Method for detecting internal delamination defects of concrete by fusing impact-echo and deep learning

Abstract

The application discloses a concrete internal layering defect detection method fusing impact echo and deep learning, and specifically comprises intelligent defect identification, defect area quantification and defect depth automatic positioning. Firstly, a one-dimensional model method combining wavelet decomposition and a GRU network is used to realize automatic diagnosis of internal defect signals of a concrete structure, then on the basis of a probability hotspot map of the automatic defect identification result, a threshold segmentation is used to extract a defect area and perform area quantification. Finally, a two-dimensional model method combining wavelet transformation and a convolutional neural network is used to realize positioning of the defect depth. The application overcomes the defects that the detection result of the impact echo method is highly subjective and a large amount of signal data of a measurement area is low in processing efficiency, and has the advantages of convenient operation, low cost, intelligence, high speed, high objectivity of the result and wide engineering application prospect.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for bridge concrete structures, specifically relating to an intelligent detection method for internal delamination defects in concrete. Background Technology

[0002] Bridges are a vital component of modern transportation networks. During the construction and long-term service of concrete bridge structures, delamination defects of varying degrees can occur due to issues such as construction techniques, material aging, steel corrosion, and environmental erosion. These delamination defects not only reduce structural strength but also lead to rapid deterioration of the bridge structure, severely impacting its safety and durability. Furthermore, the repair costs for delamination defects are high. Therefore, it is necessary to conduct regular comprehensive assessments of internal delamination defects in concrete structures to ensure the safe operation of bridges.

[0003] Choosing the appropriate detection method plays a crucial role in the effectiveness and accuracy of defect assessment. Commonly used methods for detecting internal defects in concrete include core drilling, infrared thermography, electromagnetic wave methods, and stress wave methods. Core drilling provides straightforward results, but achieving comprehensive detection of the entire test area requires numerous core drilling operations, which can cause irreversible damage to the structure, and the entire process is time-consuming and labor-intensive. Infrared thermography is effective for identifying large and shallow defects, but its effectiveness for detecting small and deep defects is not ideal. Furthermore, infrared thermography is easily affected by changes in field conditions, such as weather temperature, requiring specific test conditions for implementation, which severely limits the universality of this method for field detection. Electromagnetic radar is a typical representative of electromagnetic wave detection methods. This method can efficiently and quickly identify internal defects in surface concrete and their depth, but the detection accuracy of concrete below the reinforcing mesh is severely compromised.

[0004] Recent research indicates that stress wave detection methods are considered highly promising for defect detection, primarily including ultrasonic and shock-echo methods. Ultrasonic and shock-echo methods are two typical examples of internal defect detection. Ultrasonic detection requires one detection surface to emit a signal and another to receive it, and it necessitates applying a coupling agent to the contact area between the sensor and the detection surface. This requirement is difficult to meet for most inspection objects, such as bridge decks, base plates, and webs. Furthermore, while ultrasonic methods use high-frequency signals, resulting in high signal resolution, this also leads to significant signal attenuation and limited detection depth.

[0005] Compared to the methods mentioned above, the impact echo method has the following advantages: 1) It only requires one test surface; 2) It uses a lower frequency sound wave than ultrasound (the frequency range of the impact echo method is usually 2-50 kHz), which avoids the high signal attenuation and excessive noise interference problems encountered in ultrasound testing; 3) It does not require a coupling agent; 4) It is not sensitive to metals and can penetrate areas with dense reinforcement. Although the interpretation of impact echo signals has been enhanced by the development of signal processing techniques such as short-time Fourier transform, Gabor transform, EMD, EEMD, VMD, wavelet transform, and wavelet decomposition, this method still has two main problems: 1) The test results need to be interpreted by experienced engineers, resulting in strong subjectivity; 2) The spacing between test points is small (mostly 5-10 cm) during impact echo detection. For large-sized concrete structures, there are many test points, resulting in a large amount of data and low processing efficiency for individual data interpretation. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the deficiencies of the above-mentioned background technology by proposing a fast, convenient, intelligent and efficient intelligent detection method for internal layering defects in concrete, so as to achieve non-destructive evaluation of concrete structural performance.

[0007] To address the aforementioned problems, this invention discloses an intelligent detection method for internal delamination defects in concrete that integrates impact echo and deep learning. The method comprises three steps: internal delamination defect identification, delamination defect region quantification, and automatic defect depth localization. The specific technical solution is as follows:

[0008] The internal layered defect identification and layered defect region quantification include the following steps:

[0009] Step 1: Artificially create concrete slabs containing different types, sizes, and depths of layered defects. Use an impact echo device to collect impact echo signals from different defective and defect-free areas above the concrete slabs as the raw dataset of the signals.

[0010] Step 2: The signals collected above various layered defects are uniformly labeled as defect signals, and the signals collected above defect-free areas are labeled as sound signals.

[0011] Step 3: Perform wavelet packet decomposition on the signal described in Step 2, and decompose each signal x(t) to obtain the wavelet packet energy sequence {E}. (j,1) E (j,2) ,…, E (j,2 j )} is the feature vector of the signal, where j is the wavelet packet decomposition level.

[0012] Step 4: Input the feature vector of each signal into the gated recurrent unit network, and use the corresponding label as the network output to train and validate the network.

[0013] Step 5: Input the wave packet energy vector of a signal into the trained model, and the model will automatically give the identification label of the signal and the corresponding probability P.

[0014] Step 6: After dividing the concrete structure to be tested into a grid, obtain the echo signal of each measuring point, and then perform signal processing. Input the feature vectors of these signals into the model trained in the internal layered defect identification step to obtain the probability that each measuring point in the entire test area is identified as a defect label. Plot the defect identification probabilities of all measuring points as a heat map, and then perform binarization operation on the heat map with a defect identification probability of 0.5 as a threshold. Calculate the area of ​​the defect connected domain formed by the threshold segmentation.

[0015] Preferably, in step three, the energy sequence E of each sub-wave packet of a single signal x(t) is... (j,k) It is calculated using the following equation.

[0016]

[0017]

[0018]

[0019]

[0020] Where: N is the sampling length of the signal; k is the number of sub-bands; j is the decomposition level; d is the wavelet coefficient; and E is the energy.

[0021] Preferably, the key parameters of the GRU network can be expressed by the following simultaneous equations:

[0022]

[0023] Where: z t It updates the gate state at time t; r t It is the reset gate state at time t; It is the candidate cell state at time t; h t-1 and h t These represent the previous and current cell states, respectively. , , , These are the weight matrices for the update gate, reset gate, candidate units, and output layer, respectively; σ is the sigmoid activation function; x t and y t These are the network's input and output, respectively.

[0024] The automatic location of the depth of internal layered defects includes the following steps:

[0025] Step 1: First, classify and label the original signal dataset according to different defect depths. Then, perform wavelet transform on the signal to convert the one-dimensional time-domain waveform into a two-dimensional time-frequency domain spectrum.

[0026] The second step is to batch process the spectrograms to 224×224 pixels and use them as input images. Then, the transfer learning method based on the EfficientNet network is used to automatically learn and extract features from the input images, and the input images are used to complete the training and validation of the network model.

[0027] Step 3: Process the signals collected from the area to be tested using the methods described in Steps 1 and 2 above to obtain images that meet the model input requirements. Use the trained model to test these images to obtain the probability that each image is classified into a certain category. Mark the category corresponding to the maximum probability as the model prediction label of the signal at that test point. Draw a layered defect depth localization map based on the prediction label of each test point in the test area.

[0028] Preferably, the selected wavelet basis type is the Bump wavelet. This wavelet is relatively discrete in the time domain but more concentrated in the frequency domain, with more obvious frequency characteristics, making it easier to distinguish the frequency peaks corresponding to the defect locations. The wavelet transform result of the impact echo signal x(t) can be obtained using the wavelet transform coefficients. To indicate, Among them, wavelet basis functions t is time; a is the scale factor; τ is the translation factor.

[0029] The beneficial technical effects of the present invention compared with the prior art are as follows:

[0030] 1. This invention solves the problem of subjectivity in traditional impulse echo test results, which rely on experienced testing personnel for interpretation.

[0031] 2. In view of the problem of numerous measuring points and large amount of signal data in large-area concrete structures, the method proposed in this invention has the characteristics of intelligent, fast and efficient data processing.

[0032] 3. This invention enables the quantification of the area of ​​layered defects and the automatic location of the defect depth, and is applicable to the overall reliable assessment of layered defects inside concrete structures such as bridge decks. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the intelligent detection method for internal layering defects in concrete according to the present invention.

[0034] Figure 2 This is a schematic diagram of a defect probability identification method.

[0035] Figure 3 This is a schematic diagram of a defect region quantization method based on threshold segmentation.

[0036] Figure 4 This is a schematic diagram of the depth localization method for layered defects according to the present invention.

[0037] Figure 5 Prefabricate concrete slabs with delamination defects for the laboratory.

[0038] Figure 6 This is a heatmap showing the defect probability distribution of the test board.

[0039] Figure 7 A heat map is generated to predict the layered defect depth at each measuring point in the test board.

[0040] Figure 8 A bar chart predicting the layered defect depth at each measuring point in the test board.

[0041] Figure 9 This provides the segmentation, extraction, and area calculation results of each defect prediction region within the artificial defect contour in the test board. Detailed Implementation

[0042] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0043] This invention discloses an intelligent detection method for internal layered defects in concrete that integrates impact echo and deep learning. The method includes three steps: intelligent defect identification, defect region quantification, and automatic defect depth localization. Figure 1 As shown, firstly, a one-dimensional model method combining wavelet decomposition and GRU network is used to automatically diagnose internal defect signals in concrete structures. Then, based on the probabilistic heatmap of the automatic defect identification results, threshold segmentation is used to extract defect regions and perform area quantization. Finally, a two-dimensional model method combining wavelet transform and convolutional neural network is used to locate the defect depth.

[0044] The intelligent identification and quantification of internal layered defects includes the following steps:

[0045] Step 1: Concrete Slab Fabrication and Signal Acquisition

[0046] Concrete slabs with different types, sizes, and depths of layered defects were artificially constructed. Impact echo equipment was used to collect impact echo signals from different defective and defect-free areas above the concrete slabs as the raw dataset of the signals.

[0047] Step 2: Labeling the dataset

[0048] Signals collected above various layered defects are uniformly labeled as defect signals, while signals collected above defect-free areas are labeled as sound signals.

[0049] Step 3: Signal Processing

[0050] like Figure 2 As shown, wavelet packet decomposition is performed on the signal described in step two, and the wavelet packet energy sequence {E} obtained by decomposing each signal x(t) is... (j,1) E (j,2) ,…, E (j,2 j )} is the feature vector of the signal, where j is the wavelet packet decomposition level. The energy sequence E of each sub-wavelet packet of a single signal x(t) is... (j,k) It is calculated using the following equation.

[0051]

[0052]

[0053]

[0054]

[0055] Where: N is the sampling length of the signal; k is the number of sub-bands; j is the decomposition level; d is the wavelet coefficient; and E is the energy.

[0056] Step 4: Model Training and Validation

[0057] The feature vector of each signal is used as the input feature to the gated recurrent unit network, and the corresponding label is used as the output of the network for training and validation.

[0058] Step 5: Probabilistic Identification of Layered Defects

[0059] When the wave packet energy vector of a signal is input into a trained model, the model automatically provides the identification label of the signal and the corresponding probability P.

[0060] Step Six: Segmentation and Quantification of Defect Regions

[0061] After dividing the concrete structure under test into a mesh, the echo signal of each measuring point is acquired. Signal processing is then performed, and the feature vectors of these signals are input into the trained model to obtain the probability that each measuring point in the entire test area is identified as a defect. Figure 3As shown, the defect identification probabilities of all measuring points are plotted as a heat map. Then, the heat map is binarized with a defect identification probability of 0.5 as a threshold, and the area of ​​the defect connected region formed by the threshold segmentation is calculated.

[0062] The automatic location of the depth of internal layered defects includes the following steps:

[0063] Step 1: Obtain the two-dimensional spectrum of the signal

[0064] like Figure 4 As shown, the original signal dataset is first classified and labeled according to different defect depths. Then, the Bump wavelet is used to perform a wavelet transform on the signal, converting the signal from a one-dimensional time-domain waveform to a two-dimensional time-frequency domain spectrum. The wavelet transform result of the impact echo signal x(t) can be represented by the wavelet transform coefficients. To indicate,

[0065] Among them, wavelet basis functions t is time; a is the scale factor; τ is the translation factor.

[0066] Step 2: Image processing and model training

[0067] The spectrograms were batch-processed to 224×224 pixels and used as input images. Then, a transfer learning method based on the EfficientNet network was used to automatically learn and extract features from the input images, and the input images were used to train and validate the network model.

[0068] Step 3: Prediction of the depth of layered defects

[0069] The signals collected from the area to be tested are processed using the methods described in steps one and two above to obtain images that meet the model input requirements. These images are then tested using the trained model to obtain the probability that each image is classified into a certain category. The category corresponding to the maximum probability is marked as the model prediction label for the signal at that measurement point. A layered defect depth localization map is then drawn based on the prediction label of each measurement point in the test area.

[0070] Examples: The present invention will be further illustrated below through specific examples, but these examples should not be construed as limiting the scope of protection of the present invention.

[0071] like Figure 5 As shown, two precast concrete slabs with artificially delamination defects were poured in the laboratory. The training slab was used to acquire datasets for model training, while the test slab served as the testing object for the model's predictive ability. Signal acquisition was accomplished by striking the concrete slab with a 6mm diameter steel ball and using an accelerometer as the signal receiver. A grid of points was used for signal acquisition, with a 5cm distance between the excitation point and the signal receiver. The signal sampling frequency was 125kHz, and the sampling length was 1024.

[0072] According to the method disclosed in this invention, the defect probability distribution heatmap of the test board is as follows: Figure 6 As shown, it is evident that the proposed method can detect the vast majority of layered defects. The results of segmenting, extracting, and calculating the area of ​​these defects are as follows: Figure 9 As shown, the average area detection rate of delamination defects at different depths on the test plates was 70.58%, which is higher than the average area detection rate of 68.58% in the literature "M. Yumnam, D. Ghosh, H. Gupta, Empirical mode decomposition based techniques for imaging of shallow delamination in concrete using impact echo, Mechanical Systems and Signal Processing, 184 (2023) 109668." Moreover, the method disclosed in this invention is more intelligent and faster, and the results are more objective, verifying the feasibility and effectiveness of the method disclosed in this invention in quantifying the overall area of ​​delamination defects inside concrete structures.

[0073] Figure 7 and Figure 8 The heatmap and bar chart of the predicted location depth of layered defects at each measuring point in the test board are shown respectively. The predicted location depth is basically consistent with the depth of the preset artificial defects. Therefore, based on the defect identification probability, defect area quantification, and defect location depth prediction results of the test board, the intelligent detection method for layered defects inside concrete disclosed in this invention is a fast, convenient, intelligent, efficient, effective, and reliable method for detecting layered defects inside concrete, and can be used for non-destructive assessment of the overall condition of internal defects in concrete structures.

[0074] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for detecting internal delamination defects in concrete that integrates impact echo and deep learning, characterized in that, include: Step (1), internal layering defect identification, specifically: S101. Collect and obtain the impact echo signals corresponding to different defect locations and defect-free locations in the concrete slab to obtain the original dataset. S102. The signals collected at various layered defects are uniformly marked as defect signals, and the signals obtained above the defect-free area are marked as sound signals. S103. Perform wavelet packet decomposition on the defective signal and the intact signal described in step S102, and obtain the wavelet energy sequence {E} for each signal x(t). (j,1) E (j,2) ,…,E (j,2 j ) } is the feature vector of the signal, where j is the wavelet packet decomposition level; S104. Input the feature vector of each signal into the gated recurrent unit network (GRU) as the input feature, and the corresponding label as the output of the GRU network. Train and validate the GRU network so that the trained GRU model can automatically give the identification label and the corresponding probability P of the signal based on the wave packet energy vector of the signal. Step (II) Quantification of layered defect areas, specifically: S201. After dividing the concrete structure to be tested into a grid, the echo signal of each measuring point is obtained. Then, the signal is processed. The feature vectors of these signals are input into the GRU model trained in the internal layered defect identification step to obtain the probability that each measuring point in the entire test area is identified as a defect label. The defect identification probabilities of all measuring points are plotted as a heat map. S202. Using the pixel value corresponding to the defect identification probability P1 as the threshold, perform binarization on the heat map and calculate the area of ​​the defect connected region formed by the threshold segmentation. Step (3) Automatic location of internal layered defect depth, specifically: S301. First, classify and label the original signal dataset according to different defect depths. Then, perform wavelet transform on the signal to convert the signal from a one-dimensional time-domain waveform to a two-dimensional time-frequency domain spectrum. S302. The spectrograms are batch-processed to 224×224 pixels and used as input images. Then, the transfer learning method based on EfficientNet network is used to automatically learn and extract features from the input images. The input images are used to complete the training and validation of the EfficientNet network model. S303. Perform the operations in steps S301 and S302 above on the signals collected from the area to be tested to obtain images that meet the model input requirements. Use the trained model to test these images to obtain the probability that each image is classified into a certain category. Mark the category corresponding to the maximum probability as the model prediction label of the signal at that measurement point. Draw a layered defect depth localization map based on the prediction label of each measurement point in the measurement area.

2. The method for detecting internal delamination defects in concrete by integrating impact echo and deep learning as described in claim 1, characterized in that, S101 involves artificially creating concrete slabs with different types, sizes, and depths of layered defects. Impact echo equipment is then used to collect impact echo signals from different defective and non-defective locations above the concrete slabs, which serve as the raw dataset of the signals.

3. The method for detecting internal delamination defects in concrete by integrating impact echo and deep learning as described in claim 1, characterized in that, In the internal layered defect identification step, the energy sequence E of each sub-wave packet of a single signal x(t) is... (j,k) The following equation is used for calculation: Where: N is the sampling length of the signal; k is the number of sub-bands; j is the decomposition level; d is the wavelet coefficient; and E is the energy.

4. The method for detecting internal delamination defects in concrete by integrating impact echo and deep learning as described in claim 1, characterized in that, In step 104, the key parameters of the gated recurrent unit network are expressed by the following simultaneous equations: Where: z t It updates the gate state at time t; r t It is the reset gate state at time t; It is the candidate cell state at time t; h t-1 and h t These are the previous and current unit states, respectively; W z W r , W o These are the weight matrices for the update gate, reset gate, candidate units, and output layer, respectively; σ is the sigmoid activation function; x t and y t These are the network's input and output, respectively.

5. The method for detecting internal delamination defects in concrete by integrating impact echo and deep learning according to claim 1, characterized in that, In step S202, the pixel value corresponding to the defect recognition probability P1 being 0.5 is used as the threshold.

6. The method for detecting internal delamination defects in concrete by integrating impact echo and deep learning as described in claim 1, characterized in that, In the automatic depth localization step for internal layered defects, the selected wavelet basis type is the Bump wavelet, and the wavelet transform result of the impact echo signal x(t) is represented by the wavelet transform coefficients W(a,τ). Among them, wavelet basis functions t is time; a is the scale factor; τ is the translation factor.

Images