A method for identifying and quantifying apparent defects and diseases in concrete structures

Abstract

The present invention discloses a method for identifying and quantifying apparent defects and diseases of concrete structures. Specifically, the method comprises the following steps: obtaining a database of apparent defects and diseases of concrete structures and an associated database through an indoor test method, and predicting the depth of apparent defects and diseases using a convolutional neural network coupled with LSTM and Transformer deep learning; identifying apparent defects and diseases in a surface image of a concrete structure using a convolutional neural network coupled with LSTM and Transformer deep learning, and marking the identified apparent defects and disease areas; and estimating the area of a calibrated region of apparent defects and diseases of a concrete structure using a projection overlay method. The present invention improves the detection efficiency and accuracy of apparent defects and diseases of concrete structures and saves detection costs. It realizes the prediction of the correlation between the size characteristics of apparent defects and diseases of concrete structures, and provides very important basic data for safety assessment and maintenance decision-making of concrete structures.

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic image recognition, and in particular relates to a method for identifying and quantifying apparent defects and diseases of concrete structures. Background Art

[0002] Concrete structures are the most classic and still the most widely used structural form in the engineering field. Whether during the construction process or during the operation period, they will have various surface defects and diseases due to human or natural factors. These defects will not only damage the appearance and integrity of the concrete structure, but will also further develop, threatening the structural safety and even causing serious engineering accidents.

[0003] At present, the field of engineering inspection generally adopts manual visual inspection or simple measuring tools to carry out periodic inspections of apparent defects and diseases of concrete structures. This traditional inspection method not only requires inspectors to have high professional and technical levels and proficient equipment operation capabilities, but also requires a large amount of man-hours, with the characteristics of high cost and long construction period. Since the inspection results obtained by traditional methods are easily affected by human subjective judgment, their reliability is difficult to guarantee, and therefore the effectiveness and rationality of the countermeasures taken based on the inspection results cannot be guaranteed. In addition, manual inspection also has problems such as low accuracy, single parameters and poor ability to process large amounts of data in quantifying the apparent defects and diseases of concrete. In particular, for hidden defects, such as the depth measurement of defects (diseases), it is almost impossible to obtain their measurement results quickly and at low cost without taking destructive measures. However, this is precisely the important factor affecting the bearing capacity and safety of concrete structures.

[0004] Artificial intelligence, a computer technology that can autonomously perform complex judgments based on human judgment logic, possesses characteristics such as autonomy, adaptability, intelligent interaction, and the ability to process large amounts of data and perform parallel processing. Through sample training and adaptive learning, intelligent algorithms with advanced recognition and analysis capabilities and statistical calculation functions can be constructed. Because their judgment logic is trained on large sample data, their thinking framework and execution results are reliable. Furthermore, intelligent algorithms employ rigorous theoretical logic, effectively avoiding recognition errors caused by subjective human judgment and enabling accurate predictions for results that cannot be directly measured through associative learning. Therefore, it is clear that the use of artificial intelligence algorithms offers significant advantages in identifying surface defects and diseases in concrete structures. They not only overcome the many shortcomings of traditional manual inspection methods, but also significantly improve inspection efficiency and accuracy, reducing inspection costs. Furthermore, artificial intelligence algorithms can non-destructively predict the correlation between the dimensional characteristics of surface defects and diseases in concrete structures (including both planar and depth dimensional characteristics), providing crucial foundational data for safety assessments and maintenance decisions for existing concrete structures. Summary of the Invention

[0005] The purpose of this invention is to develop an automatic image recognition method based on the combination of convolutional neural networks and LSTM-coupled Transformer deep learning for the identification of apparent defects and diseases in concrete. At the same time, based on the combination of computer image processing technology and correlation measurement technology, a technology for defining, screening, and measuring characteristic areas is established for the quantitative characterization of the regional area of apparent defects and diseases in concrete.

[0006] To achieve the above objectives, the present invention provides a method for identifying and quantifying apparent defects and diseases of concrete structures.

[0007] The present invention provides a method for identifying and quantifying apparent defects and diseases of concrete structures, comprising the following steps:

[0008] Step 1: Through indoor test methods, obtain a database of construction quality defects and operational diseases of different types of concrete structures under different working conditions and service environments, and establish a database that associates construction factors, load conditions, and service environments with the three-dimensional characteristic dimensions of concrete structure surface defects and diseases.

[0009] Step 2: Obtain images of the concrete structure surface, preprocess the images, and establish a concrete structure surface image dataset.

[0010] Step 3: Conduct convolutional neural network training on samples from the concrete structure surface image database to establish a convolutional neural network algorithm model that can preliminarily identify structural surface defects and diseases. Then, combine the LSTM-coupled Transformer deep learning effect to optimize the parameters of the recognition model and improve its recognition and prediction accuracy for surface defects and diseases.

[0011] Step 4: Preprocess the concrete structure surface image samples and define and mark the boundaries of the apparent defects and diseased areas identified in step 2; use the unit grid covering method to cover the preprocessed sample images, and calculate the apparent characteristic size parameters by counting the number of unit grids in the apparent defects and diseased areas.

[0012] Step 5: Based on the database from step 1, conduct convolutional neural network training on data samples related to concrete structure construction factors, load conditions, and service environment, and the three-dimensional dimensions of its apparent defects and diseases. Build a prediction method for the three-dimensional dimensions of apparent defects and diseases based on these factors. Combined with the LSTM-coupled Transformer deep learning effect, optimize the parameters of the prediction model and improve its prediction accuracy for the three-dimensional dimensions of apparent defects and diseases.

[0013] Step 6: Output the three-dimensional size recognition and prediction results of concrete structure surface defects and diseases based on the combination of convolutional neural network and LSTM coupled Transformer deep learning.

[0014] Furthermore, in step 1, the construction factors include cement type, water-cement ratio, additive type, etc.; the load conditions include static load, dynamic load, and alternating load, etc.; and the three-dimensional characteristic dimensions are length, width, depth, and area.

[0015] Furthermore, in step 2, among the concrete structure surface images, intact surfaces account for 30%, images containing surface defects account for 30%, and images containing apparent diseases account for 40%.

[0016] Furthermore, image preprocessing is based on the MATLAB platform and uses Python, Java, C, C++, JavaScript, C#, Ruby, PHP or Objective-C language editing to automatically implement image flipping, scaling, sharpening, denoising and mode conversion operations.

[0017] Furthermore, step 5 is specifically as follows: after taking the preprocessed image as input, the data is first skipped through the convolution layer with the optimal shortcut, and then processed by the rectified linear unit (ReLU) activation function before output, so as to solve the gradient explosion problem of the convolutional neural network model; then, the LSTM-coupled Transformer deep learning method is used to enhance the sensitivity of the constructed network model to corrosion recognition, that is, to obtain the optimal parameters of the convolutional neural network model for recognition of apparent defects and diseases of concrete structures; for the recognition of apparent defects and diseases of concrete structures, only the categories of containing defects, containing diseases, containing both defects and diseases, and containing neither defects nor diseases are set, so as to reduce the impact of the large number of classes output by the original training model on the recognition efficiency and accuracy of corrosion; and the adaptive moment estimation is used to adjust the synaptic weights in the fully connected layer.

[0018] Furthermore, step 6 also includes: estimating the area of the apparent defects and disease calibration areas of the concrete structure using a projection coverage method.

[0019] The image with identified apparent defects and disease areas is divided using the grid coverage method, and a large number of regularly arranged unit grids are projected on the image surface. The area of the grid is denoted as ε, and its value is ε = (a×b) / N, where a is the image pixel height, that is, the ratio of the preprocessed image height to the image pixel DPI, b is the image pixel width, that is, the ratio of the preprocessed image width to the image pixel DPI, and N is the number of unit grids.

[0020] The area S of the apparent defects and diseases in the image sample can be calculated by the unit grid area ε and the number of unit grids n of the continuous closed curve of the area contour, that is, S = ε × n. S is used as a specific indicator to quantify the apparent defects and diseases of concrete structures.

[0021] The beneficial technical effects of the present invention are:

[0022] 1. The present invention solves to a certain extent the problems of the current traditional detection methods for apparent defects and diseases of concrete structures, such as large manpower input, long time, high cost and susceptibility to interference from subjective judgment.

[0023] 2. The method of combining convolutional neural network with LSTM-coupled Transformer deep learning in the present invention can effectively obtain complete characteristic size information of apparent defects and damages of concrete structures, making up for the shortcomings of existing technologies in predicting apparent defects and depth measurement, and providing a complete judgment basis for subsequent assessment of the bearing capacity of concrete structures.

[0024] 3. This invention uses LSTM coupled with Transformer deep learning to effectively improve the accuracy of prediction models under conditions of insufficient or small training sets, laying a good foundation for the engineering application and promotion of this technology.

[0025] 4. The present invention takes into account the apparent defects and depth dimensions of concrete structures under concrete construction factors and objective factors, and trains its associated data through a convolutional neural network. The constructed method for predicting the depth of apparent defects and diseases based on concrete construction factors and objective factors solves the technical difficulty of obtaining the depth of apparent defects and diseases by non-destructive means. Moreover, through precise intelligent algorithm analysis, the accuracy of the predicted depth of apparent defects and diseases can be guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the method for identifying and quantifying apparent defects and diseases of concrete structures according to the present invention.

[0027] Figure 2 Schematic diagram of the preprocessing process of images containing apparent defects and diseases of concrete structures.

[0028] Figure 3 Schematic diagram of the convolutional neural network structure.

[0029] Figure 4 Flowchart for the identification and quantification of apparent defects and diseases in concrete structure surface images.

[0030] Figure 5 Flowchart for the identification of apparent defects and diseases for the indoor test sample database.

[0031] Figure 6 Schematic diagram of the deep learning framework of LSTM coupled Transformer.

[0032] Figure 7Schematic diagram of the planar characteristic dimension measurement of apparent defects and diseased areas in concrete structures using the projection overlay method. (In the figure, 1 is the unit grid covered by the concrete structure surface image using the projection overlay method; 2 is the apparent defects and diseased areas of the concrete structure; 3 is the continuous closed curve of the outline of the apparent defects and diseased areas of the concrete structure) DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0034] The present invention is a method for identifying and quantifying apparent defects and diseases of concrete structures. The overall process is as follows: Figure 1 As shown, the following steps are included:

[0035] Step 1: Through indoor test methods, obtain a database of construction quality defects and operational diseases of different types of concrete structures under different working conditions and service environments, and establish a database that associates construction factors (cement type, water-cement ratio, additive type, etc.), load conditions (static load, dynamic load, and alternating load, etc.), and service environment with the three-dimensional characteristic dimensions (length, width, depth, and area, etc.) of the apparent defects and diseases of concrete structures.

[0036] Step 2: Obtain images of the concrete structure surface (30% of the complete surface, 30% of the images containing surface defects, and 40% of the images containing apparent defects). Images of the concrete structure's apparent defects and defects are captured using a high-performance camera with autofocus. The distance between the camera and the concrete structure is controlled between 0.5 and 1 meter to ensure high-quality images. The camera should be positioned at a straight-on, straight-down, or straight-off angle to capture the concrete structure's apparent defects (diseases). Ensure that the lens is parallel to the surface of the object to the greatest extent possible.

[0037] The images are preprocessed to establish a concrete structure surface image dataset. The preprocessing process is as follows Figure 2 In the image preprocessing stage, MATLAB is used to perform image denoising, enhancement, class conversion (converting color images to grayscale images, and grayscale images to binary images, etc.), edge detection and segmentation, histogram matching, and contour matching.

[0038] Step 3: Based on step 2, perform convolutional neural network on the concrete structure appearance image database samples (such as Figure 3 By combining the LSTM-coupled Transformer deep learning effect, the parameters of the recognition model are optimized to improve its recognition and prediction accuracy for surface defects and diseases.

[0039] Step 4: Preprocess the concrete structure surface image samples, and define and mark the boundaries of the apparent defects and disease areas identified in step 2; use the unit grid covering method to cover the preprocessed sample images, and calculate the apparent characteristic size parameters by counting the number of unit grids in the apparent defects and disease areas. The identification and quantification process of the apparent defects and diseases in the concrete structure surface images is as follows: Figure 4 shown.

[0040] First, a convolutional neural network is trained on the image sample dataset to construct an initial convolutional neural network model for the identification of apparent defects and diseases of concrete structures. Then, the LSTM-coupled Transformer deep learning method is used to improve the recognition accuracy of the initial convolutional neural network model, thereby constructing a convolutional neural network prediction model with sufficient recognition accuracy for apparent defects and diseases of concrete structures.

[0041] Step 5: Based on the database of step 1, carry out convolutional neural network training on the data samples of the association between the construction factors, load conditions and service environment of concrete structures and the three-dimensional size of their apparent defects and diseases, and build a prediction method for the three-dimensional size of apparent defects and diseases based on the construction factors, load conditions and service environment of concrete structures. At the same time, combine the LSTM coupled Transformer deep learning effect to achieve parameter optimization of the prediction model and improve its prediction accuracy for the three-dimensional size of apparent defects and diseases. The identification process of the apparent defects and diseases of the indoor test sample database is as follows: Figure 5 shown.

[0042] LSTM coupled Transformer deep learning framework such as Figure 6 As shown in the figure, after taking the preprocessed image as input, the data is first skipped through the convolution layer with the optimal shortcut, and then processed by the rectified linear unit (ReLU) activation function before output to solve the gradient explosion problem of the convolutional neural network model; then the LSTM-coupled Transformer deep learning method is used to enhance the sensitivity of the constructed network model to corrosion recognition, that is, to obtain the optimal parameters of the convolutional neural network model for identifying apparent defects and diseases of concrete structures; for the identification of apparent defects and diseases of concrete structures, only the categories of containing defects, containing diseases, containing both defects and diseases, and containing neither defects nor diseases are set to reduce the impact of the large number of classes output by the original training model on the recognition efficiency and accuracy of corrosion; and the adaptive moment estimation is used to adjust the synaptic weights in the fully connected layer.

[0043] Step 6: Output the three-dimensional size recognition and prediction results of concrete structure surface defects and diseases based on the combination of convolutional neural network and LSTM coupled Transformer deep learning.

[0044] The projection coverage method is used to estimate the area of the apparent defects and disease calibration areas of concrete structures. Figure 7 shown.

[0045] Using a grid overlay method, the image containing identified areas of surface defects and damage is divided. A large number of regularly arranged unit grids are projected onto the image surface. The area of this grid, denoted as ε, is calculated as ε = (a × b) / N, where a is the image pixel height (the ratio of the preprocessed image height to the image pixel depth per inch), b is the image pixel width (the ratio of the preprocessed image width to the image pixel depth per inch), and N is the number of unit grids. The grid overlay method projects a sufficiently small, regularly arranged unit grid onto the preprocessed image to be analyzed. This method avoids the influence of tilt errors caused by non-perpendicular image capture on the actual number of filled grids, thereby ensuring that the unit grids in the corroded areas are effectively filled.

[0046] The area S of the apparent defects and diseases in the image sample can be calculated by the unit grid area ε and the number of unit grids n of the continuous closed curve of the area contour, that is, S = ε × n. S is used as a specific indicator to quantify the apparent defects and diseases of concrete structures.

[0047] Prioritize, use indoor test methods to obtain the data set of apparent defects and diseases of concrete structures under different construction mixing, loading conditions and service environments, use the ResNet18 network algorithm model to train the associated data set, and construct a deep feature prediction model of apparent defects and diseases under the combined influence of concrete mixing technology, loading conditions and service environment. Optimize the model parameters through LSTM coupled Transformer deep learning to improve the accuracy of the prediction results.

[0048] Preferably, the photographic hardware equipment used to capture images of apparent defects and diseases of concrete structures is equipped with both automatic and manual focusing functions, specifically, unmanned equipment that is waterproof, anti-vibration, corrosion-resistant, has a long battery life, can provide lighting functions, and has a sufficiently high resolution.

[0049] Preferably, the image database collected using camera hardware must contain more than 1,000 data samples, and the ratio of data samples containing apparent defects, apparent diseases, and data samples without apparent defects or diseases should be 3:3:4.

[0050] Prior to this, the captured images of apparent defects and diseases of concrete structures are automatically and intelligently preprocessed. The preprocessing includes adjusting the angle, size, pixels, and resolution of the images. Specifically, the preprocessing is based on the MATLAB platform and is edited using languages such as Python, Java, C, C++, JavaScript, C#, Ruby, PHP, and Objective-C to automatically perform operations such as image flipping, scaling, sharpening, denoising, and mode conversion. The mode conversion here includes but is not limited to changes in bitmap, hue, color index, and number of channels.

[0051] Preferably, the grid spacing used to cover the pre-processed image containing apparent defects or diseases should be a common divisor of the pre-processed image pixel size to ensure that the segmented units are small enough and can completely cover the measurement image surface. For the grid units on the boundary line of the apparent defects and diseases of the concrete structure, their actual number is measured according to geometric methods and added to the total number of unsegmented units within the boundary line. This is used to convert the characteristic size parameters (length, width, and area) of the apparent defects and diseases of the concrete structure.

[0052] Example:

[0053] The accuracy of the present invention's method for identifying and quantifying apparent defects and diseases in concrete structures, namely the ResNet-18+LSTM-coupled Transformer deep learning model, was compared with similar technologies. All models were trained and tested 40 times during the comparative testing. Furthermore, the number of training epochs, batch size, and learning rate for the present invention's LSTM-coupled Transformer deep learning method for predicting apparent defects and diseases in concrete structures were set to 5, 10, and 0.0001, respectively. The comparison results are shown in Table 1.

[0054] Table 1 Comparison of the accuracy of the present invention and similar technologies

[0055]

[0056] Judging from the comparison results, the recognition accuracy of apparent defects and diseases of concrete structures by the present invention is at least 15.0% higher than that of existing technologies. That is, after improving the convolutional neural network model by adopting LSTM coupled Transformer deep learning, the technology proposed by the present invention is more accurate in identifying apparent defects and diseases of concrete structures.

Claims

1. A method for identifying and quantifying apparent defects and diseases of concrete structures, characterized in that: The following steps are involved: Step 1: Through indoor testing methods, a database of construction quality defects and operational defects of different types of concrete structures under different working conditions and service environments is obtained. A database is then established that correlates construction factors, load conditions, and service environments with the three-dimensional characteristic dimensions of concrete structure surface defects and defects. Step 2: Obtain concrete structure surface images, preprocess the images, and establish a concrete structure surface image dataset; Step 3: Train a convolutional neural network on samples from a database of concrete structure surface images to establish a convolutional neural network algorithm model that can initially identify structural surface defects and diseases. Then, by combining the LSTM-coupled Transformer deep learning results, we optimize the recognition model parameters and improve its accuracy in identifying and predicting surface defects and diseases. Step 4: Preprocess the concrete structure surface image samples and define and mark the boundaries of the apparent defects and disease areas identified in step 2; use the unit grid covering method to cover the preprocessed sample images, and calculate the apparent characteristic size parameters by counting the number of unit grids in the apparent defects and disease areas; Step 5: Based on the database from Step 1, a convolutional neural network is trained on data samples related to concrete structure construction factors, load conditions, and service environment, and the three-dimensional dimensions of its apparent defects and diseases. This method is used to predict the three-dimensional dimensions of apparent defects and diseases based on these factors. The LSTM-coupled Transformer deep learning approach is then combined to optimize the parameters of the prediction model and improve its accuracy in predicting the three-dimensional dimensions of apparent defects and diseases. Step 6: Output the three-dimensional size recognition and prediction results of concrete structure surface defects and diseases based on the combination of convolutional neural network and LSTM coupled Transformer deep learning.

2. A method for identifying and quantifying apparent defects and diseases of concrete structures according to claim 1, characterized in that: The construction factors in step 1 include cement type, water-cement ratio, and additive type; the load conditions include static load, dynamic load, and alternating load; The three-dimensional characteristic dimensions are length, width, depth and area.

3. A method for identifying and quantifying apparent defects and diseases of concrete structures according to claim 1, characterized in that: In the surface pictures of the concrete structure in step 2, 30% of the pictures are intact, 30% of the pictures contain surface defects, and 40% of the pictures contain apparent diseases.

4. A method for identifying and quantifying apparent defects and diseases of concrete structures according to claim 1, characterized in that: The image preprocessing is based on the MATLAB platform and uses Python, Java, C, C++, JavaScript, C#, Ruby, PHP or Objective-C language editing to automatically implement image flipping, scaling, sharpening, denoising and mode conversion operations.

5. The method for identifying and quantifying apparent defects and diseases of concrete structures according to claim 1, characterized in that: Specifically, step 5 comprises the following steps: after taking the preprocessed image as input, first skipping the convolutional layer with an optimal shortcut, and then processing the data with a rectified linear unit (ReLU) activation function before output, so as to solve the gradient explosion problem of the convolutional neural network model; then utilizing the LSTM coupled with the Transformer deep learning method to enhance the sensitivity of the constructed network model to corrosion recognition, i.e., obtaining the optimal parameters of the convolutional neural network model for identifying apparent defects and diseases of concrete structures; setting the classes for identifying apparent defects and diseases of concrete structures to only include defects, diseases, both defects and diseases, and neither defects nor diseases, so as to reduce the impact of the large number of classes output by the original training model on the recognition efficiency and accuracy of corrosion; and using adaptive moment estimation to adjust the synaptic weights in the fully connected layer.

6. The method for identifying and quantifying apparent defects and diseases of concrete structures according to claim 1, characterized in that: The step 6 further includes: estimating the area of the concrete structure apparent defects and disease calibration areas using a projection coverage method; The image containing identified apparent defects and disease areas is divided using a grid coverage method. A large number of regularly arranged unit grids are projected onto the image surface. The area of this grid is denoted as ε, and its value is ε = (a × b) / N, where a is the image pixel height, that is, the ratio of the pre-processed image height to the image pixel DPI, b is the image pixel width, that is, the ratio of the pre-processed image width to the image pixel DPI, and N is the number of unit grids. The area S of the apparent defects and diseases in the image sample can be calculated by the unit grid area ε and the number of unit grids n of the continuous closed curve of the area contour, that is, S = ε × n. S is used as a specific indicator to quantify the apparent defects and diseases of concrete structures.

Images