Self-compacting concrete characteristic lossless identification method based on image identification technology

Abstract

The invention provides a lossless identification method for self-compacting concrete characteristics based on an image identification technology. The lossless identification method comprises the following steps: determining self-compacting concrete composition and mix proportion working conditions; mixing the self-compacting concrete, and carrying out a slump experiment and a J-ring experiment; preparing concrete test pieces under different working conditions, shooting the cured concrete test pieces, and collecting self-compacting concrete pictures of the concrete test pieces under various working conditions; carrying out a compressive strength experiment and a rapid chloride ion migration experiment by utilizing the concrete test piece; carrying out data enhancement on the collected self-compacting concrete picture of the concrete test piece, carrying out standardization processing on an experiment result, and marking the picture by utilizing the standardized experiment result; determining a judgment index for evaluating the prediction precision of the neural network; and selecting a convolutional neural network, and training to obtain a network model for predicting the characteristics of the self-compacting concrete.

Description

Technical Field

[0001] The present invention belongs to the field of nondestructive testing of concrete quality, and is applied to a laboratory testing process before concrete is poured on site and after concrete mixture is sampled, and specifically relates to a nondestructive identification method of self-compacting concrete characteristics based on convolutional neural networks and image recognition. Background Art

[0002] According to the requirements of the specification, before the concrete is poured on site during the construction process, it is necessary to randomly sample and send it to a professional laboratory for strength testing. This requires making a standard concrete specimen in the laboratory, applying a specific load or exposing it to a specific environment until the specimen is destroyed. Although this method has high accuracy, it is a destructive test, which is not only complicated to operate and has a high detection cost, but also belongs to sampling detection, and cannot fully cover the concrete to be tested. At present, relatively mature concrete non-destructive testing methods include ultrasonic method and rebound hammer method, etc. These methods can realize in-situ and non-destructive testing of concrete, but require the use of special instruments, the instrument cost is high, and it is easily affected by various factors, which makes its measurement accuracy and scope of application limited. Considering the direct correlation between the macroscopic properties (such as strength and durability) of self-compacting concrete and its microstructure, and the feasibility of analyzing the concrete microstructure and its surface morphology characteristics using image recognition technology, the present invention associates the surface morphology characteristics of self-compacting concrete with its raw material composition, proportion information and macroscopic properties through image recognition technology. Convolutional neural network is a representative network structure in deep learning and has been widely used in the field of computer vision. At present, many studies have applied convolutional neural network to image recognition of concrete, but these studies focus on the surface quality detection, crack detection and damage identification of concrete, and do not associate image recognition with the macroscopic properties of concrete such as strength and durability. Summary of the invention

[0003] The purpose of the present invention is to provide an intelligent, non-destructive, convenient and highly accurate non-destructive identification method for self-compacting concrete properties, and to provide a feasible solution for intelligent supervision of concrete quality inspection. The present invention establishes a self-compacting concrete property identification model based on convolutional neural network and image recognition, which can identify the relevant properties of the concrete by inputting the surface image of the self-compacting concrete to be inspected. The specific technical solution is as follows:

[0004] A non-destructive identification method of self-compacting concrete characteristics based on image recognition technology, characterized in that the main steps include:

[0005] (1) Determine the composition and mix conditions of self-compacting concrete, including the volume fraction of coarse aggregate (α), fly ash replacement ratio (β), and water-cement ratio (W / C);

[0006] (2) Mix self-compacting concrete, conduct slump test and J-ring test, obtain the slump (SF) and J-ring expansion of self-compacting concrete under different working conditions, and calculate the gap passability performance index (PA) of self-compacting concrete;

[0007] (3) Prepare concrete specimens under different working conditions, photograph the concrete specimens after curing, and collect self-compacting concrete images of concrete specimens under different working conditions;

[0008] (4) Compressive strength test and rapid chloride ion migration test were carried out on concrete specimens to obtain the compressive strength (f cu ) and the chloride ion diffusion coefficient (D RCM );

[0009] (5) Data enhancement is performed on the collected self-compacting concrete images of concrete specimens, the experimental results are standardized, and the images are "labeled" using the standardized experimental results, and the images and experimental results are associated with each other to form a training set, a validation set, and a test set. The experimental results include seven properties of self-compacting concrete: coarse aggregate volume fraction (α), fly ash replacement rate (β), water-cement ratio (W / C), slump (SF), self-compacting concrete gap passing performance index (PA), compressive strength (f cu ), chloride ion diffusion coefficient (D RCM ) to assess the workability, strength and durability of self-compacting concrete;

[0010] (6) Determine the judgment index for evaluating the prediction accuracy of the neural network;

[0011] (7) A convolutional neural network is selected and trained to obtain a network model for predicting the properties of self-compacting concrete. Its input is a picture of self-compacting concrete, and its output is the seven properties of self-compacting concrete.

[0012] Furthermore, before step (7), it also includes: building different types of convolutional neural networks, exploring the neural network type, image type, image size and resolution, and neural network parameters to determine the operating condition combination with the best prediction accuracy.

[0013] Furthermore, a smart terminal is used to take pictures of concrete, and the pictures of concrete are input into a trained network model for predicting the properties of self-compacting concrete. The seven properties of self-compacting concrete are predicted by this network model.

[0014] Furthermore, the self-compacting concrete image of the concrete specimen collected in step (3) is an image of concrete with a polished upper surface.

[0015] Furthermore, the size of the image of the polished upper surface of concrete is 26.8 cm × 26.8 cm, and the image resolution is 1266 × 1266 pixels.

[0016] Furthermore, the optimal convolutional neural network type was determined to be ResNet18, whose output is the seven characteristics of self-compacting concrete.

[0017] Furthermore, the adopted ResNet18 convolutional neural network contains a residual block structure of 4 deep residual networks, including 22 hidden layers in total, namely 18 convolutional layers, 2 pooling layers and 2 fully connected layers. The RELU function is used as the activation function, and the 7 concrete characteristics predicted by the neural network are passed to the output layer by adding 1 fully connected layer.

[0018] Furthermore, the learning rate of the convolutional neural network is 0.1 and the dropout rate is 0.

[0019] Furthermore, in order to eliminate the differences in dimensions and magnitudes between different experimental results, the following formula is used to standardize the experimental results:

[0020]

[0021] Where: x norm is the standardized variable, x actual is the concrete properties test result, x max is the maximum value of the experimental results, x min is the minimum value of the experimental results.

[0022] Furthermore, the evaluation criteria for the neural network prediction results are based on R 2 and MBE, whose formula form is as follows:

[0023]

[0024] Where: n is the number of data, P i is the experimental value of the ith concrete property, O i is the predicted value of the i-th concrete characteristic model, is the average value of the experimental values ​​of concrete properties in the data set.

[0025] The present invention utilizes the outstanding advantages of convolutional neural networks in the field of image recognition, trains convolutional neural networks to learn the morphological features of the surface image of self-compacting concrete, and establishes a mapping relationship with the characteristics of the concrete, thereby realizing non-destructive testing of the characteristics of self-compacting concrete. It has the following advantages:

[0026] (1) The present invention identifies seven characteristics of self-compacting concrete, including the composition and mix ratio of concrete - coarse aggregate volume fraction α, fly ash replacement rate β, water-cement ratio W / C, the fluidity of concrete - slump SF and gap passing performance index PA, and the strength of concrete - compressive strength f cu , concrete durability - chloride ion diffusion coefficient D RCM , conduct a comprehensive assessment of the quality of self-compacting concrete.

[0027] (2) The present invention uses image recognition to detect the quality of self-compacting concrete. There is no need to conduct destructive experiments on concrete specimens. It only takes a picture of the polished upper surface of the concrete specimen using a smart phone. The trained convolutional neural network can be used to automatically identify the corresponding characteristics of the concrete based on the input picture. No professional equipment is required, the operation is simple, and there is no need to consume too much manpower and material resources. This provides a feasible solution for the intelligent and automated supervision of concrete quality inspection.

[0028] (3) The present invention fully explores the influence of different parameter variables on the prediction accuracy of neural network, determines the image type, image size and image resolution most suitable for self-compacting concrete quality detection, as well as the optimal convolutional neural network type and neural network parameters, and has high prediction accuracy and universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 are the original pictures of self-compacting concrete, (a) is the picture of the polished upper surface, and (b) is the cross-sectional picture;

[0030] Figure 2 These are the images of self-compacting concrete after processing for neural network training, (a) is the image of the polished upper surface, (b) is the image of the cross section;

[0031] Figure 3 This is a comparison chart of the prediction accuracy of different neural network types;

[0032] Figure 4 It is a schematic diagram of the optimal neural network type ResNet18 selected by the present invention;

[0033] Figure 5 This is a comparison chart of prediction accuracy for different image types;

[0034] Figure 6 It is a comparison chart of prediction accuracy of different neural networks. (a) is a comparison chart of prediction accuracy of different neural network discard rates; (b) is a comparison chart of prediction accuracy of different neural network learning rates;

[0035] Figure 7 It is a comparison chart of prediction accuracy of different image sizes. (a) is a comparison chart of prediction accuracy of different image sizes;

[0036] (b) is a comparison chart of prediction accuracy at different image resolutions;

[0037] Figure 8 It is a comparison chart of the predicted values ​​of 7 properties of self-compacting concrete by the neural network trained by the present invention and the experimental true values;

[0038] Fig. 9 It is a graph showing the accuracy changes of the neural network trained by the present invention during the training process of the training set and the test set. DETAILED DESCRIPTION

[0039] First, the basic scheme of the present invention is further described.

[0040] The non-destructive identification method of self-compacting concrete characteristics based on convolutional neural network and image recognition of the present invention comprises the following steps:

[0041] (1) Determine the concrete composition and mix ratio. The main variables include the coarse aggregate volume fraction α, fly ash replacement rate β, and water-cement ratio W / C. The range of the above variables should cover all possible values ​​of the tested concrete to ensure the universality of the trained model.

[0042] (2) Mix concrete according to the set working conditions, and conduct slump test and J-ring test on each group of working condition concrete to obtain the slump SF and gap passing performance index PA of each group of working condition self-compacting concrete, which are used to evaluate the fluidity of self-compacting concrete.

[0043] (3) Six standard cubic specimens and three cylindrical specimens (100 mm in diameter and 50 mm in height) were prepared for each group of concrete working conditions and cured for 28 days. Three standard cubic specimens were taken from each group and the upper surface was polished to remove the upper slurry and fully expose the aggregate. Then, the surface was photographed with a smartphone to obtain the pictures of the polished upper surface of the concrete under each working condition. Then, the standard cubic specimens with polished upper surfaces were cut at a distance of 20 mm from the upper surface to obtain the cross-sectional pictures of the concrete under each working condition.

[0044] (4) Take another three standard cubic specimens for compressive strength tests to obtain the compressive strength f of concrete under each working condition. cu ; Rapid chloride ion migration experiments were carried out using three cylindrical specimens to obtain the chloride ion diffusion coefficient D of concrete under various working conditions RCM .

[0045] (5) In order to eliminate the magnitude and dimension differences among the seven characteristic parameters of concrete and avoid their influence on the accuracy of neural network training, the seven characteristic parameters are standardized so that they are uniformly distributed between 0 and 1. In order to expand the number of pictures, the concrete pictures of different working conditions are cropped and randomly flipped to ensure the number and randomness of pictures used to train the neural network. Finally, each picture is matched with the seven characteristics of concrete, and divided into training set, validation set and test set.

[0046] (6) Determine the evaluation index of the neural network prediction results

[0047] The coefficient of determination R 2 As an evaluation indicator for exploring the optimal parameters, the mean deviation error (MBE) is used to judge the universality of the optimal convolutional neural network. The corresponding formula is as follows:

[0048]

[0049] Where: n is the number of data, P i is the experimental value of the ith concrete property, O i is the predicted value of the i-th concrete characteristic model, is the average value of the experimental values ​​of concrete properties in the data set.

[0050] (7) Explore the optimal training model: First, we built five neural network models, namely GoogleNet, ResNet18, ResNet34, ResNet50, and ShuffleNet, and determined the type of convolutional neural network with the highest training accuracy. On this basis, we selected images of the polished concrete surface and concrete cross-section to determine the optimal image type. Then, we explored the discard rates of 0, 0.1, 0.2, 0.3, and 0.4, the learning rates of 0.001, 0.01, 0.1, 0.3, and 1.0, and the image sizes of 1.7 cm × 1.7 cm, 3. The prediction accuracy of the neural network when the image resolutions are 20×20px, 40×40px, 80×80px, 316×316px and 1266×1266px respectively, the parameter combination with the best prediction accuracy is: ResNet18 neural network is selected, the image type is the image of the polished concrete surface, the image size is 26.8cm×26.8cm, the resolution is 1266×1266px, the neural network learning rate is 0.1, and the discard rate is 0. The determination coefficient R corresponding to the optimal parameter combination 2 It can reach 0.9987, which has a very high prediction accuracy.

[0051] The method of the present invention is further described below in conjunction with specific embodiments:

[0052] Example

[0053] The present invention is a non-destructive identification method of self-compacting concrete characteristics based on convolutional neural network and image recognition, which mainly includes the following steps:

[0054] S1. Determine the concrete composition and mix ratio;

[0055] The present invention determines the concrete composition and mix ratio with the coarse aggregate volume fraction α, fly ash replacement rate β, and water-cement ratio W / C as the main variables, and sets 11 groups of working conditions as shown in Table 1. The components of self-compacting concrete include cement, water, gravel (coarse aggregate), sand, fly ash, and polycarboxylate water reducer. Fly ash is used to replace cement to enhance the fluidity and resistance to chloride ion penetration of concrete; adding a water reducer can reduce the plastic viscosity of concrete, improve fluidity while ensuring its stability and anti-segregation, so as to meet the working requirements of self-compacting concrete.

[0056] Table 1 Self-compacting concrete mix proportion

[0057]

[0058] S2. Carry out slump test and J-ring test on fresh concrete;

[0059] Self-compacting property is an important index to measure the working performance of self-compacting concrete, which mainly includes filling property and passability. It is tested by slump expansion test and J-ring expansion test respectively, and is expressed as two indicators: slump expansion (SF) and the difference between slump expansion and J-ring expansion of self-compacting concrete (PA). The experimental results of the above 11 working conditions are shown in Table 2.

[0060] Table 2 Experimental results of self-compacting concrete

[0061]

[0062] S3. Prepare and cure self-compacting concrete specimens, and obtain polished upper surface images and cross-sectional images of the concrete;

[0063] In order to complete image collection and subsequent experiments, 6 standard cube specimens and 3 cylindrical specimens (100 mm in diameter and 50 mm in height) need to be prepared for each set of working conditions and cured for 28 days in accordance with the specifications.

[0064] Three of the six standard cubic specimens were taken out for image collection, and the remaining three were used for subsequent compressive strength tests. First, the upper surfaces of the three cubic specimens were polished to remove the slurry on the upper surfaces and fully expose the aggregate. Then, a smartphone was used to take a photo of the upper surface of each cube to obtain a concrete image of the polished upper surface. Then, the cubic specimen was transversely cut at a distance of 20 mm from the upper surface to obtain a cross-sectional image of the concrete. The two types of concrete images are shown in Figure 1. Figure 1 shown.

[0065] S4. Conduct compressive strength test and rapid chloride ion migration test;

[0066] The remaining three standard cubic specimens were used for compressive strength tests, and the average value of the experimental results of the three specimens was taken as the representative value of the working condition to obtain the compressive strength f of the concrete under each working condition. cu ; Rapid chloride ion migration experiment was carried out using three cylindrical concrete specimens, and the average value of the experimental results of the three specimens was taken as the representative value of the working condition to obtain the chloride ion permeability coefficient D of the concrete under each working condition RCM , the experimental results are shown in Table 2.

[0067] S5, standardize the experimental results, crop and randomly flip the images, associate the standardized concrete properties with the images, and divide the data set;

[0068] In order to unify the distribution range of the seven characteristics of concrete, eliminate their magnitude differences and dimension differences, and avoid their impact on neural network training, the following formula is used to standardize the seven characteristics so that their distribution range is between 0 and 1.

[0069]

[0070] Where: x norm is the standardized variable, x actual is the concrete properties test result, x max is the maximum value of the experimental results, x min is the minimum value of the experimental results.

[0071] In order to ensure the number of images used to train the neural network and enable the neural network to fully learn the surface morphology characteristics of concrete under different working conditions, the collected concrete images are cropped and randomly flipped. Figure 2 shown.

[0072] Each image was "labeled" using the seven properties of concrete, so that the image of self-compacting concrete was associated with its seven properties. Finally, the training set, validation set, and test set were randomly allocated in a 4:1:1 ratio to train the convolutional neural network, evaluate the best hyperparameters, and evaluate the accuracy and universality of the optimal neural network after training.

[0073] S6. Determine the evaluation index of the neural network prediction result;

[0074] The coefficient of determination R 2 As an evaluation indicator for exploring the optimal parameters, the mean bias error (MBE) is used to judge the accuracy of the optimal convolutional neural network. The corresponding formula is as follows:

[0075]

[0076] Where: n is the number of data, P i is the experimental value of the ith concrete property, O i is the predicted value of the i-th concrete characteristic model, is the average value of the experimental values ​​of concrete properties in the data set.

[0077] S7, explore the optimal training model;

[0078] The effects of convolutional neural network type, image type, discard rate, learning rate, image size and image resolution on the training accuracy of the neural network are explored successively, and the corresponding working condition settings are shown in Table 3.

[0079] (1) To explore the effect of convolutional neural network type on prediction accuracy, five neural network models were built: GoogleNet, ResNet18, ResNet34, ResNet50, and ShuffleNet. Neural network training was performed on a workstation computer with two 3.00 GHz CPUs and 131 GB of memory. These models were trained and their training time, memory usage, and prediction accuracy were obtained, as shown in Tables 4 and 5. Figure 3 As shown in Table 4 and Figure 3 It can be seen that ResNet18 has the best effect, with an average R 2 is 0.9904, which can accurately predict the values ​​of all 7 characteristics of concrete. In addition, ResNet18 has the shortest training time, accounting for about 80% of the calculation time of the lightweight model ShuffleNet. With the increase of network depth, the calculation time and memory occupied by ResNet18, ResNet34 and ResNet50 increase significantly in turn, but the recognition performance does not improve further. This is because with the increase of network depth, gradient vanishing problems or overfitting problems will occur. GoogleNet performs well in prediction accuracy and calculation time, second only to ResNet18, and can be used as an alternative model. Due to the compression and optimization of the model structure, ShuffleNet's parameters and memory usage are relatively reduced, but its prediction accuracy is not as good as other models, and it does not show obvious advantages in saving calculation time. Therefore, the ResNet18 model is selected as the optimal neural network model. Its network structure is as follows Figure 4 The ResNet18 model constructed by the present invention is based on the built-in ResNet18 convolutional neural network in the open source PyTorch library, and includes a residual block structure of 4 deep residual networks, which includes a total of 22 hidden layers, namely 18 convolutional layers, 2 pooling layers and 2 fully connected layers. The RELU function is used as the activation function, and the 7 concrete properties predicted by the neural network are passed to the output layer by adding a fully connected layer.

[0080] Table 3 Parameter settings in the optimal training model exploration

[0081]

[0082]

[0083] Table 4 Training time and memory usage of different neural network types

[0084]

[0085] (2) To explore the effect of image type on prediction accuracy, we selected images of the polished concrete upper surface and cross-section images. The prediction accuracy is as follows: Figure 5 The image recognition accuracy of the self-compacting concrete characteristics using both types of images is high, among which the accuracy of the concrete image with the polished upper surface is the highest, R 2 It can reach 0.9975. This is because the three-dimensional irregular shape and random distribution of concrete aggregates make the image information of the cross section less evenly distributed than that of the polished surface. The upper surface of the concrete is polished, and the aggregates are more clearly exposed, which makes the details of the concrete image on the polished upper surface better reflect the morphological characteristics of the self-compacting concrete surface and more suitable for feature extraction.

[0086] (3) Explore the impact of neural network parameters (discard rate, learning rate) on prediction accuracy. The discard rate refers to the ratio of randomly discarding neurons from the neural network. Randomly discarding neurons not only avoids the problem that convolutional neural networks rely on training data and are difficult to adapt to other data, but also reduces network complexity and achieves structural optimization. Figure 6 (a) As can be seen, as the dropout rate increases, the prediction accuracy of the neural network decreases. This is because the neural network has adopted the RELU activation function, which can learn quickly and obtain accurate results with only a small amount of calculation. The RELU activation function uses a more efficient gradient descent backpropagation algorithm to minimize overfitting problems, so there is no need to rely on random dropout neurons to achieve overfitting.

[0087] An appropriate learning rate can ensure that the objective function converges to a good global minimum in a short time. Figure 6(b) It can be seen that when the learning rate is 0.1, the prediction accuracy of the neural network is the highest. 2 It can reach 0.9987. When the learning rate is small, the parameter update is slow, the convergence speed is slow, and it is impossible to converge to the optimal solution within the same number of iterations; when the learning rate is too large, the step size of each parameter update is too large, which may cause the model to oscillate near the optimal solution or even fail to converge, affecting the training accuracy.

[0088] (4) Explore the impact of image size and image resolution on prediction accuracy. The prediction accuracy of the neural network is calculated for five conditions: image size of 1.7cm×1.7cm, 3.4cm×3.4cm, 6.7cm×6.7cm, 13.4cm×13.4cm, and 26.8cm×26.8cm. Figure 7 (a). The larger the image size, the smaller the difference in image information and the less likely it is to be affected by the uneven distribution of sand, aggregate, and cement. Therefore, the model using the largest image size (26.8 cm × 26.8 cm) achieves the best training accuracy, R 2 It can reach 0.9987.

[0089] In addition, five working conditions with image resolutions of 20×20px, 40×40px, 80×80px, 316×316px, and 1266×1266px were calculated. The prediction accuracy of the neural network is as follows: Figure 7 (b) As shown. Although the prediction accuracy of the neural network is greater than 0.99 under the five image resolution conditions, the prediction accuracy of the neural network decreases as the image resolution decreases. When the image resolution is 1266×1266px, the higher image resolution ensures the integrity of details such as concrete mortar joints, enabling the neural network to extract sufficient image features, thereby improving the prediction ability of the neural network model.

[0090] (5) Explore the accuracy and universality of the optimal training model

[0091] The optimal training model is used to identify the features of the test set and the R values ​​of the seven features are obtained. 2 and MBE are shown in Table 5. The average deviation errors of the seven characteristics of concrete did not exceed 0.2, and the determination coefficients were all greater than 0.99, indicating that the optimal neural network model can accurately identify the seven characteristics of self-compacting concrete.

[0092] Table 5 R of the optimal neural network training and test set 2 and MBE

[0093]

[0094] Figure 8It is a comparison between the predicted value of the optimal neural network model and the experimental value. It can be seen from the figure that the predicted value of the neural network and the experimental value are distributed near the straight line y=x, which means that the predicted value of the neural network is close to the true value, thus illustrating the accuracy of the optimal neural network model. Fig. 9 It is the change process of the training accuracy of the training set and the test set during the training process. As can be seen from the figure, the R 2 They are very close and both increase with the number of iterations, indicating that the model has strong generalization ability and there is no overfitting problem.

[0095] The embodiments disclosed above of the present invention can be easily implemented by professionals in this industry and related industries, and it is not difficult to modify many details in the embodiments. The methods defined in this article can also be appropriately modified and used in other specific embodiments.

Claims

1. A non-destructive identification method for self-compacting concrete properties based on image recognition technology, characterized in that: The main steps include: (1) Determine the composition and mix conditions of self-compacting concrete, including the volume fraction of coarse aggregate (α), fly ash replacement ratio (β), and water-cement ratio (W / C); (2) Mix self-compacting concrete, conduct slump test and J-ring test, obtain the slump (SF) and J-ring expansion of self-compacting concrete under different working conditions, and calculate the gap passability performance index (PA) of self-compacting concrete; (3) Prepare concrete specimens under different working conditions, photograph the concrete specimens after curing, and collect self-compacting concrete images of concrete specimens under different working conditions; (4) Compressive strength test and rapid chloride ion migration test were carried out on concrete specimens to obtain the compressive strength (f cu ) and the chloride ion diffusion coefficient (D RCM ); (5) Data enhancement is performed on the collected self-compacting concrete images of concrete specimens, the experimental results are standardized, and the images are "labeled" using the standardized experimental results, and the images and experimental results are associated to form a training set, a validation set, and a test set. The experimental results include seven properties of self-compacting concrete: coarse aggregate volume fraction (α), fly ash replacement rate (β), water-cement ratio (W / C), slump (SF), self-compacting concrete gap passing performance index (PA), compressive strength (f cu ), chloride ion diffusion coefficient (D RCM ) to assess the workability, strength and durability of self-compacting concrete; (6) Determine the judgment index for evaluating the prediction accuracy of the neural network; (7) A convolutional neural network is selected and trained to obtain a network model for predicting the properties of self-compacting concrete. Its input is a picture of self-compacting concrete, and its output is the seven properties of self-compacting concrete.

2. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 1 is characterized in that: Before step (7), it also includes: building different types of convolutional neural networks, exploring the neural network type, image type, image size and resolution, and neural network parameters to determine the operating condition combination with the best prediction accuracy.

3. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 1 is characterized in that: Use a smart terminal to take pictures of concrete, and input the pictures into a trained network model for predicting the properties of self-compacting concrete. This network model is used to predict the seven properties of self-compacting concrete.

4. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 3 is characterized in that: The self-compacting concrete image of the concrete specimen collected in step (3) is an image of the concrete with the upper surface polished.

5. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 4 is characterized in that: The image size of the polished upper surface concrete is 26.8cm×26.8cm, and the image resolution is 1266×1266 pixels.

6. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 1, characterized in that: The optimal convolutional neural network type determined is ResNet18, The output is 7 properties of self-compacting concrete.

7. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 6 is characterized in that: The ResNet18 convolutional neural network used contains a residual block structure of 4 deep residual networks, with a total of 22 hidden layers, namely 18 convolutional layers, 2 pooling layers and 2 fully connected layers. The RELU function is used as the activation function, and the 7 concrete characteristics predicted by the neural network are passed to the output layer by adding 1 fully connected layer.

8. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 6, characterized in that: The learning rate of the convolutional neural network is 0.1 and the dropout rate is 0.

9. The non-destructive identification method of self-compacting concrete properties based on image recognition technology according to claim 1, characterized in that: In order to eliminate the differences in dimensions and magnitudes between different experimental results, the following formula is used to standardize the experimental results: Where: x norm is the standardized variable, x actual is the concrete properties test result, x max is the maximum value of the experimental results, x min is the minimum value of the experimental results.

10. The non-destructive identification method of self-compacting concrete characteristics based on image recognition technology according to claim 1, characterized in that: The evaluation criteria for the neural network prediction results are based on R 2 and MBE, whose formula form is as follows: Where: n is the number of data, P i is the experimental value of the ith concrete property, O i is the predicted value of the i-th concrete characteristic model, is the average value of the experimental values ​​of concrete properties in the data set.

Images