Bridge crack identification method

Abstract

The invention discloses a bridge crack identification method, and relates to the field of bridge crack identification, and the method comprises the following steps: applying an alternating load through a running automobile, and carrying out the time sequence image collection; performing displacement field reconstruction on the time sequence image through a DIC algorithm, and obtaining strain field distribution; preliminarily positioning the potential position of the crack by capturing gradient mutation characteristics in the displacement field and a Gaussian distribution rule presented by virtual strain in the strain field; and constructing a dynamic state space model based on a Kalman filtering algorithm, substituting the preliminarily determined potential position of the crack into a state vector of the dynamic state space model, fusing multi-source heterogeneous data in a recursive manner, and suppressing measurement noise and single detection deviation to obtain a final crack position. The method is simple and clear in use process, higher in recognition precision and capable of recognizing micro cracks.

Description

Technical Field

[0001] The present invention relates to the field of bridge crack identification, and particularly to a bridge crack identification method. Background Art

[0002] Current crack identification is of great significance in the field of structural health monitoring. Common methods mainly include traditional image processing, machine learning, and other advanced technologies. Traditional image processing methods such as Canny and Sobel edge detection, threshold segmentation, morphological operations, and template matching detect cracks through edge recognition, grayscale binarization, and denoising, and are suitable for scenarios with simple environments or clear crack edges. Machine learning methods include support vector machine (SVM), decision tree, random forest, and K-nearest neighbor algorithm (K-NN), which classify and identify by extracting texture and shape features of cracks, adapt to data of different scales and complexities, but have large computational amounts, long time consumption, and low accuracy. Summary of the Invention

[0003] Aiming at the above deficiencies in the prior art, a bridge crack identification method provided by the present invention solves the problems of large computational amount and low accuracy of existing machine learning methods in bridge crack identification.

[0004] In order to achieve the above invention purpose, the technical solution adopted by the present invention is as follows:

[0005] Provide a bridge crack identification method, which includes the following steps:

[0006] Apply an alternating load through a moving vehicle to collect time-series images of the bridge area to be measured;

[0007] Reconstruct the displacement field of the time-series images through the DIC algorithm, and obtain the strain field distribution based on spatio-temporal correlation analysis;

[0008] Preliminarily locate the potential crack positions by capturing the gradient mutation features in the displacement field and the Gaussian distribution law presented by the imaginary strain in the strain field;

[0009] Combine the characteristics of alternating load application, conduct multi-cycle repeated measurements in the potential crack areas, and further preliminarily locate multiple groups of potential crack positions;

[0010] Construct a dynamic state space model based on the Kalman filter algorithm, substitute the preliminarily determined potential crack positions into the state vector of the dynamic state space model, and suppress measurement noise and single-detection deviation through recursive fusion of multi-source heterogeneous data to obtain the final crack positions.

[0011] The beneficial effects of the present invention are as follows: This method utilizes the alternating loads of vehicles induced by traffic flow as an external excitation source, and promotes the dynamic opening and closing behavior of cracks through stress perturbation. Based on the principle of DIC full-field strain analysis, by tracking the spatial distribution characteristics of the virtual strain field on the surface of the structure, it is found that the crack area shows a significant Gaussian distribution characteristic under the action of alternating loads, and this characteristic is used as a preliminary criterion for the crack initiation and propagation regions. To further improve the detection accuracy and eliminate the interference of measurement noise, this method introduces the Kalman filter (Federated Kalman Filter) framework and constructs a fusion mechanism for multi-source heterogeneous data. Through the data fusion of the Kalman filter, the position of micro-cracks can be accurately determined. This method is simple and clear in the using process, and has a higher recognition accuracy, and can identify micro-cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic flow chart of this method;

[0013] Figure 2 Schematic diagram of a significant gradient transformation of the displacement field at the crack. DETAILED DESCRIPTION OF THE INVENTION

[0014] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0015] As Figure 1 shown, the bridge crack identification method includes the following steps:

[0016] S1. Apply alternating loads through a moving vehicle and collect time-series images of the bridge area to be measured;

[0017] S2. Reconstruct the displacement field of the time-series images through the DIC algorithm and obtain the strain field distribution based on spatio-temporal correlation analysis;

[0018] S3. Initially locate the potential crack positions by capturing the gradient mutation characteristics in the displacement field and the Gaussian distribution law presented by the virtual strain in the strain field;

[0019] S4. Combine the loading characteristics of the alternating loads, conduct multi-cycle repeated measurements in the potential crack areas, and then initially locate multiple groups of potential crack positions;

[0020] S5. Construct a dynamic state space model based on the Kalman filter algorithm. Substitute the preliminarily determined potential positions of the cracks into the state vector of the dynamic state space model, and fuse multi-source heterogeneous data recursively to suppress measurement noise and single-detection deviation, thereby obtaining the final positions of the cracks.

[0021] The specific method for performing sequential image acquisition on the bridge area to be measured is as follows: Apply speckles to the bridge area to be measured. The size of the speckles should generally be 1 / 20 to 1 / 50 of the image resolution to ensure the clarity of the speckles and the accuracy of subsequent analysis. Then, use an unmanned aerial vehicle system with a unified coordinate system to perform sequential image acquisition.

[0022] In this embodiment, white paint and black random paint are respectively sprayed on the selected bridge area, and a high-quality speckle field is determined according to the approximate ratio of white paint to black paint of 1:1.

[0023] The specific method for reconstructing the displacement field of the sequential images by the DIC algorithm and obtaining the strain field distribution based on spatio-temporal correlation analysis in step S2 includes the following steps:

[0024] S2-1. Denoise the sequential images through median filtering to obtain the denoised sequential images. Let Z be the original image and L be the image after median filtering. For each pixel point in the image set as (i, j), the processing formula of median filtering can be expressed as:

[0025] L(i, j) = median{Z(p, c)|(p, c) ∈ N(i, j}}

[0026] Where:

[0027] L(i, j): The pixel value of the filtered image at the position (i, j);

[0028] Z(p, c): The pixel value of the original image at the position (p, c);

[0029] N(i, j): The neighborhood window centered on (i, j), usually a rectangular area with an odd size;

[0030] median{.}: Take the median of all pixel values in the neighborhood.

[0031] S2-2. Convert the denoised sequential images into a sequence of grayscale images through DIC software, and use the pixel subset containing the target measurement points in the first-frame grayscale image as the reference image; Select the region of interest in the reference image through DIC software;

[0032] S2-3. Discretize the region of interest into several independent subgraphs in the DIC software, and search for the positions of the subgraphs in the subsequent images respectively. That is, by matching the subgraphs before and after applying the alternating load, the displacement field reconstruction and strain field data calculation are completed.

[0033] The specific method for completing the displacement field reconstruction by matching the subgraphs before and after applying the alternating load includes the following steps:

[0034] With the goal of optimizing the correlation of the subgraphs before and after applying the alternating load, obtain the displacement amount of a single subgraph before and after applying the alternating load, and combine the displacement amounts of all subgraphs before and after applying the alternating load to obtain the displacement vector of the reference image after applying the alternating load, that is, obtain the displacement field.

[0035] In this embodiment, the zero-mean normalized cross-correlation function (NCC) and the sum of squared differences (SSD) are used to calculate the displacement data. The zero-mean normalized cross-correlation eliminates the influence of brightness differences during calculation, making the method more robust to illumination changes and local texture differences. The zero-mean normalized cross-correlation function j ZNCC ;

[0036]

[0037] where C ZNCC(u，v) represents the correlation of the subgraphs before and after applying the alternating load; I(x i , y i ) represents the pixel gray value at the position (x i , y i ) in the subgraph of the reference image; J(x i + u, y i + v) represents the pixel gray value at the position (x i + u, y i + v) in the image area obtained by subgraph matching; (u, v) represents the displacement vector of the image obtained by matching relative to the reference image; and respectively represent the average gray values within the subgraph of the reference image and the image area obtained by matching; N is the total number of pixels in the subgraph.

[0038] Mean and The calculation formulas are:

[0039]

[0040] The calculation formula of the sum of squared differences (SSD) is:

[0041]

[0042] Where SSD(u, v) represents the correlation of the sub - graphs before and after the application of the alternating load.

[0043] The zero - mean normalized cross - correlation function (ZNCC) is used to measure the similarity between the reference image and the deformed image. By calculating the correlation coefficient of image patches, ZNCC can effectively track the displacement and deformation of image feature points. Before performing DIC calculations, pre - process the reference image and the deformed image (the image obtained after applying the alternating load), such as denoising and enhancing contrast, to improve the accuracy of the correlation calculation. For the selected region of interest (ROI), calculate the ZNCC value between it and the corresponding region in the deformed image. According to the ZNCC value, the best - matching position can be determined, and thus the displacement and strain of the image feature points can be calculated.

[0044] The strain - field data includes the normal strain ∈ xx in the x - direction, the normal strain ∈ yy in the y - direction, and the shear strain γ xy , and their calculation expressions are respectively:

[0045]

[0046] where u represents the displacement component in the x - direction, v represents the displacement component in the y - direction; U represents the displacement field.

[0047] The gradient change near the crack position can be detected by calculating the second - order derivative of the displacement gradient or using a difference operator. In this embodiment, in step S3, the specific method for preliminarily locating the potential crack region by capturing the gradient mutation feature in the displacement field and the Gaussian distribution law presented by the virtual strain in the strain field is as follows:

[0048] S3 - 1. Capture the gradient mutation feature in the displacement field through the Laplace operator, and its expression is:

[0049]

[0050] where both Δu and Δv are the gradient mutation features in the displacement field. When Δu or Δv increases significantly at a certain position, it may indicate that there is a crack at that position. Therefore, when the increase amount of Δu or Δv at a certain position is greater than the threshold, that position is taken as the potential crack position;

[0051] As Figure 2 shown, according to the significant gradient transformation of the displacement field at the crack and the Gaussian distribution of the virtual strain generated by the DIC calculation window in the strain field and the strain concentration phenomenon that occurs in the surrounding area of the crack, which is manifested as a sharp change or an abnormally high value of the strain value.

[0052] S3 - 2. Calculate the strain feature based on the strain - field data, and its expression is:

[0053]

[0054] where ∈ total is the strain characteristic;

[0055] When ∈ total is greater than the strain threshold (the threshold is determined by the properties of the material, the tensile strain threshold of concrete is 0.001, and the compressive strain threshold is 0.002), it is determined that there may be cracks, and the crack position is identified by fitting a Gaussian function based on the dic code. Its expression is:

[0056]

[0057] where ∈ virtual (x, y) represents the fitted Gaussian function at the position (x, y); A is the maximum value of the strain in the strain field, reflecting the strain intensity at the crack; exp(.) represents the exponential with the natural constant e as the base; (x0, y0) is the mean parameter of the Gaussian distribution, that is, the crack center position; σ x and σ y are the distribution widths of the strain field, reflecting the extent of the crack influence area expansion in the strain field;

[0058] S3 - 3. Take the area where the crack center position (x0, y0) is located as the crack potential area.

[0059] Combined with the characteristics of alternating load loading, the specific method for carrying out multi - cycle repeated measurement in the crack potential area is as follows:

[0060] Obtain the relevant data of vehicles on the road. According to the mass and type of vehicles, establish the load distribution model of different types of vehicles at different time periods. Combine the structural characteristics of the bridge, calculate the load exerted on the bridge structure by the vehicle load, and then calculate the alternating stress caused by the vehicle alternating load. Carry out multi - cycle repeated measurement in the crack potential area based on the alternating stress caused by the vehicle alternating load.

[0061] The dynamic state - space model includes state variables, state - transition equations, observation equations, state - prediction equations, and covariance - prediction equations; among them:

[0062] The expression of the state variable is:

[0063]

[0064] where X k is the state variable of the crack; x k represents the crack position at time k; v k represents the crack position change rate at time k;

[0065] The expression of the state - transition equation is:

[0066] x k = F k-1 ·x k-1 + B k-1 ·u k-1 + w k-1

[0067] where F k-1 represents the state transition matrix from time k - 1 to time k; x k-1 represents the crack position at time k - 1; B k-1 is the control input matrix at time k - 1; u k-1 is the control input vector at time k - 1; w k-1 is the transition noise at time k - 1;

[0068] The expression of the observation equation is:

[0069] z k = H k ·x k + o k

[0070] where z k is the crack observation value at time k; H k is the crack observation matrix, H k = [1 0]; o k is the crack observation noise at time k;

[0071] The expression of the state prediction equation is:

[0072]

[0073] where is the predicted crack state at time k; is the predicted crack state at time k - 1;

[0074] The expression of the covariance prediction equation is:

[0075]

[0076] where is the covariance of the predicted crack state at time k; P k-1 is the covariance of the predicted crack state at time k - 1; is the transpose of F k-1 ; Q k-1 is the process noise covariance matrix at time k - 1.

[0077] In step S5, the potentially identified crack locations are input into the state vector of the dynamic state space model, and multi-source heterogeneous data are fused through recursion to suppress measurement noise and single-detection deviation. The specific method for obtaining the final crack locations includes the following steps:

[0078] S5-1. Construct the Kalman gain equation based on the observation equation and the covariance prediction equation, and its expression is:

[0079]

[0080] where K k is the Kalman gain at time k; is the transpose of H k ; R k is the observation noise covariance at time ;

[0081] S5-2. Predict the crack state based on the state prediction equation;

[0082] S5-3. Update the predicted crack state based on the Kalman gain and the crack observation value, and its expression is:

[0083]

[0084] where is the crack state updated at time k; represents the predicted measurement value;

[0085] S5-4. Update the covariance, and its expression is:

[0086]

[0087] where P k is the covariance of the crack state at time k; I is the identity matrix;

[0088] S5-5. Determine whether all measurement data have been processed. If so, use the latest updated crack state as the final crack identification result to obtain the final crack location; otherwise, continue to update the predicted crack state and covariance.

[0089] In this embodiment, the state transition matrix and the control input matrix are expressed as:

[0090]

[0091] where Δt is the sampling time interval.

[0092] In one embodiment of the present invention, since the method judges the alternating load on the bridge by statistical analysis of traffic flow, first, relevant data of vehicles on the road are obtained. According to the mass and type of vehicles, the vehicles are divided into different categories, such as extra-heavy vehicles, heavy vehicles and light vehicles. According to the statistical analysis results, a vehicle load spectrum is established, that is, a load distribution model of different types of vehicles in different time periods. Using the load spectrum and combining with the structural characteristics of the road or bridge, the load applied by the vehicle load to the structure is calculated. Finally, the alternating stress caused by the vehicle alternating load is calculated, and the cracks are identified multiple times based on the above method. During the long-term crack identification process, the device collected a large amount of crack data. These data are systematically collected and recorded for subsequent analysis and research.

[0093] In summary, the present invention takes the reinforced concrete bridge in traffic as the object, and through the alternating stress generated by the vehicle dynamic load generated by the spontaneous traffic flow, promotes the repeated opening and closing behavior of potential cracks. An unmanned aerial vehicle system with a unified coordinate system is used to collect time-series images of the bridge area to be measured, and the vehicle information of the traffic flow is recorded synchronously; subsequently, the displacement field of the time-series images after noise reduction is reconstructed by the DIC algorithm, and the strain field distribution is obtained based on the spatio-temporal correlation analysis. By capturing the gradient mutation characteristics in the displacement field and the Gaussian distribution law presented by the virtual strain in the strain field, the potential crack area is initially located. Further combining the loading characteristics of the alternating load, multi-cycle repeated measurements are carried out to obtain multiple sets of displacement-strain data sets, so as to initially determine the crack position. To improve the detection robustness, the Kalman filter algorithm is introduced to construct a dynamic state space model, and parameters such as the crack position coordinates are defined as state vectors. By recursively fusing multi-source heterogeneous data, measurement noise and single detection deviation are suppressed, and finally the millimeter-level crack positioning accuracy is achieved. This method uses traffic load as a natural excitation source, and through non-contact measurement and intelligent data fusion, significantly improves the ability to identify micro-cracks, providing an efficient and low-cost innovative solution for bridge structural health monitoring.

Claims

1. A method for identifying bridge cracks, characterized in that, It includes the following steps: Apply an alternating load through a moving vehicle to collect time-series images of the bridge area to be measured; Reconstruct the displacement field of the time-series images through the DIC algorithm, and obtain the strain field distribution based on spatio-temporal correlation analysis; Preliminarily locate the potential crack positions by capturing the gradient mutation features in the displacement field and the Gaussian distribution law presented by the virtual strain in the strain field; Combined with the loading characteristics of the alternating load, conduct multi-cycle repeated measurements in the potential crack areas, and then preliminarily locate multiple groups of potential crack positions; Construct a dynamic state space model based on the Kalman filter algorithm, substitute the preliminarily determined potential crack positions into the state vector of the dynamic state space model, and suppress measurement noise and single-detection deviation by recursively fusing multi-source heterogeneous data to obtain the final crack positions.

2. The bridge crack identification method according to claim 1, wherein, The specific method for collecting time-series images of the bridge area to be measured is as follows: Match speckles for the bridge area to be measured, and use a drone system with a unified coordinate system to collect time-series images.

3. The bridge crack identification method according to claim 1, characterized in that, The specific method for reconstructing the displacement field of the time-series images through the DIC algorithm and obtaining the strain field distribution based on spatio-temporal correlation analysis includes the following steps: Perform denoising processing on the time-series images through median filtering to obtain denoised time-series images; Convert the denoised time-series images into a grayscale image sequence through DIC software, use the pixel subset containing the target measurement points in the first-frame grayscale image as the reference image; select the region of interest in the reference image through DIC software; Discretize the region of interest into several independent sub-images in DIC software, and search for the positions of the sub-images in the subsequent images respectively, that is, complete the displacement field reconstruction and strain field data calculation by matching the sub-images before and after applying the alternating load.

4. The bridge crack identification method according to claim 3, characterized in that The specific method for completing the displacement field reconstruction by matching the sub-images before and after applying the alternating load includes the following steps: Aim at the optimal correlation of the sub-images before and after applying the alternating load, obtain the displacement amount of a single sub-image before and after applying the alternating load, combine the displacement amounts of all sub-images before and after applying the alternating load, and obtain the displacement vector of the reference image after applying the alternating load, that is, obtain the displacement field.

5. The bridge crack identification method according to claim 4, characterized in that, The strain field data includes the normal strain ∈ in the x direction xx , the normal strain ∈ in the y direction yy and the shear strain γ xy , and the calculation expressions are respectively as follows: Where u represents the displacement component in the x direction, v represents the displacement component in the y direction; U represents the displacement field.

6. The bridge crack identification method according to claim 5, wherein The specific method for preliminarily locating the potential crack areas by capturing the gradient mutation features in the displacement field and the Gaussian distribution law presented by the virtual strain in the strain field is as follows: Capture the gradient mutation features in the displacement field through the Laplace operator, and its expression is: Where Δu and Δv are both gradient mutation features in the displacement field. When the increase amount of Δu or Δv at a certain position is greater than the threshold, then take this position as the potential crack position; Calculate the strain characteristics based on the strain field data, and its expression is: where ∈ total is the strain characteristic; When ∈ total When it is greater than the strain threshold, it is determined that there may be cracks and the crack position is identified by fitting a Gaussian function. The expression is as follows: where ∈ virtual (x, y) represents the fitted Gaussian function at the position (x, y); A is the maximum value of the strain in the strain field; exp(.) represents the exponential with the natural constant e as the base; (x0, y0) is the mean parameter of the Gaussian distribution, that is, the crack center position; σ x and σ y are the distribution widths of the strain field; Take the area where the crack center position (x0, y0) is located as the potential crack area.

7. The bridge crack identification method according to claim 1, wherein The specific method for conducting multi-cycle repeated measurements in the potential crack areas combined with the loading characteristics of the alternating load is as follows: Obtain the relevant data of the vehicles on the road. According to the mass and type of the vehicles, establish a load distribution model of different types of vehicles at different time periods, combine the structural characteristics of the bridge, calculate the load exerted on the bridge structure by the vehicle load, and then calculate the alternating stress caused by the vehicle alternating load; Perform multi - cycle repeated measurements in the potential crack area based on the alternating stress caused by vehicle alternating loads.

8. The bridge crack identification method according to claim 1, characterized in that The dynamic state - space model includes state variables, state - transition equations, observation equations, state - prediction equations, and covariance - prediction equations; where: The expression of the state variables is: where X k is the state variable of the crack; x k represents the crack position at time k; v k represents the crack position change rate at time k; The expression of the state - transition equation is: x k = F k-1 · x k-1 + B k-1 · u k-1 + w k-1 Among them, F k-1 represents the state transition matrix from time k - 1 to time k; x k-1 represents the crack position at time k - 1; B k-1 is the control input matrix at time k - 1; u k-1 is the control input vector at time k - 1; w k-1 is the transition noise at time k - 1; The expression of the observation equation is: z k = H k · x k + o k where z k is the crack observation value at time k; H k is the crack observation matrix, and H k = [1 0]; o k is the crack observation noise at time k; The expression of the state - prediction equation is: wherein is the crack state at time k obtained by prediction; is the crack state at time k-1 obtained by prediction; The expression of the covariance - prediction equation is: where is the covariance of the predicted crack state at time k; P k-1 is the covariance of the predicted crack state at time k-1; is the transpose of F k-1 ; Q k-1 is the process noise covariance matrix at time k-1.

9. The bridge crack identification method according to claim 8, wherein The specific method of bringing the preliminarily determined potential crack location into the state vector of the dynamic state - space model, suppressing measurement noise and single - detection deviation through recursive fusion of multi - source heterogeneous data, and obtaining the final crack location includes the following steps: Construct the Kalman gain equation based on the observation equation and the covariance - prediction equation, and its expression is: Where K k is the Kalman gain at time k; is the transpose of H k ; R k is the observation noise covariance at time; Perform crack - state prediction based on the state - prediction equation; Update the predicted crack state based on the Kalman gain and the crack observation value, and its expression is: Among them is the crack state updated at time k; represents the predicted measured value; Update the covariance, and its expression is: where P k is the covariance of the crack state at time k; I is the identity matrix; Determine whether all measurement data have been processed. If so, take the latest updated crack state as the final crack - identification result to obtain the final crack location; otherwise, continue to update the predicted crack state and covariance.

10. The bridge crack identification method according to claim 8, wherein, The state - transition matrix and the control - input matrix are expressed as: Where Δt is the sampling time interval.

Images