Concrete 3D structure real-time reconstruction method based on multi-mode vibration feature fusion and AIGC

Abstract

The invention discloses a concrete 3D structure real-time reconstruction method based on multi-modal vibration feature fusion and AIGC. The method comprises the steps that S1, a multi-modal concrete vibration data set is constructed; s2, constructing a concrete vibration multi-mode fusion finite feature sequence; s3, constructing a concrete vibration knowledge embedding learning feature sequence; s4, establishing a concrete aggregate-bubble space topological structure; s5, obtaining a two-dimensional concrete aggregate-bubble shape sample; s6, generating a corresponding real 3D structure, and reconstructing a concrete 3D structure sample in combination with the concrete aggregate-bubble spatial topological structure; s7, performing data enhancement on the concrete 3D structure sample; s8, constructing a data set based on the steps S1-S7, and carrying out model training; based on sound-light-force-electricity-touch multi-source and multi-mode fusion concrete vibration data and vibration field knowledge, a concrete 3D structure is reconstructed in real time by an AIGC means.

Description

Technical Field

[0001] The present invention belongs to the field of concrete construction in building engineering, and specifically relates to a real-time reconstruction method of concrete 3D structure based on multimodal vibration feature fusion and AIGC. Background Art

[0002] Concrete vibration is a critical step in ensuring the quality of concrete construction, and its quality directly impacts the performance of concrete structures. However, due to the extensive and subjective nature of construction management, concrete vibration often suffers from quality issues such as under-vibration and over-vibration. Under-vibration refers to insufficient vibration time, which results in a large number of harmful bubbles remaining within the concrete and a lack of a uniform and dense three-phase distribution of aggregate, mortar, and bubbles. Over-vibration, on the other hand, refers to excessive vibration time, leading to concrete segregation and bleeding. Both under-vibration and over-vibration are concrete vibration quality issues and should be avoided during construction.

[0003] Current methods for evaluating the quality of concrete vibration construction are primarily categorized as in-process and post-process. In-process quality evaluation focuses on the vibration process of fresh concrete and primarily relies on manual observation, using the criteria for acceptable vibration as "coarse aggregate no longer significantly sinks and begins to slurry" [1,2]. However, this method is subject to significant subjectivity and labor-intensiveness. Furthermore, since fresh concrete is an opaque, heterogeneous substance, vibration operations are concealed, making it difficult to determine its internal density based solely on visual information. Therefore, this method is insufficient to fully characterize the quality of concrete vibration. Post-process quality evaluation focuses on poured, hardened concrete and primarily relies on coring the concrete structure. Core sampling, CT cross-sectional scanning, or mechanical property testing are then performed, or in-hole television imaging is performed based on the drilled holes. Core sampling, CT cross-sectional scanning, and in-hole television imaging can all provide information on the distribution of aggregate, mortar, and pore structure within the concrete, enabling qualitative or quantitative evaluation of concrete uniformity and density. However, quality inspection based on core sampling analysis has two shortcomings: first, the visual modal information based on the surface or cross-section of the core sample is mainly a two-dimensional plane image or a mosaic of two-dimensional images, which still cannot reflect the 3D structure of concrete; second, core sampling will damage the concrete structure. At the same time, it is subject to factors such as the stress conditions of the concrete structure, the size of the core drill, and embedded parts. This results in a limited number of concrete core samples, making it difficult to comprehensively evaluate the three-phase distribution of aggregate-mortar-air bubbles inside the concrete.

[0004] In summary, current concrete vibration quality assessment methods lack 3D structure-specific detection methods. Existing methods based on cross-sections of hardened concrete core samples or borehole television image data can only provide two-dimensional distribution information of concrete aggregate, mortar, and bubbles. These data are sourced from a single modality and cannot provide three-dimensional distribution information. Furthermore, due to the destructive nature of coring, sample size is limited, making it difficult to obtain a complete picture of the concrete's internal structure. Furthermore, concrete core-based testing has a significant lag. Once a problem is detected, the cast structure must be destroyed and reconstructed, making it difficult to provide a reference for vibration operations during construction. Therefore, it is necessary to fully exploit the complementarity and redundancy of multimodal vibration information based on multiple sensing methods, such as sound, light, force, electricity, and touch, to fully explore the vibration process information and reconstruct the concrete's 3D structure in real time. This allows for real-time understanding of the aggregate, mortar, and bubble distribution within the concrete during the vibration process, providing real-time decision support for vibration quality evaluation and refined vibration feedback control. Summary of the Invention

[0005] The present invention proposes a real-time reconstruction method for concrete 3D structure based on multimodal vibration feature fusion and AIGC. The technical problem to be solved is: based on the sound-light-force-electricity-touch multi-source multimodal concrete vibration feature fusion sequence, AIGC is used to reconstruct the concrete 3D structure in real time.

[0006] The present invention is implemented through the following technical solutions:

[0007] The real-time reconstruction method of concrete 3D structure based on multimodal vibration feature fusion and AIGC includes the following steps:

[0008] S1. Based on the collaborative detection of concrete vibration process by multi-source sensors of sound, light, force, electricity and touch, a multi-modal concrete vibration dataset is constructed.

[0009] S2. By extracting the feature information of different scales from the multimodal concrete vibration dataset, a finite feature sequence of multimodal fusion of concrete vibration is constructed;

[0010] S3. Extract the knowledge features of concrete vibration domain through text extractor and construct the concrete vibration knowledge embedding learning feature sequence;

[0011] S4. Inputting the concrete vibration multimodal fusion finite feature sequence and the concrete vibration knowledge embedded learning feature sequence into a deep residual network at the same time to establish a concrete aggregate-bubble space topological structure;

[0012] S5. Based on the collected concrete cross-section scanning image data, a two-dimensional concrete aggregate-bubble shape sample is obtained by instance segmentation;

[0013] S6, generating a corresponding real 3D structure based on the two-dimensional concrete aggregate-bubble shape sample, and combining the concrete aggregate-bubble spatial topological structure to reconstruct a concrete 3D structure sample;

[0014] S7, performing data enhancement on the concrete 3D structure sample based on latent pattern autonomous learning;

[0015] S8, constructing a data set based on steps S1-S7, and carrying out model training to obtain a prediction model from a multi-modal vibration limited feature sequence to a concrete 3D structure.

[0016] Further, the process of inputting the concrete vibration multi-modal fusion limited feature sequence and the concrete vibration knowledge embedding learning feature sequence into the deep residual network for training and establishing the concrete aggregate-bubble spatial topological structure in step S4 includes:

[0017] Predicting the joint distribution of the distance and angle between the aggregate and the bubble in the concrete aggregate-bubble spatial topological structure;

[0018] Taking the distance-angle joint distribution as a constraint condition, and converting the prediction problem of the distance and angle between the concrete aggregate and the bubble into a differentiable function optimization problem;

[0019] Solving the above differentiable function optimization optimization problem based on the gradient descent method to establish the concrete aggregate and bubble spatial topological structure.

[0020] The process of obtaining a two-dimensional concrete aggregate-bubble shape sample by instance segmentation on the collected concrete cross-section scan image in step S5 includes:

[0021] Selecting a certain number of cross sections in the height direction of the hardened concrete core sample at equal intervals to generate concrete cross-section scan image data by CT; cutting the concrete cross-section scan image data into small size image blocks, and adopting a parallel computing strategy to carry out instance segmentation; wherein: adopting a FocalLoss classification loss function to balance the different categories of data in the concrete cross-section scan image data.

[0022] Further, the process of generating a corresponding real 3D structure based on the two-dimensional concrete aggregate-bubble shape sample and combining the concrete aggregate-bubble spatial topological structure to reconstruct a concrete 3D structure sample in step S6 includes:

[0023] Based on the real shape of the concrete aggregate in three-dimensional laser scanning, and regarding the appearance shape as a high-dimensional distribution, learning its distribution pattern, so as to correct the generated aggregate appearance shape to make it conform to the real distribution and meet the condition of spatial collision constraint;

[0024] The learned distribution pattern is transferred to the concrete aggregate-bubble spatial topology, corresponding objects are generated at the corresponding positions of aggregates and bubbles, and rendered based on neural style transfer, thereby realizing the visualization of the internal 3D structure of concrete.

[0025] Furthermore, in step S7, data enhancement is performed on the concrete 3D structure sample based on the potential pattern autonomous learning, specifically including:

[0026] Variational autoencoders are used to pre-train the latent space of samples in a concrete 3D data reconstruction model. This allows encoding of arbitrary 3D concrete structure samples into the latent space, ensuring that redundant information in the 3D samples is compressed while improving the semantic consistency of the samples.

[0027] A three-dimensional sample simulation is used to generate classifier-free guidance, which is then distilled into a variational autoencoder as the input parameter of a concrete three-dimensional data reconstruction model.

[0028] Using skipping strategy to calculate consistency loss to improve the distillation process of variational autoencoder;

[0029] A neural style transfer network is used to improve the similarity between the generated 3D samples of concrete internal structure and the real 3D samples.

[0030] Beneficial effects

[0031] It can be seen from the above technical solutions that the present invention has the following advantages:

[0032] First, the method for real-time reconstruction of concrete 3D structures based on multimodal vibration feature fusion and AIGC, provided by the present invention, achieves real-time prediction from a multimodal finite feature sequence of concrete to an ultra-complex internal three-dimensional space based on AIGC technology. This demonstrates AIGC's outstanding creative generation and complex data processing capabilities, fills a gap in the field of real-time reconstruction of ultra-complex 3D structures in concrete vibration construction, and provides new real-time decision-making support for concrete vibration quality evaluation.

[0033] Second, the real-time reconstruction method for concrete 3D structure based on multimodal vibration feature fusion and AIGC provided by the present invention achieves in-depth exploration of the various dependencies between the finite sequence of multimodal vibration information and the concrete 3D structure based on the AIGC model. This avoids the lag and destructiveness of traditional core drilling analysis of concrete 3D structure, and also achieves an upgrade from the traditional "human + tool" primary exploration mode to the AIGC + exploration mode.

[0034] Third, the real-time reconstruction method of concrete 3D structure based on multimodal vibration feature fusion and AIGC provided by the present invention is based on the multimodal vibration information of concrete obtained by the collaborative detection of sound, light, force, electricity and touch, which fully explores the complementarity and redundancy between different modal data, and realizes the feature fusion of multi-granularity, multi-scale and heterogeneous vibration multimodal information. Compared with the single modal quality evaluation method, it has stronger generalization and robustness, and can more fully reflect the quality of concrete vibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the process for predicting the internal aggregate-air bubble topological structure of concrete provided in an embodiment of the present application;

[0036] Figure 2 Schematic diagram of the intelligent segmentation solution for aggregate and air bubbles in a concrete cross-section scan image provided in an embodiment of the present application;

[0037] Figure 3 Schematic diagram of the process of enhancing 3D samples of concrete internal structure based on autonomous learning of latent patterns provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] Currently, concrete vibration quality evaluation is based on in-process appearance evaluation and subsequent core sample image or in-hole television analysis. This relies on single-modal visual data and cannot fully reflect the three-phase distribution of aggregate, mortar, and air bubbles in the concrete during the vibration process, making it difficult to characterize the compaction performance of concrete. At the same time, the existing "human + tool" approach is difficult to achieve the exploration of concrete's 3D structure. Using AIGC+, reconstructing ultra-complex concrete 3D structures from a finite feature sequence of concrete vibration detected through multi-modal collaborative detection is expected to meet this demand. However, this technology still lacks a research foundation in the field of concrete construction. To this end, this application proposes the following technical solutions, including:

[0040] S1. Based on the multi-source sensors of sound, light, force, electricity and touch, the concrete vibration process is collaboratively detected to construct a multimodal concrete vibration dataset, where:

[0041] Acoustic-optical-force-electrical-touch refer to five categories of process signals, namely, concrete vibration sound, visual, vibration acceleration, vibrating rod current, and six-dimensional force acting on the vibrating rod, which are the sources of multimodal vibration information. Acoustic sensors include but are not limited to audio acquisition devices such as mobile phones, microphones, and voice recorders. Visual sensors include but are not limited to image acquisition devices such as mobile phones and cameras. Current sensors include but are not limited to current acquisition devices such as clamp ammeters and Hall current sensors. Vibration acceleration sensors include but are not limited to sensors based on piezoelectric integrated circuits (IEPE) and microelectromechanical systems (MEMS), which are mainly used to detect the vibration acceleration signals actively emitted by the vibrating rod. Six-dimensional force sensors are mainly used to detect the generalized force vector signals acting on the vibrating rod at the interface between the rod shell and concrete, such as the load exerted on the vibrating rod by obstacles such as aggregates and steel bars when the obstacles touch the vibrating rod. The present invention relates to the use of direct denoising methods and indirect denoising methods in the collaborative detection of concrete vibration. The direct denoising method directly constructs a filter to remove noise, including filtering methods such as mean filtering, median filtering, and Gaussian filtering for two-dimensional image data, as well as time domain filtering and frequency domain filtering methods for time series data such as sound, current, and acceleration. The indirect denoising method is based on autonomous learning of multi-source noise distribution and reversely obtains vibration information hidden in massive noise through conditional probability.

[0042] S2. By extracting the feature information of different scales from the multimodal concrete vibration dataset, a finite feature sequence of multimodal fusion of concrete vibration is constructed;

[0043] S3. Extract the domain knowledge features of concrete vibration through a text extractor and construct a concrete vibration knowledge embedding learning feature sequence. The domain knowledge of concrete vibration includes but is not limited to text descriptions of qualified vibration, under-vibration, and over-vibration in the concrete process. For example, qualified vibration corresponds to uniform and dense distribution of aggregates and bubbles in the mortar, under-vibration corresponds to unclear aggregate redistribution and a large number of harmful bubbles have not been discharged, and over-vibration corresponds to concrete segregation, with obvious stratification of concentrated aggregates at the bottom and a large amount of floating slurry at the top.

[0044] S4. The concrete vibration multimodal fusion finite feature sequence and the concrete vibration knowledge are embedded in the learning feature sequence and input into the deep residual network at the same time to establish the concrete aggregate-bubble space topological structure; the use of artificial intelligence (AIGC) in the present invention refers to the training of the deep residual network based on the artificial intelligence (AI) algorithm, and by learning and identifying the features of existing data, generating content with certain creativity and quality with appropriate generalization capabilities, wherein the AI ​​algorithm includes but is not limited to generative adversarial networks (Generative Adversarial Network), pre-trained models (Pre-Trained Model, PTM), etc. From the perspective of content development, AIGC is a new production method that uses AI technology to automatically generate content after PGC (Professional Generated Content) and UGC (User Generated Content). Among them:

[0045] Predict the joint distribution of aggregate-bubble pairwise distances and angles in the concrete aggregate-bubble spatial topology;

[0046] The distance-angle joint distribution is used as a constraint condition to transform the prediction problem of the distance and angle between concrete aggregate and air bubbles into a differentiable function optimization problem;

[0047] The optimization problem is solved based on the gradient descent method, and a prediction model of the spatial topological structure of concrete aggregate and air bubbles is established.

[0048] S5. Collect concrete cross-section scanning image data and obtain two-dimensional concrete aggregate-bubble shape samples through precise instance segmentation; including: selecting a certain number of cross-sections of the hardened concrete core sample at equal intervals along the height direction, and generating concrete cross-section scanning image data by CT; cropping the concrete cross-section scanning image data into small-sized image blocks, and using a parallel computing strategy to perform instance segmentation; wherein: the FocalLoss classification loss function is used to balance the data of different categories in the concrete cross-section scanning image data.

[0049] S6. Generating a corresponding real 3D structure based on the two-dimensional concrete aggregate-bubble shape sample, and reconstructing the concrete 3D structure sample in combination with the concrete aggregate-bubble spatial topological structure; including:

[0050] Based on the real shape of concrete aggregate in 3D laser scanning, the appearance shape is regarded as a high-dimensional distribution, and its distribution pattern is learned to modify the generated aggregate appearance shape to make it conform to the real distribution and meet the conditions of spatial collision constraints;

[0051] The learned distribution pattern is transferred to the concrete aggregate-bubble spatial topology, corresponding objects are generated at the corresponding positions of aggregates and bubbles, and rendered based on neural style transfer, thereby realizing the visualization of the internal 3D structure of concrete.

[0052] S7. Performing data enhancement on the concrete 3D structure sample based on autonomous learning of latent patterns; wherein:

[0053] Variational autoencoders are used to pre-train the latent space of samples in a concrete 3D data reconstruction model. This allows encoding of arbitrary 3D concrete structure samples into the latent space, ensuring that redundant information in the 3D samples is compressed while improving the semantic consistency of the samples.

[0054] A three-dimensional sample simulation is used to generate classifier-free guidance, which is then distilled into a variational autoencoder as the input parameter of a concrete three-dimensional data reconstruction model.

[0055] Using skipping strategy to calculate consistency loss to improve the distillation process of variational autoencoder;

[0056] S8. Based on steps S1 to S7, a data set is constructed and model training is performed to obtain a prediction model from a multimodal vibration finite feature sequence to a concrete 3D structure.

[0057] A neural style transfer network is used to improve the similarity between the generated 3D sample of the concrete internal structure and the real 3D sample. In one embodiment, the present invention aims to provide a real-time reconstruction method of concrete 3D structure based on multimodal vibration feature fusion and AIGC, comprising the following steps:

[0058] 1) Collect multi-source perception data based on acoustic-optical-mechanical-electrical-tactile multimodal collaborative detection and construct a dataset.

[0059] 1.1) Use a camera to collect concrete vibration visual data with a resolution of 1920*1080 and a frame rate of 30FPS;

[0060] 1.2) Use a microphone to collect concrete vibration sound data with a sampling rate of 22kHz;

[0061] 1.3) Use a clamp-on ammeter to collect concrete vibration current data at a sampling rate of 3 Hz;

[0062] 1.4) Use an IEPE uniaxial accelerometer to collect the vibration acceleration of the concrete vibrator and set a high sampling frequency;

[0063] 1.5) Using a six-dimensional force sensor to collect tactile information from the concrete vibrator;

[0064] 2) Based on the collaborative detection of multi-source sensors, a direct method is used to reduce data noise, multi-modal information features are extracted at multiple scales, and a feature sequence of multi-modal information fusion of concrete vibration is constructed.

[0065] 2.1) Apply data transformation methods to convert between different modal data formats and align data of different frequencies and dimensions, thereby achieving data-level fusion of multi-source cross-modal information;

[0066] 2.2) Extracting different scale features of multimodal information and constructing a finite feature sequence for vibration multimodal fusion;

[0067] 3) Using a text extractor to extract the text features of concrete vibration domain knowledge, and obtain the concrete vibration knowledge embedding learning feature sequence;

[0068] 4) If Figure 1 As shown in the figure, the spatial topological structure of concrete aggregate and bubbles is generated based on generative artificial intelligence (AIGC).

[0069] 4.1) Construct a concrete aggregate and bubble topology prediction model based on Deep Residual Network (Deep ResNet);

[0070] 4.2) Using the concrete vibration multimodal fusion finite feature sequence and the concrete vibration knowledge embedding learning feature sequence obtained in steps 2) and 3) respectively as input, predict the joint distribution of aggregate-air bubble pairwise distances and angles;

[0071] 4.3) Using the joint distribution of aggregate-bubble distances and angles as constraints, the problem of predicting the distances and angles between concrete aggregates and bubbles is converted into a differentiable function optimization problem.

[0072] 4.4) Solve the optimization model in 4.3) based on the gradient descent method to obtain the spatial topological structure of concrete aggregate and bubbles.

[0073] 5) Collecting concrete cross-section scanning image data and constructing a corresponding dataset, obtaining aggregate and bubble samples based on accurate instance segmentation of the concrete cross-section scanning image data, and reconstructing the concrete 3D structure based on the corresponding topological structure generated in step 4).

[0074] 5.1) Drill and coring the hardened concrete after vibration. Use medical spiral CT to longitudinally select and collect concrete cross-sectional scan image data and construct a corresponding dataset. The image pixel matrix size is 640*640;

[0075] 5.2) Based on the real concrete section scanning results, the two-dimensional shape distribution samples of concrete aggregates and bubbles are segmented.

[0076] 5.2.1) The concrete section scanning image is cropped into multiple smaller size image blocks, and parallel computing strategy is adopted to improve the calculation efficiency;

[0077] 5.2.2) As shown in Figure 2 , based on high-performance GPU, deep learning framework is constructed, current mainstream segmentation algorithms such as Mask R-CNN, QueryInst and Solo are improved, deep neural network for intelligent segmentation of aggregates and bubbles in concrete slice scanning image is built, and public data set is used for pre-training;

[0078] 5.2.3) Through transfer learning, fine-tuning is carried out on concrete section scanning image data set, so that it is more suitable for concrete aggregate and bubble segmentation task.

[0079] 5.3) Based on three-dimensional laser scanning, three-dimensional model of aggregate with real distribution of appearance shape is obtained.

[0080] 5.3.1) Collecting full-grade concrete aggregates at construction site, modeling the appearance and shape of aggregates by three-dimensional laser scanning;

[0081] 5.3.2) Considering the appearance shape as high-dimensional distribution, using generative adversarial network (GAN) to learn the high-dimensional joint distribution of aggregate appearance and shape, realizing arbitrary modeling of aggregate shape.

[0082] 6) Based on aggregate bubble sample and corresponding topological structure, reconstructing concrete 3D structure sample.

[0083] 6.1) Transferring the aggregate-bubble two-dimensional shape distribution pattern obtained in step 5.2) to the input topological structure, using GAN to generate corresponding three-dimensional aggregates and bubbles conforming to real distribution at their corresponding positions, and meeting the collision constraint in space;

[0084] 6.2) Based on the three-dimensional distribution of aggregates obtained in 5.3), shape correction is made to the three-dimensional aggregates generated in 6.1);

[0085] 6.3) Based on neural style transfer, rendering concrete 3D structure to realize visual display.

[0086] 7) As shown in Figure 3 , based on latent pattern autonomous learning, concrete 3D structure data enhancement is carried out to provide sufficient samples for AIGC model training.

[0087] 7.1) Using a variational autoencoder (VAE) to pre-train the latent space of highly complex concrete 3D structure samples, we can encode any 3D concrete structure sample into the latent space.

[0088] 7.2) Generate classifier-free guidance using 3D fine numerical simulation;

[0089] 7.3) Distill the classifier-free guidance as model input parameters into the latent consistency VAE model;

[0090] 7.4) Use a neural style transfer network to improve the similarity between synthetic concrete 3D structure samples and real samples.

[0091] 8) Build and train a concrete 3D structure prediction model based on the PyTorch deep learning framework.

[0092] Preferably, the concrete 3D structure prediction model trained in step 8) is deployed to the concrete vibration construction monitoring system, which can achieve the purpose of real-time reconstruction of the concrete 3D structure based on the sound-light-force-electricity-touch concrete vibration multimodal feature sequence.

[0093] Although the preferred embodiments of the present invention are described above, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, which all fall within the scope of protection of the present invention.

[0094] References

[0095] [1] SL677-2014, Specification for hydraulic concrete construction[S].

[0096] [2]DL / T 5144-2015, Specification for Construction of Hydraulic Concrete[S].

Claims

1. A real-time reconstruction method for concrete 3D structure based on multimodal vibration feature fusion and AIGC, characterized by: The following steps are involved: S1. Based on the collaborative detection of concrete vibration process by multi-source sensors of sound, light, force, electricity and touch, a multi-modal concrete vibration dataset is constructed. S2. By extracting the feature information of different scales from the multimodal concrete vibration dataset, a finite feature sequence of multimodal fusion of concrete vibration is constructed; S3. Extract the knowledge features of concrete vibration domain through text extractor and construct the concrete vibration knowledge embedding learning feature sequence; S4. Inputting the concrete vibration multimodal fusion finite feature sequence and the concrete vibration knowledge embedded learning feature sequence into a deep residual network at the same time to establish a concrete aggregate-bubble space topological structure; S5. Based on the collected concrete cross-section scanning image data, a two-dimensional concrete aggregate-bubble shape sample is obtained by instance segmentation; S6. generating a corresponding real 3D structure based on the two-dimensional concrete aggregate-bubble shape sample, and reconstructing the concrete 3D structure sample in combination with the concrete aggregate-bubble spatial topological structure; S7. performing data enhancement on the concrete 3D structure sample based on latent pattern autonomous learning; S8. Based on steps S1 to S7, a data set is constructed and model training is performed to obtain a prediction model from a multimodal vibration finite feature sequence to a concrete 3D structure.

2. The method for real-time reconstruction of concrete 3D structure based on multimodal vibration feature fusion and AIGC according to claim 1 is characterized in that: In step S4, the concrete vibration multimodal fusion finite feature sequence and the concrete vibration knowledge embedded learning feature sequence are simultaneously input into the deep residual network for training and establishing the concrete aggregate-bubble space topological structure process, including: Predict the joint distribution of aggregate-bubble pairwise distances and angles in the concrete aggregate-bubble spatial topology; The distance-angle joint distribution is used as a constraint condition to transform the prediction problem of the distance and angle between concrete aggregate and air bubbles into a differentiable function optimization problem; The above differentiable function optimization problem is solved based on the gradient descent method, and the spatial topological structure of concrete aggregate and bubbles is established.

3. The method for real-time reconstruction of concrete 3D structure based on multimodal vibration feature fusion and AIGC according to claim 1 is characterized in that: In step S5, the collected concrete cross-section scan image is segmented by instance segmentation to obtain a two-dimensional concrete aggregate-bubble shape sample; include: A certain number of cross sections are selected at equal intervals along the height direction of the hardened concrete core sample, and CT scan image data of the concrete section is generated; The concrete section scanning image data is cropped into small-sized image blocks, and instance segmentation is carried out using a parallel computing strategy; wherein: the FocalLoss classification loss function is used to balance the data of different categories in the concrete section scanning image data.

4. The method for real-time reconstruction of concrete 3D structure based on multimodal vibration feature fusion and AIGC according to claim 1 is characterized in that: The process of generating a corresponding real 3D structure based on the two-dimensional concrete aggregate-bubble shape sample and reconstructing the concrete 3D structure sample in combination with the concrete aggregate-bubble spatial topological structure in step S6 includes: Based on the real shape of concrete aggregate in 3D laser scanning, the appearance shape is regarded as a high-dimensional distribution, and its distribution pattern is learned to modify the generated aggregate appearance shape to make it conform to the real distribution and meet the conditions of spatial collision constraints; The learned distribution pattern is transferred to the concrete aggregate-bubble spatial topology, corresponding objects are generated at the corresponding positions of aggregates and bubbles, and rendered based on neural style transfer, thereby realizing the visualization of the internal 3D structure of concrete.

5. The method for real-time reconstruction of concrete 3D structure based on multimodal vibration feature fusion and AIGC according to claim 1 is characterized in that: In step S7, data enhancement is performed on the concrete 3D structure sample based on the potential pattern autonomous learning, specifically including: Variational autoencoders are used to pre-train the latent space of samples in a concrete 3D data reconstruction model. This allows encoding of arbitrary 3D concrete structure samples into the latent space, ensuring that redundant information in the 3D samples is compressed while improving the semantic consistency of the samples. A three-dimensional sample simulation is used to generate classifier-free guidance, which is then distilled into a variational autoencoder as the input parameter of the concrete three-dimensional data reconstruction model. Using skipping strategy to calculate consistency loss to improve the distillation process of variational autoencoder; A neural style transfer network is used to improve the similarity between the generated 3D samples of concrete internal structure and the real 3D samples.

Images