Method, system and device for identifying the quality of a metamorphic concrete combined with a roller compacted concrete
By combining ground-penetrating radar and the deep learning model DNFN with image recognition and self-attention mechanism, the problem of non-destructive, rapid and accurate identification of the quality of the interface between modified concrete and roller-compacted concrete was solved, achieving a highly efficient quality identification effect.
Patent Information
- Application Number
- CN202610364547.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies cannot quickly, non-destructively, and accurately identify quality defects at the interface between modified concrete and roller-compacted concrete. Traditional detection methods are inefficient, damage structures, and lack sufficient detection accuracy. Deep learning models have shortcomings in one-dimensional signal denoising and feature extraction for complex working conditions.
One-dimensional waveform data is acquired using ground-penetrating radar. Combined with the deep learning model DNFN, feature extraction is optimized using Gram matrix and self-attention mechanism through image recognition and denoising branches. This constructs a concrete bonding quality identification model, achieving non-destructive, fast, and accurate quality identification.
It enables efficient, non-destructive, and accurate quality identification of the interface between modified concrete and roller-compacted concrete, and is suitable for rapid on-site testing, improving testing accuracy and adaptability to complex environments.
Smart Images
Figure CN122109143A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of non-destructive testing technology in building engineering. More specifically, it relates to a method, system, and equipment for identifying the quality of the combination of modified concrete and roller-compacted concrete, which is particularly suitable for the pouring of roller-compacted concrete dam surfaces. Background Technology
[0002] The interface between modified concrete and roller-compacted concrete is a critical load-bearing node in engineering projects such as hydraulic dams and road subgrades. The quality of this interface directly determines the overall structural load-bearing stability and service life. Due to significant differences in the mix proportions, density, and construction techniques of the two types of concrete, the interface is highly susceptible to hidden quality defects such as honeycomb, voids, cracks, and poor bonding. If these defects are not identified promptly and accurately, they can easily lead to stress concentration, leakage, and damage, seriously threatening the structural safety of the project.
[0003] Currently, quality inspection of this joint area mainly employs traditional methods such as core drilling and ultrasonic testing. While core drilling allows for direct observation of core sample defects, it is a destructive testing method that can compromise structural integrity, potentially leading to leaks. Furthermore, it has a limited sampling range and a long testing cycle, making it unsuitable for large-scale, rapid inspections. Ultrasonic testing is susceptible to interference from factors such as aggregate content and moisture content within the concrete, resulting in insufficient signal resolution and poor identification of minute or deep defects, with highly subjective results. Both methods share common drawbacks: low efficiency, structural damage, and insufficient accuracy, making them unsuitable for the efficient, non-destructive, and precise testing requirements of engineering sites.
[0004] In recent years, deep learning technologies such as convolutional neural networks have developed rapidly in the fields of image recognition and signal detection. Compared with traditional detection methods, they have stronger feature extraction and accurate recognition capabilities. However, existing models still have shortcomings in one-dimensional signal denoising and feature extraction under complex working conditions, and their adaptability is insufficient. Ground penetrating radar (GPR), as a non-destructive testing device, can realize non-destructive detection of the internal structure of concrete based on the principle of electromagnetic wave reflection imaging. It has the advantages of high efficiency, non-destructiveness, and wide coverage. However, there is very little research on GPR detection of the interface between deformed concrete and roller-compacted concrete. Research on combining deep learning technology to process radar signals in this area and realize intelligent identification of quality defects is even more lacking.
[0005] Existing detection technologies cannot simultaneously achieve non-destructive testing, high efficiency, and accurate identification, making it difficult to meet the quality control requirements of this special joint in engineering sites. Therefore, developing an efficient and non-destructive quality identification method based on ground-penetrating radar and incorporating deep learning has significant practical engineering value for improving the quality inspection level of concrete joints and ensuring the safety of engineering structures. Summary of the Invention
[0006] To overcome the shortcomings of existing technologies and solve the problem that traditional identification methods cannot quickly, non-destructively, and accurately identify internal defects in the bonding area, this invention proposes a ground-penetrating radar-based method, system, and equipment for identifying the bonding quality of modified concrete and roller-compacted concrete. This method can quickly, accurately, and non-destructively identify whether the bonding area meets quality standards, providing reliable technical support for engineering quality control.
[0007] The objective of this invention can be achieved through the following technical solutions.
[0008] A method for identifying the bonding quality of modified concrete and roller-compacted concrete includes the following steps: S1 Dataset Construction: Ground penetrating radar was used to acquire one-dimensional raw waveform data of the joint between modified concrete and variable pressure concrete. After data preprocessing, noisy waveform data was obtained. The noisy waveform data and the raw waveform data were integrated into a dataset after unified labeling. The dataset was then divided into training set, validation set and test set. S2 Model Construction and Training: Based on the above training and validation sets, a concrete bonding quality identification model is constructed and trained using the DNFN model. S3 Quality Identification: Using the concrete bond quality identification model trained above, inference and prediction are performed on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified.
[0009] Furthermore, the data preprocessing in step S1 includes, but is not limited to, data normalization and data augmentation; the unified labeling follows the rule that unqualified data are labeled as "0" and qualified data are labeled as "1".
[0010] Furthermore, the concrete bond quality identification model described in step S2 is constructed and trained based on the existing DNFN model. The model parameters are optimized using a binary cross-entropy loss function until the performance is maximized. This model includes the following two branches: Image recognition branch: An image recognition module based on Gram matrix is added to the DNFN model. This module consists of 2D-CNN, including RSM layer, EGR layer, 2D convolutional layer, GAP layer, and Flatten layer. One-dimensional waveform data is transformed into a two-dimensional original signal matrix by the RSM layer. The EGR layer calculates the Gram matrix of the original signal matrix to obtain the Gram matrix representing the image features. After the Gram matrix is further mined by the 2D convolutional layer, it is reduced in dimensionality by the GAP layer and finally flattened by the Flatten layer to obtain a one-dimensional feature vector. Denoising Branch: After initial feature extraction by a one-dimensional convolutional layer, the one-dimensional waveform data is successively processed by a denoising module based on a self-attention mechanism and a contrastive learning module for denoising and feature optimization, resulting in a one-dimensional feature vector. The denoising module based on a self-attention mechanism is an integral part of the original multi-branch adaptive denoising module of the DNFN model. While calculating the threshold for each channel, a self-attention mechanism is introduced to calculate the weight of the threshold for each channel, and the threshold size is adaptively adjusted using this weight. The one-dimensional feature vectors of the image recognition branch and the denoising branch are concatenated by thresholding, and then feature integration and classification prediction are completed through a fully connected layer to obtain the final result.
[0011] The objective of this invention can also be achieved through the following technical solutions.
[0012] A quality identification system for the combination of modified concrete and roller-compacted concrete includes: The dataset construction module uses ground-penetrating radar to perform all-round scanning and detection of the joint between modified concrete and variable pressure concrete, and collects one-dimensional raw waveform data corresponding to this part. After data preprocessing of the collected raw waveform data, diverse noisy waveform data is obtained. After unified labeling, the noisy waveform data and the raw waveform data are integrated into a dataset, and the dataset is divided into training set, validation set and test set. The model building and training module, based on the above training set and validation set, uses the DNFN model to build and train a concrete bonding quality identification model; The quality identification module uses the trained concrete bonding quality identification model to perform inference and prediction on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified.
[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the above-described method for identifying the bonding quality of modified concrete and roller-compacted concrete.
[0014] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the above-described method for identifying the bonding quality of modified concrete and roller-compacted concrete.
[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: The present invention obtains waveform data of the combined part of deformed concrete and roller-compacted concrete through a ground penetrating radar. After data preprocessing, a concrete bonding quality recognition model (i.e., an improved DNFN model) is used to perform feature learning on the processed data and obtain a classification result (qualified or unqualified). The present invention uses advanced one-dimensional data processing methods and deep learning neural networks to achieve the task of intelligent recognition of the quality of the combined part of deformed concrete and roller-compacted concrete based on a ground penetrating radar, and provides a basis for automatically identifying the quality of the combined part by combining other tools. In addition, the recognition method proposed in the present invention is not only applicable to the quality recognition of the combined part of deformed concrete and roller-compacted concrete in roller-compacted concrete dams, but also applicable to other fields where one-dimensional data exists, such as gear fault discrimination and roller compaction quality assessment. Description of the Drawings
[0016] Figure 1 It is a flowchart of the method for identifying the combined quality of deformed concrete and roller-compacted concrete based on a ground penetrating radar according to the present invention.
[0017] Figure 2 It is an architecture diagram of the concrete bonding quality recognition model in the present invention. Detailed Embodiments
[0018] The present invention will be further described below with reference to the accompanying drawings.
[0019] The present invention provides a method for identifying the combined quality of deformed concrete and roller-compacted concrete based on a ground penetrating radar, and accurately classifies the qualified and unqualified combined parts at different positions on the construction surface of a roller-compacted concrete dam by using data processing means and deep learning methods. The present invention proposes a concrete bonding quality recognition model based on the existing DNFN model (i.e., the deep denoising filtering network model), and uses the concrete bonding quality recognition model to combine technologies such as adaptive denoising, self-attention mechanism, contrast learning, and Gram matrix to improve the recognition performance of the model under complex conditions (high noise environment). The proposed method can provide guidance for actual on-site construction and improve the construction quality of on-site construction.
[0020] The method for identifying the combined quality of deformed concrete and roller-compacted concrete based on a ground penetrating radar according to the present invention constructs a rich data set, improves the deficiencies of the original model in processing one-dimensional signals, and optimizes the model training strategy, so that the proposed concrete bonding quality recognition model can provide technical support for the task of identifying the quality of the combined part of deformed concrete and variable-pressure concrete. As Figure 1 shown, the specific implementation process of the method of the present invention is as follows: S1 Data Set Construction Ground-penetrating radar was used to perform a full-range scan of the interface between modified concrete and variable pressure concrete, and one-dimensional raw waveform data of the corresponding interface was collected. The collected raw waveform data was preprocessed to obtain diverse noisy waveform data to expand the sample size and coverage. The raw waveform data and the noisy waveform data were uniformly labeled and integrated into a dataset. The dataset was divided into training set, validation set and test set according to a preset ratio (e.g., 7:2:1).
[0021] Preferably, the present invention uses a SIR4000 ground-penetrating radar device from Geophysical Surveying Systems, Inc. for data acquisition.
[0022] Preferably, the data preprocessing includes, but is not limited to, data normalization and data augmentation, which involves adding data with different signal-to-noise ratios to the original waveform data. Before model training, data augmentation applies operations such as random pruning and Gaussian noise injection to improve the model's robustness to noise, missing data, and other interference.
[0023] Preferably, the unified labeling follows the rule that unqualified parts are labeled "0" and qualified parts are labeled "1". Qualified means that the quality of the interface between modified concrete and roller-compacted concrete meets the engineering specifications. Specifically, this is manifested in a tight bond at the interface, the absence of voids, cracks, inclusions, or other defects, and the continuous and regular waveform data from ground-penetrating radar, reflecting the uniformity and integrity of the concrete interface. Unqualified means that the interface between modified concrete and roller-compacted concrete has quality defects and does not meet the engineering specifications. For example, the interface may have voids, cracks, honeycombing, or looseness. The ground-penetrating radar waveform data will show abnormal reflection, scattering, or signal attenuation, exhibiting discontinuous and chaotic characteristics, reflecting the non-uniformity and defects of the concrete interface.
[0024] S2 Model Construction and Training Based on the aforementioned training and validation sets, a concrete bond quality identification model was constructed and trained using the existing DNFN model. Image recognition, self-attention mechanisms, and multi-branch adaptive denoising were added to the existing DNFN model to extract key features from waveform data, ensuring the new model achieves accurate results in identifying bond quality. The existing DNFN model used in this invention was applied to deep denoising of one-dimensional signals, enabling one-dimensional data with a signal-to-noise ratio (SNR) of -5 to still serve as a basis for judging whether a concrete bond is faulty, demonstrating powerful feature extraction capabilities.
[0025] The core of the concrete bonding quality identification model is to process a one-dimensional signal by dividing it into two branches: the signal itself and its corresponding two-dimensional image. This model primarily consists of multi-core 1D-CNN, multi-core 2D-CNN, contrastive learning, and a binary cross-entropy loss function. The 1D-CNN (one-dimensional neural network model) processes the one-dimensional signal, extracting its local features through convolution. The 2D-CNN (two-dimensional neural network model) processes the image corresponding to the one-dimensional signal, extracting the image's local features. The feature representations obtained from the two branches are then concatenated and input into a fully connected layer to obtain the result. Finally, the model parameters are optimized using the binary cross-entropy loss function until the highest performance is achieved.
[0026] like Figure 2 As shown, the concrete bond quality identification model mainly includes the following two branches: Image Recognition Branch: An image recognition module based on Gram matrix is added to the existing DNFN model. This module consists of a 2D-CNN, including an RSM layer (original signal matrix layer), an EGR layer (Gram matrix layer), a 2D convolutional layer, a GAP layer (global pooling layer), and a Flatten layer (flattening layer). First, the one-dimensional waveform data is transformed into a two-dimensional original signal matrix by the RSM layer. The EGR layer performs Gram matrix calculation on the original signal matrix to obtain a Gram matrix representing the image features. After further feature extraction by the 2D convolutional layer, the Gram matrix is reduced in dimensionality by the GAP layer, and finally flattened by the Flatten layer to obtain a one-dimensional feature vector. Specifically, the RSM layer calculates a corresponding pixel value for each signal point in the one-dimensional waveform data, with the value between 0 and 255. Then, the calculated points are shifted to obtain a complete two-dimensional grayscale image (i.e., the original signal matrix), which is then converted into a Gram image (i.e., a Gram matrix) through matrix operations. Through the above operations, this invention can represent a one-dimensional signal as a Gram matrix image.
[0027] Currently, deep learning is highly mature in image recognition, so converting one-dimensional signals into images for processing is one way to improve signal processing. Furthermore, when converting one-dimensional signals into two-dimensional images, the images are usually displayed in grayscale. Gram matrices, however, are displayed as color images and are better able to reveal useful features. This module consists of a 2D-CNN, deriving results through convolution operations and residual connections. The concrete quality recognition model maximizes the utilization of one-dimensional signal data by adding an image recognition branch based on Gram matrices, deeply mining its various features to address quality recognition problems in complex situations.
[0028] Denoising Branch: After initial feature extraction from a one-dimensional convolutional layer, the one-dimensional waveform data undergoes denoising and feature optimization via a self-attention-based denoising module and a contrastive learning module, resulting in a one-dimensional feature vector. The contrastive learning module compares and optimizes features from different branches, improving the model's ability to extract effective features and its generalization performance.
[0029] The original multi-branch adaptive denoising module of the DNFN model adaptively calculates the threshold for each channel, relying on soft thresholding (K=3, K=5) and hard thresholding (K=3, K=5) methods to filter data. While this method effectively filters out most noise values, its brute-force assignment approach simplifies the number of valid information values in the signal, negatively impacting model learning. To address this issue, this invention introduces a self-attention mechanism-based denoising module into the original multi-branch adaptive denoising module of the DNFN model. This self-attention mechanism calculates the weight of each channel's threshold and uses this weight to adaptively adjust the threshold value (decreasing or increasing the threshold value) to maximize the processing of irrelevant information and the retention of valid information.
[0030] The self-attention mechanism refers to adding an adaptive weight to the neural network model that can calculate the weight of each channel. By adding different weights to each channel, the effective information is improved and the error caused by redundant or noisy information is reduced.
[0031] Adaptive denoising involves adding a threshold to the neural network model that can automatically calculate the threshold for each channel. This threshold can filter out invalid or noisy information that interferes with the results.
[0032] Finally, the one-dimensional feature vectors of the image recognition branch and the denoising branch are thresholded and concatenated to achieve multi-feature fusion. Then, the feature integration and classification prediction are completed through a fully connected layer to obtain the final result.
[0033] Training the concrete bond quality identification model mainly involves parameter training and adjustment. To obtain a better model, a decay rate for the learning rate is set during the training phase, causing the learning rate to gradually decrease as the number of iterations increases, thus achieving the optimal model parameters. Furthermore, the deployment scheme design should be determined based on the actual situation. For example, training parameters might include: batch size of 16, initial learning rate of 1e-4, weight decay of 5e-4, and training cycles of 200; early stopping is used to prevent overfitting.
[0034] S3 Quality Identification: Using the concrete bond quality identification model trained above, inference and prediction are performed on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified, with 0 representing unqualified and 1 representing qualified.
[0035] For example, 4000 waveform data points were collected as a test set. Experiments show that the present invention achieves an accuracy of approximately 99% in quality identification, significantly higher than the optimized existing DNFN model. Furthermore, it also demonstrates good accuracy in noisy environments.
[0036] Based on the above-mentioned method and process principles, this invention also proposes a quality identification system combining modified concrete and roller-compacted concrete, mainly including a dataset construction module, a model construction and training module, and a quality identification module. Among them, The dataset construction module uses ground-penetrating radar to perform all-round scanning and detection of the joint between modified concrete and variable pressure concrete, and collects one-dimensional raw waveform data corresponding to this part. After data preprocessing of the collected raw waveform data, diverse noisy waveform data is obtained. After unified labeling, the noisy waveform data and the raw waveform data are integrated into a dataset, and the dataset is divided into training set, validation set and test set. The model building and training module, based on the above training set and validation set, uses the DNFN model to build and train a concrete bonding quality identification model; The quality identification module uses the trained concrete bonding quality identification model to perform inference and prediction on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified.
[0037] Based on the above principles and processes, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the above-mentioned method for identifying the bonding quality of modified concrete and roller-compacted concrete.
[0038] Based on the above principles and processes, the present invention also proposes a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the above-mentioned method for identifying the combination quality of modified concrete and roller-compacted concrete.
[0039] Although the functions and working processes of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific functions and working processes described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these are within the protection scope of the present invention.
Claims
1. A method for identifying the quality of the bond between modified concrete and roller-compacted concrete, characterized in that, Includes the following steps: S1 Dataset Construction: Ground penetrating radar was used to acquire one-dimensional raw waveform data of the joint between modified concrete and variable pressure concrete. After data preprocessing, noisy waveform data was obtained. The noisy waveform data and the raw waveform data were integrated into a dataset after unified labeling. The dataset was then divided into training set, validation set and test set. S2 Model Construction and Training: Based on the above training and validation sets, a concrete bonding quality identification model is constructed and trained using the DNFN model. S3 Quality Identification: Using the concrete bond quality identification model trained above, inference and prediction are performed on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified.
2. The method for identifying the bonding quality of modified concrete and roller-compacted concrete according to claim 1, characterized in that, The data preprocessing in step S1 includes, but is not limited to, data normalization and data augmentation; the unified labeling follows the rule that unqualified data are labeled "0" and qualified data are labeled "1".
3. The method for identifying the bonding quality of modified concrete and roller-compacted concrete according to claim 1, characterized in that, The concrete bond quality identification model described in step S2 is built and trained based on the existing DNFN model. The model parameters are optimized using a binary cross-entropy loss function until the performance is maximized. This model includes the following two branches: Image recognition branch: An image recognition module based on Gram matrix is added to the DNFN model. This module consists of 2D-CNN, including RSM layer, EGR layer, 2D convolutional layer, GAP layer, and Flatten layer. One-dimensional waveform data is transformed into a two-dimensional original signal matrix by the RSM layer. The EGR layer calculates the Gram matrix of the original signal matrix to obtain the Gram matrix representing the image features. After the Gram matrix is further mined by the 2D convolutional layer, it is reduced in dimensionality by the GAP layer and finally flattened by the Flatten layer to obtain a one-dimensional feature vector. Denoising Branch: After initial feature extraction by a one-dimensional convolutional layer, the one-dimensional waveform data is successively processed by a denoising module based on a self-attention mechanism and a contrastive learning module for denoising and feature optimization, resulting in a one-dimensional feature vector. The denoising module based on a self-attention mechanism is an integral part of the original multi-branch adaptive denoising module of the DNFN model. While calculating the threshold for each channel, a self-attention mechanism is introduced to calculate the weight of the threshold for each channel, and the threshold size is adaptively adjusted using this weight. The one-dimensional feature vectors of the image recognition branch and the denoising branch are concatenated by thresholding, and then feature integration and classification prediction are completed through a fully connected layer to obtain the final result.
4. A system for identifying the quality of the bond between modified concrete and roller-compacted concrete based on the method for identifying the bond quality of modified concrete and roller-compacted concrete according to any one of claims 1 to 3, characterized in that, include: The dataset construction module uses ground-penetrating radar to perform all-round scanning and detection of the joint between modified concrete and variable pressure concrete, and collects one-dimensional raw waveform data corresponding to this part. After data preprocessing of the collected raw waveform data, diverse noisy waveform data is obtained. After unified labeling, the noisy waveform data and the raw waveform data are integrated into a dataset, and the dataset is divided into training set, validation set and test set. The model building and training module, based on the above training set and validation set, uses the DNFN model to build and train a concrete bonding quality identification model; The quality identification module uses the trained concrete bonding quality identification model to perform inference and prediction on the test set data to determine whether the quality of the bond between modified concrete and roller-compacted concrete is qualified or unqualified.
5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for identifying the bonding quality of modified concrete and roller-compacted concrete as described in any one of claims 1 to 3.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for identifying the bonding quality of modified concrete and roller-compacted concrete as described in any one of claims 1 to 3.