Concrete chiseling roughness evaluation method based on convolutional neural network

Abstract

The invention discloses a concrete chiseling roughness evaluation method based on a convolutional neural network, and the method comprises the steps: obtaining a concrete surface chiseling picture through a depth camera, and building a concrete chiseling evaluation index, i.e., the roughening roughness according to the chiseling point cloud information obtained after preprocessing; the method comprises the following steps: establishing a relationship between roughening roughness and interlayer binding force through a splitting tensile test of layered pouring test blocks with different contact surface scabbling degrees, classifying the scabbling degrees into under scabbling, ideal scabbling and over scabbling, and finally, establishing a convolutional neural network, and on the basis of camera photos and scabbling classification, calculating the scabbling degree of the layered pouring test blocks. And a data set is established and data enhancement is carried out, and classification and evaluation can be automatically obtained by inputting a network after camera shooting. According to the method, the surface information of the joint surface is obtained in a manner of scanning and solidifying the concrete surface by the laser radar and the camera, the joint surface quality evaluation index system method is constructed, the joint surface chiseling condition is automatically detected based on deep learning, and intelligent evaluation of the concrete chiseling quality of the dam worker in the alpine region is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent management of engineering construction, and relates to a method for evaluating the roughness of concrete chiseling based on a convolutional neural network. Background Art

[0002] At present, there is no unified standard for the treatment standards and detection methods of the roughness of the bonding surface of layered concrete pouring at home and abroad. According to existing research and practice, there are three common detection methods: sand filling method, fractal dimension method and visual recognition method. The sand filling method is one of the earliest detection methods. It reflects the degree of concavity and convexity of the bonding surface by pouring a certain amount of sand on the bonding surface and then measuring the height of the sand. This method is simple to operate, but its accuracy is not high, which is not conducive to construction quality control. The fractal dimension method is a detection method based on mathematical theory. It reflects the complexity and irregularity of the bonding surface by taking a complete fractal analysis of the surface morphology of the bonding surface to obtain its fractal dimension value. This method can theoretically better describe the roughness of the bonding surface, but its equipment cost is high, the operation is complex, the data processing volume is large, and it is not suitable for engineering construction in high-altitude cold areas. The human eye recognition method is one of the commonly used methods on construction sites. Construction technicians usually guide the roughening process based on highly subjective terms such as "no milky skin on the construction joint surface, slightly exposed fine aggregate", and judge the roughening quality based on personal experience, making it impossible to finely control the construction quality of multi-layer poured concrete.

[0003] With the improvement of the resolution accuracy of optical camera equipment and the multi-dimensional development of artificial intelligence neural networks, visual recognition technology, as a technology that uses computers and artificial intelligence to analyze and understand the content and information in images, has provided new ideas for concrete construction quality inspection. In the harsh construction environment of high cold and high altitude, workers' attention and endurance are accelerated to decline, which not only significantly reduces efficiency, but also easily causes construction safety accidents. Therefore, it is necessary to give full play to digital and intelligent means to solve many adverse problems in the construction of water conservancy projects, such as large temperature difference between day and night, significant reduction in manpower efficiency, and continuous construction in ultra-long low temperature seasons, so as to improve the efficiency of safe construction of water conservancy projects in high cold and high altitude areas.

[0004] At present, the horizontal seam surface treatment adopts high-pressure water artificial roughening. Due to the restriction of the process, the roughening effect is difficult to meet the uniformity requirements, and the process qualification inspection lacks scientific and quantitative objective evaluation standards and refined detection methods. Summary of the invention

[0005] The purpose of the present invention is to provide a method for evaluating the roughness of concrete chiseling based on a convolutional neural network. According to the requirements of the intelligent management and control system for concrete construction, the present invention obtains the surface information of the bonding surface by scanning the surface of the consolidated concrete with equipment such as a laser radar and a camera, constructs a bonding surface quality evaluation index system method, and automatically detects the roughening of the bonding surface based on deep learning, thereby realizing an intelligent evaluation of the roughening quality of concrete for dam workers in high-altitude and cold regions.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for evaluating concrete chiseling roughness based on a convolutional neural network, characterized by comprising the following steps:

[0008] S1: Use the depth camera to obtain the chiseling photos and pre-process the collected pictures;

[0009] S2: Based on the pre-processed information, combined with Pearson test and multiple correlation analysis methods, the roughness suitable for the evaluation of the roughening treatment of the concrete blank surface is derived. R Calculation formula and use roughness R The calculation formula calculates the roughness of the roughening;

[0010] S3: Constructing a roughening evaluation standard based on the relationship between roughening roughness and interlayer bonding strength;

[0011] S4: Construct a chiseling evaluation classification model based on convolutional neural network;

[0012] S5: Evaluate the roughness of concrete chiseling in engineering based on the chiseling evaluation classification model.

[0013] Furthermore, in step S1, the preprocessing step specifically includes:

[0014] 1.1. Denoise the original photo point cloud. The point cloud denoising adopts a multi-gradient method. First, the area is limited by straight-through filtering, then the discrete and interference points are removed by statistical filtering, and then the noise points are removed by least squares filtering according to the distance threshold from the point to the fitting surface. Finally, the denoising result is smoothed by bilateral filtering.

[0015] 1.2. Correct the chiseled point cloud to keep it horizontal. Use matlab to implement the RANSAC algorithm to fit the best plane of the point cloud to represent the plane of the chiseled point cloud. By adjusting the plane, the purpose of keeping the chiseled point cloud horizontal is achieved. After inputting the point cloud data, after processing and calculation, the normal vector of the fitted plane is output. Adjust the normal vector to be perpendicular to the horizontal plane, and multiply the point set coordinates with the rotation leveling matrix calculated by matlab to achieve the purpose of correction.

[0016] Further, the step S2 specifically includes:

[0017] 2.1. For the scanned point cloud data, separate and store the three-phase coordinates; traverse z Coordinate sequence, get the highest point of the plane to be detected , The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , traverse the z coordinate sequence again to obtain the lowest point of the plane to be detected . Maximum height difference of the plane to be detected ; The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , and its calculation formula is:

[0018] ;

[0019] ;

[0020] In order to improve the calculation speed and reduce data redundancy, the Delaunay triangulation method is used to calculate:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] Where: Z C For all points in the detection area z coordinate, S C is the scanning area of ​​the area to be detected, V CS is the volume of the point cloud surface of the concrete surface to be detected projected onto the zero-point plane. R C The roughness of the sand pile surface to be tested is decisive. R S The roughness of the concrete surface to be tested is determined by the , and are the three vertices of the triangle after Delaunay triangulation coordinate, , and are the three vertices of the triangle after Delaunay triangulation z coordinate;

[0026] 2.2 According to RS and R C The Pearson correlation coefficient is used to determine the contribution rate of a single parameter to the roughness of the hair, and then the roughness of the hair is obtained. R The specific operation is as follows: Based on the three-dimensional point cloud data of multiple groups of random chiseling surfaces R S , R C and The Pearson test was performed using SPSS software and the results were R S and R C；

[0027] ; ; ;

[0028] Where: is the decisive roughness contribution rate of the sand pile, is the decisive roughness contribution rate of concrete, Determining roughness correlation for sand pile, is the concrete deterministic roughness correlation, R It is the roughness of hair formation.

[0029] Furthermore, in step S3, the evaluation method step specifically includes:

[0030] 3.1. Design and test the concrete mix ratio according to the concrete mix ratio of the project site, and cast the standard specimen with a volume of 150mm×150mm×150mm. In order to achieve the purpose of layered casting, the concrete specimen consists of the bottom test block A and the top test block B, and the volume of test block A and test block B is close to 150mm×150mm×75mm.

[0031] First, cast the test block A, make a mold and mark the position of 78mm in mold height as a reference, move the mixed concrete into the mold, and stop adding slurry when the central convex surface reaches the horizontal mark. Place the mold on the vibrating table and vibrate for 1 to 2 minutes, and demould after curing for 24 hours. Due to casting fluctuations, the test block height will appear between 75 and 80mm. Mark to reserve a distance of 3mm to facilitate manual roughening to 75mm to construct test blocks with different degrees of roughening.

[0032] 3.2. The top surface of the bottom test block A is roughened to different degrees, so that the volume of the test block A after roughening is close to 150mm×150mm×75mm and the top is rough; several coarsely roughened, medium-roughened and deeply roughened test pieces are made by roughening treatment; uniformly take photos to obtain the surface point cloud, and repeat steps S1 and S2 to obtain the roughening roughness.

[0033] In the above step 3.2, the top surface of the bottom test block A is roughened to different degrees by combining a hand-held punching drill and manual roughening: the roughening treatment of each bottom test block A is carried out according to the thickness of the test block A, because the bonding strength of the over-smooth and rough bonding surfaces is relatively low in the macroscopic law, and at the same time, bonding surfaces of different degrees are allowed to appear in the training set. For the convenience of production, the specific operations are as follows: the test block A with a thickness of less than 76 mm is roughened a little or not, and the thickness after treatment is close to 75 mm, simulating under-roughening, characterized by a basically flat surface or a small amount of roughening marks; the test block A with a thickness of 76-78 mm is simulated to be moderately roughened by manual chiseling, and the thickness after treatment is close to 75 mm, simulating moderate roughening, and the surface is uniformly roughened; the test block A with a thickness greater than 79 mm is simulated to be excessively roughened by manual chiseling, and the thickness after treatment is close to 75 mm, simulating over-roughening, and the surface is characterized by different depths; the volume of the test block A after roughening treatment is close to 150 mm×150 mm×75 mm; a concrete group with a smooth surface to an excessively roughened surface is formed.

[0034] 3.3. Casting test block B, rinse the roughened surface of the bottom test block A and place it on the bottom of the mold to ensure that the test block fits tightly with the bottom plate of the mold, then assemble the mold, and then pour concrete into the mold above the bottom test block A. Then place it on a flat vibrating table and vibrate for 1 to 2 minutes to form a complete 150mm×150mm×150mm standard test block, and then continue to cure the specimen to the target age.

[0035] 3.4. Carry out splitting tensile test on the specimen. The interlayer bonding strength is used as the standard to characterize the degree of roughening, and the splitting tensile strength of the test result is used to characterize the interlayer bonding strength of the roughening; specifically:

[0036] Taking the required splitting strength as the standard value, the change law of the mean splitting strength of each roughening group at 7 days and 28 days was obtained, and the roughening roughness obtained by photographing the concrete surface was corresponded to three areas: the low roughening roughness value with splitting strength lower than the standard value corresponds to the under-roughening area, the high roughening roughness value with splitting strength lower than the standard value corresponds to the over-roughening area, and the rest corresponds to the ideal roughening area. Different pictures correspond to different roughening degrees according to their roughening roughness, and a network training set with this strength standard is obtained at the same time. After training, the purpose of obtaining the roughening degree through its appearance picture without the need to obtain the splitting strength through the splitting test is achieved.

[0037] The above-mentioned standard values ​​of splitting strength that need to be achieved are design values ​​based on different requirements.

[0038] Furthermore, in step S4, the convolutional neural network construction method specifically includes:

[0039] 4.1. Construct a convolutional neural network model. Based on the Mobilenet V2 network model, the CA attention mechanism is introduced to construct a Mobilenet-CA model that is lightweight, efficient, and has high image recognition performance. The experimental results on the public dataset dogs-vs-cats show that Mobilenet-CA achieves higher accuracy in image classification tasks than Mobilenet V2, while also inheriting the feature of being easy to deploy on mobile devices. At the same time, the VGG16-BN network model, which is similar to the Mobilenet CA network, is introduced for comparison.

[0040] 4.2. Construct a concrete roughness image dataset.

[0041] Firstly, the original images scanned by the camera are used as training sets and the three partitions corresponding to the concrete roughness are marked.

[0042] Then, the tilted target content is corrected using perspective transformation, and the image with tilted perspective is projected into the horizontal perspective image through the perspective transformation matrix to support effective model training and evaluation and reduce the interference of building structure differences and environmental changes on the application.

[0043] Secondly, through data enhancement techniques such as random cropping, random rotation, brightness and saturation alienation, and mirroring, a variety of training samples were generated. After data enhancement, the constructed dataset has a total of 12,000 image data with classification labels, which can better train the model to adapt to the diverse concrete chiseling surface image input. The trained classification model for segmenting the ring knife point cloud is obtained.

[0044] 4.3. Select appropriate hyperparameters for model training through hyperparameter search. SGD and Adam optimizers were used for model training, and their loss functions and convergence speeds were compared. Different initial learning rates were also tested in model training; the results showed that for the Vgg16-BN network model, the initial learning rate of 0.01 performed best, with fast convergence speed and good model learning effect; the same was true for the Mobilenet-CA network model.

[0045] 4.4. The two network models were evaluated by multiple different indicators. The two network models were trained based on self-built data sets, and the learning rate was optimized in combination with deep learning model training theory. Finally, the model performance was evaluated by evaluating indicators such as training time, prediction time, total number of parameters, and computing unit occupancy. Although the Mobilenet CA network is slightly inferior to the Vgg16-BN network model in some accuracy indicators, its training efficiency and prediction speed are more efficient than the Vgg16-BN network model, and the number of parameters and computational complexity are also lower, making it easy to deploy on terminal devices.

[0046] Furthermore, in step S5, the photos obtained on site by the depth camera are input into the chiseling evaluation classification model, and the chiseling classification and evaluation are automatically performed through deep learning. The chiseling is classified into: under-chiselled area, ideal chiseled area, and over-chiselled area.

[0047] The benefits of the present invention:

[0048] (1) The method of the present invention aims to solve the problem that manual evaluation of the roughness of the surface of the concrete multi-layer casting construction slab is subjective and random, and mechanical detection is complicated and costly. It provides a method for quickly and accurately detecting surface roughness parameters, thereby achieving a quantitative description of the roughness of the chiseled surface.

[0049] (2) The method of the present invention adopts a non-contact method of camera shooting, which will not cause damage to the concrete surface and interior, and will not affect the strength and durability of the dam concrete. The use of digital detection not only has a fast detection speed and accurate measurement results, but can also complete large-area concrete surface detection in a short time. It has good versatility and is not limited to professional technicians with complex training, nor is it restricted by the shape and size of the concrete. Its automated detection results can be fed back to the entire intelligent roughening process quality control system, facilitating the development of intelligent and refined construction.

[0050] (3) The method of the present invention proposes roughening roughness to characterize the degree of irregularity of the concrete surface after roughening treatment. It is detected and quantified by three-dimensional space reconstruction method based on the physical sand pile method. Compared with the sand pile method, it is simpler to operate and has low accuracy interference. Through point cloud noise reduction processing, more than 90% of outliers and heterogeneous points are filtered out. The point cloud plane correction algorithm is used to solve the problem that the roughening roughness calculation model has a great influence on the deflection of the scanning equipment or the uneven settlement of the ground, thereby distorting the calculation results. The roughening roughness suitable for the evaluation of the roughening treatment of the concrete blank surface is derived. R The calculation formula for .

[0051] (4) The method of the present invention studies the interlayer bonding performance of concrete. The tensile strength measured by the splitting tensile test is used as an evaluation method for the interlayer bonding performance. As a form of indirect tensile test, it is easier to operate and more accurate results are obtained in the study of the interlayer bonding performance of concrete. At the same time, the model established based on the test is more in line with engineering practice. During the test, the casting method was changed by improving the detachable mold, so that the stress conditions of layered casting of concrete at the construction site were simulated, and the reproducibility of the test was improved. The embodiment of the present invention observed the failure form and the change law of splitting tensile strength under different roughness through experimental testing of 36 test blocks: with the increase of the roughening roughness of the bonding surface, the splitting tensile strength of the specimen gradually increased within a certain range, and when the roughening roughness was too high, the splitting tensile strength of the specimen began to decrease. That is, when the concrete is cast in layers, it is necessary to control the roughening roughness of the bonding surface within an ideal range to obtain a higher interlayer bonding strength.

[0052] (5) The method of the present invention constructs a concrete roughness image dataset suitable for the project based on the specifications that need to be followed in the data collection process, including resolution, environmental conditions, shooting distance and angle. The concrete roughness image data in the data collection process comprehensively considers data quality, diversity and collection specifications, and can support effective model training and evaluation. Through data enhancement techniques such as random cropping, random rotation, brightness and saturation alienation, and mirroring, a variety of training samples are generated. After data enhancement, the dataset constructed by the embodiment of the present invention has a total of 12,000 image data with classification labels, which can better train the model to adapt to the diverse concrete roughness surface image inputs and improve the performance and generalization ability of the model.

[0053] (6) The method of the present invention learns two network models, Mobilenet V2 network and VGG16 network, and the emerging attention mechanism CA attention. By introducing the CA attention mechanism on the Mobilenet V2 network model, a Mobilenet-CA model is constructed that can improve image recognition performance while maintaining light weight and high efficiency. In the test on the public dataset dogs-vs-cats, the accuracy of the Mobilenet-CA model is 98.25%, which is 0.53% higher than the accuracy of the Mobilenet V2 model of 97.72%. In the model training process based on the self-built dataset, the appropriate hyperparameters are selected for model training through hyperparameter search. SGDm and Adam optimizers are used in model training, and their loss functions and convergence speeds are compared, both of which have good optimization effects. At the same time, different initial learning rates are also tested in model training. The results show that for the Vgg16-BN network model, the initial learning rate of 0.01 performs best, converges quickly and has good model learning effect; and for the Mobilenet-CA network model, the initial learning rate of 0.01 can also achieve good training effect.

[0054] (7) The method of the present invention adjusts the neural network optimization method and hyperparameters according to the accuracy index and the loss of the training set and the test set, so that the neural network model can have better prediction accuracy. The accuracy of the vgg16-bn model is 99.63%, the precision is 99.64%, and the recall is 99.56%. It performs slightly better than Mobilenet-CA in different accuracy indicators, especially in accurately finding all positive examples, but requires longer prediction response time. Mobilenet-CA performs better in prediction response time, which is only 0.172s, which is 0.074s faster than vgg16-bn's 0.246s, and is suitable for scenarios with high response speed requirements. The lightweight Mobilenet-CA model has fewer parameters and FLOPs, occupies less memory resources, and therefore performs better in situations where memory and computing resources are limited.

[0055] (8) In view of the different performance characteristics of the Mobilenet-CA network model and the Vgg16BN network model, the method of the present invention created a concrete roughening linear combination model evaluation method, developed a highly accurate and easy-to-operate concrete roughness detection platform, and tested it in actual engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Comparison of concrete surface point cloud before and after denoising;

[0057] Figure 2Comparison of concrete surface point cloud before and after correction;

[0058] Figure 3 This is a schematic diagram of casting test block A and test block B;

[0059] Figure 4 This is the relationship between the roughness of the hair and the bonding force between the layers;

[0060] Figure 5 The photos are not well-cut for classification results;

[0061] Figure 6 Ideal chiseling photos for classification results;

[0062] Figure 7 Over-cut photos for classification results. DETAILED DESCRIPTION

[0063] The present invention is further described below in conjunction with specific implementation methods. The specific implementation methods are further descriptions of the principles of the present invention and do not limit the present invention in any way. Technologies that are the same or similar to the present invention do not exceed the scope of protection of the present invention.

[0064] The present invention is further described in detail below with reference to the accompanying drawings.

[0065] The invention discloses a method for evaluating the roughness of concrete chiseling based on a convolutional neural network, which is used for visual recognition and intelligent evaluation of the quality of concrete chiseling for dam workers in alpine areas.

[0066] The present invention utilizes non-contact depth camera technology to accurately capture the surface of concrete after roughening, performs point cloud denoising, leveling and other preprocessing through the two-dimensional visual features and three-dimensional information of surface depth obtained by the depth camera, and establishes the roughening roughness evaluation index of concrete roughening, which provides a scientific basis for quantitative evaluation of the roughening quality.

[0067] The present invention conducts splitting tensile tests on layered cast test blocks and uses different roughening degrees at the layering to establish the relationship between roughening roughness and interlayer bonding strength and classify the roughening degrees into under-roughening, ideal roughening and over-roughening. The subsequent intelligent evaluation of this classification provides a clear classification standard.

[0068] The present invention establishes a Mobilenet V2 convolutional neural network model that introduces the CA attention mechanism, establishes a data set and performs data enhancement based on camera photos and chiseling classification, and automatically obtains classification and evaluation by inputting the data into the network after taking a photo with the camera.

[0069] The specific steps are as follows:

[0070] Step 1: Preprocess the collected chiseling pictures;

[0071] The depth camera is set up on the roughened concrete to collect data, and a point cloud file containing the three-dimensional information of the roughened concrete surface is obtained. Affected by the environment such as light, temperature, humidity and the influence of the machine equipment itself, the point cloud file contains a large number of noise points distributed in the point cloud of the concrete surface. At the same time, the placement of the depth camera cannot always be directly opposite the concrete surface, and there is usually a slight tilt. At this time, noise reduction and leveling processing are required.

[0072] (1.1) Denoise the original photo point cloud. The point cloud denoising adopts a multi-gradient method. First, the area is limited by straight-through filtering, then the discrete and interference points are removed by statistical filtering, and then the least squares filter is used to remove noise points according to the distance threshold from the point to the fitting surface. Finally, the bilateral filter is applied to smooth the denoising result.

[0073] (1.2) Correct the roughened point cloud to keep it horizontal. Use MATLAB to implement the RANSAC algorithm to fit the best plane of the point cloud to represent the plane of the roughened point cloud. By adjusting the plane, the purpose of keeping the roughened point cloud horizontal is achieved. After inputting the point cloud data, after processing and calculation, the normal vector of the fitted plane is output. Adjust the normal vector to be perpendicular to the horizontal plane, and multiply the point set coordinates with the rotation leveling matrix calculated by MATLAB to achieve the purpose of correction.

[0074] The point cloud before and after the multi-gradient denoising and smoothing process in this embodiment is compared. Figure 1 As shown; the comparison before and after correction is as follows Figure 2 shown.

[0075] Based on the pre-processed information, the roughness value suitable for evaluating the roughening treatment of the concrete blank surface is derived. R The calculation formula for .

[0076] In order to determine the specific composition of roughening roughness, this example analyzes the deterministic roughness of sand pile based on 100 sets of random chiseled surface 3D point cloud data. R s and concrete decisive roughness R c Extreme difference with the roughened surface h c The correlation of R s and R c It represents the average depth of the roughened surface. h c Represents the maximum fluctuation of the rough surface. Pearson test was used to study the correlation R s and R c .

[0077] The Pearson test can evaluate the decisive roughness of sand and concrete with a faster calculation speed. h c The strength of the linear relationship. The Pearson test method of SPSS was used to perform data analysis, and the results are shown in Table 1.

[0078] Table 1 Pearson test analysis results:

[0079]

[0080] Descriptive Statistics R s and R c The sample size of variables involved in the partial correlation analysis is 100. R s The mean value of is 4.6098, which is R c The mean is very close. R s The standard deviation is 2.0919, and the fluctuation range is R c Table 2 is the results of correlation analysis.

[0081] Table 2 Correlation analysis results:

[0082]

[0083] R s and R c The significance level is less than 0.01, indicating R s and R c and h c are all relevant. R s and h c The Pearson correlation coefficient is 0.769, which is greater than 0.5, indicating a positive correlation and a strong correlation. R c and h c The Pearson correlation coefficient is 0.812, which is also greater than 0.5, indicating that the positive correlation strength is higher, so consider using R c It is used to characterize the roughness of the concrete surface after random chiseling, which not only retains the statistic of the average depth of the roughness of the concrete chiseled surface, but also contains the information of the extreme difference of the peaks and valleys of the chiseled surface.

[0084] When comparing the actual data, we found that Rc In most cases, the roughness is consistent with the visual observation. However, when the roughening is uneven due to the influence of the roughening tool or process, R c The actual roughness of the chiseled surface is quite different. R s and R c The deviation value of R c It cannot reflect the actual roughness of the chiseled surface. It can be seen from Table 1 and Table 2 that R s and R c The significance level sig (two-tailed) is 0.079, which is much higher than 0.05, and the Pearson correlation coefficient is 0.251, indicating that there is no obvious linear correlation between the two. In other words, R s and R c The data are independent and contain different roughness information, so it is necessary R s and R c These two data support the roughness composition of roughening.

[0085] Step 2: Combining multiple correlation analysis methods such as Pearson test, derive the roughness suitable for the evaluation of the roughening treatment of the concrete blank surface R Calculation formula.

[0086] (2.1) For the scanned point cloud data, separate and store the three-phase coordinates; traverse z Coordinate sequence, get the highest point of the plane to be detected , The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , traverse the z coordinate sequence again to obtain the lowest point of the plane to be detected ; Maximum height difference of the plane to be detected . The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , and its calculation formula is:

[0087]

[0088] ;

[0089] In order to improve the calculation speed and reduce data redundancy, Delaunay triangulation method is used.

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] Where: Z C For all points in the detection area z coordinate, S C is the scanning area of ​​the area to be detected, V CS is the volume of the point cloud surface of the concrete surface to be detected projected onto the zero-point plane. R C The roughness of the sand pile surface to be tested is decisive. R S The roughness of the concrete surface to be tested is determined by the , and The three vertices of the triangle after Delaunay triangulation coordinate, , and are the three vertices of the triangle after Delaunay triangulation z coordinate.

[0095] (2.2) According to R S and R C The Pearson correlation coefficient is used to determine the contribution rate of a single parameter to the roughness of the hair, and then the roughness of the hair is obtained. R The specific operation is as follows: Based on the three-dimensional point cloud data of multiple groups of random chiseling surfaces R S , R C and The Pearson test was performed using SPSS software and the results were R S and R C ;

[0096] ; ; ;

[0097] Where: is the decisive roughness contribution rate of the sand pile, is the decisive roughness contribution rate of concrete, Determining roughness correlation for sand pile, is the concrete deterministic roughness correlation, R It is the roughness of hair formation.

[0098] Step 3: Construct a roughening evaluation method based on the relationship between roughening roughness and interlayer bonding strength. The process is shown in the figure below. Figure 3 shown.

[0099] (3.1) The standard specimens were cast according to the concrete mix ratio design test on site, with a volume of 150mm×150mm×150mm. In order to achieve the purpose of layered casting, the concrete specimens consisted of bottom test block A and top test block B, and the volumes of A and B test blocks were close to 150mm×150mm×75mm.

[0100] First cast test block A, mark the position of 78mm in the mold, move the mixed concrete into the mold, and stop adding mortar when the central convex surface reaches the horizontal mark. Place the mold on the vibrating table and vibrate for 1 to 2 minutes, and demould after curing for 24 hours.

[0101] (3.2) The top surface of the bottom layer A test block was roughened to different degrees using a combination of a hand-held punching drill and manual roughening. The test block A was classified according to its thickness. Thickness not exceeding 76 mm was rough roughening, 76 to 78 mm was medium roughening, and 79 mm or more was deep roughening. At the same time, the volume of the test block A was made close to 150 mm × 150 mm × 75 mm. After taking photos to obtain the surface point cloud, the previous steps were repeated to calculate its roughening roughness.

[0102] (3.3) Then cast test block B, rinse the roughened surface of the bottom test block A and place it on the bottom of the mold to ensure that the test block fits tightly against the bottom plate of the mold. Then assemble the mold and pour concrete into the mold above the bottom test block A. Then place it on a flat vibrating table and vibrate for 1 to 2 minutes to form a complete 150 mm × 150 mm × 150 mm standard test block. Then continue to cure the test piece to the target age.

[0103] (3.4) The test pieces were subjected to splitting tensile tests, and the results were used to characterize the interlayer bonding strength of the roughened surface. After analysis, according to the change rules of the mean splitting tensile strength of each group at 7 days and 28 days, as shown in Figure 4 As shown, the roughening roughness obtained by photographing the concrete surface is divided into three areas: under-roughening area, ideal roughening area and over-roughening area.

[0104] Step 4: Construct a chiseling evaluation classification model based on convolutional neural network.

[0105] (4.1) Construct a convolutional neural network model. Based on the Mobilenet V2 network model, the CA attention mechanism is introduced to construct a Mobilenet-CA model that is lightweight, efficient, and has high image recognition performance. The experimental results on the public dataset dogs-vs-cats show that Mobilenet-CA achieves higher accuracy in image classification tasks than Mobilenet V2, while also inheriting the feature of being easy to deploy on mobile devices. At the same time, the VGG16-BN network model, which is similar to the Mobilenet CA network, is introduced for comparison.

[0106] (4.2) Construct a dataset of images of rough concrete. First, the original images scanned by the camera are used as the training set and the three partitions corresponding to the rough concrete are marked. Then, the tilted target content is corrected by perspective transformation, and the image with tilted perspective is projected to the horizontal perspective image through the perspective transformation matrix to support effective model training and evaluation, and reduce the interference of building structure differences and environmental changes on the application. Secondly, through data enhancement techniques such as random cropping, random rotation, brightness and saturation alienation, and mirroring, a variety of training samples are generated. After data enhancement, the constructed dataset has a total of 12,000 image data with classification labels, which can better train the model to adapt to the diverse concrete chiseled surface image inputs. A trained classification model for segmenting the ring knife point cloud is obtained.

[0107] (4.3) Select appropriate hyperparameters for model training through hyperparameter search. SGD and Adam optimizers were used for model training, and their loss functions and convergence speeds were compared. At the same time, different initial learning rates were tested in model training. The results showed that for the Vgg16-BN network model, the initial learning rate of 0.01 performed best, with fast convergence speed and good model learning effect; the same is true for the Mobilenet-CA network model.

[0108] (4.4) The two network models were evaluated using multiple different indicators. The two network models were trained based on a self-built data set, and the learning rate was optimized in combination with the deep learning model training theory. Finally, the model performance was evaluated by evaluating indicators such as training time, prediction time, total number of parameters, and computing unit occupancy. Although the Mobilenet CA network is slightly inferior to the Vgg16-BN network model in some accuracy indicators, its training efficiency and prediction speed are more efficient than the Vgg16-BN network model, and the number of parameters and computational complexity are also lower, making it easy to deploy on terminal devices.

[0109] Step 5: After the camera scan is completed, the photo is input into the deep learning model, and the chiseling situation is automatically classified and evaluated through deep learning. Figure 5 For the classification results, the photos are not well-defined. Figure 6 For ideal classification results, Figure 7 Over-cut photos for classification results.

Claims

1. A method for evaluating the roughness of concrete chiseling based on convolutional neural network, characterized in that The following steps are involved: S1: Use the depth camera to obtain the chiseling photos and pre-process the collected pictures; S2: Based on the pre-processed information, combined with Pearson test and multiple correlation analysis methods, the roughness suitable for the evaluation of the roughening treatment of the concrete blank surface is derived. R Calculation formula and use roughness R The calculation formula calculates the roughness of the roughening; S3: Constructing a roughening evaluation standard based on the relationship between roughening roughness and interlayer bonding strength; S4: Construct a chiseling evaluation classification model based on convolutional neural network; S5: Evaluate the roughness of concrete chiseling in engineering based on the chiseling evaluation classification model.

2. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 1, characterized in that: In S1, the preprocessing step includes: 1.

1. De-noise the original photo point cloud; the point cloud denoising adopts a multi-gradient method. First, the area is limited by straight-through filtering, then the discrete and interference points are removed by statistical filtering, and then the noise points are removed by least squares filtering according to the distance threshold from the point to the fitting surface. Finally, the denoising result is smoothed by bilateral filtering. 1.

2. Correct the chiseled point cloud to keep it horizontal; use MATLAB to implement the RANSAC algorithm to fit the best plane of the point cloud to represent the plane of the chiseled point cloud, and adjust the plane to keep the chiseled point cloud horizontal; after inputting the point cloud data, process and calculate, and output the normal vector of the fitted plane; adjust the normal vector to be perpendicular to the horizontal plane, and multiply the point set coordinates with the rotation leveling matrix calculated by MATLAB to achieve the purpose of correction.

3. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 1, characterized in that: The S2 specifically includes: 2.

1. For the scanned point cloud data, separate and store the three-phase coordinates; traverse z Coordinate sequence, get the highest point of the plane to be detected , The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , traverse again z Coordinate sequence, get the lowest point of the plane to be detected ; Maximum height difference of the plane to be detected ; The volume of the envelope formed by the intersection of the plane and the irregular surface formed by the point cloud , and its calculation formula is: ； ； In order to improve the calculation speed and reduce data redundancy, the Delaunay triangulation method is used; ； ； ； ； Where: Z C For all points in the detection area z coordinate, S C is the scanning area of ​​the area to be detected, V CS is the volume of the point cloud surface of the concrete surface to be detected projected onto the zero-point plane. R C The roughness of the sand pile surface to be tested is decisive. R S The roughness of the concrete surface to be tested is determined by the , and are the three vertices of the triangle after Delaunay triangulation coordinate, , and are the three vertices of the triangle after Delaunay triangulation z coordinate; 2.2 According to R S and R C The Pearson correlation coefficient is used to determine the contribution rate of a single parameter to the roughness of the hair, and then the roughness of the hair is obtained. R The specific operation is as follows: Based on the three-dimensional point cloud data of multiple groups of random chiseling surfaces R S , R C and The Pearson test was performed using SPSS software and the results were R S and R C ; ； ； ； Where: is the decisive roughness contribution rate of the sand pile, is the decisive roughness contribution rate of concrete, Determining roughness correlation for sand pile, is the concrete deterministic roughness correlation, R It is the roughness of hair formation.

4. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 2 or 3, characterized in that: In S3, the evaluation criteria construction includes: 3.

1. Cast several standard specimens according to the concrete mix ratio at the project site; the total volume of each standard specimen is 150mm×150mm×150mm, and it consists of a bottom test block A and a top test block B stacked in half; First, cast test block A; make a mold and mark the position of 78mm in height as a reference, move the mixed concrete into the mold, and stop adding slurry when the central convex surface reaches the mark; place the mold on a vibrating table and vibrate for 1 to 2 minutes, and demould after curing for 24 hours; 3.

2. The top surface of the bottom test block A is roughened to different degrees, so that the volume of the test block A after roughening is close to 150mm×150mm×75mm and the top is rough; several test pieces with shallow roughening, medium roughening and deep roughening are made by roughening treatment; the surface point cloud is obtained by taking photos uniformly, and steps S1 and S2 are repeated to obtain the roughening roughness; 3.

3. Casting test block B; After cleaning the roughened surface of the bottom test block A, place it at the bottom of the mold to ensure that the test block A fits tightly with the bottom plate of the mold, and then assemble the mold. Then pour concrete into the mold above the bottom test block A; then place it on a flat vibrating table and vibrate for 1 to 2 minutes to form a complete 150mm×150mm×150mm standard test piece, and then continue to cure the test piece to the target age; 3.

4. Carry out splitting tensile test on the specimen; use the interlayer bonding strength as the standard to characterize the degree of roughening, and use the splitting tensile strength of the test result to characterize the interlayer bonding strength of roughening; specifically: Taking the required splitting strength as the standard value, the change law of the mean splitting strength of each roughening group at 7 days and 28 days was obtained, and the roughening roughness obtained by taking pictures of the concrete surface was corresponded to three areas: the low roughening roughness value with splitting strength lower than the standard value corresponded to the under-roughening area, the high roughening roughness value with splitting strength higher than the standard value corresponded to the over-roughening area, and the rest corresponded to the ideal roughening area; each picture was corresponded to the corresponding roughening degree according to its roughening roughness, and a network training set with this strength standard was obtained; after training, it was used to obtain the roughening degree of the detection object through picture comparison.

5. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 4 is characterized in that: In the step 3.2, the roughening treatment of each bottom test block A is performed according to the thickness of the test block A, and the specific operation is as follows: the test block A with a thickness of less than 76 mm is subjected to a small amount of roughening or no roughening, and the thickness after treatment is close to 75 mm, simulating under-roughening, characterized by a smooth surface or a small amount of roughening marks; the test block A with a thickness of 76-78 mm is simulated to be moderately roughened by hand, and the thickness after treatment is close to 75 mm, simulating moderate roughening, and the characteristic is that the surface is evenly roughened; the test block A with a thickness greater than 79 mm is simulated to be excessively roughened by hand, and the thickness after treatment is close to 75 mm, simulating over-roughening, and the characteristic is that the depth of the surface is different; the volume of the test block A after the roughening treatment is close to 150 mm×150 mm×75 mm; a concrete group with a smooth surface to an excessively roughened surface is formed.

6. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 4, characterized in that: In S4, the convolutional neural network construction method specifically includes: 4.

1. Construct a convolutional neural network model. Based on the Mobilenet V2 network model, introduce the CA attention mechanism to construct the Mobilenet-CA model. At the same time, introduce the VGG16-BN network model similar to the Mobilenet CA network for comparison. 4.

2. Construct a concrete roughness image dataset; The original images of standard specimens scanned by the depth camera are used as training sets and the three partitions corresponding to the concrete roughness are marked; Use perspective transformation to correct tilted target content, and project the tilted image to the horizontal image through the perspective transformation matrix to support effective model training and evaluation, and reduce the interference of building structure differences and environmental changes on the application; Through random cropping, random rotation, brightness and saturation alienation, and mirror data enhancement, diverse training samples are generated, image data with classification labels are constructed, and the model is trained to adapt to diverse concrete chiseled surface image inputs, and a trained classification model for segmenting ring knife point cloud is obtained; 4.

3. Select appropriate hyperparameters for model training through hyperparameter search; use SGD and Adam optimizers for model training and compare their loss functions and convergence speeds; and experiment with different initial learning rates during model training; 4.

4. The two network models were evaluated using multiple different indicators. The two network models were trained based on self-built data sets, and the learning rate was optimized in combination with deep learning model training theory. The model performance was evaluated by evaluating the training time, prediction time, total number of parameters, and computing unit occupancy indicators. The Mobilenet CA network was selected to build a chiseling evaluation classification model based on a convolutional neural network.

7. The method for evaluating concrete chiseling roughness based on convolutional neural network according to claim 1, characterized in that: In S5, the photos obtained on site by the depth camera are input into the chiseling evaluation classification model, and the chiseling classification and evaluation are automatically performed through deep learning. The chiseling is classified into: under-chiselled area, ideal chiseled area, and over-chiselled area.

Images