Reinforced concrete interface debonding detection method based on prior knowledge learning

Abstract

The present application relates to a reinforced concrete interface debonding detection method based on prior knowledge learning, which comprises the following steps: Step 1: constructing a plurality of source structures with different debonding states; obtaining guided wave signals of each source structure through wave propagation by a detection device, and recording the debonding state to form an original data set; Step 2: performing prior knowledge data enhancement on the original data set to obtain a data set containing prior knowledge; Step 3: performing time-frequency domain information extraction based on continuous wavelet transform on the data set containing prior knowledge to obtain a data set containing two-dimensional frequency domain guided wave signals; Step 4: constructing a deep learning network based on prior knowledge learning, and training the deep learning network using the constructed data set containing two-dimensional frequency domain guided wave signals; Step 5: using the deep learning network for debonding diagnosis of the guided wave signal of the structure to be detected. The present application can realize stable and real-time reinforced concrete interface debonding detection without relying on artificial feature extraction.

Description

Technical Field

[0001] This invention relates to the field of interface debonding detection in reinforced concrete structures, and particularly to a method for detecting interface debonding in reinforced concrete based on prior knowledge learning. Background Technology

[0002] Reinforced concrete plays an indispensable role in bridge construction and various structural designs, its importance undeniable. However, due to the influence of loads and environmental factors, reinforced concrete is inevitably subject to structural damage, with interfacial debonding being a particularly serious problem. Interfacial debonding refers to the decrease in the bond strength between the reinforcing steel and the concrete, preventing them from working together and significantly reducing the overall performance of the structure. This phenomenon not only reduces the strength of the material but may also lead to more serious damage, posing potential safety risks to the overall structure. To ensure the long-term stability and safety of reinforced concrete, the field of structural health monitoring has conducted in-depth research on this issue, focusing particularly on interfacial debonding.

[0003] Among various technologies, the application of smart piezoelectric ceramics (PZT) has attracted considerable attention from researchers. Due to the inherent piezoelectric effect of PZT, it can be attached to the interior of structures to form structural health monitoring networks. Its electromechanical coupling characteristics allow it to function as both an exciter and a sensor within the network. Furthermore, the wave propagation technology paired with PZT, due to its high sensitivity to boundary conditions, is considered a powerful tool for studying debonding problems at reinforced concrete interfaces. Current methods include: determining the type of debonding within the structure based on the wave response of concrete at different excitation frequencies; determining the magnitude of interface damage using inductive monitoring networks composed of surface-mounted and embedded PZT; establishing a bond degradation index based on the statistical characteristics of wave propagation signals to describe bond integrity; characterizing the length of interface debonding using group velocity differences based on wave propagation dispersion characteristics; and locating interface debonding by analyzing the nonlinear characteristics of waveguides based on the interface breathing effect.

[0004] While existing technologies have achieved good results under laboratory conditions, they heavily rely on manual feature extraction by experts, posing a significant challenge for large-scale applications and real-time monitoring of critical areas. Furthermore, because existing methods only analyze certain key characteristics of the signal, their performance is significantly affected by interference during reception or transmission. Against this backdrop, to ensure the safety and durability of reinforced concrete structures in infrastructure, developing an analysis system for the debonding location and dimensional quantification of reinforced concrete interfaces with non-manual feature extraction and strong anti-interference capabilities, enabling long-term real-time tracking, detection, and monitoring of the service status of concrete structures under construction or existing structures, and timely assessment of local defects and damage, has become a critical requirement in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning. This method utilizes a distributed PZT system information source attached to the structure under test and connected to the detection device to achieve stress wave diagnosis of the entire cross section of the structure under test. The detection device is based on data science and, based on the idea of ​​prior knowledge learning in deep learning theory, constructs an artificial neural network that can automatically extract damage-sensitive features from guided wave signals, thereby achieving stable and real-time detection of debonding at reinforced concrete interfaces without relying on manual feature extraction.

[0006] The technical solution to achieve the objective of this invention is as follows: This invention includes the following steps: Step 1: Construct multiple reinforced concrete structures with different debonding states as source structures for signal analysis; each source structure is equipped with a distributed PZT system information source for collecting guided wave signals and connected to a detection device; the debonding state of each source structure is known; wave propagation is performed through the detection device to obtain the guided wave signals of each source structure, and the debonding state of each source structure is recorded. The guided wave signals and their corresponding debonding states are combined into data pairs, and multiple data pairs are used to form the original dataset. Step 2: Perform prior knowledge data augmentation on the original dataset formed in the previous step; the prior knowledge data augmentation involves randomly adding Gaussian white noise, less than 5% of the original signal amplitude, to the dataset formed by the guided wave signals of each source structure in the original dataset, to increase the distinguishing information about the effective and ineffective features of the guided wave signals in the dataset formed by the guided wave signals of each source structure; after the above prior knowledge data augmentation, a dataset containing prior knowledge is obtained; Step 3: Extract time-frequency domain information from the dataset containing prior knowledge using continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows:

[0007] in, Represents guided wave signal, ; Represents the guided wave signal in space. and time scale The transformation result is as follows; Represents the conjugate of wavelet basis functions; Fourier transform representing the guided wave signal; Fourier transform representing wavelet basis functions; Through the above continuous wavelet transform, the guided wave signals in the dataset containing prior knowledge are converted into two-dimensional frequency domain guided wave signals for feature extraction and analysis by subsequent deep learning networks, and a dataset containing two-dimensional frequency domain guided wave signals is formed. Step 4: Construct a deep learning network based on prior knowledge learning, and train the deep learning network using the constructed dataset containing two-dimensional frequency domain guided wave signals; the deep learning network performs feature learning and parameter optimization on the dataset containing two-dimensional frequency domain guided wave signals using the following formula:

[0008] in, The loss function for the classification task is the cross-entropy function. These represent the source structure guided wave signal and its corresponding structural debonding state, respectively. Representative feature extraction network; The network represents the result prediction network; K represents the total number of deadhesion states; k represents a single deadhesion state. Through the above feature learning and its own parameter optimization, a deep learning network with feature extraction and result prediction functions is obtained. Step 5: The deep learning network obtained in Step 4 will be used for debonding diagnosis of the guided wave signal of the structure to be tested; the structure to be tested is attached with a distributed PZT system information source connected to the detection device.

[0009] Furthermore, in Step 3 above: the wavelet basis function of the continuous wavelet transform uses the Morlet wavelet, which is correlated with the guided wave signal in both the time and frequency domains.

[0010] Furthermore, in Step 3 above: the continuous wavelet transform spectrum of the guided wave signal is adjusted to the same order of magnitude by writing a Python script to adapt to the feature analysis of the deep learning network.

[0011] Furthermore, in Step 4 above: the feature extraction network of the deep learning network adopts a convolutional neural network architecture adapted to use a convolutional neural network backbone layer with two-dimensional convolutional kernels; the result prediction network of the deep learning network adopts a fully connected network architecture that uses fully connected layers for feature analysis and result output.

[0012] Furthermore, the training of the aforementioned deep learning network follows a supervised learning method, minimizing the loss function based on backpropagation and gradient descent, so that the source model can fully learn the spatial distribution and mapping relationship of the source structural data.

[0013] This invention offers several advantages: it eliminates the reliance on manual analysis of guided wave signal characteristics, instead using data science to identify universal characteristics across different data distributions through nonlinear mapping between data spaces, based on the overall macroscopic signal distribution. The method described in this invention allows for automatic signal feature extraction during implementation. Furthermore, the incorporation of prior knowledge about environmental noise reduces sensitivity to environmental parameters of the analyzed structure, enhancing the feasibility of the debonding detection method in practical engineering applications and demonstrating broad prospects for engineering use. Attached Figure Description

[0014] 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... Figure 1 This is a schematic diagram of the detection system used in this invention to implement the method for detecting debonding of reinforced concrete interfaces based on prior knowledge learning; Figure 2 This is a schematic diagram illustrating the process of constructing a deep learning network based on prior knowledge learning in this invention. Figure 3 This is an experimental overview diagram of the component source structure and detection system of the present invention. Figure 4 This is a graph showing the loss variation during the deep learning network training process in this invention. Figure 5 This invention provides a time-domain diagram of the signal to be detected. Figure 6 This is a prediction diagram of the beam constructed by the deep learning network of the present invention. Detailed Implementation

[0015] The following is based on the attached diagram. Figures 1-6 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.

[0016] Please see Figure 1 The detection system of this invention, used to implement the method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning, 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 and a detection device; the detection device includes 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 detected; 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.

[0017] Please see Figures 2-5 To facilitate understanding, the following explanation uses an embodiment of the present invention applied to the above-mentioned detection system as an example. The method for detecting debonding of reinforced concrete interfaces based on prior knowledge learning in this embodiment of the present invention includes the following steps: Step 1: Construct multiple reinforced concrete structures with different debonding states as source structures for signal analysis; each source structure is equipped with a distributed PZT system information source for collecting guided wave signals and connected to a detection device; the debonding state of each source structure is known; wave propagation is performed through the detection device to obtain the guided wave signals of each source structure, and the debonding state of each source structure is recorded. The guided wave signals and their corresponding debonding states are combined into data pairs, and multiple data pairs are used to form the original dataset. Preferably, in Step 1: The waveguide signals of each source structure are obtained by wave propagation through a detection device, and the debonding state of each source structure is recorded at the same time, so as to provide features for deep learning networks.

[0018] For example, a total of 16 reinforced concrete beams were constructed, with lengths, widths, and heights of 500mm, 100mm, and 100mm respectively. Figure 3 (a) The concrete was made of fine aggregate, river sand, and Portland cement with a grade of PC 32.5, with a water-cement ratio of 0.45. A 9mm radius smooth steel bar was cast at the center of the beam section. A Tektronix AFG-3152C waveform generator was used for waveform excitation, an Agitek ATA-2032 amplifier was used to enhance the signal energy, and a Krohn-Hite Model 3384 broadband filter was used to suppress ambient noise. A Tektronix MDO-3054 digital oscilloscope was used for signal display and recording. Figure 3 (d)). In the example beam, interface debonding is achieved by pre-attaching PVC pipes to the reinforcing bars to isolate them from the concrete. Figure 3 (bc)). 15*5*1mm single-point piezoelectric ceramic PZT sheets 11 (model: PZT-5A18) were symmetrically installed on the side of the ends of the reinforcing bars. The material parameters of the reinforced concrete and the single-point piezoelectric ceramic PZT sheets 11 are shown in Tables 1 and 2, respectively. The structural characteristics and debonding features of the beam are shown in Table 3.

[0019]

[0020] Note: The analyzed reinforced concrete beams are divided into 5 unit segments, numbered 1-5. The debonding start position mentioned in the table represents the unit segment where debonding begins, and the debonding degree represents the number of debonded unit segments starting from the debonding start position.

[0021] For example, the aforementioned reinforced concrete beam underwent 28 days of standard curing, and a guided wave test was conducted after it had reached full strength.

[0022] Step 2: Perform prior knowledge data augmentation on the original dataset formed in the previous step; the prior knowledge data augmentation involves randomly adding Gaussian white noise, less than 5% of the original signal amplitude, to the dataset formed by the guided wave signals of each source structure in the original dataset, to increase the information on the distinction between effective and ineffective features of the guided wave signals in the dataset formed by the guided wave signals of each source structure; after the above prior knowledge data augmentation, a dataset containing prior knowledge is obtained; For example, each group of guided wave signals was normalized, and Gaussian white noise was injected into all the original signals using data augmentation techniques. The noise level used was 1%-3% of the maximum signal amplitude.

[0023] Step 3: Extract time-frequency domain information from the dataset containing prior knowledge using continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows:

[0024] in, Represents guided wave signal, ; Represents the guided wave signal in space. and time scale The transformation result is as follows; Represents the conjugate of wavelet basis functions; Fourier transform representing the guided wave signal; Fourier transform representing wavelet basis functions; Through the aforementioned continuous wavelet transform, the guided wave signals in the dataset containing prior knowledge are converted into two-dimensional frequency domain guided wave signals for subsequent feature extraction and analysis by the deep learning network, thus forming a dataset containing two-dimensional frequency domain guided wave signals.

[0025] Preferably, in Step 3: The wavelet basis functions of the continuous wavelet transform use Morlet wavelets, which are correlated with the guided wave signal in both the time and frequency domains.

[0026] Preferably, in Step 3: The continuous wavelet transform spectrum of the guided wave signal usually does not have the same spatial and temporal scale. Therefore, it is necessary to write a Python script to adjust the scale of the output spectrum to the same order of magnitude to adapt to the feature analysis of deep learning networks.

[0027] 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 time scale of 1000. Therefore, the wavelet energy spectrum of the guided signal is a matrix of (20, 1000). The matrix is ​​then resized using the Python built-in function `reshape`, changing it to a matrix of (200, 100).

[0028] Step 4: Construct a deep learning network based on prior knowledge learning, and train the deep learning network using the constructed dataset containing two-dimensional frequency domain guided wave signals; the deep learning network performs feature learning and parameter optimization on the dataset containing two-dimensional frequency domain guided wave signals using the following formula:

[0029] in, The loss function for the classification task is the cross-entropy function. These represent the source structure guided wave signal and its corresponding structural debonding state, respectively. Representative feature extraction network; The network represents the result prediction network; K represents the total number of deadhesion states; k represents a single deadhesion state. Through the aforementioned feature learning and its own parameter optimization, a deep learning network with feature extraction and result prediction functions is obtained.

[0030] Preferably, in Step 4: In Step 4: the feature extraction network adopts a convolutional neural network architecture adapted to use a two-dimensional convolutional kernel as the backbone layer; the result prediction network adopts a fully connected network architecture that uses fully connected layers for feature analysis and result output.

[0031] Preferably, in Step 4: The deep learning network is trained using supervised learning methods, minimizing the loss function based on backpropagation and gradient descent, so that the source model can fully learn the spatial distribution and mapping relationship of the source structural data.

[0032] For example, the neural network structure shown in Table 4 was established, using a feature extraction network and an damage analysis network to form a deep learning network for learning the data distribution characteristics.

[0033]

[0034] Where CL represents a convolutional layer; BN represents a batch normalized layer; PL represents a pooling layer; FCL represents a fully connected layer in the damage analysis network; and subscripts represent the indices of each layer.

[0035] For example, the deep learning network is used to perform feature learning and parameter optimization on a dataset of guided wave signals containing environmental noise and prior knowledge. Figure 4 The changes in the loss function during the learning process were recorded. From Figure 4 It can be seen that the network achieved good convergence during training. The loss functions of both the training and validation sets decreased to low levels and remained stable, indicating that the network did not exhibit overfitting.

[0036] Step 5: The deep learning network obtained in Step 4 will be used for debonding diagnosis of the guided wave signal of the structure to be tested; the structure to be tested is attached with a distributed PZT system information source connected to the detection device.

[0037] For example, guided wave signals containing high levels of ambient noise were constructed to test the signal analysis capability and noise immunity of the method. The ambient noise amplitudes used were 5%, 10%, 15%, and 20% of the maximum signal amplitude. Figure 5 The time-domain plots of the original signal and the signal after high-level noise injection are shown.

[0038] For example, the constructed neural network was used to analyze guided wave signals containing different levels of environmental noise. Tests were performed on 16 sets of beams established in Step 1. Figure 6 The analysis results of the deep learning network on the guided wave signals of 16 beams containing high-level noise are presented. It can be seen that the prediction results of the proposed method on all 16 beams show good agreement with the actual results. Its performance surpasses traditional methods in all three task types.

[0039] The present invention, by employing the above detection system and technical solution, has the following beneficial effects: This invention no longer relies on manual analysis of guided wave signal characteristics. Instead, it uses data science to find universal characteristics between different data distributions by nonlinear mapping between data spaces based on the overall macroscopic signal distribution. Through the method described in this invention, the debonding detection method can automatically extract signal features during implementation. Because prior knowledge of environmental noise is injected, the sensitivity to 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.

[0040] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning, characterized in that... Includes the following steps: Step 1: Construct multiple reinforced concrete structures with different debonding states as source structures for signal analysis; each source structure is equipped with a distributed PZT system information source for collecting guided wave signals and connecting to the detection device. The debonding state of each source structure is known; the guided wave signal of each source structure is obtained by wave propagation through a detection device, and the debonding state of each source structure is recorded at the same time. The guided wave signal and the corresponding debonding state are combined into a data pair, and the original dataset is formed by multiple data pairs. Step 2: Perform prior knowledge data augmentation on the original dataset formed in the previous step; the prior knowledge data augmentation involves randomly adding Gaussian white noise, less than 5% of the original signal amplitude, to the dataset formed by the guided wave signals of each source structure in the original dataset, to increase the distinguishing information about the effective and ineffective features of the guided wave signals in the dataset formed by the guided wave signals of each source structure; after the above prior knowledge data augmentation, a dataset containing prior knowledge is obtained; Step 3: Extract time-frequency domain information from the dataset containing prior knowledge using continuous wavelet transform. The calculation formula for the continuous wavelet transform is as follows: in, Represents guided wave signal, ; Represents the guided wave signal in space. and time scale The transformation result is as follows; Represents the conjugate of wavelet basis functions; Fourier transform representing the guided wave signal; Fourier transform representing wavelet basis functions; Through the above continuous wavelet transform, the guided wave signals in the dataset containing prior knowledge are converted into two-dimensional frequency domain guided wave signals for feature extraction and analysis by subsequent deep learning networks, and a dataset containing two-dimensional frequency domain guided wave signals is formed. Step 4: Construct a deep learning network based on prior knowledge learning, and train the deep learning network using the constructed dataset containing two-dimensional frequency domain guided wave signals; the deep learning network performs feature learning and parameter optimization on the dataset containing two-dimensional frequency domain guided wave signals using the following formula: in, The loss function for the classification task is the cross-entropy function. These represent the source structure guided wave signal and its corresponding structural debonding state, respectively. Representative feature extraction network; The network represents the result prediction network; K represents the total number of deadhesion states; k represents a single deadhesion state. Through the above feature learning and its own parameter optimization, a deep learning network with feature extraction and result prediction functions is obtained. Step 5: The deep learning network obtained in Step 4 will be used for debonding diagnosis of the guided wave signal of the structure to be tested; the structure to be tested is attached with a distributed PZT system information source connected to the detection device.

2. The method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning according to claim 1, characterized in that: In Step 3: the wavelet basis function of the continuous wavelet transform uses the Morlet wavelet, which is correlated with the guided wave signal in both the time and frequency domains.

3. The method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning according to claim 2, characterized in that: In Step 3: The continuous wavelet transform spectrum of the guided wave signal is adjusted to the same order of magnitude by writing a Python script to adapt to the feature analysis of the deep learning network.

4. The method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning according to claim 1, characterized in that: In Step 4: the feature extraction network of the deep learning network adopts a convolutional neural network architecture adapted to use a two-dimensional convolutional kernel in the backbone layer of the convolutional neural network; the result prediction network of the deep learning network adopts a fully connected network architecture that uses fully connected layers for feature analysis and result output.

5. The method for detecting debonding at reinforced concrete interfaces based on prior knowledge learning according to claim 1, characterized in that: The deep learning network is trained using supervised learning methods, minimizing the loss function based on backpropagation and gradient descent, so that the source model can fully learn the spatial distribution and mapping relationship of the source structural data.

Images