Quantitative prediction method of residual bearing capacity of concrete components based on image recognition

Abstract

The present invention discloses a quantitative prediction method for the residual bearing capacity of concrete components based on images, comprising the following steps: first, taking a photo of a building structural component as input, automatically identifying the damage condition through an existing structural residual performance prediction model, and then outputting three residual performance indicators of the existing structural component, namely, damage index, bearing capacity degradation and stiffness degradation, wherein the calculation of the three indicators adopts a calculation formula with clear physical meaning, and its values ​​are calibrated by experimental data and clearly defined. By directly outputting the corresponding performance based on the damaged photo of the building structural component, the method has high prediction accuracy and avoids the subjective influence of conventional human judgment.

Description

Technical Field

[0001] The invention relates to the field of residual performance prediction of existing structures, and in particular to a quantitative prediction method for residual bearing performance of concrete components based on image recognition. Background Art

[0002] The performance of building structures changes with the use time and external conditions. In order to meet the performance-based requirements in the current seismic code, it is particularly necessary to reasonably estimate the residual bearing capacity of the structure. Accurately estimating the residual bearing capacity of building structures can reduce economic costs while still meeting the requirements for structural safety. Therefore, it is necessary to develop a reliable and fast method to quantitatively predict the residual bearing capacity of building structures to achieve accurate evaluation and timely maintenance of the structure. During the reciprocating action of earthquakes, buildings or components will produce corresponding responses, such as cracking, deformation or damage, in order to consume the energy released by the earthquake. The response of each component of the building structure to the earthquake will cause its strength and stiffness to decrease, the structure will be damaged to varying degrees, cracks, fractures and damage will occur, and even the entire structure will collapse. Correspondingly, the seismic performance indicators of the structure mainly include period (stiffness), bearing capacity, ductility and accumulated hysteretic energy dissipation. The structural period determines the degree of response under elastic earthquake action. When the bearing capacity of the structure is less than the earthquake action, the structure will enter the inelastic stage and damage will occur. The size of ductility and accumulated hysteretic energy dissipation reflects the degree of damage to the structure. Ductility and cumulative hysteretic energy dissipation are the supplements to the insufficient elastic seismic resistance of the structure, but they also mean damage to the structure (component). Therefore, the present invention uses three indicators, namely stiffness, bearing capacity and damage, as the main reference for the residual performance of existing structures. The application scenarios of this method include but are not limited to the evaluation of old buildings, the design of seismic reinforcement schemes, the monitoring and operation management of building safety, etc., which provide an important basis for ensuring the reliability of the structure and extending its service life. At present, the image recognition evaluation of the residual bearing capacity of the structure mainly focuses on the qualitative level while relying on the current specifications. The quantitative evaluation of the residual bearing capacity of the structure after damage is not mature and perfect. For example, Japan's JBDPA Guidelines divide reinforced concrete components into five damage states (DS Ⅰ-Ⅴ) according to the appearance of visible damage, while the US FEMA 306 divides structural damage into five damage severity levels: insignificant, slight, moderate, severe and extreme damage. However, this hierarchical evaluation method means that there is a certain degree of rough estimation, so the quantitative evaluation of the residual bearing capacity of the structure is the trend of future development. Through quantitative assessment, we can more accurately understand the bearing capacity of the structure after damage and make more accurate predictions on the safety and reliability of the structure. This quantitative assessment method is expected to provide a more scientific and effective basis for the maintenance of the structure and the design of seismic reinforcement schemes, thereby providing strong support for ensuring the safety of the structure and extending its service life. In addition, there are many types of engineering disasters. Once they occur, they affect a wide range of areas and cause serious losses of life and property. Therefore, it is extremely important to quickly assess the post-disaster situation.However, current safety assessment procedures are time-consuming and labor-intensive, which limits the timely and accurate assessment of post-disaster conditions; such assessments also involve risks because inspectors need to observe post-disaster buildings at close range; and current assessment results are highly dependent on the expertise and experience of engineers or inspectors and may be affected by human bias. However, with the emergence and increasing maturity of computer vision and machine learning, there is a huge potential benefit in using these technologies to allow computers to perform labor-intensive and repetitive tasks. Computer vision can automatically parse and understand image data, while machine learning achieves autonomous learning and prediction capabilities through pattern recognition and data-driven methods. Introducing these technologies into post-disaster safety assessments can greatly improve the efficiency and accuracy of assessments, reduce labor costs, and reduce subjective factors in the assessment process. Therefore, it is necessary to improve the current safety assessment procedures and use the mature technologies of computer vision and machine learning to achieve rapid quantitative assessments of post-disaster conditions. This will provide a more reliable and efficient solution for post-disaster safety assessments and provide strong support for protecting people's lives and reducing property losses. Summary of the invention

[0003] The technical problem to be solved by the present invention is: in view of the low efficiency of the residual performance evaluation of existing structures in the prior art, a quantitative prediction method for the residual bearing performance of concrete components based on image recognition is provided to better improve the efficiency and accuracy of the evaluation, reduce labor costs, and reduce subjective factors in the evaluation process. In order to solve the above technical problems, the present invention provides the following technical solutions: A quantitative prediction method for the residual bearing performance of concrete components based on image recognition, comprising the following steps:

[0004] S1. Based on the building construction picture, extract the hysteresis curve data of the existing structure, perform preprocessing, and obtain the zero-point segmented hysteresis curve loop segment;

[0005] S2. Use stiffness degradation, bearing capacity degradation and damage indicators as references for residual bearing performance, extract stiffness, bearing capacity and damage at the zero point of each cycle based on hysteresis curve data, and use interpolation to determine the stiffness, bearing capacity and damage corresponding to each point;

[0006] S3. Construct a neural network for predicting residual bearing capacity, including a feature extraction network and a regression prediction network. Specifically, the feature extraction network is constructed by taking the damaged image of the building structure as input, and the texture and local features of the image as output. The regression prediction network is constructed by taking the texture and local features of the image as input, and the stiffness, bearing capacity and damage at the zero point of each cycle as output.

[0007] S4, using the building structure damage picture as input and the stiffness, bearing capacity and damage index corresponding to the picture as output, training the residual bearing capacity quantitative prediction neural network to obtain the residual bearing capacity prediction model;

[0008] S5. Collect building construction pictures and apply the residual bearing capacity prediction model to predict the damage state of the concrete structure;

[0009] Step S1 is specifically as follows: using zero point to divide the hysteresis curve into segments of hysteresis curve circles divided by zero point for the building construction picture, and selecting the maximum load value point in each segment as the flip point for secondary division and calculation;

[0010] In step S2, the stiffness at the zero point of each cycle is extracted, specifically: by drawing the curve, locating the flip point and the zero point of the hysteresis curve, according to the formula: , calculate the stiffness at the zero point of each cycle, and perform linear interpolation to obtain the stiffness index of each point, where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n, and represents the corresponding peak displacement;

[0011] In step S2, the bearing capacity at the zero point of each cycle is extracted, specifically: by drawing the curve, locating the flip point and the zero point of the hysteresis curve, according to the formula: , calculate the bearing capacity at the zero point of each period, and perform linear interpolation to obtain the bearing capacity of each point; where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n;

[0012] In step S2, the damage of each cycle is extracted, specifically: by drawing the curve, locating the flip point and zero point of the hysteresis curve, according to the formula: , calculate the damage at the zero point of each cycle, and perform cubic spline interpolation to obtain the damage index of each point; where D is the damage index; It is the maximum deformation of a component or structure under earthquake action; It is the limit deformation of a member or structure under monotonic load; is the calculated value of yield load; It is the accumulated hysteresis energy consumption; is the cyclic load influence coefficient.

[0013] Furthermore, the networks used in the feature extraction network in the aforementioned step S3 include LeNet, AlexNet, VGGNet, GoogLeNet and ResNet architectures.

[0014] Furthermore, in the aforementioned step S3, the regression prediction network adopts multi-task learning to improve the generalization performance of all tasks by utilizing the effective common information contained in related tasks.

[0015] Furthermore, in the aforementioned step S4, when training the residual carrying capacity prediction neural network, the back propagation algorithm and the stochastic gradient descent SGD optimizer are used to train the model, and an early stopping strategy is adopted to decide to stop training according to the change of the loss function on the validation set; by using the pre-training parameters, frozen training is performed; the learning rate is dynamically updated, and the cosine function is used to decrease and match the continuously optimized training process.

[0016] Furthermore, the aforementioned step S4 also includes using mean square error MSE, root mean square error RMSE, mean absolute error MAE, R 2 The coefficient of determination index is used to evaluate the test images obtained by the residual bearing capacity prediction model.

[0017] Compared with the prior art, the beneficial technical effects of the above technical solutions of the present invention are as follows: A quantitative evaluation of the residual bearing performance state of concrete structural components is achieved, and a comprehensive evaluation system is defined by introducing main parameters such as stiffness degradation, strength degradation and damage index. A comprehensive end-to-end prediction framework is proposed, which can realize rapid prediction from structural damage identification to quantitative indicators of residual bearing performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIG1 is a schematic diagram of the general process of deep learning in the present invention.

[0019] FIG. 2 is a hysteresis curve diagram corresponding to an example in the present invention.

[0020] FIG. 3 is a hysteresis loop diagram analyzed by the hysteresis curve analysis code in the present invention.

[0021] FIG. 4 is a graph showing the bearing capacity degradation curve corresponding to an example of the present invention.

[0022] FIG. 5 is a stiffness degradation curve corresponding to an example in the present invention.

[0023] FIG6 is a schematic diagram showing the principle of the energy equivalence method of the present invention.

[0024] FIG. 7 is a skeleton curve diagram corresponding to the example extracted in the present invention.

[0025] FIG8 is a damage index curve diagram corresponding to an example in the present invention.

[0026] FIG. 9 is a diagram showing a network structure in the present invention.

[0027] FIG. 10 is a diagram showing the bearing capacity degradation prediction result corresponding to the example in the present invention.

[0028] FIG. 11 is a diagram showing the stiffness degradation prediction results corresponding to the example in the present invention.

[0029] FIG. 12 is a diagram showing the damage index prediction results corresponding to the example in the present invention.

[0030] FIG. 13 is a schematic diagram of the prediction process in the present invention.

[0031] FIG14 is a prediction flow chart for a single sample example in the present invention. DETAILED DESCRIPTION

[0032] In order to better understand the technical content of the present invention, specific embodiments are cited and explained in conjunction with the accompanying drawings as follows. In the present invention, various aspects of the present invention are described with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the present invention are not limited to those described in the accompanying drawings. It should be understood that the present invention is implemented by any of the various concepts and embodiments introduced above, as well as the concepts and implementations described in detail below, because the concepts and embodiments disclosed in the present invention are not limited to any implementation. In addition, some aspects disclosed in the present invention can be used alone or in any appropriate combination with other aspects disclosed in the present invention. With reference to Figure 1, the present invention discloses an image-based quantitative prediction method for the residual bearing capacity of concrete components. The present invention selects a set of quasi-static shear test data of gypsum block walls from the open database platform of the Laboratory of Earthquake Engineering and Structural Dynamics of the Federal Institute of Technology in Lausanne as a sample to evaluate the generalization performance and robustness of the present invention. The present invention comprises the following steps:

[0033] S1. Based on the building construction picture, the hysteresis curve data of the existing structure is extracted and preprocessed to obtain the zero-point segmented hysteresis curve loop fragment; when processing the sample, the data is first cleaned and processed, including noise removal, standardization and data partitioning. Through the hysteresis curve analysis code, the hysteresis curve diagram corresponding to the selected sample data is generated, the hysteresis curve is analyzed, and the hysteresis loop schematic diagram is generated, as shown in Figure 3;

[0034] S2. Use stiffness degradation, bearing capacity degradation and damage index as references for residual bearing performance, and extract stiffness, bearing capacity and damage index at the zero point of each cycle according to the hysteresis curve data; Calculate stiffness degradation: By analyzing the hysteresis curve, first calculate the zero point on each circle, so as to achieve a division of the hysteresis curve. Then, according to the hysteresis curve segments divided by these zero points, see Figure 3, find the hysteresis loop flip point, see Figure 2. According to the formula , calculate the stiffness at the zero point of each period, and perform linear interpolation to obtain the stiffness index of each point, where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n, and Represents the corresponding peak displacement. Finally, the stiffness degradation values ​​of all discrete points are obtained by linear interpolation, and combined with the existing image data in the database, after the data screening process, the stiffness degradation curve is finally obtained, as shown in Figure 5. Calculate the bearing capacity degradation: By analyzing the hysteresis curve, first calculate the zero point on each circle to achieve a division of the hysteresis curve. Then, according to the hysteresis curve segments divided by these zero points, see Figure 3, find the hysteresis loop flip point, see Figure 2. According to , calculate the bearing capacity at the zero point of each period, and perform linear interpolation to obtain the bearing capacity at each point; where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n. And assign it to the corresponding zero point. Finally, the bearing capacity degradation value of all discrete points is obtained by linear interpolation, and combined with the existing image data in the database, after the data screening process, Figure 4 is finally obtained.

[0035] Calculate damage index: According to the energy equivalent method, see Figure 6, calculate the equivalent yield point, obtain the equivalent yield point, and generate the corresponding equivalent yield point data, as shown in Figure 7; use the trapezoidal approximation integral method to calculate the energy consumption and obtain the cumulative energy consumption of each point; use the Park-ang damage calculation model formula , calculate the bearing capacity at the zero point of each cycle, and perform linear interpolation to obtain the bearing capacity of each point. Calculate the damage index and assign it to the corresponding zero point. Finally, the bearing capacity degradation value of all discrete points is obtained by cubic spline interpolation, and combined with the existing image data in the database, after the data screening process, Figure 8 is finally obtained.

[0036] S3. Construct a neural network for predicting residual bearing performance, including a feature extraction network and a regression prediction network; Specifically: take the damaged image of the building structure as input, and the texture and local features of the image as output to construct a feature extraction network; take the texture and local features of the image as input, and the stiffness, bearing capacity, and damage corresponding to the image as output to construct a regression prediction network; Network model construction: Use the VGG16 classic network feature extraction layer as the feature extraction part, and use the multi-task learning framework to build a regression prediction network. The overall network architecture is shown in Figure 9. The network structure uses the powerful feature extraction ability of the VGG16 network in the image field and applies it to the task of predicting the residual performance of existing structures through transfer learning. The multi-task learning network can be seen in Figure 13. Through multi-task learning, multiple tasks related to the residual performance of existing structures can be processed simultaneously, and the relevant information between tasks can be fully utilized to improve the generalization performance and training efficiency of the model. By constructing this network, the present invention can effectively extract image features and realize accurate prediction of the residual performance of existing structures. The design and construction of these network structures can make full use of the advantages of pre-training parameters and transfer learning, and provide strong support for the performance and generalization ability of the model.

[0037] S4. Taking the damaged image of the building structure as input and the corresponding stiffness, bearing capacity and damage at the zero point of each cycle as output, the residual bearing capacity prediction neural network is trained to obtain the residual bearing capacity prediction model; the back propagation algorithm and the stochastic gradient descent (SGD) optimizer are used to train the model during the training process. In order to improve the training efficiency, the early stopping strategy is adopted to decide whether to stop the training according to the change of the loss function on the validation set. At the same time, the pre-training parameters are fully utilized by freezing the training. The learning rate is dynamically updated, and the cosine function is used to decrease to match the continuously optimized training process.

[0038] S5. Collect building construction pictures and apply the residual bearing capacity prediction model to predict the damage state of the concrete structure. In terms of model evaluation, the present invention uses multiple indicators to objectively evaluate the performance of the model, including mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE) and R 2Coefficient of determination, etc. Among them, MSE measures the prediction accuracy of the model by calculating the square average of the difference between the predicted value and the true value. RMSE is the square root of MSE, which is used to measure the average difference between the predicted value and the true value. MAE calculates the average of the absolute difference between the predicted value and the true value, providing an overall assessment of the prediction error. 2 It is a statistical indicator used to measure the degree to which the model explains the variability of the dependent variable. Its value ranges from 0 to 1. The closer it is to 1, the better the model fits. By comprehensively considering these evaluation indicators, we can comprehensively evaluate the accuracy and stability of the model in predicting the residual performance of the existing structure, and provide an important basis for further optimizing the model. Model tuning: During the model tuning process, the model is adjusted by modifying hyperparameters, changing the network structure, etc. to improve its performance and generalization ability. It should be noted that hyperparameter adjustments should not be made based on the test set results to avoid overfitting and leakage of experience. Model tuning aims to optimize the training process so that the model can better adapt to the data and improve prediction accuracy, thereby achieving the best performance. Result verification: The mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and R 2 The test images were evaluated by using indicators such as the coefficient of determination. For example, in the stiffness degradation prediction model, the MSE was 0.002, the RMSE was 0.045, the MAE was 0.029, and the R2 was 0.982. Similarly, the bearing capacity degradation and damage indicators can also be evaluated. In addition, the trained model has been applied to new data for prediction. An image was selected from the test set for prediction. The prediction results are shown in Figure 14. The prediction results of the entire test set are shown in Figures 10, 11, and 12. It was observed that all prediction results showed good prediction results, which verified the feasibility of this method.

[0039] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the definition of the claims.

Claims

1. A quantitative prediction method for the residual bearing capacity of concrete components based on image recognition, characterized in that: The steps include: S1. Based on the building construction picture, extract the hysteresis curve data of the existing structure, perform preprocessing, and obtain the zero-point segmented hysteresis curve loop segment; S2. Use stiffness degradation, bearing capacity degradation and damage indicators as references for residual bearing performance, extract stiffness, bearing capacity and damage at the zero point of each cycle based on hysteresis curve data, and use interpolation to determine the stiffness, bearing capacity and damage corresponding to each point; S3. Construct a residual bearing capacity prediction neural network, including a feature extraction network and a regression prediction network; specifically: take the building structure damage picture as input, the texture of the picture, and the local features as output to construct a feature extraction network, take the texture of the picture and the local features as input, and the stiffness, bearing capacity, and damage at the zero point of each cycle as output to construct a regression prediction network; S4, using the building structure damage picture as input and the stiffness, bearing capacity and damage index corresponding to the picture as output, training the residual bearing capacity quantitative prediction neural network to obtain the residual bearing capacity prediction model; S5. Collect building construction pictures and apply the residual bearing capacity prediction model to predict the damage state of the concrete structure; Step S1 is specifically as follows: using zero point to divide the hysteresis curve into segments of hysteresis curve circles divided by zero point for the building construction picture, and selecting the maximum load value point in each segment as the flip point for secondary division and calculation; In step S2, the stiffness at the zero point of each cycle is extracted, specifically: by drawing the curve, locating the flip point and the zero point of the hysteresis curve, according to the formula: , calculate the stiffness at the zero point of each cycle, and perform linear interpolation to obtain the stiffness index of each point, where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n, and represents the corresponding peak displacement; In step S2, the bearing capacity at the zero point of each cycle is extracted, specifically: by drawing the curve, locating the flip point and the zero point of the hysteresis curve, according to the formula: , calculate the bearing capacity at the zero point of each period, and perform linear interpolation to obtain the bearing capacity of each point; where, and Represents the absolute value of the maximum positive and negative load corresponding to the number of loading cycles n; In step S2, the damage of each cycle is extracted, specifically: by drawing the curve, locating the flip point and zero point of the hysteresis curve, according to the formula: , calculate the damage at the zero point of each cycle, and perform cubic spline interpolation to obtain the damage index of each point; where D is the damage index; It is the maximum deformation of a component or structure under earthquake action; It is the limit deformation of a member or structure under monotonic load; is the calculated value of yield load; It is the accumulated hysteresis energy consumption; is the cyclic load influence coefficient.

2. The method for quantitatively predicting the residual bearing capacity of concrete components based on image recognition according to claim 1 is characterized in that: The networks used in the feature extraction network in step S3 include LeNet, AlexNet, VGGNet, GoogLeNet and ResNet architectures.

3. The method for quantitatively predicting the residual bearing capacity of concrete components based on image recognition according to claim 1 is characterized in that: In step S3, the regression prediction network adopts multi-task learning to utilize the effective common information contained in related tasks to improve the generalization performance of all tasks.

4. The method for quantitatively predicting the residual bearing capacity of concrete components based on image recognition according to claim 1 is characterized in that: Step S4 also includes using mean square error MSE, root mean square error RMSE, mean absolute error MAE, R 2 The coefficient of determination index evaluates the test images obtained by the residual bearing capacity prediction model.

Images