Three-dimensional segmentation method for burst damage of RC structure after fire

Abstract

The invention relates to the technical field of concrete structure damage detection, in particular to a post-fire RC structure burst damage three-dimensional segmentation method, which comprises the following steps: firstly, determining the type and three-dimensional characteristics of post-fire RC structure member surface concrete burst damage, and constructing a corresponding three-dimensional point cloud data set for network training and verification; and then based on a KP-FCNN network structure, a KPConv layer is improved and optimized, so that the detection and segmentation precision is improved, the model size is reduced, the reasoning time is remarkably shortened, the optimal segmentation precision of 82.3% is achieved under the working conditions of different damage degrees, automatic damage segmentation of the burst damage of the concrete structure after the fire disaster is realized, and the automatic damage segmentation of the burst damage of the concrete structure after the fire disaster is realized. And technical support is provided for subsequent damage three-dimensional quantification and deployment to an unmanned aerial vehicle system.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete structure damage detection, and in particular to a three-dimensional segmentation method for burst damage of RC structures after fire. Background Art

[0002] Concrete structure damage detection technology is primarily divided into two broad categories: non-destructive testing (NDT) and destructive testing (DTT). The hammer-rebound method is a representative NDT technique, which uses a hammer to strike the surface of a component and quantifies the extent of concrete damage based on the rebound value. However, this method has significant limitations. Firstly, it strictly requires a high degree of flatness on the surface being tested, making it difficult to effectively detect areas of concrete cracking and spalling, which are common after fires. Secondly, the point-by-point testing method results in low efficiency, making it difficult to meet the timeliness requirements of large-scale post-disaster assessments.

[0003] Core sampling is widely used in destructive testing. This technique, which involves drilling concrete core samples for laboratory analysis, can obtain precise damage parameters, but its inherent drawbacks are significant. The sampling process not only causes secondary damage to damaged structures, but also takes two to three weeks to complete, making it difficult to meet the time-sensitive requirements of post-disaster emergency decision-making. Furthermore, the dust and vibration generated by on-site core drilling pose a potential threat to the safety of test personnel, further limiting the widespread application of this technology.

[0004] While deep learning-based image detection technology has made recent progress in intelligent detection, it still faces fundamental limitations in post-fire concrete damage assessment scenarios. Existing image segmentation algorithms commonly suffer from the following technical bottlenecks: 1) Sensitivity to environmental factors such as lighting conditions and camera angles, resulting in significant fluctuations in detection accuracy; 2) 2D image information lacks depth data, making it impossible to accurately quantify the volumetric parameters and spatial distribution characteristics of burst damage; 3) Model generalization capabilities are insufficient, resulting in poor adaptability to damage morphologies in diverse fire scenarios; and 4) high computational complexity, hindering real-time detection and large-scale application. These limitations hinder the effective acquisition of 3D damage information using existing technologies, hindering the development of temperature field inversion analysis and structural safety assessment.

[0005] In summary, there are still deficiencies in the research on the segmentation of concrete structure damage after fire. Therefore, it is necessary to propose a lightweight 3D reconstruction and 3D target segmentation method for damage to corroded concrete structures after fire that integrates 3D point cloud and deep learning to detect fire damage defects and provide theoretical and data support for subsequent reinforcement and repair. Summary of the Invention

[0006] This invention provides a three-dimensional segmentation method for burst damage in RC structures after fire. It addresses the following technical issues: First, detection efficiency is low. For example, the hammer-rebound method is time-consuming and cumbersome for inspecting damaged components. Second, it can easily cause secondary damage to damaged components. For example, when using core sampling, since the sampling targets are often the most severely damaged components, the sampling process can exacerbate component damage and pose a threat to the safety of on-site inspectors. Finally, the ability to interpret three-dimensional information is insufficient. Image segmentation boundaries are fuzzy, resulting in an error rate of 15%-20% in identifying burst damage areas. The inability to obtain three-dimensional morphology and volume parameters of damage leads to a lack of key input for temperature field inversion calculations. The model also suffers from weak generalization capabilities. For example, image-based neural networks have low segmentation accuracy and are unable to determine the three-dimensional shape and volume of damage. To address these technical challenges, this invention proposes an innovative detection framework that integrates three-dimensional point clouds and deep learning. It offers the following breakthrough advantages: 1. Efficient nondestructive testing; 2. Accurate three-dimensional damage segmentation; and 3. Intelligent feature extraction for multi-scenario adaptation.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] A three-dimensional segmentation method for burst damage of RC structures after fire includes the following steps:

[0009] Step 1: Construct a 3D point cloud database of RC structure damage after fire;

[0010] Step 2: Use KP-FCNN as the baseline network to build a concrete structure 3D point cloud segmentation network;

[0011] Step 3: Use average intersection-over-union (IoU) to systematically evaluate the performance of the 3D point cloud segmentation network and conduct ablation experiments.

[0012] Preferably, the step 1 includes the following specific methods: the first method is to use an ordinary high-definition camera to take photos of concrete burst damage. During the shooting, it is important to take photos of the same damaged component from different angles. In order to obtain a better 3D reconstruction point cloud effect based on the 2D pictures, it is necessary to ensure that each two adjacent pictures have at least 50% overlapping areas. Then all the pictures taken are imported into the colmap software for 3D point cloud reconstruction. The second method is to use the iPhone's built-in laser radar and camera in conjunction with the 3D Scanner The app scans the damaged components and automatically generates a three-dimensional reconstructed point cloud map; the third method is to use a depth camera to take pictures of the fire-damaged components and then directly obtain a point cloud map; the fourth method is to use an R8 lidar to scan the fire-damaged components and then obtain a point cloud map; in addition, it also includes a method of synthesizing data, first modeling the fire-damaged components through three-dimensional modeling software, and then importing the exported obj file into the cloudcompare software for upsampling operation, thereby obtaining a high-density three-dimensional point cloud map; all the data sets obtained by the above methods are imported into the CloudCompare software for point cloud segmentation and annotation, after outlining the edge contour of the burst damage, set the label to "1", and then set the background label to "0"; after all are annotated, export the txt format file, expand the original annotated data set through data enhancement, and then divide the expanded data set into training set, validation set and test set in a ratio of 7:2:1 to facilitate subsequent network training.

[0013] Preferably, in step 2, the improvement of the baseline network specifically includes:

[0014] (21) Introducing multi-scale dilated convolution to achieve a dynamic, multi-level receptive field expansion mechanism; this mechanism forms a multi-scale feature perception capability by combining convolution kernels with different dilation rates, thereby improving the ability to capture complex geometric structures while maintaining parameter efficiency;

[0015] (22) The cross-attention mechanism is introduced to achieve a dynamic balance between global semantic association and local feature enhancement; by dynamically establishing the semantic association of long-distance point cloud features, the local receptive field limitation is overcome, and the intelligent suppression of noise interference is achieved with the help of an adaptive weight allocation strategy;

[0016] (23) The linear deformable convolution module is introduced into the KPConv layer to implement a dynamic and adaptive kernel adjustment mechanism. The dynamic offset is learned for each convolution kernel in a data-driven manner, so that it can adaptively adjust the shape and size according to the local geometric features of the input point cloud.

[0017] Preferably, in step 2, the cross-attention module is deeply fused with the dilated convolution layer to form a "local-global" collaborative feature enhancement paradigm, that is, the dilated convolution expands the local perception range through sparse sampling, and the cross-attention further aggregates cross-region contextual information on this basis.

[0018] Beneficial effects of the three-dimensional segmentation method for burst damage of RC structures after fire provided by the present invention:

[0019] 1. The present invention provides a method for expanding a three-dimensional point cloud damage dataset based on a synthetic three-dimensional model: Given that neural network training requires a large amount of data, it is difficult to meet the training requirements by relying solely on real data. To this end, the present invention proposes a method for synthesizing data. First, the fire-damaged component is modeled using three-dimensional modeling software (such as SketchUp Pro) to generate a virtual damaged component. Then, the exported obj file is imported into the CloudCompare software for upsampling to obtain a high-density three-dimensional point cloud map. This method can fundamentally expand the amount of original data and improve the adequacy and accuracy of neural network training.

[0020] 2. The present invention proposes a dynamic receptive field expansion mechanism based on multi-scale dilated convolution: To address the problem of insufficient multi-scale defect capture caused by the fixed receptive field of KPConv, a multi-level dilated convolution collaborative optimization method is proposed. By combining convolution kernels with dilation rates of 1, 3, and 5, a three-level feature extraction network is constructed to achieve dynamic perception and cross-scale fusion of local details (dilation rate 1), mesoscale (dilation rate 3), and large-scale (dilation rate 5) features. While maintaining parameter efficiency, this mechanism expands the effective receptive field to more than 25 times that of traditional convolution, significantly improving the multi-scale characterization capability of concrete burst damage.

[0021] 3. This paper proposes a cross-attention mechanism and a local-global collaborative optimization paradigm: This paper embeds the cross-attention mechanism into the KPConv network architecture and breaks through the field of view limitations of local convolution by dynamically modeling the semantic associations of long-range point cloud features. Specifically, it includes: global semantic association modeling: explicitly establishing the spatial dependencies of cross-region burst damage to capture the overall pattern of damage distribution; noise suppression and adaptive weight allocation: dynamically filtering key features through attention weights to suppress the interference of scanning noise and sparse areas; local-global feature collaboration: deeply integrating the expanded receptive field of dilated convolution with the global information aggregation of cross-attention, forming a dual optimization mechanism that takes into account both local geometric details and global spatial layout.

[0022] 4. The present invention provides a geometric adaptive kernel adjustment technology for linear deformable convolution: a data-driven linear deformable convolution (LDConv) module is proposed, which realizes adaptive adjustment of kernel shape and coverage by dynamically learning the offset of the convolution kernel. It can achieve the following effects: geometric deformation sensitive feature extraction: for irregular deformation of post-disaster concrete (such as spalling boundaries), the kernel sampling position is dynamically adjusted to enhance the characterization ability of complex topological structures; multi-dimensional attention collaborative optimization: kernel deformation perception is realized in the spatial dimension, which complements the channel attention mechanism and improves the robustness to heterogeneous features such as aggregate distribution and reinforcement differences; lightweight plug-and-play design: adopts a low-parameter modular structure and seamlessly integrates into the KPConv framework to support efficient training and real-time processing.

[0023] 5. This invention provides a multi-module collaboratively optimized 3D point cloud segmentation network architecture: Based on the aforementioned technologies, the KPConvDef-DC and KP-G-FCNN network architectures systematically integrate multi-scale dilated convolution, cross-attention, and linear deformable convolution modules to form a collaborative technology system of "dynamic receptive field expansion, global semantic association, and geometric adaptive optimization." This architecture overcomes the limitations of traditional 3D segmentation models in complex post-disaster scenarios, significantly improving the segmentation accuracy of blast damage (MIoU reaching 82.31%), noise robustness, and multi-scale adaptability, providing a highly reliable solution for post-fire concrete structure safety assessment.

[0024] 6. The present invention has the following advantages in practical applications: a. Seamless adaptability of multi-source heterogeneous devices: supports instant access to multi-modal point cloud acquisition terminals such as mobile phones, radars, and depth cameras, and realizes standardized pre-processing and real-time reasoning of cross-platform image data through lightweight model architecture design, meeting the flexible detection needs in complex environments at fire scenes. b. Engineering adaptability for dynamic deployment: provides a multi-level deployment solution from embedded devices to server clusters, and develops a dedicated reasoning acceleration interface specifically for drone inspection systems to achieve an automated closed-loop operation of aerial point cloud acquisition-damage identification-result return. c. Data foundation for damage quantification analysis: The output point cloud segmentation results are stored in a structured data format to accurately locate key damage features such as the area, depth, and volume of the concrete burst area, providing high-precision input data for subsequent damage level assessment, structural bearing capacity calculation, and temperature field calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the steps of the method of the present invention.

[0026] Figure 2It is the KP-FCNN network architecture (Note: KP-FCNN: point cloud segmentation network; points: point cloud; features: features; classes: categories; KPConv: Kernel Point Convolution kernel point convolution; strided KPConv: strided kernel point convolution; 1conv: 1 convolution; Neur: Ups.+Concat: upsampling + connection in neural network; Skip link: skip link).

[0027] Figure 3 It is the KPConv layer network architecture (Note: KPConv layer: KPConv layer; Features dimension: feature dimension; Maxpool: maximum pooling; ReLU: rectified linear unit, an activation function widely used in deep learning; BN: (Batch Normalization) is a technology in deep learning, mainly used to accelerate the training process of neural networks and improve the stability of the model).

[0028] Figure 4 It is the KPConvDef network architecture (Note: PointConv: point convolution; offset field: refers to the field used to represent the offset of the pixel position of the convolution kernel in the deformable convolution operation; offsets: offset; Deformableconvolution: deformable convolution; input feature map: input feature map; out feature map: output feature map).

[0029] Figure 5 It is the Cross-Attention network architecture (Note: Cross-Attention: cross attention mechanism; Embedding size: embedding dimension; Number of tokens: number of tokens; input: input; output: output; new: new; weights: weights; Attention matrix: attention matrix; Softmax: a classification function).

[0030] Figure 6 It is the KPConvDef-DC network architecture (Note: Bias+Weight: displacement + weight; Offset(x,y)andweight coefficients: bias(x,y) and weight coefficients).

[0031] Figure 7It is the LDConv network architecture (Note: Generating Algorithm: Generating algorithm; Based on the size of N: Based on the size of N; Generating initial sampled shapes: Generating initial sampled shapes; Original Coordinate: Original coordinate; Modified Coordinate: Modified coordinate; Resample: Resampling; Initial sampled shapes are adjusted by offsets: Initial sampled shapes are adjusted by offsets; Norm: Normalization; SiLU: A new type of activation function).

[0032] Figure 8 It is the KPConv-G network architecture.

[0033] Figure 9 It is the KP-G-FCNN network architecture.

[0034] Figure 10 It is a prediction effect diagram.

[0035] Figure 11 It is the Loss value curve.

[0036] Figure 12 It is the Miou value curve graph. DETAILED DESCRIPTION

[0037] The following describes in detail the implementation methods of the present invention in a step-by-step manner. This description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0038] In the description of the present invention, it should be noted that the terms "up", "down", "left", "right", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings. They are only for describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, and a specific orientation structure and operation. Therefore, they cannot be understood as limiting the present invention.

[0039] Example 1

[0040] A three-dimensional segmentation method for burst damage of RC structures after fire includes the following steps:

[0041] Step 1: Construct a 3D point cloud database of RC structure damage after fire;

[0042] Step 2: Use KP-FCNN as the baseline network to build a concrete structure 3D point cloud segmentation network;

[0043] Step 3: Use average intersection-over-union (IoU) to systematically evaluate the performance of the 3D point cloud segmentation network and conduct ablation experiments.

[0044] Example 2

[0045] Based on Example 1, this embodiment discloses:

[0046] Post-fire concrete structure burst damage data has long been scarce in the field of civil engineering because its collection process faces multiple technical bottlenecks and safety risks. First, the structural safety hazards at high-temperature disaster sites greatly limit the frequency of field surveys, resulting in an extremely short window for collecting original data. Second, the spatiotemporal distribution of damage samples has significant heterogeneity, and existing detection methods are difficult to fully capture the three-dimensional damage characteristics. Third, such special working condition data in public databases are highly fragmented, and the data formats and collection standards between different institutions lack uniformity, which seriously restricts the generalization ability of damage models. In response to this technical dilemma, this paper summarizes and concludes Four methods were used to construct a 3D point cloud database of RC structure damage after fire that can be used to train 3D point cloud segmentation models. Specifically, the first method is to use an ordinary high-definition camera to take photos of concrete burst damage. During the shooting, it is important to take photos of the same damaged component from different angles. In order to obtain a better 3D reconstruction point cloud effect based on 2D images, it is necessary to ensure that each adjacent photo has at least 50% overlapping area. Then all the photos taken are imported into the colmap software for 3D point cloud reconstruction. The second method is to use the iPhone's built-in lidar and camera and cooperate with 3D The Scanner App scans the damaged component and automatically generates a 3D reconstructed point cloud map. Method 3 uses a depth camera to take a picture of the fire-damaged component and then directly obtains a point cloud map. Method 4 uses an R8 lidar to scan the fire-damaged component and then obtain a point cloud map. In addition, since a large amount of data is required to achieve good results in neural network training, it is not enough to train the neural network with real data alone. In order to fundamentally expand the amount of original data, the present invention proposes a method for synthesizing data. First, the fire-damaged component is modeled using 3D modeling software (such as Sketch Up Pro), and then the exported obj file is imported into the CloudCompare software for upsampling, thereby obtaining a high-density 3D point cloud map. The specific processes of the five methods are as follows: Figure 1 As shown in the figure, the comprehensive application of these methods can provide more abundant and accurate data support for the study of burst damage of concrete structures after fire, and promote technological progress and application development in this field.

[0047] In this embodiment, the number of point cloud models of the constructed concrete structure fire damage point cloud dataset is 1000, including 700 real data and 300 synthetic data; all of them are imported into CloudCompare software for point cloud segmentation and annotation, and after outlining the edge contour of the burst damage, the label is set to "1", and the background label is set to "0"; after all are annotated, a txt format file is exported, and the original annotated dataset is expanded through data enhancement methods. Then, the expanded dataset is divided into training set, validation set and test set in a ratio of 7:2:1 to facilitate subsequent network training.

[0048] Example 3

[0049] Based on Example 2, this example discloses the specific implementation of step 2:

[0050] 2.1 About the baseline network:

[0051] KP-FCNN is a fully convolutional network designed for complex scene segmentation. Its network architecture is as follows: Figure 2 As shown in the figure, its core innovation lies in the integration of the advantages of Kernel Point Convolution (KPConv) and Fully Convolutional Network (FCN); in the task of 3D point cloud segmentation of concrete structure explosion after fire, the use of KP-FCNN as the baseline network has significant necessity and technical advantages, which are mainly reflected in the following three aspects: First, accurate modeling of complex geometric structures. After a fire, concrete structures often undergo irregular deformations such as cracking and spalling due to high temperatures. The point cloud data exhibits uneven distribution, strong noise interference, and complex local geometric features. KP-FCNN introduces the Kernel Point Convolution (KPConv) module, which uses learnable kernel points to dynamically capture local geometric structures and effectively adapt to the irregular shapes of concrete burst fragments. Secondly, it deeply integrates multi-scale contextual information. The semantic classification of fire-damaged areas (such as cracked areas, undamaged areas, and exposed steel bars) requires a combination of local details and global context. The fully convolutional U-shaped structure of KP-FCNN achieves multi-level feature fusion through skip connections. Finally, the flexibility and scalability of 3D point cloud processing. Point cloud data of RC structures after a fire often suffers from missing points, noise, and scale changes. The fully convolutional characteristics of KP-FCNN enable it to directly process raw point clouds without the need for preprocessing such as voxelization or projection, thus avoiding information loss. At the same time, the network structure has good scalability.

[0052] KPConv (Kernel Point Convolution) is a point cloud semantic segmentation model based on convolution operation. A new point convolution operation is used in the KPConv network. By defining the convolution kernel point in Euclidean space and performing convolution operation with the points in the point cloud, the deformable convolution can be extended to the point convolution operation. Its network architecture is as follows: Figure 3 However, in the concrete structure burst segmentation task, it has the following limitations:

[0053] (1) The local receptive field is fixed and the multi-scale feature capture is insufficient: KPConvDef (such as Figure 4 ) performs feature aggregation within a fixed radius and cannot adaptively capture geometric structures of different scales (such as tiny cracks and large-scale damaged areas in concrete bursts). For complex, multi-scale burst defects in concrete blocks (such as differences in crack width and depth), a single-scale receptive field may miss key details.

[0054] (2) Weak long-distance dependency modeling capability: KPConvDef mainly relies on local neighborhood information and lacks the ability to model the global association of long-distance point cloud features. For example, the burst area in a concrete block may span a large spatial range (such as a through-going crack), and local convolution is difficult to capture these long-distance dependencies.

[0055] (3) Insufficient robustness to noise and sparse areas: In concrete point cloud data, scanning noise, occlusion, or sparsely sampled areas may lead to incomplete local features; the fixed kernel point strategy of KPConvDef is sensitive to noise and cannot effectively fuse multi-scale contextual information to enhance robustness.

[0056] 2.2 Network Improvements:

[0057] 2.2.1 Introduction of Multi-scale Dilated Convolution:

[0058] Multi-scale dilated convolution enables a dynamic, multi-level receptive field expansion mechanism. This mechanism combines convolution kernels with different dilation rates to form multi-scale feature perception capabilities, significantly improving the ability to capture complex geometric structures while maintaining parameter efficiency. Introducing the multi-scale dilated convolution module into KPConvDef overcomes its fixed local receptive field limitation, enabling it to adapt to the multi-scale characteristics of concrete defects, from tiny cracks to large cracks and cracks of varying depths. The specific effects are as follows:

[0059] (1) Enhanced multi-scale feature fusion capabilities:

[0060] Multi-level receptive field collaboration: A three-level feature extraction network is constructed using convolution kernels with dilation rates of 1, 3, and 5. Convolution with a dilation rate of 1 focuses on local details, while convolutions with dilation rates of 3 and 5 cover medium- and large-scale areas (such as large-scale crack surfaces), respectively. Features from the three branches are fused across scales through channel splicing to form a complete crack defect representation.

[0061] (2) Parameter efficiency optimization:

[0062] Advantages of the dilated convolution structure: Compared with traditional stacked convolution, the dilated convolution's dilated sampling mode can expand the effective receptive field to more than 5 times the radius (for example, the coverage area is expanded 25 times when the dilation rate is 5) without increasing the number of parameters. This allows the model to capture a wider range of contextual information at a lower computational cost.

[0063] (3) Breakthrough of geometric structure limitations:

[0064] Dynamically adapt to defect morphology: Through the combination of multi-level expansion rates, the contradiction that the fixed radius kernel of KPConvDef is difficult to capture both millimeter-level and centimeter-level spalling is resolved, which can effectively improve the accuracy of the model in detecting shallow cracks.

[0065] 2.2.2 Introducing Cross-Attention Mechanism:

[0066] The Cross-Attention mechanism (such as Figure 5 ) is introduced into the KPConvDef network architecture, which can achieve a dynamic balance between global semantic association and local feature enhancement; this improved mechanism effectively breaks through the local receptive field limitation of traditional convolution kernels by dynamically establishing the semantic association of long-distance point cloud features, and at the same time realizes intelligent suppression of noise interference with the help of adaptive weight allocation strategy; specifically, the cross-attention module explicitly models the long-distance dependencies between different regions in the point cloud, enabling the model to capture the global distribution pattern of burst damage in concrete structures after fire, such as the spatial correlation between cross-regional burst propagation paths and damage-intensive areas.

[0067] At the feature optimization level, the present invention deeply integrates the cross-attention module with the dilated convolution layer to form a "local-global" collaborative feature enhancement paradigm; the dilated convolution expands the local perception range through sparse sampling, while the cross-attention further aggregates cross-region contextual information on this basis; this dual-perspective feature optimization mechanism enables the model to not only finely depict the local geometric details of the damage area (such as the burst shape and size), but also accurately understand the overall spatial layout of the damage pattern.

[0068] In summary, the improved scheme of "multi-scale dilated convolution + cross attention mechanism" proposed in the KPConvDef network architecture can break through the limitation of fixed receptive field and enhance the ability to capture the multi-scale geometric structure of burst defects. It can also improve long-distance dependency modeling and noise robustness through global correlation modeling and dynamic weight allocation. The improved KPConvDef-DC network architecture is as follows: Figure 6 shown.

[0069] 2.2.3Introducing the Linear Deformable Convolution module:

[0070] The present invention introduces the Linear Deformable Convolution module into the KPConv layer. LDConv can implement a dynamic and adaptive kernel adjustment mechanism. Its network structure is as follows: Figure 7 As shown in the figure, this mechanism breaks through the fixed geometric constraints of traditional convolution kernels and learns dynamic offsets for each convolution kernel in a data-driven manner, enabling it to adaptively adjust its shape and size based on the local geometric features of the input point cloud. This dynamic kernel adjustment mechanism demonstrates significant advantages in addressing the complex characteristics of the post-fire concrete structure burst damage segmentation task:

[0071] (1) Improve feature extraction capabilities:

[0072] While traditional KPConv can effectively process 3D point clouds, its fixed kernel shape has limitations when dealing with irregular deformations of post-disaster concrete. By introducing the linear deformable convolution module, the model can adaptively adjust the sampling position and coverage of the convolution kernel based on the local topological structure of the damaged area (such as spalling boundaries). This mechanism significantly enhances contextual awareness and effectively improves the capture of cross-scale damage features.

[0073] (2) Multi-dimensional extended attention mechanism:

[0074] Linear deformable convolution effectively complements the traditional channel attention mechanism through dynamic kernel adjustment in the spatial dimension. The model not only focuses on areas of significant damage but also adaptively responds to geometric heterogeneity in concrete structures (such as aggregate distribution and reinforcement differences). This multi-dimensional optimization strategy enables feature representation to incorporate both spatial deformation sensitivity and semantic discriminability, demonstrating enhanced robustness in complex post-disaster scenarios.

[0075] (3) Easy to integrate and optimize:

[0076] This module adopts a lightweight design and is seamlessly integrated into the KPConv framework as a plug-and-play component. This not only avoids large-scale modifications to the original network structure, but also simplifies the model optimization process, reduces the consumption of computing resources, and accelerates the training process. This efficient integration feature provides technical support for the real-time processing of large-scale point cloud data after disasters.

[0077] In summary, the introduction of the linear deformable convolution module enables the KPConv layer to form a synergistic mechanism of "geometric adaptation-feature optimization-efficient training", providing a new solution for the refined segmentation of burst damage in concrete structures after fire; the improved KPConv-G layer network architecture is as follows Figure 8 As shown, it replaces the original network architecture KPConv layer to form the KP-G-FCNN network architecture, whose architecture is as follows Figure 9 shown.

[0078] Example 4

[0079] Based on Example 3, this example discloses an implementation method for systematically evaluating the performance of the KP-G-FCNN network model, as follows:

[0080] 1. Training strategies and evaluation indicators:

[0081] The network was trained on a custom workstation equipped with an RTX 3090 graphics card, which has 24GB of VRAM and uses CUDA to accelerate network training. In addition, the workstation integrates two E52699 v3 CPUs, each operating at 2.3GHz, with 64 cores and 128 threads. The network architecture was developed using the Python-based PyCharm integrated development environment and the PyTorch library for deep learning applications.

[0082] The KP-G-FCNN network uses the AdamW optimizer (β1 = 0.9, β2 = 0.999, weight decay = 0.01), adopts a piecewise constant decay strategy, and the initial learning rate is set to 0.01, which decays by 0.1 times every 5 epochs. Through the learning rate warm-up technology, the first two epochs use 0.1 times the initial value to alleviate the instability of the optimizer in the initial stage. The data enhancement methods used include geometric transformation (random rotation (±15°), scaling (0.8-1.2 times) and translation (±0.2 meters) are applied to the point cloud to improve the robustness of the model to spatial transformation) and neighborhood perturbation. (randomly adding Gaussian noise (standard deviation 0.01 meters) in the local neighborhood to simulate the measurement error of the real scene), density adjustment (enhancing the model's adaptability to different point cloud densities through random downsampling (retaining 80%-100% of the points)), etc.; in order to prevent feature loss caused by excessive deformation and prevent overfitting during network training, a regularization strategy of kernel point offset constraint and weight decay is adopted; the training batch is set to 16, the training cycle is 300 rounds, and the loss function adopts cross entropy loss (CrossEntropyLoss) combined with label smoothing (LabelSmoothing) technology, and the smoothing parameter is set to 0.1.

[0083] This paper systematically evaluates the performance of the KP-G-FCNN network model using the Mean Intersection over Union (MIoU). MIoU is used as the core segmentation accuracy indicator. The average overlap between the predicted area and the true labeled area in each category is calculated to quantitatively characterize the model's positioning accuracy for the fire damage area. The calculation formulas are shown in Equations (1)-(4).

[0084]

[0085]

[0086]

[0087]

[0088] In formula (1), the numerator is the sum of all correctly predicted pixels, the denominator is the total number of pixels in the image, and FN (False Negative), FP (False Positive), TN (True Negative), and TP (True Positive) are the values in the confusion matrix.

[0089] 2. Ablation test results:

[0090] This paper presents an improved method for 3D point cloud segmentation models based on multi-module collaborative optimization. Systematic ablation experiments verify the effectiveness and synergistic advantages of each technical module. The experiment uses the KP-FCNN model as a benchmark and gradually introduces the Dilated Conv module, the CA attention mechanism, and the LD Conv module to quantitatively analyze the impact of each module on the mean over (MIoU) segmentation accuracy. The experimental results are shown in Table 1. Based on the baseline model KP-FCNN (MIoU 69.21%), the effects of the introduction and combination strategies of each module on segmentation accuracy are as follows:

[0091] (1) Single module analysis:

[0092] The Dilated Conv module (No. 2) increases MIoU to 76.45%, indicating that it effectively enhances the ability to extract local features by expanding the receptive field;

[0093] After the introduction of the CA attention mechanism (No. 3), the MIoU reaches 74.67%, verifying the optimization effect of channel attention on feature screening;

[0094] The LD Conv module (No. 4) performs best, with an MIoU of 78.64%, demonstrating the significant advantages of its layered dynamic convolution in modeling complex spatial structures.

[0095] (2) Dual-module synergy effect:

[0096] The combination of Dilated Conv and CA (No. 5) increases MIoU to 79.41%, which is a significant gain compared to a single module, reflecting the complementarity of feature enhancement and attention screening.

[0097] The combination of CA and LD Conv (No. 7) achieves an MIoU of 81.85%, making it the optimal dual-module solution. This indicates that the collaborative optimization of the attention mechanism and dynamic convolution can significantly improve the efficiency of global-local feature fusion.

[0098] The MIoU of the combination of Dilated Conv and LD Conv (No. 6) is 78.48%, which is slightly lower than that of the single LD Conv module. It is speculated that this is because both focus on local feature expansion and have a redundant effect.

[0099] (3) Full module integration optimization:

[0100] When all modules are integrated (No. 8), the MIoU reaches 82.31%, which is 0.46% higher than the optimal dual-module combination (No. 7), verifying the cumulative gain effect of multi-module collaborative optimization. Although the marginal benefit is reduced, the scheme maximizes the segmentation accuracy through comprehensive feature enhancement, attention guidance and dynamic convolution mechanism.

[0101] Table 1 Ablation test results

[0102] Serial number KP-FCNN DilatedConv CA LDConv MIoU 1 √ 69.21％ 2 √ √ 76.45％ 3 √ √ 74.67％ 4 √ √ √ 78.64％ 5 √ √ √ 79.41％ 6 √ √ √ 78.48％ 7 √ √ √ 81.85％ 8 √ √ √ √ 82.31％

[0103] The above results show that the multi-module collaborative optimization method proposed in this invention significantly improves the performance of the three-dimensional point cloud segmentation model through the organic combination of differentiated technical paths, among which the synergistic effect of the CA attention mechanism and the LD Conv module is particularly prominent.

[0104] The loss function images of KP-G-FCNN and KP-FCNN networks are as follows Figure 11 As shown in the figure, the results of MIoU values with the number of training rounds are as follows Figure 12 As shown. Figure 11 、 12 It can be seen that:

[0105] (1) For the loss value curve, both the KP-G-FCNN and KP-FCNN models show a rapid decrease in loss value in the first 30 time steps. Subsequently, the rate of decrease in loss value slows down between time steps 30 and 100, and finally stabilizes after 250 training time steps until the training process is completed. It is worth noting that no overfitting is observed during training or validation.

[0106] (2) As for the miou value curve, in the first 50 time steps, both the KP-G-FCNN and KP-FCNN models show a rapid increase in loss value. Subsequently, the KP-G-FCNN model continues to rise steadily to 0.828 and then stops rising, while the KP-FCNN model rises to 0.692 and then stops rising. This shows that the improved model has improved the miou value by 13.6%.

[0107] The model prediction results of the present invention are shown in the following figure: Figure 10 As shown in the figure, the model achieves accurate segmentation for both original data and synthetic data, with good results, and can accurately segment burst damage of different sizes.

Claims

1. A three-dimensional segmentation method for burst damage in RC structures after fire, comprising the following steps: Step 1: Construct a 3D point cloud database of RC structure damage after fire; Step 2: Use KP-FCNN as the baseline network to build a concrete structure 3D point cloud segmentation network; Step 3: Use average intersection-over-union (IoU) to systematically evaluate the performance of the 3D point cloud segmentation network and conduct ablation experiments.

2. A method for 3D segmentation of burst damage in RC structures after fire as described in claim 1, characterized in that step 1 comprises the following specific methods: The first method is to use an ordinary high-definition camera to take photos of the concrete burst damage, and pay attention to photographing the same damaged component from different angles during the shooting. In order to obtain a good 3D point cloud reconstruction effect based on the 2D images, it is necessary to ensure that each adjacent photo has at least 50% overlap. Then, all the photos are imported into the colmap software for 3D point cloud reconstruction; the second method is to use the iPhone's built-in laser radar and camera in conjunction with 3D Scanner The app scans the damaged component and automatically generates a 3D reconstructed point cloud. A third method uses a depth camera to take photos of the fire-damaged component and directly obtain a point cloud. A fourth method uses an R8 lidar to scan the fire-damaged component and obtain a point cloud. In addition, a method for synthesizing data is also included. First, the fire-damaged component is modeled using 3D modeling software, and then the exported .OBJ file is imported into CloudCompare software for upsampling to obtain a high-density 3D point cloud. All datasets obtained by the above methods are imported into CloudCompare software for point cloud segmentation and annotation. After outlining the edge of the burst damage, the label is set to "1", and the background label is set to "0". After all the annotations are completed, the file is exported in txt format. The original annotated dataset is expanded through data augmentation, and the expanded dataset is divided into training, validation, and test sets in a ratio of 7:2:1 to facilitate subsequent network training.

3. A three-dimensional segmentation method for burst damage of RC structures after fire according to claim 2, characterized in that: In step 2, the baseline network is improved, specifically including: (21) Introducing multi-scale dilated convolution to achieve a dynamic, multi-level receptive field expansion mechanism; this mechanism forms a multi-scale feature perception capability by combining convolution kernels with different dilation rates, thereby improving the ability to capture complex geometric structures while maintaining parameter efficiency; (22) The cross-attention mechanism is introduced to achieve a dynamic balance between global semantic association and local feature enhancement; by dynamically establishing the semantic association of long-distance point cloud features, the local receptive field limitation is overcome, and the intelligent suppression of noise interference is achieved with the help of an adaptive weight allocation strategy; (23) The linear deformable convolution module is introduced into the KPConv layer to implement a dynamic and adaptive kernel adjustment mechanism. The dynamic offset is learned for each convolution kernel in a data-driven manner, so that it can adaptively adjust the shape and size according to the local geometric features of the input point cloud.

4. A three-dimensional segmentation method for burst damage of RC structures after fire according to claim 3, characterized in that: In step 2, the cross-attention module is deeply fused with the dilated convolution layer to form a "local-global" collaborative feature enhancement paradigm, that is, the dilated convolution expands the local perception range through sparse sampling, while the cross-attention further aggregates cross-region contextual information on this basis.

Images