Deep learning admittance feature-based steel bar corrosion damage identification method and system

Abstract

The invention discloses a steel bar corrosion damage identification method and system based on deep learning admittance characteristics, and relates to the technical field of concrete structure detection, and the method comprises the steps: carrying out the frequency sweeping in a preset frequency band, and obtaining admittance signals of a reinforced concrete structure in different corrosion rate states; gaussian noise is added to the obtained admittance signal data, each data point is added to the random number conforming to Gaussian distribution, and a new admittance signal is obtained; the measured admittance signals and the newly generated admittance signals form training data, and a one-dimensional convolutional neural network is trained; and inputting an admittance signal of the reinforced concrete in an unknown corrosion state into the trained one-dimensional convolutional neural network to obtain a predicted corrosion rate. The admittance signal is an indication standard of structural integrity change caused by concrete reinforcement corrosion, the admittance signals of the reinforced concrete structure in different corrosion rate states are acquired, and the acquired data are processed and trained to obtain a network, so that the unknown corrosion state of the reinforced concrete can be accurately identified.

Description

Technical Field

[0001] The present application relates to the technical field of concrete structure detection, and specifically to a method and system for identifying steel bar corrosion damage based on deep learning admittance features. Background Art

[0002] As the world's most widely used building material, reinforced concrete has many advantages such as high strength, high rigidity, good durability, and mature construction technology. Therefore, it is widely used in various engineering structures. In the process of coordinated work, reinforced concrete has strong corrosion resistance, and because cement contains When the steel bar contacts with it, a dense oxide film will be formed on the surface, which prevents oxygen from contacting the steel bar and protects the steel bar from rusting. This is why the durability of reinforced concrete is often ignored. However, with the extensive use of reinforced concrete in various projects, durability issues are constantly emerging. In the actual use of structures, due to the carbonization of concrete and possible ion erosion, it is easy to cause corrosion of steel bars in concrete, which may cause certain safety hazards to the entire structure. Once corrosion begins, it will develop rapidly. The corrosion products on the surface of the steel bars are continuously produced, resulting in a corresponding reduction in the cross-sectional area of ​​the original steel bars. Moreover, since the volume of the steel bar corrosion products is much larger than the steel bars themselves, this causes the surrounding concrete to crack or peel under tension. As a result, the steel bars are directly exposed to the environment, which in turn aggravates their own corrosion. If the rusted parts of the steel bars are not repaired in time, the bearing capacity of the structure and components will be greatly weakened, which will be difficult to repair at the least, and the collapse of the structure and casualties will occur at the worst. Therefore, by monitoring the corrosion damage of steel bars in reinforced concrete, it is possible to understand the structural state before the structural corrosion can be observed by the naked eye, and then provide effective early warning information to avoid destructive behavior of the structure and provide reference information for the maintenance and maintenance of the later structure.

[0003] In traditional technology, the identification of steel corrosion damage uses piezoelectric impedance technology to obtain the impedance spectrum of reinforced concrete structures under different corrosion rates. The differences in different impedance spectra are analyzed to verify the feasibility of using the impedance spectrum of reinforced concrete under different corrosion rates to identify corrosion damage. Piezoelectric ceramics have positive and negative piezoelectric effects. The positive piezoelectric effect makes it a sensor, and its negative piezoelectric effect makes it an exciter. The piezoelectric impedance technology can be realized by using these two functions at the same time. The piezoelectric sensor is pasted on the surface of the structure or buried in it, and the wire is welded and connected to the impedance analyzer. The piezoelectric sheet is driven in the AC electric field, deformed and drives the structure to vibrate. The vibration deformation of the structure causes the deformation of the piezoelectric sheet coupled with it, and this deformation causes new current to be generated in the piezoelectric sheet. The ratio of the input voltage to the measured current is the measured impedance, which can be measured by the impedance analyzer. The damage to the structure will cause changes in the structural characteristics, resulting in changes in the structural response and impedance. Therefore, the damage of the structure can be effectively inferred by the information of the impedance obtained. Generally, the state of the structure before rusting is used as the baseline reference state, and then the impedance signals under different corrosion rates are obtained according to the same method to compare the difference with the baseline state and identify the local damage of the structure, thereby realizing the identification of different corrosion rates of the structure and judging its corrosion state.

[0004] However, this method of identifying corrosion damage to reinforced concrete also has its problems. Although this method can use the changes in the impedance spectrum obtained by measuring the piezoelectric film to reflect the different degrees of corrosion of the structure, the judgment of the degree of corrosion mostly remains at the qualitative level, and it is difficult to achieve quantitative analysis and characterization of the corrosion rate of reinforced concrete structures; and because the amount of data obtained by piezoelectric ceramics is large, it takes a lot of time and energy for humans to process the data. Summary of the invention

[0005] In order to solve the above technical problems, this application proposes the following technical solutions: In a first aspect, an embodiment of the present application provides a method for identifying steel bar corrosion damage based on deep learning admittance features, comprising: The admittance signal of the reinforced concrete structure under different corrosion rates is obtained by sweeping the frequency in the preset frequency band; Add Gaussian noise to the obtained admittance signal data, add each data point to a random number that conforms to the Gaussian distribution, and obtain a new admittance signal; The measured and newly generated admittance signals are combined into training data to train a one-dimensional convolutional neural network; The admittance signal of reinforced concrete under unknown corrosion state is input into the trained one-dimensional convolutional neural network to obtain the predicted corrosion rate.

[0006] In a possible implementation, the admittance signal of the reinforced concrete structure under different corrosion rate states obtained by frequency sweeping within a preset frequency band includes: Set up a PZT piezoresistor on the surface of the reinforced concrete structure; Welding a wire with a clamp on the PZT piezoelectric impedance and connecting it to an impedance analyzer; The dynamic impedance information of the PZT piezoresistive impedance is collected by the impedance analyzer, and the dynamic impedance information is fed back as an admittance signal of the reinforced concrete structure.

[0007] In a possible implementation, collecting the dynamic impedance information of the PZT piezoresistive impedance by the impedance analyzer includes: The PZT piezoresistor is set in a reinforced concrete structure to simplify it into a single-degree-of-freedom spring-mass-damper system; According to the piezoelectric equation, vibration equation and boundary conditions, the mechanical impedance piezoelectric coupling equation is obtained: in: and They are the piezoelectrical impedance of PZT and the mechanical impedance of reinforced concrete structure; is an imaginary unit; is the angular frequency of the driving voltage; , and are the width, length and thickness of the PZT piezoresistance; is the complex Young's modulus of the PZT piezoresistive material under zero electric field; is the coupled piezoelectric constant at zero stress in the z and x directions; is the complex dielectric constant of the PZT piezoresistance at zero stress; and Represent the mechanical loss factor and dielectric loss factor of PZT piezoresistance respectively; is related to the excitation frequency The relevant wave number; The expression is as follows: in: is the material density of the structure; Complex electromechanical admittance of PZT piezoresistance can be interpreted as , which contains real and imaginary parts, the real part is called conductance G, and the imaginary part is called susceptance B.

[0008] In a possible implementation, adding Gaussian noise to the obtained admittance signal data, adding each data point to a random number that conforms to the Gaussian distribution, and obtaining a new admittance signal includes: Adding Gaussian noise to the obtained admittance signal data; Each data point is added with a random number that conforms to the Gaussian distribution to obtain new generated data to achieve admittance signal data enhancement.

[0009] In a possible implementation, the probability density of the Gaussian noise obeys Gaussian distribution, and its probability density function is: Add each data point to a random number that conforms to the Gaussian distribution to get new generated data: Where: is the element in the original data, is the element after data augmentation.

[0010] In a possible implementation, the generated admittance data is obtained by selecting sensitive frequency bands to calculate the RMSD and MAPD indicators between the generated data under different working conditions to evaluate the effectiveness of the generated data. The RMSD and MAPD evaluation indicators of the generated data are stable and are used for the training and testing of the one-dimensional convolutional neural network. The calculation formulas of the RMSD and MAPD are as follows: Where: , They are the electrical admittance signal values ​​of the structure in the non-corrosion state and after corrosion respectively; n is the number of frequency points collected.

[0011] In a possible implementation, the measuring and newly generated admittance signals are combined into training data to train a one-dimensional convolutional neural network, including: The obtained admittance signal and the newly generated admittance signal are combined into training data, and Z-score normalization is performed to obtain one-dimensional input data: The Z-Score normalization formula is as follows: in: Where: n is the number of vector elements, is a specific data in the vector; Label the processed data according to the working conditions, and divide the data into training set, validation set and test set according to the preset ratio; The normalized data is sent as input data to the one-dimensional convolutional neural network for training; With the goal of minimizing the error between the corrosion rate predicted by the one-dimensional convolutional neural network and the actual corrosion rate of the admittance signal, the trained one-dimensional convolutional neural network is trained until convergence.

[0012] In a possible implementation, the loss function used to train the one-dimensional convolutional neural network is a cross entropy function, and the formula is: in, Takes 0 or 1, if the category is predicted to be the correct category, then ,otherwise , there are C categories in total, Indicates The probability that the class is predicted correctly, is the label smoothing parameter.

[0013] In a possible implementation, the training to convergence to obtain a trained one-dimensional convolutional neural network with the goal of minimizing the error between the corrosion rate predicted by the one-dimensional convolutional neural network and the actual corrosion rate of the admittance signal includes: Use the training set to train the neural network, and the validation set to verify the loss and accuracy of the neural network training; Then, based on the results of the validation set, the hyperparameters are continuously modified to improve the efficiency of the neural network, reduce the loss, improve the accuracy, and make its convergence and stability meet the requirements; Finally, the performance of the neural network is tested through the test set, so that its accuracy reaches the preset requirements, the loss decreases and tends to be stable, and a trained one-dimensional convolutional neural network is obtained.

[0014] In a second aspect, the embodiment of the present application provides a steel bar corrosion damage identification system based on deep learning admittance features, including: A data acquisition module, used to obtain the admittance signal of the reinforced concrete structure under different corrosion rate conditions by frequency sweeping within a preset frequency band; A data processing module is used to add Gaussian noise to the obtained admittance signal data, and add each data point to a random number that conforms to the Gaussian distribution to obtain a new admittance signal; A network training module, used to form training data from measured and newly generated admittance signals to train a one-dimensional convolutional neural network; The corrosion damage identification module is used to input the admittance signal of reinforced concrete in an unknown corrosion state into the trained one-dimensional convolutional neural network to obtain the predicted corrosion rate.

[0015] In the embodiment of the present application, the admittance signal is an indicator of changes in the structural integrity caused by corrosion of concrete reinforcement. By acquiring the admittance signal of the reinforced concrete structure under different corrosion rate states, the trained network after processing the acquired data can realize accurate identification of unknown corrosion states of reinforced concrete. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of a process for identifying steel bar corrosion damage based on deep learning admittance features provided in an embodiment of the present application; Figure 2 A schematic diagram of the electromechanical coupling between a piezoelectric sheet and a structure provided in an embodiment of the present application; Figure 3 A schematic diagram of a PZT piezoelectric sheet provided in an embodiment of the present application; Figure 4 The admittance measurement system of the reinforced concrete structural member provided in the embodiment of the present application; Figure 5 A comparison diagram of the admittance curves of the working condition 2 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Figure 6 A comparison diagram of the admittance curves of the working condition 3 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Figure 7 A comparison diagram of the admittance curves of the working condition 4 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Figure 8 A comparison diagram of the admittance curves of the working condition 5 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Fig. 9 A comparison diagram of the admittance curves of the working condition 6 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Fig.10 A comparison diagram of the admittance curves of the working condition 7 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Fig.11 A comparison diagram of the admittance curves of the working condition 8 and the non-corrosion working condition (working condition 1) provided in the embodiment of the present application; Fig.12 The RMSD value between the generated data under different working conditions provided in the embodiment of the present application; Fig.13 The MAPD value between the generated data under different working conditions provided in the embodiment of the present application; Fig.14 A one-dimensional convolutional neural network training set accuracy curve diagram provided in an embodiment of the present application; Fig.15 A one-dimensional convolutional neural network training set loss curve diagram provided in an embodiment of the present application; Fig.16A one-dimensional convolutional neural network verification set accuracy curve diagram provided in an embodiment of the present application; Fig.17 A one-dimensional convolutional neural network validation set loss curve diagram provided in an embodiment of the present application; Fig.18 A one-dimensional convolutional neural network test set accuracy curve diagram provided in an embodiment of the present application; Fig.19 A one-dimensional convolutional neural network test set loss curve diagram provided in an embodiment of the present application; Fig. 20 A confusion matrix diagram output by a one-dimensional convolutional neural network test set provided in an embodiment of the present application; Fig.21 A schematic diagram of a steel bar corrosion damage identification system based on deep learning admittance features provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present solution is described below in conjunction with the accompanying drawings and specific implementation methods.

[0018] See also Figure 1 , the steel bar corrosion damage identification method based on deep learning admittance feature provided by the embodiment cannot be applied, including: S101, the admittance signal of the reinforced concrete structure under different corrosion rate states is obtained by sweeping the frequency within a preset frequency band.

[0019] A PZT piezoresistor is set on the surface of a reinforced concrete structure; a wire with a clamp is welded on the PZT piezoresistor to connect it to an impedance analyzer; dynamic impedance information of the PZT piezoresistor is collected by the impedance analyzer, and the dynamic impedance information is fed back as an admittance signal of the reinforced concrete structure.

[0020] The impedance analyzer used in the example of this application is the Agilent 4294A precision impedance analyzer, and the test frequency range is: 40Hz~110MHz. The fixture used is 16047E. During use, the wire welded on the PZT is connected to the fixture to connect to the analyzer. The fixture must be calibrated before use, and then the measurement is performed. Since the admittance data collected in the frequency band exceeding 500kHz will be more sensitive to environmental changes rather than damage to the structure itself, the preset frequency band for measurement is selected as 40-500kHz, the number of measurement points is 801 points, and the voltage is 1V. The collected data is connected to the computer via a network cable, and the local area network is set up using Agilent Connection Expert. The analyzer and the computer are in the same local area network, and the results are output to Excel using the 4294A Data Transfer Program program to facilitate subsequent data processing and analysis. During the experimental data collection process, the ambient temperature under different working conditions is guaranteed to differ by no more than 5°C, avoiding measurement errors caused by changes in ambient temperature.

[0021] like Figure 2 and Figure 3 As shown in the figure, the piezoelectric impedance is the reflection of the dynamic impedance information of the mechanical structure on the PZT piezoelectric impedance. The slight change of the mechanical structure causes the PZT impedance to change. Conversely, the change of the mechanical structure can be known from the impedance information. We can simplify the PZT bonded to the structure into a single degree of freedom spring-mass-damper system as shown in the figure. According to the piezoelectric equation, vibration equation and boundary conditions, the mechanical impedance piezoelectric coupling equation is obtained: in: and are the mechanical impedances of the PZT and the main structure, respectively; is an imaginary unit; is the angular frequency of the driving voltage; , and are the width, length and thickness of the PZT sheet, respectively; is the complex Young’s modulus of the PZT material under zero electric field; is the coupled piezoelectric constant at zero stress in the z and x directions; is the complex dielectric constant of PZT at zero stress; and Represent the mechanical loss factor and dielectric loss factor of PZT respectively; is related to the excitation frequency The relevant wave number; The expression is as follows: in: is the material density of the structure.

[0022] Complex electromechanical admittance of PZT sheets can be interpreted as , which contains real and imaginary parts, the real part is called conductance (G) and the imaginary part is called susceptance (B). The piezoelectric coupling equation of mechanical impedance shows that the electromechanical admittance is directly related to the impedance of the main structure. The change of PZT electromechanical admittance within a certain frequency range is similar to the frequency response function of the structure, and this function contains key information about the health status of the structure. Since all other PZT characteristics remain unchanged, the structural mechanical impedance (Z) is the only characteristic that determines the overall admittance. Therefore, any change in the electromechanical admittance signal may be an indication of a change in structural integrity caused by the presence of structural damage. It can be concluded through experiments that the damage index calculated by the conductance signal has a better effect on the damage detection of the structure, so the present invention can accurately reflect the damage state of the structure by processing and analyzing the conductance signal.

[0023] The accelerated corrosion test of reinforced concrete beams adopts the veneer impressed current accelerated corrosion method, using three Huayi HY3005B DC power supplies to connect the reinforced concrete beam steel bars and the 304 stainless steel plate (400×200×3mm) placed on the surface for electrification corrosion. The corrosion rate of the steel bars can be controlled by adjusting the impressed current and the power-on time, and can be quantitatively described according to Faraday's law in theory. The corrosion time expression obtained according to Faraday's law is: in, is the rusting time after power on (s); To set the steel bar corrosion rate; is the Faraday constant, which is 96320C / mol in the present embodiment; is the radius of the steel bar to be corroded (cm); and are the density and molar mass of steel bars, respectively. Here, they are taken according to the density and molar mass of iron; To set the applied current density (A / cm2).

[0024] like Figure 4 As shown in the figure, the power is turned off after the predetermined corrosion degree is reached, and the admittance signal is measured using an Agilent 4294A commercial impedance meter after the reinforced concrete beam is dried. Considering the actual engineering situation, 8 groups of working conditions are pre-set for the structural test, namely: 0%, 1%, 2%, 3%, 4%, 6%, 8%, and 10%. Under each group of working conditions, 3 groups of impedance spectra are measured using an impedance meter. The impedance spectrum measurement frequency band is selected as 40Hz~500kHz, and 801 measurement points are selected.

[0025] like Figure 5-Figure 11 As shown in Figure 2, by comparing the admittance data of each working condition and the lossless working condition, it can be found that: (1) Figure 5 and Figure 6 The shapes of the admittance data curves of working conditions 2 and 3 are similar, with only minor local differences; when comparing working conditions 2, 3 and working condition 1 (non-destructive), the shape change rules are difficult to grasp, and the changes are so small that they cannot be identified by human experience alone. (2) Figures 7 to 11 The corresponding admittance data of working conditions 4 to 8 have similar shapes. As the corrosion rate increases, the resonance peak of the admittance curve shifts to the left and upward, but the difference between the working conditions is very small, and it is difficult to distinguish the difference between them with the naked eye. (3) The reason for the difference between the admittance diagrams of working conditions 2 to 3 and working conditions 4 to 8 is that in working conditions 2 to 3, corrosion has just occurred and the passive film has not been destroyed. At this time, corrosion only causes a slight decrease in structural stiffness. In working conditions 4 to 8, the passive film is destroyed, the corrosion rate is accelerated, and the structural stiffness changes significantly, resulting in changes in the admittance data, resulting in a large difference in the admittance diagrams of the two working conditions. Both working conditions lack reliable indicators to quantify the mapping relationship between the admittance data and the corrosion rate conditions, and thus lack an effective method to identify the corrosion rate through the admittance data.

[0026] S102, adding Gaussian noise to the obtained admittance signal data, adding each data point to a random number that conforms to the Gaussian distribution, and obtaining a new admittance signal.

[0027] In practical engineering applications, it may be very difficult and expensive to obtain a large amount of labeled data. When preparing training data, the present invention adds Gaussian noise to the obtained admittance signal data, adds each data point to a random number that conforms to the Gaussian distribution, and obtains new generated data, which effectively achieves the data enhancement effect. Gaussian noise is added to the obtained admittance data to generate new admittance data. A total of 20 groups of data are added to the measured data under each working condition for training and testing the neural network. Gaussian noise refers to a type of noise whose probability density function obeys a Gaussian distribution (i.e., normal distribution). Mathematically, Gaussian noise is a type of noise generated by adding normally distributed random values ​​with a mean of zero and a standard deviation (σ) to the input data.

[0028] The probability density of Gaussian noise obeys Gaussian distribution (normal distribution), and its probability density function is: Specifically, each data point is added with a random number that conforms to the Gaussian distribution to obtain new generated data: Where: is the element in the original data, is the element after data augmentation.

[0029] The generated admittance data was selected in the sensitive frequency band of 150kHz-300kHz, and the RMSD and MAPD indicators between the generated data under different working conditions were calculated to evaluate the effectiveness of the generated data. Fig.12 and Fig.13 As shown in the figure, the RMSD and MAPD evaluation indicators of the generated data are stable and can be used for the training and testing of one-dimensional convolutional neural networks. The calculation formulas of RMSD and MAPD are as follows: Where: , They are the admittance signal values ​​of the structure in a non-corroded state and after corrosion, respectively; n is the number of frequency points collected, which is specifically 801 in the example of the present invention.

[0030] S103, the measured and newly generated admittance signals are combined into training data to train a one-dimensional convolutional neural network.

[0031] In this embodiment, the basic structure of the one-dimensional convolutional neural network is: six convolutional layers and two fully connected layers.

[0032] The six convolutional layers perform convolution operations in sequence. The input data of the first convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 16×398×1; the input data of the second convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 16×196×1; the input data of the third convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 32×95×1; the input data of the fourth convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 32×45×1; the input data of the fifth convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 64×20×1; the input data of the sixth convolutional layer is 1 convolution kernel of size 7, step size 1, and the output result is 64×7×1; the ReLU activation function is nonlinearly transformed after the convolutional layer, and the output size remains unchanged at 64×7×1.

[0033] It is followed by two fully connected layers. The convolutional layer output data is first expanded to obtain 1×448 one-dimensional data. The first fully connected layer first performs a linear transformation on the one-dimensional data output by the convolutional layer, transforming the 1×448 data into 1×256 data, and then performs a nonlinear transformation using the ReLU activation function. The second fully connected layer only performs a linear transformation, transforming the 1×256 data into 8×1 data, and then outputs the final result through the Softmax function transformation.

[0034] The obtained measured data and generated data are combined into training data, and Z-score normalization is performed to obtain one-dimensional input data of 1×801. Correspondingly, the data Z-Score normalization formula is as follows: in: Wherein: n is the number of vector elements, which is specifically 801 in the example of the present invention; is a specific data in the vector.

[0035] The processed data are labeled with values ​​from 0 to 7 according to the working conditions, and the data are divided into training set, validation set and test set in the ratio of 7:3:10. The normalized data are sent as input data to the one-dimensional convolutional neural network for training. The training is carried out until convergence to obtain a trained one-dimensional convolutional neural network with the goal of minimizing the error between the corrosion rate predicted by the one-dimensional convolutional neural network and the actual corrosion rate of the admittance signal.

[0036] Specifically, the loss function for training a one-dimensional convolutional neural network in the embodiment of the present application is a cross entropy function, in which the cross entropy is mainly used to judge the degree of proximity between the actual output and the expected output, and it combines the two functions of LogSoftmax and NLLLoss. The cross entropy function is very effective in training neural networks for multi-classification tasks. At the same time, label smoothing is used to reduce the risk of overfitting of the model and improve the stability of model training. The loss function of the one-dimensional convolutional neural network during training is: in, Takes 0 or 1, if the category is predicted to be the correct category, then ,otherwise , there are C categories in total, Indicates The probability that the class is predicted correctly, is the label smoothing parameter, which is taken as 0.1 in this model.

[0037] The training of a one-dimensional convolutional neural network includes: Design a convolutional neural network according to the classification problem in supervised learning. Use the training set to train the neural network; verify the loss and accuracy of the neural network training with the validation set, and then continuously modify the hyperparameters based on the results obtained from the validation set to improve the efficiency of the neural network, reduce the loss, improve the accuracy, and make its convergence and stability meet the requirements. Finally, test the performance of the neural network with the test set, so that its accuracy reaches 100%, the loss decreases and tends to be stable, and the one-dimensional convolutional neural network model that meets the requirements is the final model.

[0038] The normalized data is sent as input data to the one-dimensional convolutional neural network for training. Fig.14 and Fig.15 It can be observed that: the accuracy curve of the training set increases rapidly with the increase of the number of training rounds, and then stabilizes to 100%, which indicates that the neural network has completely learned the data features of the training set; and the loss of the training set decreases rapidly and reaches stability; the accuracy and loss of the training set both decrease rapidly during the training process within 50 rounds and then stabilize, indicating that the one-dimensional convolutional neural network has strong learning ability and stable performance. Fig.16 and Fig.17 The accuracy and loss curves of the validation set are basically consistent with those of the training set, indicating that the neural network does not overfit and is also very robust to new data. Fig.18 and Fig.19 The accuracy and loss in represent the final performance of the one-dimensional convolutional neural network, which is basically consistent with the training set and validation set, and meets our requirements for the one-dimensional convolutional neural network.

[0039] S104, inputting the admittance signal of the reinforced concrete in the unknown corrosion state into the trained one-dimensional convolutional neural network to obtain the predicted corrosion rate.

[0040] After the network training is completed, the one-dimensional conductivity data of the unknown corrosion rate condition is input, and the one-dimensional convolutional neural network can output the recognition result of the corrosion rate. Fig. 20 The confusion matrix in the figure shows the relationship between the true label and the predicted label of the data in the test set. All 10 groups of data in each of the 8 working conditions were successfully predicted, indicating that the model has good performance and can achieve the goal of quantitatively and accurately identifying the corrosion rate.

[0041] Corresponding to the method for identifying steel corrosion damage based on deep learning admittance features provided in the above embodiment, the present application also provides an embodiment of a system for identifying steel corrosion damage based on deep learning admittance features.

[0042] See also Fig.21 The steel bar corrosion damage identification system 20 based on deep learning admittance features in this embodiment includes: The data acquisition module 201 is used to obtain the admittance signal of the reinforced concrete structure under different corrosion rate states by frequency sweeping within a preset frequency band.

[0043] The data processing module 202 is used to add Gaussian noise to the obtained admittance signal data, and add each data point to a random number that conforms to the Gaussian distribution to obtain a new admittance signal.

[0044] The network training module 203 is used to combine the measured and newly generated admittance signals into training data to train a one-dimensional convolutional neural network.

[0045] The corrosion damage identification module 204 is used to input the admittance signal of the reinforced concrete in an unknown corrosion state into a trained one-dimensional convolutional neural network to obtain a predicted corrosion rate.

[0046] In the embodiments of the present application, "at least one" refers to one or more, and "plurality" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and subsequent associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0047] The above is only a specific implementation of the present application. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. The protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for identifying steel bar corrosion damage based on deep learning admittance features, characterized in that: include: The admittance signal of the reinforced concrete structure under different corrosion rates is obtained by sweeping the frequency in the preset frequency band; Add Gaussian noise to the obtained admittance signal data, add each data point to a random number that conforms to the Gaussian distribution, and obtain a new admittance signal; The measured and newly generated admittance signals are combined into training data to train a one-dimensional convolutional neural network; The admittance signal of reinforced concrete under unknown corrosion state is input into the trained one-dimensional convolutional neural network to obtain the predicted corrosion rate.

2. The steel bar corrosion damage identification method based on deep learning admittance feature according to claim 1 is characterized in that: The admittance signal of the reinforced concrete structure under different corrosion rate conditions obtained by frequency sweeping within the preset frequency band includes: Set up a PZT piezoresistor on the surface of the reinforced concrete structure; Welding a wire with a clamp on the PZT piezoelectric impedance and connecting it to an impedance analyzer; The dynamic impedance information of the PZT piezoresistive impedance is collected by the impedance analyzer, and the dynamic impedance information is fed back as an admittance signal of the reinforced concrete structure.

3. The steel bar corrosion damage identification method based on deep learning admittance feature according to claim 1 is characterized in that: The collecting the dynamic impedance information of the PZT piezoresistive impedance by the impedance analyzer includes: The PZT piezoresistor is set in a reinforced concrete structure to simplify it into a single-degree-of-freedom spring-mass-damper system; According to the piezoelectric equation, vibration equation and boundary conditions, the mechanical impedance piezoelectric coupling equation is obtained: in: and They are the PZT piezoresistance and the mechanical impedance of the reinforced concrete structure; is an imaginary unit; is the angular frequency of the driving voltage; , and are the width, length and thickness of the PZT piezoresistance, respectively; is the complex Young's modulus of the PZT piezoresistive material under zero electric field; is the coupled piezoelectric constant at zero stress in the z and x directions; is the complex dielectric constant of the PZT piezoresistance at zero stress; and Represent the mechanical loss factor and dielectric loss factor of PZT piezoresistance respectively; is related to the excitation frequency The relevant wave number; The expression is as follows: in: is the material density of the structure; Complex electromechanical admittance of PZT piezoresistance can be interpreted as , which contains real and imaginary parts, the real part is called conductance G, and the imaginary part is called susceptance B.

4. The steel bar corrosion damage identification method based on deep learning admittance feature according to claim 1 is characterized in that: The step of adding Gaussian noise to the obtained admittance signal data and adding each data point to a random number that conforms to the Gaussian distribution to obtain a new admittance signal includes: Adding Gaussian noise to the obtained admittance signal data; Each data point is added with a random number that conforms to the Gaussian distribution to obtain new generated data to achieve admittance signal data enhancement.

5. The method for identifying steel bar corrosion damage based on deep learning admittance features according to claim 4 is characterized in that: The probability density of the Gaussian noise obeys Gaussian distribution, and its probability density function is: Add each data point to a random number that conforms to the Gaussian distribution to get new generated data: Where: is the element in the original data, is the element after data augmentation.

6. The steel bar corrosion damage identification method based on deep learning admittance feature according to claim 5 is characterized in that: The generated admittance data obtained are selected from sensitive frequency bands to calculate the RMSD and MAPD indicators between the generated data under different working conditions to evaluate the effectiveness of the generated data. The RMSD and MAPD evaluation indicators of the generated data are stable and are used for the training and testing of the one-dimensional convolutional neural network. The calculation formulas of the RMSD and MAPD are as follows: Where: , They are the electrical admittance signal values ​​of the structure in the non-corrosion state and after corrosion respectively; n is the number of frequency points collected.

7. The method for identifying steel bar corrosion damage based on deep learning admittance features according to claim 1 is characterized in that: The measured and newly generated admittance signals are combined into training data to train a one-dimensional convolutional neural network, including: The obtained admittance signal and the newly generated admittance signal are combined into training data, and Z-score normalization is performed to obtain one-dimensional input data: The Z-Score normalization formula is as follows: in: Where: n is the number of vector elements, is a specific data in the vector; Label the processed data according to the working conditions, and divide the data into training set, validation set and test set according to the preset ratio; The normalized data is sent as input data to the one-dimensional convolutional neural network for training; With the goal of minimizing the error between the corrosion rate predicted by the one-dimensional convolutional neural network and the actual corrosion rate of the admittance signal, the trained one-dimensional convolutional neural network is trained until convergence.

8. The steel bar corrosion damage identification method based on deep learning admittance feature according to claim 1 is characterized in that: The loss function used in training the one-dimensional convolutional neural network is the cross entropy function, and the formula is: in, Takes 0 or 1, if the category is predicted to be the correct category, then ,otherwise , there are C categories in total, Indicates The probability that the class is predicted correctly, is the label smoothing parameter.

9. The method for identifying steel corrosion damage based on deep learning admittance features according to claim 7 is characterized in that: The method aims to minimize the error between the corrosion rate predicted by the one-dimensional convolutional neural network and the actual corrosion rate of the admittance signal, and trains until convergence to obtain a trained one-dimensional convolutional neural network, including: Use the training set to train the neural network, and the validation set to verify the loss and accuracy of the neural network training; Then, based on the results of the validation set, the hyperparameters are continuously modified to improve the efficiency of the neural network, reduce the loss, improve the accuracy, and make its convergence and stability meet the requirements; Finally, the performance of the neural network is tested through the test set, so that its accuracy reaches the preset requirements, the loss decreases and tends to be stable, and a trained one-dimensional convolutional neural network is obtained.

10. A steel bar corrosion damage identification system based on deep learning admittance features, characterized in that: include: A data acquisition module, used to obtain the admittance signal of the reinforced concrete structure under different corrosion rate conditions by frequency sweeping within a preset frequency band; A data processing module is used to add Gaussian noise to the obtained admittance signal data, and add each data point to a random number that conforms to the Gaussian distribution to obtain a new admittance signal; A network training module, used to form training data from measured and newly generated admittance signals to train a one-dimensional convolutional neural network; The corrosion damage identification module is used to input the admittance signal of reinforced concrete in an unknown corrosion state into the trained one-dimensional convolutional neural network to obtain the predicted corrosion rate.

Images