An AI vision-based concrete component crack identification and positioning method

Abstract

The application discloses a kind of concrete member crack identification and positioning method based on AI vision, comprising: collecting concrete member surface image data;Image is preprocessed and window segmentation;Variational autoencoder model is constructed, and crack candidate area is identified based on unsupervised learning preliminary;Displacement field analysis is carried out to crack candidate area, and crack position is accurately positioned by calculating the displacement field and strain field of strain before and after image;Crack parameters are calculated and combined with sensor data to carry out hierarchical evaluation;Identification result and positioning information are visualized and output.The application combines unsupervised learning and displacement field analysis, avoids dependence on a large number of labeled data, and also realizes high-precision positioning of cracks, with the advantages of high recognition accuracy, accurate positioning and strong environmental adaptability, and can be widely applied to health monitoring of concrete structures such as bridges, tunnels and buildings.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring technology, specifically to a method for identifying and locating cracks in concrete components based on AI vision. Background Technology

[0002] Cracking of concrete is one of the fundamental causes of structural failure in buildings. The initiation and propagation of concrete cracks generally occur on the structural surface, manifesting as localized discontinuities in the surface material. Timely and accurate identification and location of cracks on the surface of concrete components are of significant engineering importance for assessing structural safety and guiding repair and reinforcement.

[0003] Currently, crack detection in concrete structures mainly relies on manual visual inspection or ultrasonic testing. Manual inspection is inefficient, highly subjective, and struggles to detect minute cracks. While ultrasonic testing can measure crack depth and width, it typically only detects cracks visible to the naked eye, failing to identify minute cracks or automatically locate crack locations. In recent years, deep learning-based crack recognition methods have been extensively studied, primarily categorized into supervised and unsupervised learning. Supervised learning methods require a large number of labeled samples for training, but in practical engineering, obtaining sufficient high-quality labeled data is costly, and the imbalance between positive and negative samples often limits recognition accuracy. While some existing AI visual detection methods have improved recognition accuracy to some extent, most have failed to effectively solve the problem of precise crack location and still lack adaptability to environmental conditions and multi-scale crack recognition capabilities.

[0004] Therefore, there is an urgent need for an automated detection method that can balance crack identification accuracy and location accuracy, while having low dependence on data. Summary of the Invention

[0005] 1. Purpose of the invention

[0006] The present invention aims to overcome the shortcomings of the prior art and provide a method for identifying and locating cracks in concrete components based on AI vision, so as to solve the problems of low crack identification accuracy, difficulty in location, and strong dependence on labeled data in the prior art.

[0007] 2 Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying and locating cracks in concrete components based on AI vision, characterized by comprising the following steps:

[0009] Step 1: Image Data Acquisition. Real-time image data of the concrete component surface is acquired using a multispectral camera, while simultaneously obtaining stress-strain sensor monitoring data corresponding to the component in the same time and space.

[0010] Step 2: Image preprocessing. The acquired image data is denoised and enhanced, and the preprocessed image is divided into several sub-image blocks using the sliding window method to form a sliding window image sequence;

[0011] Step 3: Preliminary identification of crack regions based on unsupervised learning (coarse screening):

[0012] 3.1 Construct a variational autoencoder model, wherein the variational autoencoder is trained using images of crack-free concrete surfaces as training samples;

[0013] 3.2 Input the sliding window image sequence obtained in step 2 into the trained variational autoencoder, and output the corresponding reconstructed sliding window image;

[0014] 3.3 Calculate the difference map between the original sliding window image and the reconstructed sliding window image, and preliminarily identify sub-image blocks containing crack candidate regions through threshold segmentation;

[0015] Step 4: Precise crack location based on displacement field (precision calculation):

[0016] 4.1 For the candidate crack regions identified in step 3, extract image sub-blocks before and after strain;

[0017] 4.2 Correlation analysis was performed on the image sub-blocks before and after strain, and the displacement field of each pixel in the image after strain relative to the image before strain was calculated using the cubic surface fitting method.

[0018] 4.3 The strain field distribution is calculated using the local least squares method, and the precise location of the crack is determined based on the local abrupt change characteristics of the strain field;

[0019] Step 5: Crack parameter calculation and classification assessment:

[0020] 5.1 Extract the crack skeleton from the precisely located crack area and calculate the crack length, average width, maximum width, and orientation angle parameters;

[0021] 5.2 Construct a crack feature vector by combining stress-strain sensor monitoring data;

[0022] 5.3 Input the crack feature vector into the pre-trained random forest classification optimization module to perform graded evaluation of cracks and output the crack risk level;

[0023] Step 6: Result visualization and location output: Map the crack identification results and location information back to the original image coordinate system, mark the crack location, outline and risk level on the original image, and output the spatial location information of the crack in the concrete component coordinate system.

[0024] 3. Technical Effects

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. High recognition accuracy: By combining unsupervised learning variational autoencoder with displacement field-based correlation analysis, it avoids the dependence of supervised learning on a large amount of labeled data and can accurately locate cracks, effectively improving the detection rate of micro-cracks.

[0027] 2. Accurate positioning: The displacement and strain fields are calculated using cubic surface fitting and local least squares method, which can identify sub-pixel level displacement changes and achieve high-precision crack positioning;

[0028] Strong environmental adaptability: The fusion of multispectral image data and stress-strain sensor data improves the robustness of detection under different lighting and surface conditions;

[0029] 3. High level of intelligence: The random forest classifier is used to classify and evaluate cracks, providing a quantitative basis for structural safety assessment and realizing full automation from crack identification to risk assessment. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the method of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the principle of unsupervised crack identification using a variational autoencoder.

[0032] Figure 3 This is a schematic diagram of a method for precise crack location based on displacement field.

[0033] Figure 4 This is a flowchart for crack parameter calculation and classification assessment. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] This embodiment provides a method for identifying and locating cracks in concrete components based on AI vision, specifically including the following steps:

[0037] Step 1: Image Data Acquisition

[0038] An image acquisition system was deployed on the underside of a concrete box girder of a bridge. A multispectral camera (including visible and near-infrared bands) was used to acquire images of the concrete surface at predetermined intervals, while simultaneously acquiring monitoring data from stress-strain sensors located at key locations. Sufficient overlap was ensured during image acquisition to facilitate subsequent image stitching and positioning.

[0039] Step 2: Image Preprocessing

[0040] The acquired images were denoised (using median filtering) and subjected to illumination unevenness correction (using histogram equalization). Then, the preprocessed images were divided into 64×64 pixel sub-image blocks using a sliding window method, with a sliding window step size of 32 pixels, forming a sliding window image sequence.

[0041] Step 3: Preliminary identification of crack regions based on unsupervised learning

[0042] A variational autoencoder (VAE) model is constructed, comprising an encoder network and a decoder network. The encoder maps the input image to mean and variance parameters in a latent space, while the decoder samples from the latent space and reconstructs the image. The model is trained using images of crack-free concrete surfaces as training samples, enabling it to accurately reconstruct crack-free images.

[0043] The sliding window image sequence obtained in step 2 is input into the trained variational autoencoder, and the model outputs a reconstructed image. Since the model is only trained on crack-free images, for images containing cracks, the reconstruction result treats the cracks as "anomalies" and smooths them, thus generating a crack-free reconstructed image. The pixel-level difference between the original sliding window image and the reconstructed sliding window image is calculated to generate a difference map. Adaptive thresholding is performed on the difference map; pixels with a difference value exceeding the threshold are considered crack candidate points, and sub-image patches containing a sufficient number of crack candidate points are identified as crack candidate regions.

[0044] Step 4: Precise crack location based on displacement field

[0045] For the candidate crack regions identified in step 3, image sub-blocks of these regions before and after the application of external load are extracted as pre-strain and post-strain images. Digital image correlation methods are used to perform correlation analysis on the pre-strain and post-strain images: the pre-strain image is divided into several computational sub-regions, and the matching position with the highest correlation coefficient to the sub-region is searched in the post-strain image. A cubic surface fitting method is used to fit the correlation coefficient distribution to obtain sub-pixel level displacement measurement accuracy. The displacement field of the entire region is constructed based on the displacement of each computational sub-region. Then, the local least squares method is used to differentiate the displacement field to obtain the strain field distribution. Since strain concentration occurs at cracks, the precise location of the crack can be accurately determined by detecting local abrupt peaks in the strain field.

[0046] Step 5: Crack parameter calculation and classification assessment

[0047] For precisely located crack areas, an image thinning algorithm is used to extract the crack skeleton lines. The crack width is calculated along the normal direction of the crack skeleton lines, and the average and maximum crack widths are statistically obtained. The pixel length of the skeleton lines is calculated and converted into the actual physical length based on the image resolution. The principal direction of the crack skeleton lines is calculated as the crack direction angle.

[0048] The crack parameters (length, average width, maximum width, and orientation angle) are combined with synchronously acquired stress and strain sensor data (strain values ​​and stress values) to construct a crack feature vector. This feature vector is then input into a pre-trained random forest classification optimization module, which is trained based on historical crack data and can classify cracks into three risk levels: "minor," "moderate," and "severe."

[0049] Step 6: Result Visualization and Location Output

[0050] The crack identification results and location information are mapped back to the original image coordinate system through coordinate transformation. Crack areas of different risk levels are marked with different colors on the original image, crack skeleton lines are drawn, and crack parameters and risk levels are labeled. At the same time, combined with the camera pose parameters during image acquisition and the 3D model of the component, the crack location is transformed into the component coordinate system, and the spatial positioning coordinates of the crack on the concrete component are output to generate an inspection report.

[0051] Example 2

[0052] This embodiment, based on Embodiment 1, further improves the variational autoencoder model in step 3 by replacing the fully connected network with a convolutional variational autoencoder (CVAE) to better extract spatial features from the image. The encoder part uses four convolutional layers, each followed by batch normalization and ReLU activation functions; the decoder part uses four deconvolutional layers. The latent space dimension is set to 128 dimensions. During training, mean squared error is used as the reconstruction loss, combined with KL divergence as a regularization term.

[0053] Experiments show that the method in this embodiment can identify micro-cracks as small as 0.05 mm in width in the detection of cracks in the bottom slab of concrete bridges, with a positioning error of less than 0.5 mm and a crack classification accuracy of over 92%.

Claims

1. A method for identifying and locating cracks in concrete components based on AI vision, characterized in that, Includes the following steps: Step 1: Acquire real-time image data of the concrete component surface; Step 2: Preprocess the image data and segment it into a sliding window image sequence; Step 3: Construct an unsupervised learning model, using images of crack-free concrete surfaces as training samples; input the sliding window image sequence into the trained model, and identify sub-image blocks containing crack candidate regions by calculating the difference map between the original image and the reconstructed image; Step 4: For the candidate crack region, extract image sub-blocks before and after strain, perform correlation analysis to construct a displacement field, and determine the precise location of the crack based on the strain field characteristics obtained by differentiating the displacement field. Step 5: Extract the geometric parameters of the crack after precise location, combine them with external monitoring data for graded assessment, and output the risk level and spatial location information.

2. The method for identifying and locating cracks in concrete components based on AI vision according to claim 1, characterized in that: In step 1, a multispectral camera is used to acquire images, and the multispectral range includes the visible light band and the near-infrared band.

3. The method for identifying and locating cracks in concrete components based on AI vision according to claim 1, characterized in that: The unsupervised learning model mentioned in step 3 is a variational autoencoder model or a convolutional variational autoencoder model.

4. The method for identifying and locating cracks in concrete components based on AI vision according to claim 3, characterized in that: During the reconstruction process, the variational autoencoder model smooths out abnormal regions containing cracks and extracts pixels exceeding a threshold in the difference map as crack candidate points through adaptive threshold segmentation.

5. The method for identifying and locating cracks in concrete components based on AI vision according to claim 1, characterized in that: In step 4, the correlation coefficient distribution is fitted using a cubic surface fitting method to obtain sub-pixel level displacement measurement accuracy.

6. The method for identifying and locating cracks in concrete components based on AI vision according to claim 1, characterized in that: In step 4, the local least squares method is used to differentiate the displacement field to obtain the strain field, and the crack is located based on the local abrupt peak value in the strain field.

7. The method for identifying and locating cracks in concrete components based on AI vision according to claim 1, characterized in that: The external monitoring data mentioned in step 5 is the stress and strain sensor monitoring data of the component acquired synchronously in the corresponding time and space.

8. The method for identifying and locating cracks in concrete components based on AI vision according to claim 7, characterized in that: In step 5, the crack geometric parameters and stress-strain data are used to construct a crack feature vector, which is then input into a pre-trained random forest classification optimization module for risk classification.

Images