A method for detecting internal cracks in concrete based on image-impact echo

Abstract

The present invention relates to the technical field of concrete crack detection, specifically a method for detecting internal cracks in concrete based on image-impact echoes. The method comprises the following steps: acquiring and labeling a concrete crack dataset and preprocessing it using data augmentation techniques; utilizing an RFB network module to capture the multi-scale features of cracks and train a target detection model; applying the trained model to detect target concrete, obtaining the crack extent and coordinates, and drawing a two-dimensional crack map to delineate crack regions; then, arranging monitoring points based on the crack regions, recording the concrete echo signals using impact echo technology, extracting crack features, and inputting them into a machine learning model to obtain a trained model for analyzing internal cracks in the target concrete; and expanding the two-dimensional crack map into a three-dimensional crack map to obtain the actual crack extent. This method not only improves detection efficiency but also provides visualization, providing an effective technical means for crack monitoring in concrete structures.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete crack detection, in particular to a method for detecting internal cracks in concrete based on image-impact echo. Background Art

[0002] With the development of modern architecture and infrastructure, concrete, as a primary building material, has been widely used in various projects. However, over time, concrete often develops internal cracks due to factors such as load fluctuations, environmental impacts, and material aging. These cracks can pose serious structural safety risks and even lead to accidents. Therefore, accurately detecting internal cracks in concrete is crucial to ensuring the safety and durability of buildings.

[0003] Traditional crack detection methods, such as drilling and cutting, often cause physical damage to concrete structures, increasing repair costs and potentially compromising structural safety and stability. While some methods have recently leveraged image detection technology to improve detection efficiency, many still rely on a single technique and lack comprehensive analysis of multi-source information. Furthermore, fixed detection devices limit flexibility and adaptability, preventing comprehensive coverage of concrete structures and resulting in insufficient assessment efficiency.

[0004] Therefore, a method for detecting internal cracks in concrete based on image-impact echo is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting internal cracks in concrete based on image-impact echo. The method involves acquiring and labeling a concrete crack dataset and preprocessing it using data augmentation techniques. The method then uses an RFB network module to capture the multi-scale characteristics of cracks and train a target detection model. The trained model is then used to detect the target concrete, determine the crack extent and coordinates, and create a two-dimensional crack map to delineate the crack region. Monitoring points are then arranged within the crack region, and the concrete echo signals are recorded using impact echo technology. Crack features are extracted and input into a machine learning model to generate a trained model for analyzing internal cracks in the target concrete. The two-dimensional crack map is then expanded into a three-dimensional crack map to determine the actual crack extent. This method improves detection efficiency and reliability, provides visualization, and thus offers an effective technical approach for concrete structure health monitoring.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting internal cracks in concrete based on image-impact echo, comprising:

[0008] Acquire a concrete crack dataset, mark the crack degrees of the concrete crack dataset to obtain a second dataset, and preprocess the second dataset using data enhancement to obtain a first dataset;

[0009] Introducing an RFB network module into the target detection model, wherein the RFB network module is used to capture multi-scale features of cracks; and training the target detection model based on the first data set;

[0010] Acquire target concrete data, input the target concrete data into the trained target detection model, obtain the crack degree and crack coordinates, draw a two-dimensional crack map based on the crack degree and the crack coordinates, and divide the crack area;

[0011] Layout monitoring points of the target concrete according to the crack area, detect the monitoring points using impact echo, and record the target concrete echo signal;

[0012] Extracting crack features from echo signals using digital signal processing technology from the concrete crack dataset, and training a machine learning model based on the crack features;

[0013] Inputting the target concrete echo signal into the trained machine learning model to obtain the internal crack results of the target concrete;

[0014] The two-dimensional crack map is expanded into a three-dimensional crack map according to the internal crack results of the target concrete, and the overall crack defect area of ​​the target concrete is determined according to the three-dimensional crack map to obtain the actual crack extent.

[0015] Furthermore, the crack degree acquisition process includes:

[0016] Measuring the concrete cracks using image processing software according to the concrete crack data set, and obtaining the crack size according to the starting point and the ending point of the concrete crack;

[0017] Obtain the RGB value of the concrete crack using a color analysis tool, and obtain the brightness and saturation based on the concrete background color;

[0018] The texture features of the concrete cracks are extracted using a Gabor filter to obtain the crack complexity;

[0019] The crack degrees of the concrete crack dataset are divided and marked according to the crack size, the brightness, the saturation, the crack complexity and expert experience, where the crack degrees include obvious cracks, slight cracks and no cracks.

[0020] Furthermore, the pre-processing process includes:

[0021] The data enhancement includes geometric changes, color transformations, and perspective transformations. After enhancing the second data set N times, a newly generated image is obtained. The newly generated image is merged with the image in the concrete crack data set to obtain a third data set.

[0022] resizing the third data set, wherein the resizing includes: if the third data set is larger than a preset requirement, cropping it according to the preset requirement using a sliding window method, and filling the area of ​​the cropped third data set that is smaller than the preset requirement with white; otherwise, not cropping it and directly filling the area of ​​the third data set that is smaller than the preset requirement with white;

[0023] The first data set is obtained according to the adjusted third data set.

[0024] Furthermore, the target detection model training step includes:

[0025] Step S20: inputting the first data set into the target detection model;

[0026] Step S21: introducing the feature extraction layer of the target detection model into the RFB network module, the RFB network module processes the first data set in parallel to obtain a crack feature map, and performing feature fusion on the crack feature map to obtain a crack multi-scale feature map;

[0027] Step S22: using a detection layer to output the crack coordinates, crack confidence, and crack extent according to the crack multi-scale feature map;

[0028] Step S23: Calculate the gap between the predicted value and the true label using a loss function according to the crack coordinates, the crack confidence, and the crack degree to obtain a crack loss value;

[0029] Step S24: Calculating parameter gradients using the chain method according to the crack loss value, and updating the target detection model according to the parameter gradients;

[0030] Step S25: If the number of iterations reaches the first threshold, the training is terminated; otherwise, steps S20 to S25 are repeated.

[0031] Furthermore, the two-dimensional crack map establishment process includes:

[0032] Taking the upper left corner of the target concrete as the origin of the coordinate system, a two-dimensional coordinate system is established according to the actual size of the target concrete. The two-dimensional coordinate system includes an X-axis and a Y-axis, wherein the X-axis extends horizontally to the right and the Y-axis extends vertically downward.

[0033] Check whether the crack coordinates are the clipped coordinates. If the crack coordinates are the clipped coordinates, add the sliding window offset to the crack coordinates to obtain the first crack coordinates; otherwise, directly assign the crack coordinates to the first crack coordinates.

[0034] Actual coordinates are calculated based on the first crack coordinates using a scaling ratio, and the two-dimensional crack map is obtained based on the actual coordinates.

[0035] Furthermore, the crack area division process includes:

[0036] Calculating the center coordinates of the crack according to the crack coordinates;

[0037] Calculating the distance between the cracks using the Euclidean distance formula according to the center coordinates of the cracks, dividing the cracks whose distance is less than a second threshold into the same area, and obtaining a clustering result;

[0038] According to the clustering results, the number of cracks is calculated. If the number of cracks in the area exceeds a third threshold or at least one crack is an obvious crack, a surface crack area is obtained and marked as a defect area; otherwise, it is marked as a non-defective area.

[0039] Furthermore, the monitoring point layout includes:

[0040] According to the size of the defect area and the number of cracks, the number of monitoring points in the defect area is obtained;

[0041] According to the size of the defect-free area and the number of cracks, the number of monitoring points in the defect-free area is obtained;

[0042] Arranging monitoring points at the junction of the defective area and the non-defective area to obtain the number of monitoring points at the junction;

[0043] The monitoring point spacing is calculated based on the number of monitoring points in the defective area, the number of monitoring points in the non-defective area and the number of monitoring points at the junction, and the monitoring point positions are marked based on the monitoring point spacing.

[0044] Furthermore, the crack feature acquisition process includes:

[0045] Acquiring echo signal data according to the concrete crack data set;

[0046] Using a high-pass filter to remove background noise according to the echo signal data to obtain a first echo signal;

[0047] performing signal segmentation according to the first echo signal using an overlapping segmentation method to obtain a second echo signal;

[0048] Calculating the signal peak value and the time when the impact reaches the monitoring point according to the second echo signal to obtain a time domain feature;

[0049] Extracting frequency amplitude using Fourier transform according to the second echo signal to obtain frequency domain features;

[0050] Extracting a signal using wavelet transform according to the second echo signal to obtain a time-frequency feature;

[0051] The time domain feature, the frequency domain feature and the time-frequency feature are combined to obtain the crack feature.

[0052] Furthermore, the machine learning model training process includes:

[0053] Acquire historical echo signals and historical echo signal labels according to the concrete crack dataset, wherein the historical echo signal labels include internal crack depth and width;

[0054] The crack features are input into the machine learning model, and a first loss function is used to obtain the trained machine learning model based on the crack features and the historical echo signal labels.

[0055] Furthermore, the actual crack degree acquisition process includes:

[0056] extracting crack information according to the three-dimensional crack map;

[0057] The surface crack cross-sectional area and the internal crack cross-sectional area are calculated based on the crack information, and the actual crack degree is obtained by performing weighted summation based on the surface crack cross-sectional area and the internal crack cross-sectional area.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. To improve the efficiency and accuracy of concrete surface crack detection, especially in complex environments, data enhancement technology is used to preprocess images. Combined with the RFB network module to capture the multi-scale characteristics of cracks, the detection model's sensitivity and accuracy to cracks can be improved. The target detection model accurately assesses the crack extent and locates the crack coordinates. Based on this, a two-dimensional crack map is drawn, clearly showing the crack distribution and characteristics. This achieves accurate demarcation of surface crack areas, thereby improving the efficiency and reliability of crack detection.

[0060] 2. When inspecting inside concrete, traditional impact echo detection technology cannot quickly and accurately locate deep cracks. By arranging monitoring points in different regions of the target concrete, precise monitoring of crack areas can be achieved, ensuring comprehensive and flexible detection coverage. Furthermore, the use of digital signal processing technology to extract crack features from echo signals not only improves feature extraction accuracy but also effectively removes background noise. By inputting the extracted crack features into a machine learning model, an efficient and accurate model can be trained, enabling accurate identification of cracks inside concrete and improving the efficiency and reliability of crack detection.

[0061] 3. Accurately delineating the defective area of ​​concrete cracks is crucial for structural inspection and maintenance. By expanding the 2D crack map into a 3D crack map, the three-dimensional characteristics of internal concrete cracks can be fully displayed, enabling more accurate assessment of crack defect areas. Furthermore, the 3D crack map helps improve the efficiency of delineating the overall crack defect area. The internal and external crack area areas are calculated based on the internal crack depth, width, and crack coordinates, providing an accurate estimate of the actual crack extent, thus improving the efficiency and reliability of crack detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic flow chart of a method for detecting internal cracks in concrete based on image-impact echo provided by the present invention;

[0063] Figure 2 A schematic diagram of a crack degree classification structure provided by an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of the target detection model training process provided by an embodiment of the present invention;

[0065] Figure 4 A schematic diagram of the monitoring point layout provided by an embodiment of the present invention;

[0066] Figure 5 A schematic diagram of the overall crack defect area provided by an embodiment of the present invention.

[0067] In the figure: 1. Non-defective area; 2. Monitoring point; 3. Crack; 4. Defective area; 5. First crack area; 6. Second crack area. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] In modern industrial environments, concrete is the primary building material, and its structural safety is crucial to plant operations. Over time, concrete structures are susceptible to various factors, such as temperature changes, humidity, chemical corrosion, and mechanical loads, which can lead to the formation of internal cracks. Traditional concrete inspection methods, such as manual inspection and acoustic testing, although effective, have drawbacks such as low efficiency, limited coverage, and potential damage to the structure. Although some methods have been proposed to improve inspection efficiency, they still rely on fixed inspection devices, limiting flexibility and adaptability, and are also demanding in terms of capital and time. Therefore, there is an urgent need to develop a novel, efficient, and non-destructive inspection method to further improve inspection efficiency in order to more comprehensively assess the safety of the internal structure of concrete.

[0070] Example 1

[0071] For example, a factory inspects concrete blocks. Comprehensive concrete testing promptly identifies potential structural issues, allowing appropriate maintenance measures to ensure safe operation and production efficiency. This method not only improves inspection efficiency but also provides valuable data support for long-term monitoring and management of concrete.

[0072] See also Figures 1 to 5 The present invention provides a method for detecting internal cracks in concrete based on image-impact echo. The technical solution is as follows:

[0073] like Figure 1 As shown, a method for detecting internal cracks in concrete based on image-impact echo includes:

[0074] Step S1: Acquire a concrete crack dataset, mark the crack degrees of the concrete crack dataset to obtain a second dataset, and preprocess the second dataset using data enhancement to obtain a first dataset.

[0075] The concrete crack dataset includes concrete photo sets and concrete echo signal data.

[0076] Specifically, a variety of concrete structures were selected at the factory, and high-definition cameras were used to capture images of cracks under different conditions. This ensured that different types of cracks (such as horizontal, vertical, and staggered) and interference factors (such as plant branches, graffiti, and water flow marks) were included to increase the diversity of the dataset. Furthermore, a signal exciter and a high-frequency sensor were used. The exciter was brought close to the concrete surface and applied an impact force. The sensor collected the reflected wave signal and recorded the time and intensity of the reflected wave. Computer software and expert analysis were used to calculate the path of the sound wave from the surface to the crack and back, thereby determining the crack depth and width.

[0077] Furthermore, the crack degree acquisition process includes:

[0078] The concrete cracks are measured using image processing software based on the concrete crack dataset. Based on the starting point and the ending point of the concrete crack, the software can calculate the distance between the two points to obtain the crack size. For example, in ImageJ, the "line segment tool" can be used to mark the crack, and then the length data can be obtained in the "measure" function.

[0079] Obtain the RGB value of the concrete crack using a color analysis tool, and obtain the brightness and saturation based on the concrete background color. For example, in Photoshop, the RGB value of the selected area can be viewed using the "Info" panel, and the brightness calculation formula is used to calculate the brightness value based on the RGB value. The saturation of the background color is then calculated using the HSV color space conversion.

[0080] The Gabor filter can capture texture information of different frequencies and directions, and the texture features of the concrete cracks are extracted using the Gabor filter to obtain the crack complexity;

[0081] According to the crack size, the brightness, the saturation, the crack complexity and expert experience, the crack degree of the concrete crack dataset is divided and marked, such as Figure 2 As shown, the crack degrees include obvious cracks, slight cracks and no cracks.

[0082] The conditions for division are:

[0083] Obvious cracks: length greater than 15 cm; width greater than 3 mm; brightness less than or equal to 150; saturation less than or equal to 30; height irregular, with obvious peeling or falling off;

[0084] Minor cracks: length less than or equal to 15 cm and greater than or equal to 5 cm; width less than or equal to 3 mm and less than or equal to 1 mm; brightness greater than 150 and less than or equal to 200; saturation greater than 30 and less than or equal to 60; blurred boundaries and branching;

[0085] No cracks: length less than 5 cm, width less than 1 mm; brightness greater than 200; saturation greater than 60; regular texture and clear boundaries.

[0086] Mark the category of each crack in the table and draw a rectangular box of the crack on the image to generate a labeling file.

[0087] By systematically measuring and classifying concrete cracks, accurate assessment of crack status is achieved. In addition, the generated annotation file records the location, type, size and characteristics of the cracks, providing a comprehensive data foundation for subsequent analysis, thereby improving the efficiency and reliability of crack detection.

[0088] Furthermore, the pre-processing process includes:

[0089] The data enhancement includes geometric changes, color transformations, and perspective transformations. After enhancing the second data set N times, a newly generated image is obtained. The newly generated image is merged with the image in the concrete crack data set to obtain a third data set.

[0090] Among them, N can be set to 5 to 10 times, and the generated enhanced images are randomly checked to ensure that the crack characteristics of the enhanced images are still obvious and consistent with the actual situation.

[0091] Specifically, the geometric transformation involves randomly rotating, translating, and scaling the image to simulate cracks from different perspectives. The color transformation involves randomly changing the brightness and hue of the image to simulate different lighting conditions. The perspective transformation simulates the effects of different shooting angles on crack characteristics to ensure that the model can handle cracks from different perspectives.

[0092] resizing the third data set, wherein the resizing includes: if the third data set is larger than a preset requirement, cropping it according to the preset requirement using a sliding window method, and filling the area of ​​the cropped third data set that is smaller than the preset requirement with white; otherwise, not cropping it and directly filling the area of ​​the third data set that is smaller than the preset requirement with white;

[0093] The first data set is obtained according to the adjusted third data set.

[0094] Specifically, assuming the captured image size is 1024×768 pixels and the preset requirement is 608×608 pixels, the step size is set to 64 pixels, and multiple 608×608 pixel areas are cropped from left to right and from top to bottom. After cropping, areas smaller than the preset requirement are filled with white. For each cropped image, the bounding box coordinates of the crack need to be updated to ensure its correspondence with the new image. For example, if the bounding box coordinates of a crack in the original image are (100, 150, 200, 300), and the position after cropping becomes (0, 0, 608, 608), the position of the bounding box in the new image needs to be updated and recorded.

[0095] By implementing the above preprocessing process, the quality and diversity of the concrete crack dataset are improved. Data augmentation technology is used to effectively generate diverse samples and enrich the training data. At the same time, resizing is used to ensure that the image retains its original features without distortion and can accurately capture details, thereby maximizing the retention of crack feature information and improving the efficiency and reliability of crack detection.

[0096] Step S2: introducing an RFB network module into the target detection model, wherein the RFB network module is used to capture multi-scale features of cracks; and training the target detection model according to the first data set.

[0097] Furthermore, if Figure 3 As shown, the target detection model training steps include:

[0098] Step S20: inputting the first data set into the target detection model;

[0099] Step S21: The feature extraction layer of the target detection model is introduced into the RFB network module. The RFB (Receptive Field Block) network module processes the first data set in parallel to obtain a crack feature map. The crack feature map is subjected to feature fusion to obtain a multi-scale crack feature map. For example, different convolution kernel sizes (such as 3×3, 5×5, and 7×7) are used and the output feature maps are fused to form a multi-scale crack feature map, which is expressed as:

[0100]

[0101] Among them, F 3×3 、F 5×5 and F 7×7 They are the feature maps obtained by different convolution kernels, Represents the feature fusion operation.

[0102] Step S22: Using the detection layer to output the crack coordinates, crack confidence, and crack extent according to the crack multi-scale feature map, expressed as:

[0103] output={(x min ,y min ,x max ,y max ,p,v)};

[0104] Among them, (x min ,y min ) is the coordinate of the upper left corner of the crack, (x max ,y max ) is the coordinate of the lower right corner of the crack, p is the confidence level, and v is the category of the crack.

[0105] Step S23: Calculate the gap between the predicted value and the true label using a loss function according to the crack coordinates, the crack confidence, and the crack degree to obtain a crack loss value;

[0106] Among them, the loss function is expressed as:

[0107]

[0108] Among them, y i is the true label, is the predicted value, N is the number of samples, and i is the sample number.

[0109] Step S24: Calculate parameter gradients using the chain method according to the crack loss value, and update the target detection model according to the parameter gradients, which can be expressed as:

[0110]

[0111] Among them, c is the model parameter, a is the learning rate, which can be set to 0.001. is the gradient of the loss function with respect to the parameters.

[0112] Step S25: If the number of iterations reaches the first threshold, the training is terminated; otherwise, steps S20 to S25 are repeated.

[0113] Specifically, for example, YOLOv5 was selected as the object detection model, using a first dataset containing 1,000 images of concrete cracks. At the feature extraction layer, the model was able to process some images of obvious cracks, such as cracks up to 20 cm in length and 3 mm in width. Based on the multi-scale feature map of the cracks, the detection layer output the crack coordinates as (150, 300, 155, 320) with a confidence level of 0.95, labeling them as "obvious cracks." After the first iteration, the loss value calculated using the loss function was 0.05, indicating that the model's predictions were relatively accurate. After 800 iterations, the loss value gradually decreased to 0.02. The object detection model accurately captured the multi-scale features of cracks, improving the accuracy and efficiency of crack detection. When identifying obvious cracks, the confidence level reached as high as 0.98, demonstrating that the model can stably identify cracks of varying degrees in practical applications. Furthermore, the reduction in the loss value indicates that the model is learning effectively, further improving the efficiency and reliability of crack detection.

[0114] Step S3: Obtain target concrete data, input the target concrete data into the trained target detection model, obtain the crack degree and crack coordinates, draw a two-dimensional crack map based on the crack degree and the crack coordinates, and divide the crack area.

[0115] Furthermore, the two-dimensional crack map establishment process includes:

[0116] Taking the upper left corner of the target concrete as the origin of the coordinate system, a two-dimensional coordinate system is established according to the actual size of the target concrete. The two-dimensional coordinate system includes an X-axis and a Y-axis, wherein the X-axis extends horizontally to the right and the Y-axis extends vertically downward.

[0117] Check whether the crack coordinates are the clipped coordinates. If the crack coordinates are the clipped coordinates, add the sliding window offset to the crack coordinates to obtain the first crack coordinates; otherwise, directly assign the crack coordinates to the first crack coordinates.

[0118] Actual coordinates are calculated based on the first crack coordinates using a scaling ratio, and the two-dimensional crack map is obtained based on the actual coordinates.

[0119] Specifically, assuming that the actual size of the target concrete is 100 cm × 50 cm, the established X-axis extends from 0 to 100 cm, and the Y-axis extends from 0 to 50 cm; assuming that the detected crack coordinates are (15, 25, 35, 45), since these coordinates are in the cropped image, if they are offset by a sliding window (offset is (5, 5)), the new upper left corner coordinates are (15+5, 25+5)=(20, 30), and the lower right corner coordinates are (35+5, 45+5)=(40, 50); assuming that the scaling ratio is 2, the actual coordinates are (40, 60, 80, 100); based on the actual coordinates, the position of the crack in the two-dimensional image is drawn, represented as a rectangular area from (40, 60) to (80, 100); the crack shape in the rectangular area is converted into a binary image, including black pixels and white pixels, and the widest and longest values ​​of the black pixels are recorded as the surface crack width and length, and are drawn in the two-dimensional image in equal proportion. By continuously drawing crack rectangular areas, the specific location and size of cracks on the target concrete can be clearly indicated, providing a detailed visual reference for subsequent structural assessment and maintenance, and improving the efficiency and reliability of crack detection.

[0120] Furthermore, the crack area division process includes:

[0121] The center coordinates of the crack are calculated based on the crack coordinates, which can be expressed as:

[0122]

[0123] Among them, (C x ,C y ) is the center coordinate of the crack.

[0124] The distance between the cracks is calculated using the Euclidean distance formula according to the center coordinates of the cracks, and the cracks with a distance less than the second threshold are divided into the same area to obtain a clustering result, which is expressed as:

[0125]

[0126] Where D is the distance between the center coordinates of the two slits, (C x1 ,C y1 ) and (C x2 ,C y2 ) are the center coordinates of the two cracks; in addition, the second threshold can be set to 30 mm.

[0127] According to the clustering results, the number of cracks is calculated. If the number of cracks in the area exceeds a third threshold or at least one crack is an obvious crack, a surface crack area is obtained and marked as a defect area; otherwise, it is marked as a non-defective area.

[0128] The third threshold may be set to 3.

[0129] Specifically, assuming three cracks have coordinates (10, 20, 30, 40), (32, 22, 50, 42), and (70, 80, 90, 100), the calculated center coordinates are (20, 30), (41, 32), and (80, 90). Based on these center coordinates, the distances between cracks 1 and 2 are calculated to be 21.1, 84.9, and 69.9, respectively. Because the distance between cracks 1 and 2 is less than 30, they are divided into region A and crack 3 into region B. Since crack 1 is a significant crack and crack 3 is a minor crack within region A, region A is marked as a defective region and region B as a non-defective region. This process effectively identifies and marks defective regions in concrete structures, enhancing detection accuracy. Furthermore, the set thresholds simplify the correlation between cracks, providing a reliable basis for the subsequent setting of impact echo monitoring points, thereby improving the efficiency and reliability of crack detection.

[0130] Step S4: arranging monitoring points for the target concrete according to the crack area, detecting the monitoring points using impact echoes, and recording target concrete echo signals.

[0131] Furthermore, the monitoring point layout includes:

[0132] According to the size of the defect area and the number of cracks, the number of monitoring points in the defect area is obtained;

[0133] According to the size of the defect-free area and the number of cracks, the number of monitoring points in the defect-free area is obtained;

[0134] Arranging monitoring points at the junction of the defective area and the non-defective area to obtain the number of monitoring points at the junction;

[0135] The monitoring point spacing is calculated based on the number of monitoring points in the defective area, the number of monitoring points in the non-defective area and the number of monitoring points at the junction, and the monitoring point positions are marked based on the monitoring point spacing.

[0136] Specifically, if Figure 4 As shown in the figure, assuming the actual dimensions of the target concrete are 100 cm × 50 cm, defective area 4 measures 40 cm × 20 cm, there are three cracks 3, and the defect-free area 1 is the remaining portion with zero cracks 3. Each monitoring point 2 covers an area of ​​20 cm × 10 cm. The calculated number of monitoring points 2 in defect-free area 1 is 16, and in defective area 4 is 4. Based on the number of cracks 3, 12 monitoring points 2 are added in defective area 4, 1 cm to 2 cm above, below, and to the left and right of each crack 3. Four monitoring points 2 are placed at the junction of defective area 4 and defect-free area 1. If the distance between monitoring points 2 does not exceed 5 cm, no additional monitoring points 2 are added. The uniform grid layout in defect-free area 1 ensures basic monitoring of the entire concrete structure and provides global data support. Furthermore, increasing the density of monitoring points 2 in defective area 4 allows for more detailed monitoring of defective area 4, enabling timely detection of potential problems and improving the efficiency and reliability of crack 3 detection.

[0137] Step S5: Using digital signal processing technology to extract crack features of the echo signal from the concrete crack dataset, and training a machine learning model based on the crack features.

[0138] Furthermore, the crack feature acquisition process includes:

[0139] Acquiring echo signal data according to the concrete crack data set;

[0140] Using a high-pass filter to remove background noise according to the echo signal data to obtain a first echo signal;

[0141] performing signal segmentation according to the first echo signal using an overlapping segmentation method to obtain a second echo signal;

[0142] Calculating the signal peak value and the time when the impact reaches the monitoring point according to the second echo signal to obtain a time domain feature;

[0143] Extracting frequency amplitude using Fourier transform according to the second echo signal to obtain frequency domain features;

[0144] Extracting a signal using wavelet transform according to the second echo signal to obtain a time-frequency feature;

[0145] The time domain feature, the frequency domain feature and the time-frequency feature are combined to obtain the crack feature.

[0146] Specifically, assuming the original echo signal is [0.1, 0.3, 0.5, 0.4, 0.2, 0.1], after high-pass filtering, the first echo signal is [0.2, 0.4, 0.3]. Assuming the signal is divided into 10ms segments, the resulting second echo signal may be Segment 1 [0.2, 0.4] and Segment 2 [0.4, 0.3]. For Segment 1, the peak is 0.4 and the time is 0ms. After Fourier transform, the frequency domain features obtained are [0.5, 0.3], and the time-frequency features are [0.6, 0.4]. Combining the time domain features, frequency domain features, and time-frequency features, the final crack feature may be [0.4, 0, 0.5, 0.6]. This method can effectively improve the signal-to-noise ratio of the signal and retain useful high-frequency components. The resulting crack feature contains multidimensional information in the time domain, frequency domain, and time-frequency domain, which facilitates the effective training of subsequent machine learning models and improves the model's ability and accuracy in identifying concrete cracks.

[0147] Furthermore, the machine learning model training process includes:

[0148] Acquire historical echo signals and historical echo signal labels according to the concrete crack dataset, wherein the historical echo signal labels include internal crack depth and width;

[0149] Specifically, assuming there are 100 historical echo signal data, for signal 1, the depth is 5 cm and the width is 2 mm, for signal 2, the depth is 10 cm and the width is 3 mm, the label format is expressed as:

[0150]

[0151] Among them, Lable is a label, and each row represents the depth and width of a historical echo signal;

[0152] The previously extracted fracture features are organized into a format suitable for input into the machine learning model, which can be expressed as:

[0153]

[0154] Among them, Feature is a feature matrix, and each row represents a set of crack features;

[0155] The crack features are input into the machine learning model, and a first loss function is used to obtain the trained machine learning model based on the crack features and the historical echo signal labels.

[0156] Among them, machine learning models can choose linear regression, support vector machine, random forest or neural network, etc.

[0157] Specifically, assuming that random forest is selected as the machine learning model, the mean square error is selected as the first loss function, which is expressed as:

[0158]

[0159] Among them, MSE is the mean square error, n is the total number of labels, i is the label number, is the depth and width predicted by the model, e i is the historical echo signal label;

[0160] Specifically, the crack features and corresponding labels of 80 historical echo signals were used for training, the number of iterations was set to 1000, and the model parameters were gradually adjusted to reduce the loss value. The crack features and corresponding labels of 20 historical echo signals were then used for verification, and the mean square error was 0.15, indicating that the machine learning model has good predictive performance on unseen data, can effectively capture the characteristics of cracks and perform accurate evaluation, and improve the efficiency and reliability of crack detection.

[0161] Step S6: Obtain internal crack results of the target concrete using the trained machine learning model according to the target concrete echo signal.

[0162] Step S7: Expanding the two-dimensional crack map into a three-dimensional crack map according to the internal crack results of the target concrete, determining the overall crack defect area of ​​the target concrete according to the three-dimensional crack map, and obtaining the actual crack extent.

[0163] Furthermore, the actual crack degree acquisition process includes:

[0164] The surface crack coordinates and the depth and width of the internal cracks are input into the modeling software to generate a 3D crack map. During this process, the cracks can be assumed to be uniform in depth and width, or adjusted according to the specific situation.

[0165] Extracting crack information from the three-dimensional crack map, including the coordinates, depth, and width of the cracks; marking the area containing the cracks as an overall crack defect area to facilitate subsequent analysis;

[0166] According to the overall crack area, the surface crack and the internal crack extending downward are marked as the first crack area, and all the overall crack areas are marked in this way;

[0167] The surface crack cross-sectional area of ​​the first crack region and the internal crack cross-sectional area corresponding to the first crack region are calculated based on the crack information, and the actual crack degree is obtained by weighted summation of the surface crack cross-sectional area and the internal crack cross-sectional area.

[0168] Specifically, if Figure 5As shown, for the first crack region 5, assuming that the surface crack width is 10 mm and the length is 100 mm, the cross-sectional area of ​​the surface crack is 1000 mm. 2 , assuming the depth is 20mm and the width is 5mm, the corresponding cross-sectional area of ​​the internal crack is 100mm 2 The surface crack cross-sectional area and the internal crack cross-sectional area are weighted and summed, assuming weights of 0.3 and 0.7, resulting in an actual crack severity of 370. Crack severity levels of 500 or less are marked as overall minor crack areas, and the same marking process is repeated for the second crack area 6. Visualizing 3D crack maps provides a scientific basis for repair and reinforcement decisions, enabling timely detection of crack problems. Furthermore, prioritizing the overall severe crack areas helps optimize resource allocation, reducing costs and improving the efficiency and reliability of crack detection.

[0169] In summary, by acquiring and labeling a concrete crack dataset and combining it with data augmentation techniques, we can efficiently preprocess the data and generate a high-quality first dataset. By using the RFB network module to extract the multi-scale features of cracks and train a target detection model, we can accurately identify the extent and coordinates of cracks in the target concrete, thereby creating a detailed two-dimensional crack map and demarcating the crack area. This provides a scientific basis for the subsequent layout of monitoring points. By detecting and recording the impact echo signal, we can further extract crack features and apply them to the machine learning model, making the detection of internal crack results more accurate. By expanding the two-dimensional crack map into a three-dimensional crack map, we can determine the overall crack defect area and thus obtain the actual crack extent, improving the accuracy and efficiency of crack detection and providing reliable support for the health management of concrete structures.

[0170] Example 2

[0171] Taking a factory as an example, we examined four walls within the concrete for crack detection. The walls are 4 meters high, 10 meters wide, and 0.3 meters thick. Located in a humid environment, they are subject to long-term moisture erosion and temperature fluctuations, resulting in multiple cracks in the walls, impacting the safety and service life of the structure. Using an image-echo method for detecting internal cracks in concrete, we aim to improve the efficiency and reliability of crack detection, thereby efficiently and accurately verifying the safety of the walls.

[0172] A method for detecting internal cracks in concrete based on image-impact echo, comprising:

[0173] A concrete crack dataset is obtained, including 400 wall images at different angles and lighting conditions and 200 historical echo signal samples. The crack severity of the concrete crack dataset is marked, including obvious cracks (60), minor cracks (90), and no cracks (40), to obtain a second dataset. The second dataset is preprocessed using data enhancement to obtain a first dataset.

[0174] Introducing an RFB network module into the target detection model, wherein the RFB network module is used to capture multi-scale features of cracks; and training the target detection model based on the first data set;

[0175] Acquire target concrete data, input the target concrete data into the trained target detection model, obtain the crack degree and crack coordinates, draw a two-dimensional crack map based on the crack degree and the crack coordinates, and divide the crack area;

[0176] Layout monitoring points of the target concrete according to the crack area, detect the monitoring points using impact echo, and record the target concrete echo signal;

[0177] Extracting crack features from echo signals using digital signal processing technology from the concrete crack dataset, and training a machine learning model based on the crack features;

[0178] Inputting the target concrete echo signal into the trained machine learning model to obtain the internal crack results of the target concrete;

[0179] The two-dimensional crack map is expanded into a three-dimensional crack map according to the internal crack results of the target concrete, and the overall crack defect area of ​​the target concrete is determined according to the three-dimensional crack map to obtain the actual crack extent.

[0180] Furthermore, the crack degree acquisition process includes:

[0181] Measuring the concrete cracks using image processing software according to the concrete crack data set, and obtaining the crack size according to the starting point and the ending point of the concrete crack;

[0182] Obtain the RGB value of the concrete crack using a color analysis tool, and obtain the brightness and saturation based on the concrete background color;

[0183] The texture features of the concrete cracks are extracted using a Gabor filter to obtain the crack complexity;

[0184] The crack degrees of the concrete crack dataset are divided and marked according to the crack size, the brightness, the saturation, the crack complexity and expert experience, where the crack degrees include obvious cracks, slight cracks and no cracks.

[0185] Furthermore, the pre-processing process includes:

[0186] The data enhancement includes geometric changes, color transformations, and perspective transformations. After enhancing the second dataset three times, 1,200 newly generated images are obtained. The newly generated images are merged with the images in the concrete crack dataset to obtain a third dataset.

[0187] resizing the third data set, wherein the resizing includes: if the third data set is larger than a preset requirement, cropping it according to the preset requirement using a sliding window method, and filling the area of ​​the cropped third data set that is smaller than the preset requirement with white; otherwise, not cropping it and directly filling the area of ​​the third data set that is smaller than the preset requirement with white;

[0188] For easier operation, the photo size is selected as 608×608, so there is no need to crop or fill.

[0189] The first data set is obtained according to the adjusted third data set.

[0190] Furthermore, the target detection model training step includes:

[0191] Step S20: inputting the first data set into the target detection model;

[0192] Step S21: introducing the feature extraction layer of the target detection model into the RFB network module, the RFB network module processes the first data set in parallel to obtain a crack feature map, and performing feature fusion on the crack feature map to obtain a crack multi-scale feature map;

[0193] Step S22: using a detection layer to output the crack coordinates, crack confidence, and crack extent according to the crack multi-scale feature map;

[0194] Step S23: Calculate the gap between the predicted value and the true label using a loss function according to the crack coordinates, the crack confidence, and the crack degree to obtain a crack loss value;

[0195] Step S24: Calculating parameter gradients using the chain method according to the crack loss value, and updating the target detection model according to the parameter gradients;

[0196] Step S25: If the number of iterations reaches the first threshold, the training is terminated; otherwise, steps S20 to S25 are repeated.

[0197] Specifically, YOLOV5 was used as the target detection model with the first threshold of 500. After 500 iterations, the loss value was reduced to 0.12. After the training was completed, 60 obvious cracks, 90 minor cracks and 32 no cracks were finally identified.

[0198] Furthermore, the two-dimensional crack map establishment process includes:

[0199] Taking the upper left corner of the target concrete as the origin of the coordinate system, a two-dimensional coordinate system is established according to the actual size of the target concrete. The two-dimensional coordinate system includes an X-axis and a Y-axis, wherein the X-axis extends horizontally to the right and the Y-axis extends vertically downward.

[0200] Check whether the crack coordinates are the clipped coordinates. If the crack coordinates are the clipped coordinates, add the sliding window offset to the crack coordinates to obtain the first crack coordinates; otherwise, directly assign the crack coordinates to the first crack coordinates.

[0201] Actual coordinates are calculated based on the first crack coordinates using a scaling ratio, and the two-dimensional crack map is obtained based on the actual coordinates.

[0202] Furthermore, the crack area division process includes:

[0203] Calculating the center coordinates of the crack according to the crack coordinates;

[0204] Calculating the distance between the cracks using the Euclidean distance formula according to the center coordinates of the cracks, dividing the cracks whose distance is less than a second threshold into the same area, and obtaining a clustering result;

[0205] The second threshold is set to 50 mm.

[0206] According to the clustering results, the number of cracks is calculated. If the number of cracks in the area exceeds a third threshold or at least one crack is an obvious crack, a surface crack area is obtained and marked as a defect area; otherwise, it is marked as a non-defective area.

[0207] The third threshold is set to 5.

[0208] Furthermore, the monitoring point layout includes:

[0209] According to the size of the defect area and the number of cracks, the number of monitoring points in the defect area is obtained;

[0210] According to the size of the defect-free area and the number of cracks, the number of monitoring points in the defect-free area is obtained;

[0211] Arranging monitoring points at the junction of the defective area and the non-defective area to obtain the number of monitoring points at the junction;

[0212] The monitoring point spacing is calculated based on the number of monitoring points in the defective area, the number of monitoring points in the non-defective area and the number of monitoring points at the junction, and the monitoring point positions are marked based on the monitoring point spacing.

[0213] Specifically, wall 4 has more cracks than the other walls, and there are two obvious cracks. The number of monitoring points set on the four walls are 25, 20, 24 and 35 respectively. A total of 6 additional monitoring points are set at the junction of the defective area and the non-defective area, totaling 110 monitoring points.

[0214] Furthermore, the crack feature acquisition process includes:

[0215] Acquiring echo signal data according to the concrete crack data set;

[0216] Using a high-pass filter to remove background noise according to the echo signal data to obtain a first echo signal;

[0217] performing signal segmentation according to the first echo signal using an overlapping segmentation method to obtain a second echo signal;

[0218] Calculating the signal peak value and the time when the impact reaches the monitoring point according to the second echo signal to obtain a time domain feature;

[0219] Extracting frequency amplitude using Fourier transform according to the second echo signal to obtain frequency domain features;

[0220] Extracting a signal using wavelet transform according to the second echo signal to obtain a time-frequency feature;

[0221] The time domain feature, the frequency domain feature and the time-frequency feature are combined to obtain the crack feature.

[0222] Furthermore, the machine learning model training process includes:

[0223] Acquire historical echo signals and historical echo signal labels according to the concrete crack dataset, wherein the historical echo signal labels include internal crack depth and width;

[0224] The crack features are input into the machine learning model, and a first loss function is used to obtain the trained machine learning model based on the crack features and the historical echo signal labels.

[0225] Specifically, a decision tree is used as the machine learning model with a maximum depth of 5, a minimum number of sample splits of 10, and a minimum number of sample leaves of 5. The mean square error is used as the first loss function, and 200 historical echo signal samples are used for training and testing. The mean square error during testing is 0.13, which can accurately predict the depth and width of the crack, thereby improving the efficiency and reliability of crack detection.

[0226] Furthermore, the actual crack degree acquisition process includes:

[0227] extracting crack information according to the three-dimensional crack map;

[0228] The surface crack cross-sectional area and the internal crack cross-sectional area are calculated based on the crack information, and the actual crack degree is obtained by performing weighted summation based on the surface crack cross-sectional area and the internal crack cross-sectional area.

[0229] Specifically, for Wall 4, which has significant surface and internal cracks, weights of 0.4 and 0.6 were assigned to these cracks, respectively. The resulting actual crack severity was 528, exceeding 500, marking it as a severely cracked area. Calculating the actual crack severity allows for better safety assessments of concrete structures and helps determine whether reinforcement or repair is necessary, thereby improving the efficiency and reliability of crack detection.

[0230] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting internal cracks in concrete based on image-impact echo, characterized in that: include: Acquire a concrete crack dataset, mark the crack degrees of the concrete crack dataset to obtain a second dataset, and preprocess the second dataset using data enhancement to obtain a first dataset; An RFB network module is introduced into the target detection model, and the RFB network module is used to capture multi-scale features of cracks; The target detection model is trained according to the first data set, and the target detection model training steps include: step S20: inputting the first data set into the target detection model; step S21: introducing the feature extraction layer of the target detection model into the RFB network module, the RFB network module processes the first data set in parallel to obtain a crack feature map, and performs feature fusion on the crack feature map to obtain a crack multi-scale feature map; step S22: outputting the crack coordinates, crack confidence and crack degree using the detection layer according to the crack multi-scale feature map; step S23: calculating the crack loss value using the loss function according to the crack coordinates, the crack confidence and the crack degree; step S24: calculating the parameter gradient using the chain method according to the crack loss value, and updating the target detection model according to the parameter gradient; step S25: terminating if the number of iterations reaches a first threshold. otherwise, repeating steps S20 to S25; acquiring target concrete data, inputting the target concrete data into the trained target detection model to obtain crack extent and crack coordinates, drawing a two-dimensional crack map and dividing crack areas according to the crack extent and crack coordinates; arranging monitoring points for the target concrete according to the crack areas, detecting the monitoring points using impact echoes, and recording target concrete echo signals; extracting crack features from echo signals using digital signal processing technology for the concrete crack data set, and training a machine learning model based on the crack features; inputting the target concrete echo signals into the trained machine learning model to obtain internal crack results for the target concrete; expanding the two-dimensional crack map into a three-dimensional crack map based on the internal crack results for the target concrete, determining the overall crack defect area of ​​the target concrete based on the three-dimensional crack map, and obtaining the actual crack extent.

2. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The crack degree acquisition process includes: measuring concrete cracks using image processing software based on the concrete crack dataset, and obtaining crack sizes based on the starting and ending points of the concrete cracks; obtaining RGB values ​​of the concrete cracks using a color analysis tool, and obtaining brightness and saturation based on the concrete background color; extracting texture features of the concrete cracks using a Gabor filter to obtain crack complexity; and dividing and labeling the crack degrees of the concrete crack dataset based on the crack size, brightness, saturation, crack complexity, and expert experience, where the crack degrees include obvious cracks, slight cracks, and no cracks.

3. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The preprocessing process includes: data enhancement including geometric changes, color transformations, and perspective transformations; enhancing the second data set N times to obtain a newly generated image; merging the newly generated image with images in the concrete crack data set to obtain a third data set; resizing the third data set; the resizing includes: if the third data set is larger than a preset requirement, cropping it according to the preset requirement using a sliding window method, and filling the area of ​​the cropped third data set that is smaller than the preset requirement with white; otherwise, no cropping is performed and the area of ​​the third data set that is smaller than the preset requirement is directly filled with white; and obtaining the first data set based on the adjusted third data set.

4. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The two-dimensional crack map establishment process includes: taking the upper left corner of the target concrete as the origin of the coordinate system, establishing a two-dimensional coordinate system according to the actual size of the target concrete, the two-dimensional coordinate system including an X-axis and a Y-axis, the X-axis extending horizontally to the right, and the Y-axis extending vertically downward; checking whether the crack coordinates are clipped coordinates, if the crack coordinates are the clipped coordinates, adding a sliding window offset to the crack coordinates to obtain the first crack coordinates; otherwise, directly assigning the crack coordinates to the first crack coordinates; using a scaling ratio to calculate the actual coordinates based on the first crack coordinates, and obtaining the two-dimensional crack map based on the actual coordinates.

5. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The crack area division process includes: calculating the center coordinates of the cracks based on the crack coordinates; calculating the distance between the cracks using the Euclidean distance formula based on the center coordinates of the cracks, dividing the cracks with a distance less than a second threshold into the same area, and obtaining a clustering result; calculating the number of cracks based on the clustering result, and if the number of cracks in the area exceeds a third threshold or at least one crack is an obvious crack, obtaining a surface crack area and marking the surface crack area as a defect area; otherwise, marking it as a non-defective area.

6. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The monitoring point layout includes: obtaining the number of monitoring points in the defective area according to the size of the defective area and the number of cracks; obtaining the number of monitoring points in the non-defective area according to the size of the non-defective area and the number of cracks; arranging monitoring points at the junction of the defective area and the non-defective area to obtain the number of monitoring points at the junction; calculating the monitoring point spacing according to the number of monitoring points in the defective area, the number of monitoring points in the non-defective area and the number of monitoring points at the junction, and marking the monitoring point positions according to the monitoring point spacing.

7. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The crack feature acquisition process includes: acquiring echo signal data according to the concrete crack data set; using a high-pass filter to remove background noise according to the echo signal data to obtain a first echo signal; using an overlapping segmentation method to segment the signal according to the first echo signal to obtain a second echo signal; calculating the signal peak and the impact arrival time of the monitoring point according to the second echo signal to obtain a time domain feature; using Fourier transform to extract the frequency amplitude according to the second echo signal to obtain a frequency domain feature; using wavelet transform to extract the signal according to the second echo signal to obtain a time-frequency feature; and combining the time domain feature, the frequency domain feature and the time-frequency feature to obtain the crack feature.

8. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The machine learning model training process includes: obtaining historical echo signals and historical echo signal labels based on the concrete crack dataset, wherein the historical echo signal labels include internal crack depth and width; inputting the crack characteristics into the machine learning model, and using a first loss function to obtain the trained machine learning model based on the crack characteristics and the historical echo signal labels.

9. The method for detecting internal cracks in concrete based on image-impact echo according to claim 1, characterized in that: The actual crack degree acquisition process includes: extracting crack information based on the three-dimensional crack map; calculating the surface crack cross-sectional area and the internal crack cross-sectional area based on the crack information, and performing weighted summation based on the surface crack cross-sectional area and the internal crack cross-sectional area to obtain the actual crack degree.

Images