x j Represents x i A point in the neighborhood; l i and l j Represents point x i and x j The true label of f i 、f j Represents point x i 、x j The features of; d() is the distance function; τ represents the temperature hyperparameter in contrastive learning; The category to which the point belongs, A vector representing the probability distribution of the model prediction, n represents the nth sampling stage, λ is the loss weight; N i represents the center point set.

4. A 3D printed concrete wall quality monitoring system based on point cloud segmentation, characterized in that: Executing the monitoring method of claim 1, the system comprises: Point cloud acquisition module, used to obtain point cloud data during the construction process of 3D printed concrete walls; Point cloud preprocessing module, used to preprocess point cloud data to obtain standardized and high-quality point cloud data; Layer height detection module, used to monitor in real time whether the layer height during the construction process meets the printing requirements; The early warning adjustment module is used to warn of abnormal stability detected by the layer height detection module, and adjust the control unit of the 3D printing equipment to adjust the 3D printing parameters, including speeding up or slowing down the movement speed of the printer nozzle, lowering or raising the nozzle height, speeding up or slowing down the material discharge speed; and at the same time, reminding the staff to check the material status; A display module is used to display the monitoring results output by the layer height detection module during real-time printing; A storage module, used to store the collected data and the data in the monitoring process; The point cloud boundary segmentation model is used to perform point cloud boundary segmentation and obtain the boundary point cloud of the printed concrete wall construction layer.

Images