Concrete structure internal defect identification method based on two-stage unsupervised deep learning algorithm

Abstract

The invention discloses a concrete structure internal defect identification method based on a two-stage unsupervised deep learning algorithm, and the method only needs to measure the signal response of a complete structure, and then inputs a signal into a first-stage unsupervised network, namely, a depth auto-encoder with physical constraints, for reconstruction. And eight differential damage sensitive features between the original signal and the reconstructed signal are extracted. And inputting the extracted damage features into a second-stage unsupervised network single-classification support vector machine for training again, and outputting an offset distance. The training models of the two stages can be used for detecting other concrete structures with internal defects, and finally defect positioning and damage degree identification are carried out on the concrete structures by taking the output offset distance as a damage evaluation index. The result shows that the detection precision can be effectively improved by utilizing the detection method, and a reliable basis is provided for an engineer to accurately grasp the position of the internal defect of the concrete structure and repair the internal defect.

Description

Technical Field

[0001] The present invention belongs to the technical field of using deep learning technology to detect internal defects of concrete, and in particular relates to a method for locating internal defects and identifying the degree of damage of a concrete structure based on a two-stage unsupervised deep learning algorithm. Technical Background

[0002] Concrete structures are prone to internal debonding and voids due to loose grouting and long-term loads. Therefore, it is of great significance to quickly and accurately locate internal defects and identify the degree of damage in concrete. Timely disclosure of potential structural dangers can help timely repair and reinforce the structure, reduce maintenance and repair costs, and even avoid major accidents.

[0003] In recent years, the method of analyzing impact echo signals based on deep learning has provided an effective means for defect detection and has received widespread attention from experts and scholars. However, most of the existing deep learning networks are data-driven and require the collection of data from intact areas and other damaged areas of the structure for training, but collecting sufficient damage data is often difficult to achieve. At the same time, the data-driven model has poor interpretability, and the training data needs to be manually labeled, which greatly reduces the efficiency of training and detection. In order to solve these problems, the present invention proposes a method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm. It only needs to measure the acceleration response signal of the intact structure for training. The results show that the proposed method can effectively improve the detection accuracy of internal defects in concrete structures. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm. In the first stage, the method uses an acceleration sensor to collect the impact echo signal of a non-destructive concrete structure. Secondly, the signal is input into a deep autoencoder with physical constraints (PC-DAE) for reconstruction, and the damage-sensitive features between the reconstructed signal and the original signal are extracted. In the second stage, the damage-sensitive features are input into a single-class support vector machine (OC-SVM) for training to obtain the offset distance as an indicator for damage evaluation. In the test stage, the vibration signals of other structures to be tested are first input into the PC-DAE training model for signal reconstruction, and then the damage-sensitive features are input into the OC-SVM training model for damage identification. According to the output offset distance, the internal defects of the concrete can be located and the degree of damage can be evaluated.

[0005] In order to achieve the above object, the technical solution of the present invention comprises the following steps:

[0006] Step 1: Collect the impact echo vibration signal of the intact and undamaged concrete structure.

[0007] Step 2: Construct the first stage of the unsupervised algorithm, introduce the motion equation of the single degree of freedom system as a physical constraint in the autoencoder, and reconstruct the original impact echo signal collected by the lossless structure.

[0008] Step 3: Extract the eight differences between the original signal input in step 2 and the reconstructed signal output and the original signal as damage-sensitive features. For normal data, the reconstruction error is usually small, but for abnormal data (such as structural damage), the reconstruction error will increase significantly.

[0009] Step 4: Construct the second-stage unsupervised algorithm, construct the eight damage-sensitive features into a one-dimensional sequence in order, and input them into OC-SVM for training. The training data will be clustered in the same area as much as possible. This area can be regarded as the "normal" area of ​​the data. Finally, the output offset distance is used as the damage evaluation index.

[0010] Step 5: Use the two-stage training model to detect other concrete with internal defects. The test signal will first pass through the first-stage training model to obtain damage-sensitive features, and then pass through the second-stage training model to obtain the offset distance. Finally, the internal defects in the concrete can be located and the damage degree can be evaluated based on the predicted offset distance.

[0011] Preferably, the step 2 is specifically:

[0012] A deep autoencoder is an unsupervised network consisting of an encoder and a decoder. It can simplify high-dimensional data into a low-dimensional form, and then output the reconstructed vector with the same dimension as the original vector. Adding physical constraints (motion equations of a single-degree-of-freedom system) to the deep autoencoder to form PC-DAE enhances the interpretability of the model and allows a more robust and reliable model to be learned from the data.

[0013] The network uses the gradient descent method to minimize the total loss to obtain the optimal network. The total loss of the network includes the data loss J D and physical loss P Two parts.

[0014] The total loss J is defined as:

[0015] J=α D J D +α P J P

[0016] Where α D and α P are the weights of data loss and physical loss respectively, and their sum is 1

[0017] When the PC-DAE model that only trains lossless data is used to test the data in the damaged area of ​​the structure, the changes in the damping and stiffness of the damaged area of ​​the structure affect the dynamic characteristics of the structure, which leads to certain differences between the data collected in the damaged area and the data collected in the lossless area. Secondly, because the training model is trained in the direction of lossless data, the output data when the abnormal data is reconstructed using the model is closer to lossless data. Therefore, when the damage data is input into the model, it will have a large difference from its reconstructed signal. According to this principle, the damaged area can be identified.

[0018] Preferably, in step 3, eight differences between the original signal input in step 2 and the reconstructed signal output and the original signal are extracted as damage sensitive features, specifically:

[0019] Step 3-1: Calculate the mean square error as feature 1. MSE is used to quantify the difference between the original input data and the reconstructed output data. Specifically, it measures the average square difference between the two data.

[0020] Step 3-2: Calculate the Pearson correlation coefficient as feature 2, and use the Pearson correlation coefficient to measure the linear correlation between the reconstructed data and the original data.

[0021] Step 3-3: Calculate the energy ratio as feature 3. Similar to the ratio of signal strength to background noise strength, the energy ratio of the original signal and the reconstructed signal can be used to represent the energy change of the signal.

[0022] Step 3-4: Calculate the power density ratio as feature 4. The power density ratio first uses discrete Fourier transform to convert the time domain signal to the frequency domain, and then calculates the power density ratio of all frequency components of the original signal and the reconstructed signal in the frequency domain. This indicator can effectively measure the power difference between the reconstructed signal and the original signal in the frequency domain.

[0023] Step 3-5: Calculate the frequency domain centroid ratio as feature 5. The frequency domain centroid of the signal reflects the center position of the signal's frequency domain energy distribution. To calculate the frequency domain centroid, each frequency component of the signal is used as a weight and is weighted and summed with its corresponding amplitude. Finally, the amplitudes of all frequency components are normalized.

[0024] Step 3-6: Calculate the main frequency ratio as feature 6. When analyzing the frequency characteristics of a signal, the main frequency is a key indicator and is often used to identify abnormal data. The time domain signal is converted into a frequency domain signal to reveal the amplitude of different frequency components in the signal. The larger the amplitude, the greater the energy of the signal at that frequency. The main frequency is the frequency component with the strongest amplitude.

[0025] Step 3-7: Calculate the total spectrum power ratio as feature 7. The spectrum area represents the total power of the signal in a specific frequency interval. The power in a certain frequency interval is obtained by integrating the power spectrum density in the interval. Calculating the total power of the signal in a specific frequency interval can analyze abnormal signals. The power of abnormal data will be different from that of normal data.

[0026] Step 3-8: Calculate the standard deviation of the energy change in each frequency band of the signal as feature 8. The enhancement and suppression effects of the signal collected in the structural damage area and the signal collected in the lossless area in each frequency band are also different. Therefore, analyzing the energy change of the reconstructed signal and the original input signal in each frequency band is the basis for judging whether damage occurs.

[0027] Preferably, the step 4 is specifically:

[0028] The second-stage unsupervised algorithm is constructed, and the eight damage-sensitive features are sequentially constructed into a one-dimensional sequence and input into the OC-SVM for training. Since only the damage-sensitive features of the data in the damage-free area are trained, the training data will be concentrated in the same area as much as possible, and this area can be regarded as the "normal" area of ​​the data. The optimal hyperplane is solved in the feature space. After the training is completed, the hyperplane separates the training data from the origin with the maximum boundary, and finally the offset distance is obtained as the damage evaluation index.

[0029] Preferably, the step 5 is specifically:

[0030] The two-stage training model is used to detect damage to other concretes with internal defects. The test signal will first pass through the first-stage training model to obtain damage sensitivity features, and then pass through the second-stage training model to obtain the offset distance. When the offset distance is less than 0, it is identified as an abnormal data point, because the damage sensitivity of the original data in the damaged area has changed significantly after reconstruction by the PC-DAE model. When the offset distance is less than 0, it indicates that the difference between the measured data and the non-destructive data is greater, and the degree of damage in the measured area is higher. Finally, the internal defects in the concrete can be located and the degree of damage can be evaluated based on the predicted offset distance.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. PC-DAE was constructed to reconstruct the impact echo signal of concrete, so that the data can be extracted and trained along the direction of physical guidance, which enhances the interpretability of the model;

[0033] 2. Eight new damage-sensitive features of the reconstructed signal and the original signal were extracted as the input of the second-stage OC-SVM, eliminating the interference of redundant information and improving the training efficiency;

[0034] 3. The entire training process only requires measuring the signal of the complete structure for training, and the training data does not need to be manually labeled, which solves the problem of difficulty in collecting sufficient damage data for training;

[0035] 4. Based on the offset distance output by OC-SVM, defects and damage extent of other concrete structures can be located. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flowchart of the invention.

[0037] Figure 2 This is a defect imaging diagram based on the method described in the embodiment of the present invention.

[0038] Figure 3 This is a defect damage degree assessment diagram based on the method described in the embodiment of the present invention. DETAILED DESCRIPTION

[0039] The principle of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments for explaining the content of the present invention and are not intended to limit the scope of the present invention.

[0040] Step 1: Collect the impact echo vibration signal of the intact and undamaged concrete structure.

[0041] Step 2: Construct the first stage of the unsupervised algorithm, introduce the motion equation of the single degree of freedom system as a physical constraint in the autoencoder, and reconstruct the original impact echo signal collected by the lossless structure.

[0042] Step 3: Extract the eight differences between the original signal input in step 2 and the reconstructed signal output and the original signal as damage-sensitive features. For normal data, the reconstruction error is usually small, but for abnormal data (such as structural damage), the reconstruction error will increase significantly.

[0043] Step 4: Construct the second-stage unsupervised algorithm, construct the eight damage-sensitive features into a one-dimensional sequence in order, and input them into OC-SVM for training. The training data will be clustered in the same area as much as possible. This area can be regarded as the "normal" area of ​​the data. Finally, the output offset distance is used as the damage evaluation index.

[0044] Step 5: Use the two-stage training model to detect other concrete with internal defects. The test signal will first pass through the first-stage training model to obtain damage-sensitive features, and then pass through the second-stage training model to obtain the offset distance. Finally, the internal defects in the concrete can be located and the damage degree can be evaluated based on the predicted offset distance.

[0045] The above steps are further described below.

[0046] The impact echo vibration signal of the intact and undamaged concrete structure is collected, and then the signal is input into the first-stage unsupervised network PC-DAE for reconstruction. The network consists of an encoder and a decoder, which can simplify high-dimensional data to a low-dimensional form, and then output the reconstructed vector with the same dimension as the original vector. PC-DAE introduces the motion equation of a single-degree-of-freedom system as a physical constraint, which enhances the interpretability of the model and can learn a more robust and reliable model from the data.

[0047] The network uses the gradient descent method to minimize the total loss to obtain the optimal network. The total loss of the network includes the data loss J D and physical loss P Two parts.

[0048] The total loss J is defined as:

[0049] J=α D J D +α P J P

[0050] Where α D and α P are the weights of data loss and physical loss respectively, and their sum is 1

[0051] When the PC-DAE model trained only on lossless data is used to test the data in the damaged area of ​​the structure, the changes in the damping and stiffness of the damaged area of ​​the structure affect the dynamic characteristics of the structure, which leads to certain differences between the data collected in the damaged area and the data collected in the lossless area. Secondly, because the training model is trained in the direction of lossless data, the output data when the abnormal data is reconstructed using the model is closer to lossless data. Therefore, when the damaged data is used as the input of the model, it will have a large difference from its reconstructed signal. According to this principle, the damaged area can be identified.

[0052] The difference between the reconstructed signal and the original signal output in PC-DAE can be used as a damage-sensitive feature. For normal data, the reconstruction error is usually small, while for abnormal data (such as structural damage), the reconstruction error will increase significantly. In step 3, the present invention selects 8 features between the reconstructed signal and the original signal as the input of OC-SVM, and calculates them according to the following steps.

[0053] Step 3-1: Calculate the mean square error as feature 1. MSE is used to quantify the difference between the original input data and the reconstructed output data. Specifically, it measures the average square difference between the two data.

[0054] Step 3-2: Calculate the Pearson correlation coefficient as feature 2, and use the Pearson correlation coefficient to measure the linear correlation between the reconstructed data and the original data.

[0055] Step 3-3: Calculate the energy ratio as feature 3. Similar to the ratio of signal strength to background noise strength, the energy ratio of the original signal to the reconstructed signal can be used to represent the energy change of the signal.

[0056] Step 3-4: Calculate the power density ratio as feature 4. The power density ratio first uses discrete Fourier transform to convert the time domain signal to the frequency domain, and then calculates the power density ratio of all frequency components of the original signal and the reconstructed signal in the frequency domain. This indicator can effectively measure the power difference between the reconstructed signal and the original signal in the frequency domain.

[0057] Step 3-5: Calculate the frequency domain centroid ratio as feature 5. The frequency domain centroid of the signal reflects the center position of the signal's frequency domain energy distribution. To calculate the frequency domain centroid, each frequency component of the signal is used as a weight and is weighted and summed with its corresponding amplitude. Finally, the amplitudes of all frequency components are normalized.

[0058] Step 3-6: Calculate the main frequency ratio as feature 6. When analyzing the frequency characteristics of a signal, the main frequency is a key indicator and is often used to identify abnormal data. The time domain signal is converted into a frequency domain signal to reveal the amplitude of different frequency components in the signal. The larger the amplitude, the greater the energy of the signal at that frequency. The main frequency is the frequency component with the strongest amplitude.

[0059] Step 3-7: Calculate the total spectrum power ratio as feature 7. The spectrum area represents the total power of the signal in a specific frequency interval. The power in a certain frequency interval is obtained by integrating the power spectrum density in the interval. Calculating the total power of the signal in a specific frequency interval can analyze abnormal signals. The power of abnormal data will be different from that of normal data.

[0060] Step 3-8: Calculate the standard deviation of the energy change in each frequency band of the signal as feature 8. The enhancement and suppression effects of the signal collected in the structural damage area and the signal collected in the lossless area in each frequency band are also different. Therefore, analyzing the energy change of the reconstructed signal and the original input signal in each frequency band is the basis for judging whether damage occurs.

[0061] The second-stage unsupervised algorithm is constructed, and the eight damage-sensitive features are sequentially constructed into a one-dimensional sequence and input into the OC-SVM for training. Since only the damage-sensitive features of the data in the damage-free area are trained, the training data will be concentrated in the same area as much as possible, and this area can be regarded as the "normal" area of ​​the data. The optimal hyperplane is solved in the feature space. After the training is completed, the hyperplane separates the training data from the origin with the maximum boundary, and finally the offset distance is obtained as the damage evaluation index.

[0062] A two-stage training model is used to detect damage to other concretes with internal defects. The test signal will first pass through the first-stage training model to obtain damage-sensitive features, and then pass through the second-stage training model to obtain the offset distance. When the offset distance is less than 0, it is identified as an abnormal data point, because the damage sensitivity of the original data in the damaged area has changed significantly after reconstruction by the PC-DAE model. The smaller the offset distance is than 0, the greater the difference between the measured data and the non-destructive data, and the higher the degree of damage in the measured area. Finally, the internal defects in the concrete can be located and the degree of damage can be assessed based on the predicted offset distance. Since the smaller the offset distance output by OC-SVM is than 0, the more serious the damage is, here we can add a negative sign to the output offset distance to make it a positive value for easy analysis. The minimum offset distance is 0, and other values ​​are increased or decreased accordingly, and then normalized as the final evaluation indicator. Defect imaging of the concrete structure is performed based on the processed offset distance. Figure 2 This is a schematic diagram of the black and white pixel processing of the structural defect image, where the gray box is the actual defect location and the black area inside the box is the predicted defect range. The damage degree of the defect area is further evaluated based on the offset distance. Figure 3 This is the damage degree assessment diagram after normalization of the offset distance. The results show that the offset distance is positively correlated with the damage degree (V1>V2>V3).

[0063] The present invention proposes a method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm. The method only needs to be trained on the data of the complete structure, which solves the problem of difficulty in collecting a large amount of damage data for training, and the introduction of physical constraints increases the interpretability of the model. The present invention uses the offset distance to locate internal defects of concrete and assess the degree of damage, which can provide a reliable basis for engineering and technical personnel to repair defects accurately and timely, avoiding large-scale structural damage.

Claims

1. A method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm, characterized in that The method comprises the following steps: Step 1: Collect the impact echo vibration signal of the intact and undamaged concrete structure. Step 2: Construct the first stage of the unsupervised algorithm, introduce the motion equation of the single degree of freedom system as a physical constraint in the autoencoder, and reconstruct the original impact echo signal collected by the lossless structure. Step 3: Extract the eight differences between the original signal input in step 2 and the reconstructed signal output and the original signal as damage-sensitive features. For normal data, the reconstruction error is usually small, but for abnormal data (such as structural damage), the reconstruction error will increase significantly. Step 4: Construct the second-stage unsupervised algorithm, construct the eight damage-sensitive features into a one-dimensional sequence in order, and input them into the single-class support vector machine (OC-SVM) for training. The training data will be clustered in the same area as much as possible. This area can be regarded as the "normal" area of ​​the data. Finally, the output offset distance is used as the damage evaluation index. Step 5: Use the two-stage training model to detect other concrete with internal defects. The test signal will first pass through the first-stage training model to obtain damage-sensitive features, and then pass through the second-stage training model to obtain the offset distance. Finally, the internal defects in the concrete can be located and the damage degree can be evaluated based on the predicted offset distance.

2. According to claim 1, a method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm is characterized in that: The step 2 includes the following process: The first phase of the unsupervised algorithm is constructed, and physical constraints (motion equations of single-degree-of-freedom systems) are added to the deep autoencoder to form PC-DAE to learn a more robust and reliable model from the data, thereby enhancing the interpretability of the model. The network uses the gradient descent method to minimize the total loss to obtain the optimal network. The total loss of the network includes the data loss J D and physical loss P Two parts. The total loss J is defined as: J=a D J D +a P J P Where α D and α P are the weights of data loss and physical loss respectively, and their sum is 1.

3. The method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm according to claim 1 is characterized in that: The step 3 includes the following process: Step 3-1: Calculate the mean square error as feature 1. MSE is used to quantify the difference between the original input data and the reconstructed output data. Specifically, it measures the average square difference between the two data. Step 3-2: Calculate the Pearson correlation coefficient as feature 2, and use the Pearson correlation coefficient to measure the linear correlation between the reconstructed data and the original data. Step 3-3: Calculate the energy ratio as feature 3. Similar to the ratio of signal strength to background noise strength, the energy ratio of the original signal and the reconstructed signal can be used to represent the energy change of the signal. Step 3-4: Calculate the power density ratio as feature 4. The power density ratio first uses discrete Fourier transform to convert the time domain signal to the frequency domain, and then calculates the power density ratio of all frequency components of the original signal and the reconstructed signal in the frequency domain. This indicator can effectively measure the power difference between the reconstructed signal and the original signal in the frequency domain. Step 3-5: Calculate the frequency domain centroid ratio as feature 5. The frequency domain centroid of the signal reflects the center position of the signal's frequency domain energy distribution. To calculate the frequency domain centroid, each frequency component of the signal is used as a weight and is weighted and summed with its corresponding amplitude. Finally, the amplitudes of all frequency components are normalized. Step 3-6: Calculate the main frequency ratio as feature 6. When analyzing the frequency characteristics of a signal, the main frequency is a key indicator and is often used to identify abnormal data. The time domain signal is converted into a frequency domain signal to reveal the amplitude of different frequency components in the signal. The larger the amplitude, the greater the energy of the signal at that frequency. The main frequency is the frequency component with the strongest amplitude. Step 3-7: Calculate the total spectrum power ratio as feature 7. The spectrum area represents the total power of the signal in a specific frequency interval. The power in a certain frequency interval is obtained by integrating the power spectrum density in the interval. Calculating the total power of the signal in a specific frequency interval can analyze abnormal signals. The power of abnormal data will be different from that of normal data. Step 3-8: Calculate the standard deviation of the energy change in each frequency band of the signal as feature 8. The enhancement and suppression effects of the signal collected in the structural damage area and the signal collected in the lossless area in each frequency band are also different. Therefore, analyzing the energy change of the reconstructed signal and the original input signal in each frequency band is the basis for judging whether damage occurs.

4. The method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm according to claim 1 is characterized in that: The step 4 includes the following process: The second-stage unsupervised algorithm is constructed, and the eight damage-sensitive features are sequentially constructed into a one-dimensional sequence and input into the OC-SVM for training. Since only the damage-sensitive features of the data in the damage-free area are trained, the training data will be concentrated in the same area as much as possible, and this area can be regarded as the "normal" area of ​​the data. The optimal hyperplane is solved in the feature space. After the training is completed, the hyperplane separates the training data from the origin with the maximum boundary, and finally the offset distance is obtained as the damage evaluation index.

5. The method for identifying internal defects of concrete structures based on a two-stage unsupervised deep learning algorithm according to claim 1 is characterized in that: The step 5 includes the following process: The two-stage training model is used to detect damage to other concretes with internal defects. The test signal will first pass through the first-stage training model to obtain damage sensitivity features, and then pass through the second-stage training model to obtain the offset distance. When the offset distance is less than 0, it is identified as an abnormal data point, because the damage sensitivity of the original data in the damaged area has changed significantly after reconstruction by the PC-DAE model. When the offset distance is less than 0, it indicates that the difference between the measured data and the non-destructive data is greater, and the degree of damage in the measured area is higher. Finally, the internal defects in the concrete can be located and the degree of damage can be evaluated based on the predicted offset distance.

Images