A method for identifying concrete damage by fusing time-frequency dual branches of nonlinear ultrasonic waves
By employing a nonlinear ultrasonic time-frequency dual-branch modal fusion identification method, which combines time-domain and frequency-domain feature extraction, the problem of insufficient accuracy in concrete microcrack detection in existing technologies is solved, achieving efficient and accurate concrete damage identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-03-16
- Publication Date
- 2026-07-24
AI Technical Summary
Existing concrete testing methods lack accuracy in detecting microcracks. In particular, ultrasonic testing, acoustic emission, and radar testing have limitations in penetration and surface detection, making it difficult to effectively identify concrete damage information. Deep learning methods are computationally expensive and inefficient in data preprocessing.
A time-frequency dual-branch modal fusion identification method using nonlinear ultrasound is adopted. Wavelet multi-scale periodic identification and segmentation are performed through time-domain branching. High-dimensional time-domain features are extracted by combining a Transformer encoder. In the frequency-domain branch, a lightweight convolutional neural network and SE attention module are used to extract frequency-domain features. The time-domain and frequency-domain features and nonlinear coefficients are fused to achieve damage identification.
It improves the accuracy and efficiency of concrete damage identification, enabling more precise identification of micro-damage in concrete while reducing computational costs and memory usage.
Smart Images
Figure CN122448990A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nondestructive testing technology, specifically to a nonlinear ultrasonic time-frequency dual-branch mode fusion method for identifying concrete damage. Background Technology
[0002] Concrete is a commonly used material in civil engineering and hydraulic engineering. During long-term use, concrete often develops cracks. While ordinary concrete structures can generally tolerate cracks under normal operating conditions, the presence of microcracks is of great concern in critical structures such as hydraulic dams and nuclear power plant containment structures. As cracks propagate, the concrete structure becomes unstable. Accurate detection at the microcrack stage can ensure the safe and stable operation of engineering structures, requiring precise and effective detection methods.
[0003] Current detection methods have certain limitations. For example, ultrasonic testing is only sensitive to macroscopic damage, acoustic emission requires stress conditions, radar detection has weak penetration, and infrared detection is more focused on surface detection. Especially in the detection of microcracks in concrete, the accuracy of these methods is insufficient. Nonlinear ultrasonic testing methods developed to address these issues can meet the required accuracy for detecting microcracks.
[0004] The damage information carried by ultrasonic waves during their propagation in concrete is difficult to analyze, and identifying concrete damage based on the detected signals is also challenging. Combining deep learning algorithms with damage analysis of the detected signals can improve the accuracy of identification. However, most existing deep learning-based damage identification methods rely on image processing, such as time-frequency plots, for training, resulting in significant time spent on data preprocessing, high memory consumption, and high computational costs. Furthermore, relying solely on time-domain, frequency-domain, or nonlinear ultrasonic damage indicators cannot fully reveal the damage information carried by the detected signals, leaving the physical interpretation of damage identification without a solid foundation.
[0005] Therefore, a new solution is needed to address the above problems. Summary of the Invention
[0006] The purpose of this invention is to provide a nonlinear ultrasonic time-frequency dual-branch modal fusion method for identifying concrete damage, in order to solve the technical problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound, comprising the following steps:
[0008] S1: Perform time-domain branching processing, identify and segment the nonlinear ultrasound time-domain signal using wavelet multi-scale periodicity, and extract high-dimensional time-domain features through local embedding, position coding and Transformer encoder;
[0009] For nonlinear ultrasonic detection signals, wavelet transform is used to identify the main period of the time domain information as a whole, and the time domain information is divided into multiple time domain blocks at different scales;
[0010] Local and global extractors are used in each time-domain block to extract long-time, medium-time, and short-time features from the time-domain block;
[0011] At the same time, the temporal relationship is maintained by encoding the position;
[0012] The Transformer encoder is used to integrate temporal information at each scale, and a multi-head self-attention mechanism is used to capture global temporal dependencies.
[0013] In each Transformer layer, attention weights between query, key, and value vectors are calculated. The training process is stabilized through residual connections and layer normalization. Finally, a nonlinear transformation is performed through a feedforward neural network.
[0014] Repeat the process multiple times to finally output a high-dimensional feature vector.
[0015] S2: Perform frequency domain branching processing, and extract frequency domain features from the frequency domain signal using a lightweight one-dimensional convolutional neural network combined with the SE attention module;
[0016] For the frequency domain, a lightweight convolutional neural network (CNN) combined with a Squeeze-and-Excitation (SE) attention module is used.
[0017] Local frequency domain patterns are extracted step by step through four convolutional layers. Each convolutional layer undergoes batch normalization, corrected linear unit (ReLU) and downsampling (MaxPool) to capture frequency structure features at different scales while reducing the length of the frequency domain sequence.
[0018] After each convolutional layer, an SE attention module is connected, which uses global mean pooling to adaptively calculate the importance of different channels, enabling the network to automatically enhance key frequency features and suppress noise or weakly correlated components.
[0019] After multiple convolutions and attention enhancements, adaptive average pooling is used to compress the frequency domain features into a fixed-length vector, which is then mapped to a 128-dimensional frequency domain feature representation through two fully connected layers.
[0020] S3: Nonlinear coefficient processing, adaptive weighting of various concrete damage nonlinear coefficients to obtain weighted nonlinear coefficients;
[0021] The nonlinear coefficients employ a variety of nonlinear coefficients, including classical nonlinear coefficients, tapping nonlinear coefficients, and hysteresis nonlinear coefficients. Weights of the nonlinear coefficients are assigned through a trainable parameter matrix.
[0022] S4: Feature fusion, which concatenates time-domain features, frequency-domain features and weighted nonlinear coefficients, and then inputs them into the classifier after secondary weighting by the SE attention module to achieve concrete damage identification.
[0023] The high-dimensional temporal structure features extracted by the time-domain branch, the spectral morphology features extracted by the convolutional neural network in the frequency-domain branch, and the physical mechanism features carried by the nonlinear coefficients are uniformly concatenated. After concatenation, the SE attention module is introduced to perform secondary weighting on the overall features.
[0024] The SE attention module enables the network to automatically learn the relative importance of temporal structure, spectral distribution, and nonlinear mechanism under different damage levels or categories through global channel compression and recalibration, thereby enhancing the response to key damage features and suppressing noise and invalid features.
[0025] The final output is fed into a fully connected classifier, enabling the proposed model to jointly identify signal structure features, spectral energy evolution features, and nonlinear mechanism features based on materials physics, thereby achieving concrete damage identification.
[0026] Compared with the prior art, the beneficial effects of the present invention are:
[0027] This invention identifies micro-damage in concrete by using a multimodal deep learning model that integrates time-domain, frequency-domain dual-branch, and nonlinear coefficients.
[0028] In the time domain branch, wavelet energy analysis and multi-scale time domain block time series analysis are combined to jointly model the time domain characteristics and time-series dependencies of nonlinear ultrasound signals.
[0029] In the frequency domain branch, a lightweight convolutional neural network is used and an attention mechanism is introduced to achieve effective extraction and adaptive weighting of spectral features;
[0030] Meanwhile, multiple nonlinear coefficients are incorporated into the feature fusion stage to enhance damage characterization capabilities;
[0031] In summary, this invention can improve the efficiency of nonlinear ultrasonic identification of concrete damage while ensuring the accuracy of concrete damage identification. Attached Figure Description
[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1This is a schematic diagram of the nonlinear ultrasonic time-frequency dual-branch mode fusion identification of concrete damage according to the present invention;
[0034] Figure 2 This is a schematic diagram of the static nonlinear ultrasonic testing test for precast concrete cracks according to the present invention;
[0035] Figure 3 This is a schematic diagram of the dynamic nonlinear ultrasonic testing test for uniaxial loading of concrete according to the present invention;
[0036] Figure 4 This is a diagram of the temporal branching algorithm structure of the present invention;
[0037] Figure 5 This is a diagram of the frequency domain branching algorithm structure of the present invention;
[0038] Figure 6 This is a schematic diagram illustrating the evaluation index of the model calculated using the confusion matrix in this invention. Detailed Implementation
[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0040] This invention provides a nonlinear ultrasonic time-frequency dual-branch multimodal fusion method for identifying concrete damage. Specifically, it proposes a multi-dimensional, multimodal intelligent identification method for damage identification based on nonlinear ultrasonic detection data of concrete. Based on nonlinear ultrasonic detection data of concrete micro-damage, the proposed time-frequency domain dual-branch network is used to simultaneously learn and train the detection data in both the time and frequency domains. Finally, the features of multiple nonlinear coefficients are fused to achieve concrete damage identification.
[0041] Specifically as follows:
[0042] Example 1:
[0043] See Figure 1 A nonlinear ultrasonic time-frequency dual-branch mode fusion method for identifying concrete damage includes the following steps:
[0044] S1: Perform time-domain branching processing, identify and segment the nonlinear ultrasound time-domain signal using wavelet multi-scale periodicity, and extract high-dimensional time-domain features through local embedding, position coding and Transformer encoder;
[0045] As a further aspect of the present invention, the specific steps of S1 include:
[0046] S101: Let the time-domain input signal be after Z-score normalization as follows: Where B is the batch size, C is the number of channels (set to 1), and L is the sequence length;
[0047] S102: Perform multi-scale period extraction, calculate the main period in the time domain using the energy spectrum generated by wavelet transform, and then perform X... t Perform patch segmentation, and let P key periods {T1,…,T1} be obtained. p}, the length of the corresponding patch is p i =T i / s, where s is the downsampling step size, then the number of patches at the i-th scale is:
[0048]
[0049] S103: During local embedding, for the i-th scale, X... t Cut into N i Segments, each segment having a length of p i The result after cutting , Through the same linear mapping matrix , where d e Given the dimension of the embedding vector, the local embedding is obtained as follows:
[0050]
[0051] The matrix is assembled as follows:
[0052]
[0053] in, .
[0054] S104: Global dependency modeling, using the Transformer encoder. Self-attention processing:
[0055]
[0056] in, The shape remains unchanged;
[0057] S105: Perform pooling within each scale, concatenate across different scales, and use global average pooling to combine each... Pressed The vectors are then concatenated along P scales, as follows:
[0058]
[0059]
[0060] Through this process, the shape changes to ;
[0061] S106: Flatten the vector and fully connect it. Linear mapping to dimension This facilitates dimension alignment during subsequent fusion. Therefore, the following result is obtained:
[0062]
[0063] in, .
[0064] S2: Perform frequency domain branching processing, and extract frequency domain features from the frequency domain signal using a lightweight one-dimensional convolutional neural network combined with the SE attention module;
[0065] As a further aspect of the present invention, the specific steps of S2 include:
[0066] S201: Let the frequency domain signal be... The frequency domain branch uses four layers of one-dimensional convolutions (1-D convolutions), with a Squeeze-and-Excitation (SE) attention module introduced after each layer to adaptively recalibrate the channel feature responses. This changes the number of channels, resulting in the following outcome:
[0067]
[0068] Where Conv1d represents the input to the previous layer. Perform a one-dimensional convolution operation; This represents the learnable parameters (weights and biases) of the current layer's convolution kernel; BN performs batch normalization on the convolution result; ReLU applies the ReLU activation function to the batch normalized result, introducing non-linearity into the model; MaxPool performs maximum pooling on the activated features, which reduces the size of the feature map (shortens the sequence length), enhances the translation invariance of the features, and reduces the computational cost.
[0069] The SE attention module learns the importance weights of each channel (feature dimension) and then adaptively recalibrates (weights) the features of different channels, making the model focus more on important information.
[0070] S202: The frequency domain characteristics obtained after processing are as follows: .
[0071] in, This is the weight matrix.
[0072] S3: Nonlinear coefficient processing, adaptive weighting of various concrete damage nonlinear coefficients to obtain weighted nonlinear coefficients;
[0073] As a further aspect of the present invention, the specific steps of S3 include:
[0074] The nonlinear coefficient vector read is The nonlinear coefficients are weighted, and the weight vector is... The weighting process is as follows:
[0075]
[0076] Obtain the weighted nonlinear coefficients .
[0077] S4: Feature fusion, which concatenates time-domain features, frequency-domain features and weighted nonlinear coefficients, and then inputs them into the classifier after secondary weighting by the SE attention module to achieve concrete damage identification.
[0078] As a further aspect of the present invention, the specific steps of S4 include:
[0079] S401: The results of fusing the time domain branch, the frequency domain branch, and the nonlinear coefficients are concatenated and merged as follows:
[0080]
[0081] S402: Using the SE attention module for secondary weighting, the following result is obtained:
[0082]
[0083] GlobalAvgPool is a global average pooling method that "squeezes" the global spatial / temporal information of each channel into a value that represents the overall activation intensity of that channel, providing a global view for subsequent "stimulation" operations; MLP is a multilayer perceptron. The Sigmoid activation function maps each value of the MLP output to the (0,1) interval, normalizing the learned weights to between 0 and 1.
[0084] The output value can be understood as the "attention weight" or "importance score" of each channel; a value close to 1 indicates that the channel is very important and should be greatly enhanced; a value close to 0 indicates that the channel is not important and should be suppressed.
[0085] S403: The output of the classification is:
[0086]
[0087] Dropout is a regularization process used to prevent overfitting. The weight matrix maps the features output by ReLU to the class space, where the dimensions correspond to the number of classes; the Softmax function transforms the linearly transformed values into a probability distribution for all classes.
[0088] S404: Forms a compact end-to-end notation, making If all learnable parameters are given, then:
[0089]
[0090] Example 2:
[0091] This embodiment proposes a specific application experiment based on the above embodiments.
[0092] Test specimens: The concrete specimens were made of POI42.5 cement at 350 kg / m³. 3 River sand 618 kg / m 3 1256 kg / m³ of crushed stone 3 175 kg / m³ of water 3 .
[0093] The mold size for pouring concrete is 100mm × 100mm × 100mm. To pre-cast cracks in the concrete, custom-made thin steel plates need to be inserted during the initial setting stage of the concrete pouring.
[0094] Precast concrete cracks are divided into length cracks and angle cracks. Length cracks range from 1cm to 7cm, with 1cm intervals; angle cracks range from 0° to 90°, with 15° intervals. Figure 2 As shown.
[0095] Testing instruments: The instruments used for nonlinear ultrasonic testing include a signal generator (DG1022U), a signal amplifier (ATA-2042), and an oscilloscope (TBS1072B).
[0096] Parameter calculation: After obtaining the nonlinear ultrasonic testing data of concrete, the time-domain and frequency-domain graphs are analyzed and further substituted into the damage index formula, such as... Figure 2 As shown, the classical, impact, and hysteresis nonlinear coefficients are obtained, providing a data source for the subsequent nonlinear coefficient fusion of time-domain and frequency-domain dual-branch structure identification of concrete micro-damage.
[0097] Example 3:
[0098] This embodiment proposes another specific application experiment based on the above embodiments.
[0099] Test specimens: Concrete specimens were selected with dimensions of 100mm × 100mm × 100mm, using the same material as the precast cracked concrete specimens. Concrete strength is related to the water-cement ratio. To analyze the nonlinear ultrasonic characteristics of damage during concrete loading, a variety of concrete specimens were used. Therefore, concrete specimens with different water-cement ratios were poured, ranging from 0.3 to 0.7, with intervals of 0.05. These specimens were then subjected to nonlinear ultrasonic testing under uniaxial loading. Figure 3 As shown.
[0100] Testing Instruments: The instruments used for nonlinear ultrasonic testing are the same as those in Example 1. To facilitate differentiation and identification, the crack propagation process of concrete specimens under uniaxial loading is divided into four stages: microcrack compaction, elasticity, stable crack propagation, and unstable crack propagation, abbreviated as CS, ES, SCPS, and UCPS, respectively. Based on this, after obtaining the nonlinear ultrasonic testing data of the concrete, the method proposed in this study can be applied to process and identify concrete damage.
[0101] Parameter calculation: Same as in Example 1.
[0102] After obtaining the test data, the method proposed in this invention is used to sequentially perform time-domain branching, frequency-domain branching, nonlinear coefficient processing, and feature fusion. The specific operation steps are as follows:
[0103] Temporal branch: such as Figure 4 As shown, the spliced matrix obtained by the time-domain branch through main period identification, linear mapping, and local embedding is as follows.
[0104]
[0105] By using Transformer for global dependency modeling and performing pooling and concatenation across different scales, the following results were obtained.
[0106]
[0107]
[0108] Flatten the resulting vector and fully connect it, then linearly map it to the dimension. The result is:
[0109]
[0110] Frequency domain branching: such as Figure 5 As shown, the frequency domain branch goes through four layers of one-dimensional convolution, and an SE attention module is introduced after each layer, resulting in the following result.
[0111]
[0112] The obtained frequency domain features are .
[0113] Nonlinear coefficients: The weighted results of the nonlinear coefficients are as follows.
[0114]
[0115] Feature fusion: The time-domain branch, frequency-domain branch, and nonlinear coefficients are concatenated and fused, and then weighted twice using the SE attention module. The result is as follows:
[0116]
[0117] The classification output is as follows.
[0118]
[0119] During model training, the accuracy values for the training and test sets are calculated as follows.
[0120]
[0121] Where II is the indicator function, and These are the predicted and true labels for sample i, respectively. N is the number of samples in the current subset.
[0122] The quantification of the model's damage identification performance relies on the confusion matrix and derived metrics, namely accuracy, precision, recall, and F1 score, which are calculated in detail below:
[0123]
[0124] The above indicators are calculated directly from the confusion matrix, such as... Figure 6 As shown.
[0125] In summary, the nonlinear ultrasonic time-frequency dual-branch modal fusion method for identifying concrete damage of this invention has the characteristics of high detection accuracy and precise damage identification in the field of concrete damage detection and identification.
[0126] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound, characterized in that: Includes the following steps: S1: Perform time-domain branching processing, identify and segment the nonlinear ultrasound time-domain signal using wavelet multi-scale periodicity, and extract high-dimensional time-domain features through local embedding, position coding and Transformer encoder; S2: Perform frequency domain branching processing, and extract frequency domain features from the frequency domain signal using a lightweight one-dimensional convolutional neural network combined with the SE attention module; S3: Nonlinear coefficient processing, adaptive weighting of various concrete damage nonlinear coefficients to obtain weighted nonlinear coefficients; S4: Feature fusion, which concatenates time-domain features, frequency-domain features and weighted nonlinear coefficients, and then inputs them into the classifier after secondary weighting by the SE attention module to achieve concrete damage identification.
2. The method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound according to claim 1, characterized in that: S1 includes the following steps: For nonlinear ultrasonic detection signals, wavelet transform is used to identify the main period of the time domain information as a whole, and the time domain information is divided into multiple time domain blocks at different scales; Local embedding is used to extract long-term, medium-term, and short-term features in each time-domain block using local and global extractors. At the same time, the temporal relationship is maintained by encoding the position; The Transformer encoder is used to integrate temporal information at each scale, and a multi-head self-attention mechanism is used to capture global temporal dependencies. In each Transformer layer, attention weights between query, key, and value vectors are calculated. The training process is stabilized through residual connections and layer normalization. Finally, a nonlinear transformation is performed through a feedforward neural network. Repeat the process multiple times to finally output a high-dimensional feature vector.
3. The method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound according to claim 2, characterized in that: S2 includes the following steps: For the frequency domain, a lightweight convolutional neural network combined with an SE attention module is used; Local frequency domain patterns are extracted step by step through four convolutional layers. Each convolutional layer undergoes batch normalization, linear unit correction, and downsampling processing to capture frequency structure features at different scales while reducing the length of the frequency domain sequence. After each convolutional layer, an SE attention module is connected, which uses global mean pooling to adaptively calculate the importance of different channels, enabling the network to automatically enhance key frequency features and suppress noise or weakly correlated components. After multiple convolutions and attention enhancements, adaptive average pooling is used to compress the frequency domain features into a fixed-length vector, which is then mapped to a 128-dimensional frequency domain feature representation through two fully connected layers.
4. The method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound according to claim 3, characterized in that: The nonlinear coefficients employ a variety of nonlinear coefficients, including classical nonlinear coefficients, tapping nonlinear coefficients, and hysteresis nonlinear coefficients.
5. The method for identifying concrete damage using time-frequency dual-branch mode fusion of nonlinear ultrasound according to claim 4, characterized in that: S4 includes the following steps: The high-dimensional temporal structure features extracted by the time-domain branch, the spectral morphology features extracted by the convolutional neural network in the frequency-domain branch, and the physical mechanism features carried by the nonlinear coefficients are uniformly concatenated. After concatenation, the SE attention module is introduced to perform secondary weighting on the overall features. The final output is fed into a fully connected classifier, enabling the proposed model to jointly identify signal structure features, spectral energy evolution features, and nonlinear mechanism features based on materials physics, thereby achieving concrete damage identification.