A Domain-Adaptive Method and System for Debonding Detection of Reinforced Concrete Interface Across Structures

Abstract

This invention relates to a domain-adaptive cross-structure reinforced concrete interface debonding detection method and system. The detection method of the system includes the following steps: Step 1: Constructing a source structure and a target structure of reinforced concrete; Step 2: Extracting time-frequency domain information from the guided wave signals of the source and target structures based on continuous wavelet transform; Step 3: Constructing a domain-adaptive artificial neural network; for the artificial neural network, firstly, feature learning and parameter optimization are performed on the source model; Step 4: Based on the adversarial alignment and statistical alignment theory of a pair of data spaces in the domain-adaptive domain, guided adversarial adaptive training is performed on the target model; Step 5: Combining the result prediction network of the source structure obtained in Step 3 and the feature extraction network corresponding to the target structure obtained in Step 4, the debonding diagnosis result of the reinforced concrete structure to be detected is obtained through comparison calculation. This invention can realize the detection of cross-structure reinforced concrete interface debonding under conditions of structural parameter fluctuations.

Description

Technical Field

[0001] This invention relates to the field of debonding detection at reinforced concrete interfaces, and particularly to a domain-adaptive cross-structure reinforced concrete interface debonding detection method and system. Background Technology

[0002] Interface debonding in reinforced concrete structures has long been a major factor leading to structural failure and even large-scale catastrophic events, making it a key focus in the field of structural health monitoring (SHM). In recent years, guided wave propagation (GWP) technology for detecting interface debonding in reinforced concrete structures has attracted considerable attention. Due to the inherent sensitivity of guided waves to their waveguide boundary conditions, the phenomenon of interface debonding in reinforced concrete can be clearly reflected in the guided wave signal. Therefore, many scholars have devoted their efforts to exploring the relationship between guided wave signal characteristics and the debonding state of the structure. Existing research results can be mainly divided into two categories based on the analytical mechanism: statistical model methods based on signal statistical characteristics and physical model methods based on guided wave propagation laws.

[0003] Statistical modeling methods primarily establish a statistical fitting relationship between signal characteristics and structural damage by identifying the statistical features of guided wave signals under different damage conditions. For example, statistical methods such as empirical mode decomposition and wavelet packet transform are used to decompose the signal into a set of multiple dimensionless parameters. These dimensionless parameters are then converted into a Damage Index (DI) with no clearly defined physical meaning, and a fitting relationship is constructed between this DI and the structural damage state, thereby enabling the analysis of the guided wave signal. Physical modeling methods focus on the propagation behavior of guided waves within the structure. Through theoretical calculations of waveguide dispersion, suitable guided wave excitation frequencies or modes for damage analysis are identified. This selection emphasizes the anomalous behavior of waves under certain modes and is used as the basis for structural damage analysis. For instance, the difference in guided wave group velocity between reinforced and reinforced concrete structures is used to calculate the size of the debonded segment on the structure using the Time of Flight (TOF) of different wave packets.

[0004] Both of these methods exhibit high stability and effectiveness under specific structures. However, when these methods are actually applied in engineering, various problems or limitations arise that render them unsuitable for real structures. This is because existing methods are overly sensitive to material, structural, and environmental parameters. Even a slight difference between a parameter of the applied structure and the parameter of the developed method can cause the method to fail. For statistical modeling methods, a change in a structural parameter renders the original model fitting inaccurate in describing the relationship between signal and damage characteristics in the new structure, leading to a significant decrease in method performance. Physical model-based methods rely primarily on anomalous features (wave velocity or propagation attenuation rate of a specific mode) in waveguide dispersion curves from theoretical analysis for structural damage analysis. Changes in relevant parameters alter the dispersion patterns, rendering the original analytical rules inapplicable, thus reducing the method's accuracy or even yielding completely erroneous results. Furthermore, these problems are difficult to solve by re-experimenting on new structures and constructing new method parameters. This is because, firstly, laboratory conditions—including material selection, sample preparation standards, and testing environments—are far superior to real-world conditions, and the integrity and uniformity of the structures differ from real infrastructure. On the other hand, the performance differences between different batches of raw materials and the unavoidable differences during mixing and casting will further lead to differences in different parts of the structure. Therefore, the key to solving the above problems lies in developing an analytical method that is less sensitive to material, size, and environmental parameters and can achieve cross-structure reinforced concrete interface debonding detection. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the first objective of this invention is to provide a domain-adaptive method for detecting debonding at the interface of reinforced concrete across structures. This method is based on data science and utilizes the concept of domain adaptation in deep learning theory to construct an artificial neural network capable of extracting universal features from guided wave signals across structures, thereby enabling the detection of debonding at the interface of reinforced concrete across structures under fluctuating structural parameters.

[0006] The technical solution to achieve the first objective of this invention is: the domain-adaptive cross-structure reinforced concrete interface debonding detection method of this invention includes the following steps:

[0007] Step 1: Construct a source structure of reinforced concrete, which is used to provide diagnostic knowledge for debonding detection; set the reinforced concrete structure to be tested as the target structure;

[0008] Step 2: Extract time-frequency domain information from the guided wave signals of the source and target structures based on continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows:

[0009]

[0010] Where f(t) represents the guided wave signal, f(t)∈L 2 (R); L 2 (R) denotes the set of all square-integrable functions defined on the real number field R; W f (μ,τ) represents the transformation result of the guided wave signal at the spatial scale μ and the time scale τ;

[0011] Ψ represents the conjugate of the wavelet basis functions; F(ω) represents the Fourier transform of the guided wave signal; μ,τ (ω) represents the Fourier transform of the wavelet basis functions;

[0012] Step 3: Construct a domain-adaptive artificial neural network; for the artificial neural network, first perform feature learning and parameter optimization on the source model; the optimization objective of the source model is:

[0013]

[0014] Among them, L CE (·) represents the loss function for a classification task, typically using the cross-entropy function; X s ,Y s These represent the source structure guided wave signal and its corresponding structural debonding state, respectively; E s P represents the feature extraction network corresponding to the source structure. s The result prediction network represents the source structure; K represents the number of types of debonding states of the structure; x s y s , k represents a set of guided wave signal features, the artificial neural network's prediction of the sample's state, and the actual debonding state of the structure to which the sample belongs.

[0015] Step 4: Based on the adversarial alignment and statistical alignment theory of a pair of data spaces in the domain adaptation domain, the target model is subjected to guided adversarial adaptive training; the optimization objective of the target model is:

[0016]

[0017] Among them, L ADA (·) represents the loss function for adversarial alignment; L DC (·) represents the statistical alignment loss function, where the maximum mean difference is used as the loss function; X t The guided wave signal representing the target structure; E t D represents the feature extraction network corresponding to the target structure; D represents the domain discriminator, used to implement E. s and E t Adversarial training; λ represents the weights of statistical alignment; Represents the maximum mean difference loss function; xs and x t These are waveguide signal samples representing the source structure and the target structure, respectively.

[0018] Step 5: Combine the result prediction network of the source structure obtained in Step 3 with the feature extraction network corresponding to the target structure obtained in Step 4, and obtain the debonding diagnosis result of the reinforced concrete structure to be detected by comparison and calculation.

[0019] Preferably, the source structure is a structure in a known debonding state; when the provided source structure can be used to perform debonding detection diagnostics, a wave propagation test is performed on the structure in a known debonding state to obtain a guided wave signal, and the guided wave signal and the debonding state of the source structure are recorded simultaneously.

[0020] Since the exact debonding state of the target structure is unknown, the wave propagation test only records the guided wave signal.

[0021] Preferably, in Step 1 above, debonding diagnostic knowledge of the source structure in multiple debonding states is constructed; the debonding diagnostic knowledge provided by the source structure is more than the debonding diagnostic knowledge required by the target structure.

[0022] Preferably, in Step 2 above: the wavelet basis functions of the continuous wavelet transform use Morlet wavelets, and their types are all correlated with the guided wave signal in the time and frequency domains.

[0023] Preferably, in Step 2 above, the spectrum of the continuous wavelet transform of the guided wave signal is adjusted by writing a Python script to make the spatial and temporal scales of the output spectrum the same order of magnitude, so as to accommodate the subsequent feature analysis of the neural network.

[0024] Preferably, the above-mentioned source model consists of two parts: a feature extraction network and a result prediction network. The feature extraction network adopts a convolutional neural network architecture and is adapted to the convolutional neural network skeleton layer using two-dimensional convolutional kernels. The result prediction network adopts a fully connected network architecture and is used for feature analysis and result output in the fully connected layer.

[0025] Preferably, the optimization of the source model follows a supervised learning method, minimizing the loss function based on backpropagation and gradient descent. This allows the source model to fully learn the spatial distribution and mapping relationships of the source structure data.

[0026] Preferably, the feature extraction network of the target model in Step 4 above adopts the same architecture as the feature extraction network of the source model in Step 3, and its initial network parameters are preset to the parameters of the source model network that has already been trained during training.

[0027] Preferably, in Step 4 above: the training of the target model is performed using an unsupervised learning method. The training of the target model has two objectives: a. To construct a domain discriminator to distinguish the domain labels of the feature extraction networks of the source structure and the target structure. By confusing the domain labels, the feature extraction network of the target structure is forced to have a data distribution structure that is highly consistent with the feature extraction network of the source structure, which is adversarial alignment; b. To construct the maximum mean difference between the outputs of the feature extraction networks of the source structure and the target structure. This item mainly describes the maximum value of the average difference between the two sets of distributions at each order in space. By minimizing this difference using the gradient descent method, the features extracted by the network of the target structure are forced to have statistical similarity with the features of the source structure, which is statistical alignment.

[0028] A second objective of this invention is to provide a detection system for implementing the aforementioned domain-adaptive cross-structure reinforced concrete interface debonding detection method. This detection system can progressively improve its detection accuracy through learning and training.

[0029] The technical solution to achieve the second objective of this invention is as follows: The domain-adaptive cross-structure reinforced concrete interface debonding detection system of this invention includes a distributed PZT system information source, a data acquisition system, and a domain-adaptive guided wave data and analysis system; the distributed PZT system information source is a collection of multiple single-point piezoelectric ceramic PZT sheets distributed on the reinforced concrete structure to be detected; the data acquisition system is used to emit continuous wavelet signals to the source structure and the target structure and to collect and process guided wave signals from the distributed PZT system information source; the domain-adaptive guided wave data and analysis system includes a memory, a processor, and a program stored in the memory that can implement steps Step 2 to Step 5 of claim 1 under the operation of the processor.

[0030] This invention offers several advantages: The debonding detection system and method for reinforced concrete interfaces no longer rely on theoretical analysis of a specific feature point or manually calculated fitting relationships to construct the correlation between guided wave signals and structural debonding. Instead, it uses data science to identify universal characteristics between different data distributions through nonlinear mapping between data spaces, based on the overall macroscopic signal distribution. This method reduces the sensitivity of the debonding detection method to the material, size, and environmental parameters of the analyzed structure, improving its feasibility for practical engineering applications and demonstrating broad prospects for engineering use. Attached Figure Description

[0031] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein...

[0032] Figure 1This is a schematic diagram of the debonding detection system for cross-structure reinforced concrete interfaces based on domain adaptation, according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the artificial neural network constructed by the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0034] Figure 3 This is a flowchart illustrating the training method of the source model in the artificial neural network constructed by the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0035] Figure 4 This is a flowchart illustrating the training method of the target model in the artificial neural network constructed by the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0036] Figure 5 The model diagram of a three-dimensional reinforced concrete beam based on the commercial finite element software Abaqus is shown in the embodiment of the present invention for the domain-adaptive cross-structure reinforced concrete interface debonding detection system.

[0037] Figure 6 The diagram shows the target structure and test scheme prepared in the laboratory for the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0038] Figure 7 The T-SNE diagram shows the original source structure data distribution of the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0039] Figure 8 The T-SNE diagram shows the data distribution of the source structure after training and source model feature mapping in the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0040] Figure 9 The T-SNE diagram shows the distribution of the original target structure data in the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0041] Figure 10 The T-SNE diagram shows the data distribution of the target structure after training and source model feature mapping in the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0042] Figure 11 The T-SNE diagram shows the data distribution of the target structure after training and feature mapping of the target model in the domain-adaptive cross-structure reinforced concrete interface debonding detection system according to an embodiment of the present invention.

[0043] Figure 12 This is a comparison chart showing the debonding detection performance of the domain-adaptive cross-structure reinforced concrete interface debonding detection system of this invention with that of traditional methods in this example. Detailed Implementation

[0044] The following is based on the attached diagram. Figures 1 to 12 The present invention provides preferred embodiments and describes them in detail to enable a better understanding of the functions and features of the present invention.

[0045] Please see Figure 1 The detection system of this invention, used to implement a domain-adaptive cross-structure reinforced concrete interface debonding detection method, is applied to, for example... Figure 1 In the health monitoring environment of the reinforced concrete structure shown, the detection system includes a distributed PZT system information source 1, a data acquisition system 2, and a domain-adaptive guided wave data and analysis system 3. The distributed PZT system information source 1 is a collection of multiple single-point piezoelectric ceramic PZT sheets 11 distributed on the reinforced concrete structure to be tested. The data acquisition system 2 is used to emit continuous wavelet signals to the source structure and the target structure and to collect and process guided wave signals from the distributed PZT system information source 1. The domain-adaptive guided wave data and analysis system 3 includes a memory, a processor, and a program stored in the memory that can implement the following steps Step 2 to Step 5 under the operation of the processor.

[0046] Please see Figures 2 to 12 To facilitate understanding, the following example uses the domain-adaptive cross-structure reinforced concrete interface debonding detection method of this invention applied to the above-mentioned detection system to explain the detection method of this invention. The domain-adaptive cross-structure reinforced concrete interface debonding detection method of this invention includes the following steps:

[0047] Step 1: Construct a source structure of reinforced concrete, which is used to provide diagnostic knowledge for debonding detection; set the reinforced concrete structure to be tested as the target structure;

[0048] Considering that all debonding detection methods based on guided wave propagation require a set of original structures with different debonding states as the basis for constructing signal analysis methods, this system needs to collect a set of such structures to provide diagnostic knowledge for debonding detection. In addition, it is also necessary to collect data on the reinforced concrete structures to be tested.

[0049] Preferably, in Step 1:

[0050] The source structure is collected by performing wave propagation experiments on a structure in a known debonding state to obtain guided wave signals, while simultaneously recording the guided wave signals and the debonding state of the structure; however, the target structure, as the structure to be tested, has an unknown exact debonding state, so only the guided wave signals are recorded during the wave propagation experiment.

[0051] Preferably, in Step 1:

[0052] The debonding diagnostic knowledge provided by the source structure must be greater than the debonding diagnostic knowledge required by the target structure. Specifically, the debonding state of the target structure must appear in the set of source structures so that the diagnostic knowledge of the source structure can cover the diagnostics of the target structure.

[0053] For example, such as Figure 5 As shown, based on numerical simulation algorithms, a set of three-dimensional reinforced concrete beams with different interface debonding states were established in the commercial finite element software Abaqus. The guided wave signal was solved using a "dynamic implicit" algorithm. The simulated beam is 500 mm long, with a square cross-section of 100 mm in length and width, and a 18 mm diameter steel bar is placed at the center of the cross-section. The material parameters of the beam are shown in Table 1. This set of structures contains a total of 10 different debonding states, as shown in Table 2. This set of structures is used as the source structure to provide debonding diagnosis knowledge.

[0054] Table 1. Material parameters of the reinforced concrete beam model

[0055] Material Elastic modulus (GPa) Poisson's ratio <![CDATA[Density (kg / m 3 )]]> mass damping factor Stiffness Damping Factor Unit type concrete 30.05 0.2 2360 0.01 <![CDATA[1.5×10 -7 ]]> C3D8R steel 207 0.3 7894 0.01 <![CDATA[1.5×10 -8 ]]> C3D8R

[0056] Table 2. Debonding status and damage labels of reinforced concrete beams

[0057]

[0058]

[0059] For example, such as Figure 6 As shown, a set of real reinforced concrete beams 4 were also cast in the laboratory, with initial values ​​for material and dimensional parameters consistent with the aforementioned virtual beams. However, parameter fluctuations are unavoidable, and the real reinforced concrete beams 4 cannot be exactly the same as the virtual beams. The debonding state of the 10 real reinforced concrete beams 4 was consistent with that of the corresponding numbered virtual beams. Guided wave propagation tests were also conducted on this set of real reinforced concrete beams 4, and their guided wave signals were recorded. The real reinforced concrete beams 4 will be used as the target structure for debonding detection.

[0060] Step 2: Extract time-frequency domain information from the guided wave signals of the source and target structures based on continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows:

[0061]

[0062] Where f(t) represents the guided wave signal, f(t)∈L 2 (R); L 2 (R) denotes the set of all square-integrable functions defined on the real number field R; W f (μ,τ) represents the transformation result of the guided wave signal at the spatial scale μ and the time scale τ; Ψ represents the conjugate of the wavelet basis functions; F(ω) represents the Fourier transform of the guided wave signal; μ,τ (ω) represents the Fourier transform of the wavelet basis functions;

[0063] Preferably, in Step 2:

[0064] The mother wavelet of the continuous wavelet transform uses the Morlet wavelet, which is correlated with the guided wave signal in both the time and frequency domains.

[0065] Preferably, in Step 2:

[0066] The continuous wavelet transform spectrum of the guided wave signal typically does not have spatial and temporal scales of the same order of magnitude. Therefore, a Python script is needed to adjust the scale of the output spectrum to be of the same order of magnitude to accommodate the subsequent feature analysis of the neural network. For example, continuous wavelet transforms are performed on both the source and target structures. Here, the Morlet wavelet is chosen as the mother wavelet, with a spatial scale of 20 and a temporal scale of 1000. Therefore, the wavelet energy spectrum of the guided wave signal is a matrix of (20, 1000). The built-in Python function `reshape` is then used to redefine the size of the matrix, modifying it to a matrix of (200, 100).

[0067] Step 3: Construct a domain-adaptive artificial neural network; for the artificial neural network, first perform feature learning and parameter optimization on the source model; the optimization objective of the source model is:

[0068]

[0069] Among them, L CE (·) represents the loss function for a classification task, typically using the cross-entropy function; X s ,Y s These represent the source structure guided wave signal and its corresponding structural debonding state, respectively; E s P represents the feature extraction network corresponding to the source structure. s The result prediction network represents the source structure; K represents the number of types of debonding states of the structure; x s y s, k represent a set of guided wave signal features, the neural network's predicted state of the sample, and the actual debonding state of the structure to which the sample belongs.

[0070] Preferably, in Step 3:

[0071] The source model consists of two parts: a feature extraction network and a result prediction network. The feature extraction network adopts a convolutional neural network architecture, which can be adapted to the backbone layer of a convolutional neural network using two-dimensional convolutional kernels. The result prediction network adopts a fully connected network architecture, using fully connected layers for feature analysis and result output.

[0072] Preferably, in Step 3:

[0073] The optimization of the source model follows a supervised learning method, minimizing the loss function based on backpropagation and gradient descent, enabling the source model to fully learn the spatial distribution and mapping relationships of the source structure data. The training steps of the source model are as follows: Figure 3 As shown.

[0074] For example, a neural network structure as shown in the table was established. A feature extraction network and a damage analysis network were used to form the source model to learn the data distribution characteristics of the source structure. The T-SNE method was used to flatten the high-dimensional data distribution onto a two-dimensional plane for visualization. Figure 7 and Figure 8 The data distribution of guided wave data with different debonding lengths of the source structure before and after feature mapping by the source model are shown respectively. It can be seen that the data distribution after feature mapping by the source model is extremely regular, indicating that the source model has fully learned the data distribution of the source structure.

[0075] Table 3. Structural details of the constructed artificial neural network for interface debonding analysis

[0076]

[0077]

[0078]

[0079] Where C represents a convolutional layer; BN represents a batch normalization layer; P represents a pooling layer; DP represents a dropout layer; FC represents a fully connected layer in the damage classification network; FD represents a fully connected layer in the domain discriminator; and the subscripts represent the indices of each layer.

[0080] Step 4: Based on the adversarial alignment and statistical alignment theory of a pair of data spaces in the domain adaptation domain, the target model is subjected to guided adversarial adaptive training; the optimization objective of the target model is:

[0081]

[0082] Among them, L ADA (·) represents the loss function for adversarial alignment; L DC (·) represents the statistical alignment loss function, where the maximum mean difference is used as the loss function; X t The guided wave signal representing the target structure; E t D represents the feature extraction network corresponding to the target structure; D represents the domain discriminator, used to implement E. s and E t Adversarial training; λ represents the weights of statistical alignment; Represents the maximum mean difference loss function; x s and x t These are guided wave signal samples representing the source structure and the target structure, respectively.

[0083] Preferably, in Step 4:

[0084] The feature extraction network of the target model adopts the same architecture as the feature extraction network of the source model in Step 3, and its initial network parameters are preset to the parameters of the source model network that has already been trained during training.

[0085] Preferably, in Step 4:

[0086] The target model is trained using an unsupervised learning method. The training has two objectives: a) To construct a domain discriminator to distinguish the domain labels of the feature extraction networks of the source and target structures. By confusing these domain labels, the target network's extracted features are forced to have a data distribution structure highly consistent with the source network's, i.e., adversarial alignment; b) To construct the maximum mean difference between the outputs of the source and target feature extraction networks. This term mainly describes the maximum average difference between the two sets of distributions at various spatial intervals. Minimizing this difference using gradient descent forces the features extracted by the target network to have statistical similarity to the source features, i.e., statistical alignment. The training steps of the target model are as follows: Figure 4 As shown.

[0087] For example, according to Figure 4 The steps involve training the target model using the T-SNE method. Figure 9 The original data distribution of the target structure is shown. Figure 10 and Figure 11The data distribution after feature mapping of the target structure data using the source model and the target model are shown respectively. It can be seen that because the source model cannot perform cross-structure feature learning, it cannot distinguish the target structure data; therefore, the data distribution of the target structure remains chaotic after mapping using the source model. However, after training with adversarial alignment and statistical alignment, the target model has the ability to detect cross-structure decoupling and can distinguish the target structure data to a certain extent.

[0088] Step 5: Combine the result prediction network of the source structure obtained in Step 3 with the feature extraction network corresponding to the target structure obtained in Step 4, and obtain the debonding diagnosis result of the reinforced concrete structure to be detected by comparison and calculation.

[0089] Preferably, the source structure is a structure in a known debonding state; when the provided source structure can be used to perform debonding detection diagnostics, a wave propagation test is performed on the structure in a known debonding state to obtain a guided wave signal, and the guided wave signal and the debonding state of the source structure are recorded simultaneously.

[0090] For example, the constructed neural network was used to test three types of debonding detection tasks at reinforced concrete interfaces: debonding identification, debonding localization, and debonding size quantification. Simultaneously, based on traditional debonding detection methods, a supervised learning-based convolutional neural network was constructed to perform the same three tasks. The accuracy rates achieved by the two methods on these three tasks are as follows: Figure 12 As shown in the figure, it can be seen that the method proposed in this invention outperforms traditional methods in all three types of tasks.

[0091] The present invention, by employing the above detection system and technical solution, has the following beneficial effects:

[0092] The reinforced concrete interface debonding detection system described in this invention no longer relies on a specific feature point in theoretical analysis or a manually calculated fitting relationship to construct the correlation between guided wave signals and structural debonding. Instead, it is based on data science, using nonlinear mapping between data spaces to find universal characteristics between different data distributions from the overall macroscopic signal distribution. Through this method, the sensitivity of the debonding detection method to the material, size, and environmental parameters of the analyzed structure is reduced, improving the feasibility of the debonding detection method in practical engineering applications and demonstrating broad prospects for engineering applications.

[0093] The present invention has been described in detail above with reference to the accompanying drawings and embodiments. Those skilled in the art can make various modifications to the present invention based on the above description. Therefore, certain details in the embodiments should not be construed as limiting the present invention, and the scope of protection of the present invention shall be defined by the appended claims.

Claims

1. A domain-adaptive method for detecting debonding at the interface of reinforced concrete across structures; characterized in that... Includes the following steps: Step 1: Construct a source structure of reinforced concrete, which is used to provide diagnostic knowledge for debonding detection; set the reinforced concrete structure to be tested as the target structure; Step 2: Extract time-frequency domain information from the guided wave signals of the source and target structures based on continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows: Where f(t) represents the guided wave signal, f(t)∈L 2 (R); L 2 (R) denotes the set of all square-integrable functions defined on the real number field R; W f (μ,τ) represents the transformation result of the guided wave signal at the spatial scale μ and the time scale τ; Ψ represents the conjugate of the wavelet basis functions; F(ω) represents the Fourier transform of the guided wave signal; μ,τ (ω) represents the Fourier transform of the wavelet basis functions; Step 3: Construct a domain-adaptive artificial neural network; for the artificial neural network, first perform feature learning and parameter optimization on the source model; the optimization objective of the source model is: Among them, L CE (·) represents the loss function for a classification task, typically using the cross-entropy function; X s ,Y s These represent the source structure guided wave signal and its corresponding structural debonding state, respectively; E s P represents the feature extraction network corresponding to the source structure. s The result prediction network represents the source structure; K represents the number of types of debonding states of the structure; x s y s , k represents a set of guided wave signal features, the artificial neural network's prediction of the sample's state, and the actual debonding state of the structure to which the sample belongs. Step 4: Based on the adversarial alignment and statistical alignment theory of a pair of data spaces in the domain adaptation domain, the target model is subjected to guided adversarial adaptive training; the optimization objective of the target model is: Among them, L ADA (·) represents the loss function for adversarial alignment; L DC (·) represents the statistical alignment loss function, where the maximum mean difference is used as the loss function; X t The guided wave signal representing the target structure; E t D represents the feature extraction network corresponding to the target structure; D represents the domain discriminator, used to implement E. s and E t Adversarial training; λ represents the weights of statistical alignment; Represents the maximum mean difference loss function; x s and x t These are waveguide signal samples representing the source structure and the target structure, respectively. Step 5: Combine the result prediction network of the source structure obtained in Step 3 with the feature extraction network corresponding to the target structure obtained in Step 4, and obtain the debonding diagnosis result of the reinforced concrete structure to be detected by comparison and calculation.

2. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: The source structure is a structure in a known debonding state; when the provided source structure can be used for debonding detection diagnostics, a wave propagation test is performed on the structure in a known debonding state to obtain a guided wave signal, and the guided wave signal and the debonding state of the source structure are recorded simultaneously.

3. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: In Step 1, debonding diagnostic knowledge of the source structure in various debonding states is constructed; the source structure provides more debonding diagnostic knowledge than the target structure needs.

4. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: In Step 2: the wavelet basis functions of the continuous wavelet transform use Morlet wavelets, and their types are all correlated with the guided wave signal in the time and frequency domains.

5. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: In Step 2, the spectrum of the continuous wavelet transform of the guided wave signal is adjusted by writing a Python script to make the spatial and temporal scales of the output spectrum the same order of magnitude.

6. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: The source model consists of two parts: a feature extraction network and a result prediction network. The feature extraction network adopts a convolutional neural network architecture and is adapted to the convolutional neural network skeleton layer using two-dimensional convolutional kernels. The result prediction network adopts a fully connected network architecture and is used for feature analysis and result output in the fully connected layer.

7. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: The optimization of the source model follows a supervised learning approach, minimizing the loss function based on backpropagation and gradient descent.

8. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: In Step 4, the feature extraction network of the target model adopts the same architecture as the feature extraction network of the source model in Step 3, and its initial network parameters are preset to the parameters of the source model network that has already been trained during training.

9. The domain-adaptive cross-structure reinforced concrete interface debonding detection method according to claim 1, characterized in that: In Step 4: The target model is trained using an unsupervised learning method. The training of the target model has two objectives: a. To construct a domain discriminator to distinguish the domain labels of the feature extraction networks of the source structure and the target structure. By confusing the domain labels, the feature extraction network of the target structure is forced to have a data distribution structure that is highly consistent with that of the feature extraction network of the source structure, which is adversarial alignment; b. To construct the maximum mean difference between the outputs of the feature extraction networks of the source structure and the target structure. This item mainly describes the maximum value of the average difference between the two sets of distributions at each order in space. By minimizing this difference using the gradient descent method, the features extracted by the network of the target structure are forced to have statistical similarity with the features of the source structure, which is statistical alignment.

10. A detection system for implementing the domain-adaptive cross-structure reinforced concrete interface debonding detection method as described in claim 1, characterized in that: The system includes a distributed PZT system information source, a data acquisition system, and a domain-adaptive guided wave data and analysis system. The distributed PZT system information source is a collection of multiple single-point piezoelectric ceramic PZT sheets distributed on the reinforced concrete structure to be tested. The data acquisition system is used to emit continuous wavelet signals to the source and target structures and to collect and process guided wave signals from the distributed PZT system information source. The domain-adaptive guided wave data and analysis system includes a memory, a processor, and a program stored in the memory that can implement steps Step 2 to Step 5 of claim 1 under the operation of the processor.

Images