Intelligent inversion method for concrete density based on improved transformer and bimodal fusion
By using an improved Transformer and dual-modal fusion method, combined with digital twin technology, the problems of lag in concrete vibration compaction detection and poor generalization of inversion models were solved, achieving efficient and accurate defect location and traceability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NO 3 ENG COMPANY LTD OF CCCC FIRST HARBOR ENG COMPANY
- Filing Date
- 2026-06-23
- Publication Date
- 2026-07-21
AI Technical Summary
In existing technologies, the detection of concrete compaction is lagging, the data dimension is limited, the inversion model has poor generalization ability, and it fails to deeply integrate vibration-process dual-modal data, making it difficult to accurately trace the source of defects.
An improved Transformer and dual-modal fusion method is adopted. By deploying sensors at the head of the vibrating rod to collect vibration signals and process parameters, and combining them with digital twin technology, spatial characterization and full-cycle traceability of quality status are realized, including data preprocessing, feature extraction, dual-modal fusion and digital twin modeling.
It achieves accurate inversion of concrete density, improves detection accuracy, reduces costs, and enables centimeter-level defect location and full-chain traceability.
Smart Images
Figure CN122433019A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent construction technology in building engineering, and in particular to an intelligent inversion method for concrete density based on an improved Transformer and dual-modal fusion. Background Technology
[0002] The compaction density of concrete is a core indicator determining the safety and durability of an engineering structure. Existing technologies have the following key shortcomings: First, the quality inspection methods are seriously outdated, relying on post-inspection methods such as core drilling and spring rebound, and by the time defects are discovered, it is no longer possible to rectify them at low cost. Second, process monitoring relies heavily on a single vibration signal, ignoring the decisive influence of key process parameters such as pouring rate and vibration spacing on quality, resulting in a single data dimension and extremely poor generalization of the inversion model. Third, existing machine learning methods (such as CNN and LSTM) are not good at capturing the features of long-term, strongly disturbed vibration signals, and they do not introduce attention mechanisms to focus on key features. Fourth, digital twin technology in this field is mostly limited to visualization and display, failing to be deeply coupled with bimodal quality data, and thus unable to achieve accurate traceability of defects.
[0003] Currently, no technology has organically combined "improved Transformer timing learning," "vibration-process dual-modal fusion," and "quality-oriented digital twin." This invention aims to fill this technological gap and provide a precise and efficient construction quality inversion solution. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention proposes an intelligent inversion method for concrete density based on an improved Transformer and dual-modal fusion. This method integrates temporal and static dual-modal data, utilizes an improved Transformer and attention mechanism to improve feature extraction accuracy, and combines digital twins to achieve spatial representation and full-cycle traceability of quality status, thus solving the problems of difficulty in discovering, locating, and tracing hidden defects.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion is characterized by the following steps: S1: Dual-modal heterogeneous construction data acquisition; specifically including: S11. A MEMS triaxial vibration sensor and a UWB Beidou high-precision positioning module are installed at the head of the insertable vibrator to synchronously collect time-series modal data: triaxial acceleration signal of vibration stress wave, three-dimensional spatial coordinates of the vibrator, insertion depth, moving speed, and spatial trajectory of UWB positioning. S12. Collect static process modal data through on-site sensors and construction records: concrete pouring rate, vibrator insertion spacing, formwork support pressure, concrete pouring temperature and slump. S13. Construct a dual-mode raw dataset of temporal vibration and static process.
[0006] S2: Data preprocessing and format adaptation; specifically including: S21. Perform Butterworth filtering, denoising, and Min-Max normalization on the time-series vibration signal, slice it into fixed lengths, and construct a four-dimensional input matrix. The normalization formula is as follows: ; In the formula: The normalized vibration signal value is reduced to a new value of 0 to 1. The original vibration signal value; , These are the minimum and maximum values of the signal segment, respectively.
[0007] S22. Standardize the static process parameters, fill in missing values, and construct a static feature vector; S23. Ensure accurate alignment of bimodal data in terms of timestamps and spatial locations.
[0008] S3: Based on the improved Transformer for temporal feature extraction, an improved Transformer encoder is built. The multi-head attention structure is optimized to address the non-stationary and high-noise characteristics of the vibration signal. Finally, through the stacking of multiple improved Transformer encoders, a 256-dimensional deep temporal feature Fourier transform (FT) of the vibration signal is output. The core operation of the encoder is a multi-head self-attention mechanism, calculated using the following formula: ; In the formula: MultiHead(Q,K,V): Output features of the multi-head attention module; Q, K, and V: are the query matrix, key matrix, and value matrix obtained by linear transformation of the time-series characteristics of the input vibration signal, respectively. Concat(·): The operation of concatenating h single-head attention outputs along the channel dimension; head1,…,headh: The outputs of the first to the hth single-head attention, each single-head attention independently captures the temporal dependence features of the vibration signal at different time scales; W O Output projection weight matrix, used to map the concatenated multi-head features into 256-dimensional deep temporal features (FT).
[0009] The improved Transformer encoder includes a feature filtering module for removing environmental noise interference from the vibration signal and optimizing the multi-head self-attention mechanism's ability to capture weak features, including: Position coding layer: adopts sine and cosine position coding to preserve the temporal sequence of vibration operations; Improved multi-head self-attention layer: Introduces a feature screening mechanism based on energy threshold to eliminate environmental noise interference with amplitude below a set threshold in vibration signals and optimize the multi-head self-attention mechanism's ability to capture weak features; Feedforward network layer: Completes the learning of high-dimensional abstract features.
[0010] S4: Dual-modal attention feature fusion; designing a dual-branch attention fusion network to achieve dynamic fusion of temporal and static features; specifically including: S41, Branch A: Temporal Branch: Deep temporal feature F obtained from input S3. T ; S42, Branch B: Process Branch: Features static process parameters are mapped using a fully connected layer, outputting process features F. P ; S43, Attention Mechanism Layer: Calculates the mutual attention weights of the two features, automatically focuses on features that significantly affect density, and achieves dynamic weighted fusion. Fusion formula: ; In the formula: F fusion The comprehensive quality feature is obtained after fusing dual-modal features; F T This refers to the deep temporal characteristics of the vibration signal during tamping; F P These are characteristics of static process parameters; α is the temporal feature F T Attention weights; β represents the process characteristic F P Attention weights; The attention weights satisfy α+β=1 and are generated by the model through adaptive learning based on data features.
[0011] The features that significantly affect compaction include vibration spacing, vibration attenuation coefficient, concrete pouring temperature, slump, pouring rate, vibration duration, vibrator insertion depth, peak vibration acceleration, and vibration frequency offset.
[0012] S5: Quality-oriented digital twin modeling; specifically including: S51. Divide the structure to be poured into a 5cm×5cm×5cm three-dimensional voxel grid and couple it with the BIM model in coordinates. S52, merging the fusion feature F output from S4 fusion According to the spatial location, each element is bound to the corresponding voxel mesh unit to construct a digital twin of concrete quality. This model is not only a visualization carrier, but also a three-dimensional database of quality data.
[0013] S6: Density Inversion and Defect Grading; specifically including: S61, Fusing Feature F fusion Input an adaptive fully connected regression network to invert the concrete density ρ (0~1) of each voxel element; the core inversion formula is: ; In the formula: ρ is the concrete density value of a single three-dimensional voxel element; σ is the Sigmoid activation function, used to map the network output to the [0,1] interval, matching the range of density values; W1 is the weight matrix of the first fully connected layer; b1 is the bias vector of the first fully connected layer; RELU is a modified linear unit activation function used to introduce nonlinear transformations and improve the feature representation ability of the network. W2 is the weight matrix of the second fully connected layer; b2 is the bias vector of the second fully connected layer; F fusion This refers to the comprehensive quality feature obtained after fusing dual-modal features.
[0014] S62. Setting grading standards: Excellent: ρ≥0.90; Weakly dense: 0.70 ≤ ρ < 0.90; Defect: ρ < 0.70.
[0015] The defects include vibration leakage, voids, honeycombing, and pitting.
[0016] S7: Real-time early warning, tracing, and model iteration; specifically including: S71. Early warning and positioning: Highlight defect areas with different colors in the digital twin model to achieve centimeter-level positioning and trigger audio-visual early warning; S72. End-to-end traceability: Automatically associates dual-modal data corresponding to defect voxels to generate a traceability report of defect-parameter anomaly-responsible work team; S73. Online Incremental Learning: Using measured data such as core drilling as labels, the inversion model is iterated online to continuously improve accuracy.
[0017] The present invention has the following beneficial effects: 1. The pioneering nature of dual-modal fusion: This invention integrates the vibration time sequence signal of tamping with the static parameters of the casting process, and improves the accuracy by more than 15% compared with the single vibration signal inversion method under the same engineering test set.
[0018] 2. Improved technical adaptability: The Transformer network has been customized for construction scenarios, and a new feature filtering module has been added to enable it to effectively handle engineering time series data with strong interference. This has high technical barriers.
[0019] 3. Enhanced functionality of digital twins: The constructed digital twin achieves deep binding of data, features, quality, and space, breaking through the limitations of traditional digital twins that are only used for visualization, and realizing dynamic representation and precise traceability of quality.
[0020] 4. Highly practical for engineering applications: No new dedicated hardware is required, existing sensors can be reused, the algorithm model is lightweight and easy to implement in engineering projects, significantly reducing the transformation cost of smart construction. Attached Figure Description
[0021] Figure 1 This is a flowchart of the inversion method of the present invention; Figure 2 This is a schematic diagram of an improved Transformer and a dual-modal fusion structure. Figure 3 A schematic diagram of a three-dimensional voxelized mesh and digital twin model of concrete; Figure 4 This is a visualization of the density inversion and defect location. Detailed Implementation
[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0023] Example 1
[0024] A smart inversion method for concrete density based on an improved Transformer and dual-modal fusion, such as Figure 1 As shown, it includes the following steps: S1: Dual-modal heterogeneous construction data acquisition; specifically including: S11. A MEMS triaxial vibration sensor and a UWB Beidou high-precision positioning module are installed at the head of the insertable vibrator to synchronously collect time-series modal data: triaxial acceleration signal of vibration stress wave, three-dimensional spatial coordinates of the vibrator, insertion depth, moving speed, and spatial trajectory of UWB positioning. S12. Collect static process modal data through on-site sensors and construction records: concrete pouring rate, vibrator insertion spacing, formwork support pressure, concrete pouring temperature and slump. S13. Construct a dual-mode raw dataset of temporal vibration and static process.
[0025] S2: Data preprocessing and format adaptation; specifically including: S21. Perform Butterworth filtering, denoising, and Min-Max normalization on the time-series vibration signal, slice it into fixed lengths, and construct a four-dimensional input matrix. The normalization formula is as follows: ; In the formula: The normalized vibration signal value is reduced to a new value of 0 to 1. The original vibration signal value; , These are the minimum and maximum values of the signal segment, respectively.
[0026] S22. Standardize the static process parameters, fill in missing values, and construct a static feature vector; S23. Ensure accurate alignment of bimodal data in terms of timestamps and spatial locations.
[0027] S3: Based on the improved Transformer for temporal feature extraction, an improved Transformer encoder is built. The multi-head attention structure is optimized to address the non-stationary and high-noise characteristics of the vibration signal. Finally, through the stacking of multiple improved Transformer encoders, a 256-dimensional deep temporal feature Fourier transform (FT) of the vibration signal is output. The core operation of the encoder is a multi-head self-attention mechanism, calculated using the following formula: ; In the formula: MultiHead(Q,K,V): Output features of the multi-head attention module; Q, K, and V: are the query matrix, key matrix, and value matrix obtained by linear transformation of the time-series characteristics of the input vibration signal, respectively. Concat(·): The operation of concatenating h single-head attention outputs along the channel dimension; head1,…,headh: The outputs of the first to the hth single-head attention, each single-head attention independently captures the temporal dependence features of the vibration signal at different time scales; W O Output projection weight matrix, used to map the concatenated multi-head features into 256-dimensional deep temporal features (FT).
[0028] The improved Transformer encoder includes a feature filtering module for removing environmental noise interference from the vibration signal and optimizing the multi-head self-attention mechanism's ability to capture weak features, including: Position coding layer: adopts sine and cosine position coding to preserve the temporal sequence of vibration operations; Improved multi-head self-attention layer: Introduces a feature screening mechanism based on energy threshold to eliminate environmental noise interference with amplitude below a set threshold in vibration signals and optimize the multi-head self-attention mechanism's ability to capture weak features; Feedforward network layer: Completes the learning of high-dimensional abstract features.
[0029] S4: Dual-modal attention feature fusion; designing a dual-branch attention fusion network to achieve dynamic fusion of temporal and static features; specifically including: S41, Branch A: Temporal Branch: Deep temporal feature F obtained from input S3. T ; S42, Branch B: Process Branch: Features static process parameters are mapped using a fully connected layer, outputting process features F. P ; S43, Attention Mechanism Layer: Calculates the mutual attention weights of the two features, automatically focuses on features that significantly affect density, and achieves dynamic weighted fusion. Fusion formula: ; In the formula: F fusion The comprehensive quality feature is obtained after fusing dual-modal features; F T This refers to the deep temporal characteristics of the vibration signal during tamping; F P These are characteristics of static process parameters; α is the temporal feature F T Attention weights; β represents the process characteristic F P Attention weights; The attention weights satisfy α+β=1 and are generated by the model through adaptive learning based on data features.
[0030] The features that significantly affect compaction include vibration spacing, vibration attenuation coefficient, concrete pouring temperature, slump, pouring rate, vibration duration, vibrator insertion depth, peak vibration acceleration, and vibration frequency offset.
[0031] Figure 2 This is a schematic diagram of the improved Transformer and dual-modal fusion structure.
[0032] S5: Quality-oriented digital twin modeling; such as Figure 3 As shown, it specifically includes: S51. Divide the structure to be poured into a 5cm×5cm×5cm three-dimensional voxel grid and couple it with the BIM model in coordinates. S52, merging the fusion feature F output from S4 fusion According to the spatial location, each element is bound to the corresponding voxel mesh unit to construct a digital twin of concrete quality. This model is not only a visualization carrier, but also a three-dimensional database of quality data.
[0033] S6: Density Inversion and Defect Grading; specifically including: S61, Fusing Feature F fusion Input an adaptive fully connected regression network to invert the concrete density ρ (0~1) of each voxel element; the core inversion formula is: ; In the formula: ρ is the concrete density value of a single three-dimensional voxel element; σ is the Sigmoid activation function, used to map the network output to the [0,1] interval, matching the range of density values; W1 is the weight matrix of the first fully connected layer; b1 is the bias vector of the first fully connected layer; RELU is a modified linear unit activation function used to introduce nonlinear transformations and improve the feature representation ability of the network. W2 is the weight matrix of the second fully connected layer; b2 is the bias vector of the second fully connected layer; F fusion This refers to the comprehensive quality feature obtained after fusing dual-modal features.
[0034] S62. Setting grading standards: Excellent: ρ≥0.90; Weakly dense: 0.70 ≤ ρ < 0.90; Defect: ρ < 0.70. For example... Figure 4 As shown.
[0035] The defects include vibration leakage, voids, honeycombing, and pitting.
[0036] S7: Real-time early warning, tracing, and model iteration; specifically including: S71. Early warning and positioning: Highlight defect areas with different colors in the digital twin model to achieve centimeter-level positioning and trigger audio-visual early warning; S72. End-to-end traceability: Automatically associates dual-modal data corresponding to defect voxels to generate a traceability report of defect-parameter anomaly-responsible work team; S73. Online Incremental Learning: Using measured data such as core drilling as labels, the inversion model is iterated online to continuously improve accuracy.
[0037] Example 2
[0038] The following example demonstrates the implementation of the method in Implementation Example 1: (This example uses the concrete pouring project of the shear wall of a super high-rise building as an example.) 1. Install MEMS triaxial vibration sensors and UWB Beidou high-precision positioning modules at the head of the immersion vibrator, install a flow meter at the pouring pump to collect the pouring rate, and deploy temperature and humidity sensors on site.
[0039] 2. Simultaneously acquire vibration waves, time-series modes of vibrator trajectory, and pouring rate at 0.8m. 3 / h, vibration spacing 40cm (static process mode).
[0040] 3. The improved Transformer extracts vibration features, and the attention mechanism identifies "excessive vibration spacing" as a key influencing factor, assigning it a higher weight.
[0041] 4. The digital twin model showed a density of 0.65 in a certain area, which was identified as a void defect. The system automatically traced back to the construction period in that area, during which the pouring rate suddenly increased to 1.2m. 3 / h, resulting in insufficient vibration.
[0042] 5. Immediately stop pouring on site and implement vibration repair measures. Input the core drilling data of this area into the model to complete incremental learning.
[0043] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be included within the scope of protection of the present invention.
Claims
1. A smart inversion method for concrete density based on improved Transformer and dual-modal fusion, characterized in that, Includes the following steps: S1: Dual-modal heterogeneous construction data acquisition; collect vibration timing signals and static parameters of pouring process to construct a dual-modal heterogeneous construction dataset; S2: Data preprocessing and format adaptation; preprocessing the bimodal data separately to achieve precise alignment of the time and spatial dimensions; S3: Based on the improved Transformer for temporal feature extraction, an improved Transformer encoder is built. The multi-head attention structure is optimized to address the non-stationary and high-noise characteristics of the vibration signal. Finally, through the stacking of multiple improved Transformer encoders, a 256-dimensional deep temporal feature Fourier transform (FT) of the vibration signal is output. The core operation of the encoder is a multi-head self-attention mechanism, calculated using the following formula: ； In the formula: MultiHead(Q,K,V): Output features of the multi-head attention module; Q, K, and V: are the query matrix, key matrix, and value matrix obtained by linear transformation of the time-series characteristics of the input vibration signal, respectively. Concat(·): The operation of concatenating h single-head attention outputs along the channel dimension; head1,…,headh: The outputs of the first to the hth single-head attention, each single-head attention independently captures the temporal dependence features of the vibration signal at different time scales; W O Output projection weight matrix, used to map the concatenated multi-head features into 256-dimensional deep temporal features (FT); S4: Dual-modal attention feature fusion; Design a dual-branch attention fusion network to achieve dynamic fusion of temporal and static features; Perform dynamic weighted fusion of deep temporal features and process parameter features to output comprehensive quality features; S5: Quality-oriented digital twin modeling; constructing a 3D voxelized digital twin model with 5cm accuracy, and binding comprehensive quality features to the corresponding spatial mesh; S6: Compaction Inversion and Defect Classification; Based on comprehensive quality characteristics, concrete compaction is inverted and defect classification is performed using an adaptive regression network. S7: Real-time early warning, tracing and model iteration; achieves centimeter-level defect location, real-time early warning and full-link tracing based on digital twin model.
2. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, S1 specifically includes: S11. A MEMS triaxial vibration sensor and a UWB Beidou high-precision positioning module are installed at the head of the insertable vibrator to synchronously collect time-series modal data: triaxial acceleration signal of vibration stress wave, three-dimensional spatial coordinates of the vibrator, insertion depth, moving speed, and spatial trajectory of UWB positioning. S12. Collect static process modal data through on-site sensors and construction records: concrete pouring rate, vibrator insertion spacing, formwork support pressure, concrete pouring temperature and slump. S13. Construct a dual-mode raw dataset of temporal vibration and static process.
3. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, S2 specifically includes: S21. Perform Butterworth filtering, denoising, and Min-Max normalization on the time-series vibration signal, slice it into fixed lengths, and construct a four-dimensional input matrix. The normalization formula is as follows: ； In the formula: The normalized vibration signal value is reduced to a new value of 0 to 1. The original vibration signal value; , These are the minimum and maximum values of the signal segment, respectively. S22. Standardize the static process parameters, fill in missing values, and construct a static feature vector; S23. Ensure accurate alignment of bimodal data in terms of timestamps and spatial locations.
4. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, The improved Transformer encoder described in S3 includes a feature filtering module for removing environmental noise interference from the vibration signal and optimizing the multi-head self-attention mechanism's ability to capture weak features, including: Position coding layer: adopts sine and cosine position coding to preserve the temporal sequence of vibration operations; Improved multi-head self-attention layer: Introduces a feature screening mechanism based on energy threshold to eliminate environmental noise interference with amplitude below a set threshold in vibration signals and optimize the multi-head self-attention mechanism's ability to capture weak features; Feedforward network layer: Completes the learning of high-dimensional abstract features.
5. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, S4 specifically includes: S41, Branch A: Temporal Branch: Deep temporal feature F obtained from input S3. T ; S42, Branch B: Process Branch: Features static process parameters are mapped using a fully connected layer, outputting process features F. P ; S43, Attention Mechanism Layer: Calculates the mutual attention weights of the two features, automatically focuses on features that significantly affect density, and achieves dynamic weighted fusion. Fusion formula: ； In the formula: F fusion The comprehensive quality feature is obtained after fusing dual-modal features; F T This refers to the deep temporal characteristics of the vibration signal during tamping; F P These are characteristics of static process parameters; α is the temporal feature F T Attention weights; β represents the process characteristic F P Attention weights; The attention weights satisfy α+β=1 and are generated by the model through adaptive learning based on data features.
6. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 5, characterized in that, The features that significantly affect compaction include vibration spacing, vibration attenuation coefficient, concrete pouring temperature, slump, pouring rate, vibration duration, vibrator insertion depth, peak vibration acceleration, and vibration frequency offset.
7. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, S5 specifically includes: S51. Divide the structure to be poured into a 5cm×5cm×5cm three-dimensional voxel grid and couple it with the BIM model in coordinates. S52, merging the fusion feature F output from S4 fusion According to the spatial location, each element is bound to the corresponding voxel mesh unit to construct a digital twin of concrete quality. This model is not only a visualization carrier, but also a three-dimensional database of quality data.
8. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, S6 specifically includes: S61, Fusing Feature F fusion Input an adaptive fully connected regression network to invert the concrete density ρ (0~1) of each voxel element; the core inversion formula is: ； In the formula: ρ is the concrete density value of a single three-dimensional voxel element; σ is the Sigmoid activation function, used to map the network output to the [0,1] interval, matching the range of density values; W1 is the weight matrix of the first fully connected layer; b1 is the bias vector of the first fully connected layer; RELU is a modified linear unit activation function used to introduce nonlinear transformations and improve the feature representation ability of the network. W2 is the weight matrix of the second fully connected layer; b2 is the bias vector of the second fully connected layer; F fusion The comprehensive quality feature is obtained after fusing dual-modal features; S62. Setting grading standards: Excellent: ρ≥0.90; Weakly dense: 0.70 ≤ ρ < 0.90; Defect: ρ < 0.
70.
9. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, The defects include vibration leakage, voids, honeycombing, and pitting.
10. The intelligent inversion method for concrete density based on improved Transformer and dual-modal fusion as described in claim 1, characterized in that, Specifically, S7 includes: S71. Early warning and positioning: Highlight defect areas with different colors in the digital twin model to achieve centimeter-level positioning and trigger audio-visual early warning; S72. End-to-end traceability: Automatically associates dual-modal data corresponding to defect voxels to generate a traceability report of defect-parameter anomaly-responsible work team; S73. Online Incremental Learning: Using measured data such as core drilling as labels, the inversion model is iterated online to continuously improve accuracy.