A method for quantitatively detecting cracks in a railway tunnel portal wall based on a drone

Abstract

The application discloses a kind of railway tunnel portal wall crack quantitative detection methods based on unmanned aerial vehicle, first, by unmanned aerial vehicle carries high-resolution camera to tunnel portal wall structure and slope protection structure are inspected, the data collected are handled, establish a set of data set based on tunnel portal wall structure and slope protection structure crack, detection model is established by the method of deep learning, subsequently based on actual situation is debugged, improve accuracy, using detection model can realize tunnel portal wall and the automatic identification of protection structure crack and statistical data analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic detection, more particularly to a railway tunnel portal wall crack quantitative detection method based on a UAV. BACKGROUND

[0002] There are many railway tunnels, and the safety of the tunnels is an important part of railway safety. In order to ensure the safe, punctual and uninterrupted operation of trains, the tunnel should have a stable foundation, a solid and durable structure, and should be kept in good condition at all times. The engineering geology of tunnels, especially long and large tunnels, is extremely complex. In order to master the technical state of the tunnel and the portal slope, and to discover and analyze the causes of the disease in a timely manner, a careful special inspection of the tunnel portal slope should be carried out to avoid major geological disasters.

[0003] With the advancement of UAV platforms and sensor technologies, UAVs are helping to make railway inspection intelligent and automated. UAVs support high-precision autonomous flight and are not constrained by terrain and environment, and can survey areas that are difficult for humans to access. Through UAV positioning and autonomous cruising, and by combining UAV video and orthographic images with satellite digital maps, hidden dangers such as railway tunnel portals and slopes can be efficiently and accurately investigated, and manual precision monitoring of railway tunnel portals and slopes can be replaced to achieve the purpose of early warning of safety hazards or emergency accidents, thereby strengthening the prevention and control of sudden disasters along the railway.

[0004] Tunnel portal walls are mainly built with concrete. Due to factors such as material, temperature difference, corrosion and external force, the tunnel walls may shrink or expand unpredictably, resulting in cracks. Cracks not only affect the overall aesthetics of the tunnel, but once the cracks develop to a deeper level, they may develop into destructive spalling or deep cracks, affecting the safety and stability of the tunnel. The main hazards of deterioration of the portal wall structure and slope protection structure are wall cracks and support structure cracking. When cracks appear in the wall and support, continuous detection of the cracks is needed, and the growth of the cracks needs to be quantitatively analyzed. Therefore, it is particularly important to use UAV technology to regularly detect cracks in the tunnel portal wall and monitor the development trend of the cracks in real time.

[0005] Currently, the results of UAV inspection are mainly analyzed and judged by humans, which is low in efficiency and prone to misjudgment and omission. Therefore, the focus of current crack identification research has gradually shifted to crack detection algorithms based on image processing and crack detection algorithms based on machine learning. In addition, current detection technologies are mainly applied to the detection of regular building surfaces, and are less applied to irregular wall surfaces such as tunnel portal walls.

[0006] The original tunnel crack image obtained under the prior art often contains many non-crack noise textures, which increases the difficulty of crack extraction, so that the error of intelligent detection method is often large when quantitative analysis of cracks is carried out, and the intelligent degree of manual analysis is low. The existing image analysis method based on deep learning relies too much on the data set, but the existing data set is not suitable for tunnel wall cracks, and the existing data samples are few and cannot form an effective data set for training. The existing railway data of full-line survey is stored after post-processing, only the key areas and sensitive areas are labeled, and the local is selected for judgment when necessary, which is low in efficiency, and there is no set of rapid analysis process for the deterioration of the wall structure and slope protection structure of the railway tunnel portal.

[0007] Therefore, how to realize the rapid and efficient disease detection of the tunnel portal wall is a problem that those skilled in the art need to solve. SUMMARY

[0008] Therefore, the present application provides a railway tunnel portal wall crack quantitative detection method based on a unmanned aerial vehicle, which can expand the analysis data, improve the detection accuracy, and realize real-time detection and rapid analysis of the tunnel portal wall crack.

[0009] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0010] A railway tunnel portal wall crack quantitative detection method based on a unmanned aerial vehicle, comprising the following steps:

[0011] Step 1: collecting tunnel portal wall and surrounding environment images;

[0012] Step 2: denoising and preprocessing the images;

[0013] Step 3: combining the GAN model to expand the processed image data, and obtaining the original data set;

[0014] Step 4: using a labeling tool to label the wall cracks in the original data set, and establishing a training data set;

[0015] Step 5: using YOLOv7 reparameterization module (RepConvN module) according to the characteristics of the wall cracks in Unet model to perform reparameterization, and obtaining an improved Unet model;

[0016] The RepVGG network model reparameterization idea is adopted, the RepConvN module of YOLOv7 is introduced for reparameterization, the WCE loss loss function is used to replace the original Dice loss function to accelerate the model convergence speed, the long connection form of the encoder and decoder of the traditional Unet model is replaced by the residual connection form, then a 1*1 convolution branch is added in parallel in the residual branch part, then the three models are combined into a single model, the 1*1 convolution is expanded to a 3*3 convolution with 0 around, the identity layer is equivalent to a convolution layer with special weights, the current channel convolution kernel parameter is 1 and the remaining channel convolution kernel parameter is 0, which is equivalent to a 1*1 convolution layer, and then the convolution kernel conversion operation is performed to obtain a 3*3 convolution layer; the combination of conv and relu is completed through quantization operation; finally, the three parallel 3*3 convolution layers are fused to form a single path to form a RepConvN module; thus, the feature of the encoder is spliced with the corresponding feature of the decoder, and an improved Unet model is obtained, and the parameters of the branch channel are reparameterized to the main branch to complete the reparameterization during deployment.

[0017] Step 6: iteratively training the improved Unet model using the training data set to obtain a wall and structure deterioration detection model capable of automatically identifying cracks in the wall of the tunnel portal;

[0018] Step 7: inputting the wall image to be identified into the wall and structure deterioration detection model, extracting the cracks, and calculating the crack width and crack length based on the four-way orthogonal skeleton line method.

[0019] Preferably, the preprocessing includes enhancing the low-frequency components of the image based on the transform domain blur enhancement method, and enhancing the high-frequency components and local gradient information of the image using a nonlinear enhancement method.

[0020] The above technical solution has the following technical effects: the low-frequency components of the image are enhanced based on the transform domain blur enhancement method, the contrast of the image is improved without distorting the image, and the image noise is suppressed; the high-frequency components and local gradient information of the image are enhanced using a nonlinear enhancement method, the boundary information in the image is highlighted, the crack edge information is strengthened, and the noise influence is removed.

[0021] Preferably, the specific process of step 3 is as follows:

[0022] Step 31: selecting a plurality of images in the image data that meet the wall crack characteristics, and performing rotation, cropping, PS, etc. on the crack area to obtain first expanded image data;

[0023] Step 32: performing secondary data expansion on the first expanded image data using a generative adversarial network to obtain an original data set.

[0024] In view of the problem that the collected sample data is less and cannot be effectively used for deep learning model training, the original crack pictures need to be data augmented by means of rotation, cropping, PS and the like; then, in view of the problem of less samples, the data is further augmented by means of a generative adversarial network (GAN).

[0025] Preferably, the cracks are labeled by means of a combination of Labelimg and Photoshop and the like to obtain crack position files and images and establish a crack training data set.

[0026] Preferably, the RepVGG network model reparameterization idea is adopted for a traditional Unet model, and a RepConvN module of YOLOv7 is introduced for reparameterization, the backbone network of the Unet model is improved, an improved Unet model is obtained, and the improved Unet model is trained as a wall structure deterioration detection model.

[0027] Preferably, the specific process of obtaining the improved Unet model in step 5 is as follows:

[0028] Step 51: the long connection form of the encoder and the decoder of the traditional Unet model is replaced by a residual connection form, then a 1x1 convolution branch is added in parallel in the residual branch part, and then the down-sampling of the Unet model, the residual branch and the added convolution branch are combined into a single path to obtain a RepConvN module;

[0029] When combined, the 1x1 convolution is expanded into a 3x3 convolution with 0 around the four sides, the linear activation layer (identity layer) is equivalent to a convolution layer with a set weight (special weight) with the current channel convolution kernel parameter being 1 and the remaining channel convolution kernel parameter being 0, which is equivalent to a 1x1 convolution layer, and then a 3x3 convolution layer is obtained through the above convolution kernel conversion operation;

[0030] Step 52: the convolution layer (Conv layer) and the batch normalization layer (BN layer) are fused;

[0031] The convolution layer formula is as follows:

[0032] Conv(x) = W(x) + b

[0033] Wherein, x is the input; W(x) is the weight function; b is the bias parameter;

[0034] The batch normalization layer formula is as follows:

[0035]

[0036] Wherein, gamma and beta are both parameters learned by back propagation in the model; mean is the mean function, var is the variance function, both are functions in the pytorch platform numpy package;

[0037] The expression obtained after substituting the convolution layer formula into the batch normalization layer formula is:

[0038]

[0039] Can be converted to:

[0040]

[0041] Let:

[0042]

[0043]

[0044] The final fusion result is:

[0045] BN(Conv(x))=W fused (x)+B fused ;

[0046] Step 53: The combination of convolution layer (Conv layer) and nonlinear activation layer (relu layer) is completed by quantization operation; thus the feature splicing of the encoder and the corresponding feature of the decoder is completed;

[0047] Step 54: Replace the original loss function Dice loss function (1-1) with WCE loss function to speed up the learning process, as follows:

[0048]

[0049] Wherein, N is the total number of samples; w is the discriminant coefficient of positive samples; r n Indicates the label of sample n, positive class is 1 and negative class is 0; p n Indicates the probability of sample n being predicted as a positive class;

[0050] Finally, the improved Unet model is obtained.

[0051] Preferably, the specific process of calculating the crack width and in step 7 is:

[0052] Step 71: After obtaining the crack from the wall structure deterioration detection model, the singular value decomposition method in the four-way orthogonal skeleton line algorithm is used to calculate and obtain the crack skeleton line;

[0053] Step 72: Calculate the normal vector of each position of the crack in the length direction of the crack skeleton line based on the crack skeleton line;

[0054] Step 73: Calculate the crack width and crack length according to the crack edge points and normal vectors of each position.

[0055] Through the above technical solutions, compared with the prior art, the present application provides a railway tunnel portal wall crack quantitative detection method based on a unmanned aerial vehicle, a crack detection method based on Unet semantic segmentation, which is suitable for long-term evaluation of a specific line, including but not limited to comparison and analysis of only two data; in the data processing link, a generative adversarial network (GAN) is used to assist in making a data set, expand the data, and at the same time meet the accuracy requirements; the Unet model is improved, and the RepConvN module proposed by YOLOv7 is introduced to improve the crack extraction effect; parameter setting ranges are proposed for each link in the comparative analysis of the railway surrounding environment, which is beneficial to selection when processing different environments. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0057] Figure 1 The accompanying drawings are flow charts of the railway tunnel portal wall crack quantitative detection method based on a unmanned aerial vehicle provided by the present application;

[0058] Figure 2 The accompanying drawings are improved Unet model diagram structure schematic diagrams provided by the present application;

[0059] Figure 3 The accompanying drawings are improved Unet model diagram structure schematic diagrams provided by the present application;

[0060] Figure 4 The accompanying drawings are data set expansion schematic diagrams provided by the present application;

[0061] Figure 5 The accompanying drawings are data annotation schematic diagrams provided by the present application;

[0062] Figure 6 The accompanying drawings are data set establishment schematic diagrams provided by the present application;

[0063] Figure 7 The accompanying drawings are model training result schematic diagrams provided by the present application;

[0064] Figure 8 The accompanying drawings are model training process schematic diagrams provided by the present application;

[0065] Figure 9 The drawing is a schematic diagram of test set verification results provided by the present application.

[0066] Figure 10 The drawing is a schematic diagram of crack identification provided by the present application.

[0067] Figure 11 The drawing is a schematic diagram of segmentation effect provided by the present application.

[0068] Figure 12 The drawing is a schematic diagram of identification results provided by the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0070] The railway tunnel portal wall crack quantitative detection method based on a UAV disclosed in the embodiments of the present application is as follows: since the images collected by the UAV are affected by the distance and lighting conditions, they often contain a lot of noise, and the crack part often has poor definition and unclear outline and details; and the crack detection has a higher requirement for image quality, therefore, first, the image is denoised and preprocessed to enhance the image quality; second, the data samples are expanded in combination with the traditional data expansion method and a deep learning generative adversarial network (GAN), and the expansion effect is as shown in Figure 4 ; third, the crack area is labeled by using Labelimg, Photoshop and other software, a wall and protection structure deterioration training data set is established, the labeling process is as shown in Figure 5 , and the generated data set result is as shown in Figure 6 ; fourth, the Unet network model is improved according to the image characteristics and detection requirements, the RepVGG network model reparameterization idea is absorbed and the RepConvN module of YOLOv7 is introduced for reparameterization, so that it is applicable to other deep learning models, the module is used for structure reparameterization, the multi-path model is converted into a single-path model, the memory is saved while the inference speed of the model is greatly improved, and the expanded data set is used for model training and testing to obtain a detection model with better effect, the model training result and the model training process are as shown in Figure 7-8 , the test set verification result is as shown in Figure 9 , and the crack identification result is as shown in Figure 10 ; the problem of slow detection speed and insufficient accuracy of the original Unet model is changed, the expansion of the tunnel portal wall crack and the rapid and accurate identification of the wall structure deterioration are realized, and the crack extraction segmentation effect is as shown inFigure 11 Finally, the crack width and length were calculated by using the four-way orthogonal skeleton line algorithm on the crack extracted by the improved Unet model, and the results are shown in FIG. 8. Figure 12 The specific process is as follows:

[0071] S1: Collecting images of the tunnel portal wall and the surrounding environment;

[0072] S2: Denoising and preprocessing the images; the preprocessing includes enhancing the low-frequency components of the images based on the transform domain blur enhancement method, and enhancing the high-frequency components and local gradient information of the images by using the nonlinear enhancement method;

[0073] S3: Combining the GAN model to expand the processed image data to obtain the original data set;

[0074] S31: Selecting a number of images in the image data that meet the characteristics of the wall cracks, and rotating, cropping, and PSing the crack area to obtain the first expanded image data;

[0075] S32: Expanding the first expanded image data using the generative adversarial network to obtain the original data set;

[0076] S4: Labeling the wall cracks in the images in the original data set using a labeling tool to establish a training data set;

[0077] S5: According to the characteristics of the wall cracks, using the YOLOv7 reparameterization module (RepConvN module) in the Unet model to obtain an improved Unet model;

[0078] S51: Replacing the long connection form of the encoder and decoder of the traditional Unet model with a residual connection form, then adding a 1x1 convolution branch in parallel in the residual branch part, and then merging the three models of down-sampling, residual branch and added convolution branch of the Unet model into a single model to obtain the RepConvN module;

[0079] When merging, the 1x1 convolution is expanded to a 3x3 convolution with 0 on the four sides, and the linear activation layer (identity layer) is equivalent to a convolution layer with a set weight (special weight) with the current channel convolution kernel parameter being 1 and the remaining channel convolution kernel parameter being 0, which is equivalent to a 1x1 convolution layer, and then a 3x3 convolution layer is obtained through the above convolution kernel conversion operation;

[0080] S52: Fusing the convolution layer (Conv layer) and the batch normalization layer (BN layer);

[0081] The formula of the convolution layer is:

[0082] Conv(x) = W(x) + b

[0083] wherein x is an input; W(x) is a weight function; and b is a bias parameter;

[0084] The batch normalization layer formula is:

[0085]

[0086] wherein both γ and β are parameters learned by backpropagation; mean is a mean function; and var is a variance function;

[0087] The expression obtained after substituting the convolution layer formula into the batch normalization layer formula is:

[0088]

[0089] which can be converted into:

[0090]

[0091] Let:

[0092]

[0093]

[0094] The final fusion result is:

[0095] BN(Conv(x))=W fused (x)+B fused ;

[0096] S53: The combination of the convolution layer (Conv layer) and the nonlinear activation layer (relu layer) is completed through a quantization operation; thus, the feature splicing of the encoder and the corresponding feature of the decoder is completed;

[0097] S54: The original loss function Dice loss function is replaced by the WCE loss function to accelerate the learning process, as follows:

[0098]

[0099] wherein N is the total number of samples; w is the discriminant coefficient of the positive sample; r n represents the label of sample n, with 1 for the positive class and 0 for the negative class; p n represents the probability of sample n being predicted as the positive class;

[0100] The improved Unet model is finally obtained;

[0101] S6: The improved Unet model is iteratively trained using the training data set to obtain a wall and structure deterioration detection model capable of automatically identifying tunnel portal wall cracks;

[0102] S7: input the tunnel portal wall image to be identified into the wall and structure deterioration detection model, extract the cracks, and calculate the crack width and crack length based on the four-way orthogonal skeleton line method;

[0103] S71: after obtaining the cracks from the wall structure deterioration detection model, the singular value decomposition method in the four-way orthogonal skeleton line algorithm is used to calculate the crack skeleton line;

[0104] S72: calculate the normal vector of each position of the crack in the length direction of the crack skeleton line based on the crack skeleton line;

[0105] S73: calculate the crack width and crack length according to the crack edge points and normal vectors of each position.

[0106] In order to further optimize the above technical scheme, the data collected by the unmanned aerial vehicle inspection is preprocessed, the image low frequency component is enhanced by using the fuzzy enhancement method based on the transform domain, the image contrast is improved without distortion, and the image noise is suppressed; the high frequency component and local gradient information of the image are enhanced by using the nonlinear enhancement method, the boundary information in the image is highlighted, the crack edge information is strengthened, and the noise influence is removed.

[0107] In order to further optimize the above technical scheme, in the process of image labeling and data set establishment, for the pictures collected by the unmanned aerial vehicle, a plurality of pictures meeting the crack characteristics of the wall structure are selected to crop the region; then, for the collected sample data, which is less and cannot be effectively trained by the deep learning model, the original crack pictures also need to be data augmented by using rotation, cropping, PS and other means. Then, in view of the problem of less samples, the generated adversarial network GAN is further used for data augmentation. Due to the particularity of crack detection, the image labeling is difficult, therefore, the software Labelimg and Photoshop are combined for labeling, calculation and evaluation are carried out, the number of labeling points required on the 1 cm unit crack length is determined, the crack position information in the wall and protective structure orthographic image is extracted along the edge in the software, the crack position file and image are obtained, and the crack data set is established.

[0108] In order to further optimize the above technical scheme, the Unet semantic segmentation model based on lightweight network is optimized and trained. Since the tunnel portal wall and the protection structure region are large, the number of positive images to be detected each time is large, the data quantity is large, and the traditional Unet semantic segmentation model is slow, therefore, according to the analysis needs, the Unet model is improved, the RepConv module in the VGG network model and the RepConvN module proposed by YOLOv7 are introduced for re-parameterization, the RepConvN module is used to replace part of the conv, the model re-parameterization idea is used, only the multi-branch training is used during training, and after the training is completed, the weight is transferred to the inference network, so that the inference speed is greatly improved, the backbone network of the Unet model is optimized, and the detection efficiency is improved; after the model is established, the wall and protection structure degradation image dataset is input into the improved Unet semantic segmentation model for iterative training, and a wall and protection structure degradation detection model is obtained.

[0109] The improved Unet model structure is as shown in Figure 12 The re-parameterization idea is added, and the RepconvN module is introduced. Compared with the U-NET network, the output result of each layer is first normalized (bantch normalization layer) during downsampling, and then activated by an activation function.

[0110] The RepConvN module in the improved Unet model is as shown in Figure 3 The upper part is a training process schematic diagram, and the lower part is a corresponding inference process schematic diagram. The training process of RepConv is different from the inference process. During training, the outputs of different branches are added, and during inference, the parameters of the branches are re-parameterized as weights to the main branch. The fusion process of the convolution layer is specifically as follows: 1x1 convolution is equivalent to a special (convolution kernel with many 0s) 3x3 convolution, and the identity mapping is a special (1x1 convolution with a unit matrix as the convolution kernel) 1x1 convolution, so it is also a special 3x3 convolution. The parallel 3x3 convolutions are directly added to become a serial convolution structure; the serial structure greatly saves the calculation time compared with the parallel structure.

[0111] In order to further optimize the above technical scheme, the crack feature extraction based on the four-way orthogonal skeleton line algorithm. After the wall and protection structure cracks are obtained by using the improved Unet semantic segmentation model, first, the singular value decomposition module in the four-way orthogonal skeleton line algorithm is used to calculate the crack skeleton line; then, taking the skeleton line as a reference, the normal vector of each position of the crack on the skeleton line length is calculated; finally, the width and length of the crack are calculated and spliced by using the crack edge point at each position and the normal vector.

[0112] Embodiment

[0113] The embodiment proposes a disease detection method for tunnel portal wall structure and slope surface protection structure deterioration based on unmanned aerial vehicle detection data.

[0114] The test method is as follows:

[0115] 1. Applicability

[0116] It is suitable for evaluating the quantitative detection of wall body lines and the development trend monitoring of cracks of railway tunnel portals before and after the flood season. The digital image data of the whole line survey can be converted into indoor identification through post-processing to generate orthographic images, which reduces a large amount of field work and improves work efficiency. In the processing, the key areas of crack occurrence and crack appearance areas can be marked for later attention. In the identification and comparison process, the image data of the local section can also be selected as needed for further analysis of the causes, combined with ground artificial review, to further evaluate the safety of the work equipment and ensure environmental safety. It has strong applicability.

[0117] 2. Test instruments

[0118] 2.1. Unmanned aerial vehicle

[0119] The unmanned aerial vehicle should at least meet the following requirements: (1) the maximum flight time is greater than 95 min; (2) the maximum take-off altitude should be greater than the local altitude and the flight height; (3) the maximum wind speed that can be tolerated should be higher than 15 m / s (7-level wind); (4) the RTK accuracy should be not less than 1 cm+1ppm (horizontal), 1.5 cm+1ppm (vertical); (5) the obstacle sensing range should be as large as possible, and should not be less than 0.5-40 m in front, back, left and right, and 0.5-30 m in up and down; (6) the working environment temperature should be between-30℃ and 70℃.

[0120] 2.2. The unmanned aerial vehicle needs to carry an airborne radar device (to realize automatic navigation, automatic obstacle avoidance, and determine relative position information), and also carries a high-definition camera. The coverage rate of the high-definition camera for aerial photography should be as high as possible as that of the airborne radar device.

[0121] 2.3. Environmental conditions: for steep mountainous areas, in order to avoid shadows, photography should be carried out around noon. At the same time, it is not suitable to use the unmanned aerial vehicle for inspection in heavy fog or rainy season. Heavy fog and rainy season threaten the safety of line operation and seriously affect the accuracy of the results. The ground wind direction determines the direction of take-off and landing of the unmanned aerial vehicle, and the wind direction in the air has a great influence on the stability of the flight platform. Photography and aerial survey should be carried out as much as possible when the wind is small.

[0122] 2.4. Flight height: the flight height is determined according to the project requirements and the terrain and building height of the survey area; the lower the flight height, the higher the resolution. Note: the flight height should be at least 60 meters higher than the measured object.

[0123] The scale of a photograph is defined as the ratio of a line segment on the photograph to the corresponding horizontal line segment on the ground: 1 / m = f / H

[0124] In the formula, H is the height of the relative survey area average level, and f is the distance from the camera center to the image plane vertical distance, that is, the focal length. The selection of the aerial survey scale depends on the mapping scale, which is generally comparable to the mapping scale. After selecting the camera and scale, the flight height can be calculated according to the formula. During flight, the aircraft should fly at the predetermined flight height, and the height difference between each station in the same flight line should not be greater than 40m.

[0125] 3 Preparation

[0126] 3.1 Outdoor investigation stage: ① Find the location of each workshop of the maintenance section to facilitate the take-off and landing of the UAV and charging; ② Use a small UAV to survey the terrain and confirm the height of the mountain within 2km of the flight area, the location of the power tower, and the higher buildings of the villagers; ③ Inquire about the no-fly zone of the relevant city area and estimate the flight height of the UAV.

[0127] 3.2 Indoor preparation stage: ① Plan the flight route using the Waypoint Master simulation flight software: set the lateral overlap to 45% and the forward overlap to 65%.(The forward overlap is generally specified as 60%, the minimum is not less than 43%, and the maximum is not greater than 74%; the lateral overlap is generally specified as 30%, the minimum is not less than 14%, and the maximum is not greater than 40%.) After the flight route is planned, enter the three-dimensional satellite map to verify the feasibility of the flight route. ② Check the equipment: before the operation, check whether there is damage to each part of the UAV body, whether the signal connection between the body and the gimbal is normal, and whether the battery power and fuel are sufficient, etc.

[0128] 3.3 Data collection: ① Base station erection, collection of static data, and acquisition of fixed point coordinate data; ② Equipment mounting and power-on inspection; ③ Flight route execution for high-definition image shooting.

[0129] 3.4 Data arrangement: ① After completing the flight, stand still in an open area for 5 minutes; if a camera is configured, observe the camera indicator light, and after the indicator light is extinguished, power off and shut down the equipment; ② Check whether the POS data and radar data are normally stored; ③ After the on-board equipment is turned off for 5 minutes, stop the static data collection and copy the static observation data; ④ After completing the on-site data collection, complete the data numbering arrangement immediately on the same day to prepare for data processing.

[0130] 3.5 Four principles of special flight cases: ① Crossing routes: In principle, each adjacent parallel route needs one or two crossing routes to vertically cross, which is used to ensure the connection accuracy of the route in the later data processing. ② Small cross: In order to accurately correct the data of each flight, a small cross flight is required before or after the formal route data collection of each flight. ③ Supplementary flight: There may be a small amount of abnormal situation in the data acquisition process, which may cause the data acquisition of a certain area to be missing. For the missing data of the route, supplementary flight is required, and the two ends of the supplementary flight route should be appropriately extended, so that the data obtained twice can be well connected. ④ Data anomaly: In the flight process, there may be a small amount of satellite signal instantaneous loss phenomenon, or due to the relatively poor flight conditions, such as large wind or updraft, which causes the attitude of the aircraft to change quickly, and the satellite signal is not good. For the POS data caused by this kind of situation, the effectiveness of the route data should be determined according to whether the data abnormal time period is on the formal route and the POS data accuracy of the whole flight. For the invalid data route or flight, supplementary flight will be carried out. Each flight of this project needs to obtain valid data.

[0131] 4. Crack identification

[0132] (a) The original data is enhanced, and the data generated by the generative adversarial network needs to be processed twice;

[0133] (b) The iteration number of the improved Unet model is at least 30 times, and the optimal is between 30-50 times;

[0134] (c) The iteration error threshold should be set as small as possible, and 1x10 -8 It is OK;

[0135] (d) According to the characteristics of the cracks of the tunnel wall, the model is continuously trained using multi-period detection results, and the model parameters are optimized. The actual recognition success rate should be ≥95%, and the range considering special cases should also be ≥85%, and the detection accuracy is 2mm.

[0136] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between each embodiment can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the related parts can be referred to the method part.

[0137] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for quantitatively detecting cracks in a railway tunnel portal wall based on a UAV, characterized in that, The method comprises the following steps: Step 1: collecting a tunnel portal image; Step 2: performing noise reduction processing and preprocessing on the image; Step 3: combining a GAN model to expand the processed image data to obtain an original data set; Step 4: using a labeling tool to label the wall cracks in the original data set to establish a training data set; Step 5: according to the characteristics of the wall cracks, using a YOLOv7 reparameterization module in the Unet model to perform reparameterization to obtain an improved Unet model; the specific process of obtaining the improved Unet model is as follows: Step 51: replace the long connection form of the encoder and decoder of the Unet model with a residual connection form, add a 1x1 convolution branch in parallel in the residual branch part of the Unet model, combine the down-sampling, residual branch and added convolution branch of the Unet model into a single path to obtain a RepConvN module; When combining, expand the 1x1 convolution into a 3x3 convolution with 0 on the four sides, use a linear activation layer as a 1x1 convolution layer with a set weight, and then perform a convolution kernel conversion operation to obtain a 3x3 convolution layer; Step 52: fuse the convolution layer and the batch normalization layer; The formula of the convolution layer is: Conv(x) = W(x) + b Wherein, x is the input; W(x) is the weight function; b is the bias parameter; The formula of the batch normalization layer is: Wherein, γ and β are both parameters learned by back propagation; mean is the mean function; var is the variance function; The expression obtained by substituting the convolution layer formula into the batch normalization layer formula is: Unfolded as: Let: Then the final fusion result is: BN (Conv(x)) = W fused (x) + B fused ; Step 53: merge the convolution layer and the nonlinear activation layer through quantization operation; Step 54: use a WCE loss loss function for the loss function of the Unet model to accelerate the learning process, as follows: where N is the total number of samples; w is the discriminant coefficient of positive samples; r n represents the label of sample n, 1 for positive class and 0 for negative class; p n represents the probability that sample n is predicted as a positive class; Obtain the improved Unet model; Step 6: iteratively train the improved Unet model using the training data set to obtain a wall and structure deterioration detection model; Step 7: input the tunnel portal wall image to be recognized into the wall and structure deterioration detection model, extract the cracks, and calculate the crack width and crack length based on the four-way orthogonal skeleton line method.

2. The unmanned aerial vehicle-based railway tunnel portal wall crack quantitative detection method according to claim 1, characterized in that, The preprocessing includes enhancing the low-frequency components of the image based on the transform domain blur enhancement method, and enhancing the high-frequency components and local gradient information of the image using a nonlinear enhancement method.

3. The unmanned aerial vehicle-based railway tunnel portal wall crack quantification detection method according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 31: select a number of images in the image data that meet the characteristics of wall cracks, and rotate and crop the crack area to obtain primary expanded image data; Step 32: perform secondary data expansion on the primary expanded image data using a generative adversarial network to obtain an original data set.

4. The unmanned aerial vehicle-based railway tunnel portal wall crack quantification detection method according to claim 1, characterized in that, The specific process of calculating the crack width and crack length in step 7 is as follows: Step 71: calculate the crack skeleton line according to the crack using the singular value decomposition method in the four-way orthogonal skeleton line algorithm; Step 72: calculate the normal vector of each position of the crack in the length direction of the crack skeleton line according to the crack skeleton line; Step 73: Calculate the crack width and crack length according to the crack edge points and normal vectors of each position.

Images