Suction bucket uplift soil damage mode prediction method based on GAN-BP algorithm

Abstract

The invention provides a suction bucket up-pull soil damage mode prediction method based on a GAN-BP algorithm. The method comprises the steps that a suction anchor up-pull visual experiment device is established; carrying out a suction anchor pull-up experiment under different working conditions by using the experiment device, and obtaining laser speckle image original videos of the suction anchor in different loading rate processes; processing and calculating the original video of the laser speckle image to obtain a displacement contour map of the soil body around the suction bucket when the upward pulling force is maximum; measuring the width and depth of a land fracture surface through a land displacement contour map, and establishing a land'factor-damage size 'database; establishing a GAN-BP machine learning algorithm, and performing small sample data prediction; according to the method, a GAN-BP machine learning algorithm is established in consideration of a small sample database, a suction bucket uplift soil mass failure mode prediction model is established based on the algorithm, actual measurement data are obtained through an experimental device, small sample expansion is performed, and therefore the influence rule of the length, the diameter and the uplift rate of a suction bucket on the size of a soil mass fracture surface is obtained.

Description

Technical Field

[0001] The invention relates to the technical field of suction bucket foundations, and in particular to a method for predicting a failure mode of soil lifted by a suction bucket based on a GAN-BP algorithm. Background Art

[0002] Offshore wind farms will develop in deeper waters in the future, and floating wind turbines are an inevitable choice for future development. Offshore wind power will inevitably move from offshore to offshore, from shallow sea to deep sea. Floating wind turbines are connected to deep-sea suction anchors through anchor chains to withstand the uplift force caused by wind and waves. The marine environment is complex and changeable, often accompanied by severe typhoons and complex sea conditions. When a storm surge comes, the overturning moment of the offshore wind power foundation is very large, and the corresponding anchor cable transmits a large uplift force to the suction anchor. Therefore, it is very necessary to study and determine the uplift limit bearing capacity of the suction anchor, which provides a theoretical basis for the installation and design of wind power foundations in the deep sea. It is of considerable engineering significance to study the mechanical properties of deep-sea suction anchors under storm surges. Although a lot of research has been done on the uplift bearing characteristics of suction anchors, few people have focused on studying the deformation of the soil inside and outside the suction anchor during the uplift process and the size of the rupture surface. Summary of the invention

[0003] The object of the present invention is to provide a prediction method which can focus on studying the deformation of soil inside and outside the suction anchor and the size of the fracture surface during the pulling-up process.

[0004] To achieve the above object, the present invention proposes a method for predicting the failure mode of soil body lifted by suction bucket based on GAN-BP algorithm, which comprises the following steps:

[0005] S1: Establish a visual experimental device for suction anchor pull-out;

[0006] S2: Using the experimental device, a suction anchor pull-up experiment is performed under different working conditions to obtain the original video of the laser speckle image of the suction anchor during different loading rates;

[0007] S3: Processing and calculating the original video of the laser speckle image to obtain a displacement contour map of the soil around the suction bucket when the maximum upward pulling force is applied;

[0008] S4: measuring the width and depth of the land rupture surface through the land displacement contour map, and establishing a land "factor-damage size" database;

[0009] S5: Establish a GAN-BP machine learning algorithm to predict small sample data;

[0010] S6: Based on the prediction results, analyze the impact trend of different factors on the foundation failure surface.

[0011] Furthermore, in step S1, the suction anchor pulling visualization experimental device includes a suction anchor system, an image acquisition system and a control mechanism: the suction anchor system includes a hollow barrel suction anchor with a closed top and an open bottom, and a drainage and exhaust hole with a valve is provided on the top anchor cover of the hollow barrel to discharge the water and air trapped in the anchor during the sinking process. Before pulling up, the valve is closed to seal the drainage and exhaust hole. A pull rod is connected to the center of the anchor cover, and the pull rod is connected to the drive mechanism; a tension sensor is installed on the pull rod; the hollow barrel is fixed in a glass box, and the glass box is filled with transparent sand; the glass box is set in the reaction frame; the drive mechanism, the tension sensor and the reaction frame constitute a loading system; the control mechanism is connected to the loading system and the image acquisition system by signal.

[0012] Furthermore, the image acquisition system includes a camera mechanism and a light source mechanism, the light source mechanism includes a laser emitter, a straight line lens and a laser power supply group, the laser power supply group is connected to the laser emitter cable, the laser emitter is facing the suction anchor through a straight line lens, and generates a fan-shaped laser speckle image; the camera mechanism is arranged outside the reaction frame and facing the side of the glass box, and is used to shoot the laser speckle image; the image acquisition system is connected to the control mechanism, and the control mechanism includes an image processing system, and the image processing system uses PIVlab software to process images in an incremental sequence manner. After the image is processed, all incremental image sequence result cloud maps before the anchor reaches the maximum pull-out force are superimposed to obtain the displacement field of the suction anchor at the moment of maximum pull-out force.

[0013] Furthermore, the reaction frame is welded from square steel pipes, which plays a role in stabilizing and transmitting the reaction force; the driving mechanism is a linear servo electric cylinder, which is used to stably provide different speeds to drive the pull rod to move, the pull rod lifting range is 0-300mm, the tension range is 0-1000N, and the moving speed range is 1-500mm / s; the control mechanism is a computer, and the moving speed and lifting position can be adjusted through the touch panel on the control box (computer). A threaded hole is provided at the center of the anchor cover, and the pull rod is connected through the threaded hole.

[0014] Furthermore, step S2 is specifically as follows:

[0015] S2.1: Design working condition, the parameters that can be changed in the working condition include the inner diameter of the suction bucket, the outer diameter of the suction bucket, the length of the suction bucket and the upward pull force loading rate;

[0016] S2.2: Fill the prepared transparent sand into the plexiglass box;

[0017] S2.3: Use the driving mechanism to penetrate the suction anchor into the transparent sand in the glass box. Specifically, first, open the suction anchor drainage valve and adjust the position of the bottom glass box so that the suction anchor is at the center of the plane of the glass box; second, use the servo electric cylinder to penetrate the suction anchor into the specified position in the transparent sand;

[0018] S2.4: After the suction anchor penetrates the transparent sand, let it stand for a while;

[0019] S2.5: Arrange the land visualization system, start and adjust the image acquisition system, and complete the acquisition preparation work;

[0020] S2.6: Pull the suction anchor upwards, and record the pulling-up process of the suction anchor in real time, and record the original laser speckle image video and tension sensor readings of the suction anchor at different loading rates;

[0021] Furthermore, in step S2.5, the land visualization system is specifically arranged as follows: the position of the transparent glass box is adjusted so that the suction anchor is directly opposite to the center of the plane of the model box; the distance between the laser emitter and the glass box is arranged and adjusted so that the laser can completely illuminate the interior of the transparent sand after passing through the straight optical prism, generating a stable fan-shaped laser speckle image. The lens aperture size and ISO value of the camera mechanism are adjusted so that the obtained image is clear and visible.

[0022] Furthermore, in step S2.5, the collection preparation work is specifically as follows: start the camera mechanism to record the video, prepare to record the laser speckle image of the suction anchor during different loading rates; connect the data acquisition instrument of the tension sensor to the computer, start the computer, start the data acquisition system, and prepare to record the pulling force on the top of the suction anchor during the pulling process in real time.

[0023] Furthermore, in step S2.5, the method of pulling out the suction anchor is displacement controlled, and the pulling-up speed is controlled by a driving mechanism, and the suction anchor is pulled up at different loading rates, so that the suction anchor is pulled up at different loading rates respectively, and the reading of the tension sensor is recorded by a microcomputer.

[0024] Furthermore, step S3 is specifically as follows:

[0025] S3.1: After converting the original laser speckle image video to black and white, batch extract each frame of the video;

[0026] S3.2: Import the frames of the laser speckle slices extracted at different times and working conditions into the image processing and analysis software for image processing and calculation;

[0027] S3.3: After exporting the calculated data, the velocity field, displacement field and displacement vector diagram of the soil during the pull-up process are plotted, and then the displacement contour map of the land around the suction anchor cylinder at the maximum pull-up force is obtained through programming calculation.

[0028] Furthermore, in step S4, the factors of the "factor-failure size" database include the suction bucket length, the suction bucket diameter and the pull-up rate; the failure size of the "factor-failure size" database includes the rupture surface width and the rupture surface depth.

[0029] Furthermore, in step S5, the GAN-BP machine learning algorithm includes a GAN module and a BP module; the method for establishing the GAN-BP machine learning algorithm model is:

[0030] S5.1: Generate high-quality data samples through the GAN module, so that the generated data is close to the real data distribution;

[0031] S5.2: Combine the high-quality data samples generated by the GAN module with the original data to form a new combined training set;

[0032] S5.3: Use the combined training set to train the BP module and establish a GAN-BP hybrid model.

[0033] Furthermore, the GAN module is a generative adversarial network, and the BP module is a BP neural network;

[0034] Furthermore, in step S5.1, the method for generating high-quality data samples through the GAN module is:

[0035] S5.11: Define the generator G(z; θ G ) and the discriminator D(x;θ D );

[0036] Generator G(z;θ G ): Input random noise z~p z (z), generate sample x fake =G(z);

[0037] Discriminator D(x;θ D ): Input data x, the probability that the output data is true D(x)

[0038] S5.12: Train the discriminator D and the generator G by optimizing the loss function until the samples generated by the generator G are difficult to distinguish;

[0039] The goal of the GAN module is to maximize the ability of the discriminator D to distinguish between real data and generated data, while minimizing the probability of the generator G being recognized by the discriminator; therefore, its objective function is:

[0040]

[0041] By optimizing the above loss function, train D and G until the samples generated by the generator are difficult to distinguish;

[0042] S5.13: After training, use the generator G to input random noise z to generate augmented samples x augmented , the expanded sample x augmented That is, it is the expanded data sample of the “factor-damage size” database obtained in step S4.

[0043] The training method of the BP module is:

[0044] S5.31: Prepare the dataset, X train =[X real ,X augmented ],Y train =[Y real ,Y augmented ];

[0045] S5.32: define a BP neural network, wherein the BP neural network comprises an input layer, a hidden layer and an output layer;

[0046] enter:

[0047] Output:

[0048] Parameters: weight matrix W and bias b;

[0049] S5.33: Forward propagation, through the activation function f, calculate the values ​​of the hidden layer and the output layer:

[0050] h=f(W hidden x+b hidden )

[0051]

[0052] Where, f is the activation function, which is used to introduce nonlinearity, such as Sigmoid and ReLU; h is the output value of the hidden layer; W is hidden : The weight matrix from the input layer to the hidden layer; b hidden : Bias vector of the hidden layer; W output : The weight matrix from the hidden layer to the output layer; b output : Bias vector of the output layer;

[0053] S5.34: Use mean square error MSE as the loss function:

[0054]

[0055] In the formula, Loss function; N: the total number of training data samples; y i : The true value of the i-th sample; The predicted value of the i-th sample;

[0056] S5.35: Backpropagation, optimizing network parameters by gradient descent.

[0057]

[0058] Where, η: learning rate, controlling the step size of parameter update; The partial derivative of the loss function with respect to the weight matrix W; The partial derivative of the loss function with respect to the bias vector b.

[0059] Furthermore, in step S6, the influence trend of different factors on the foundation damage surface is analyzed, and the development law of the foundation damage envelope surface is comprehensively judged by image analysis and machine learning methods.

[0060] Compared with the prior art, the advantages of the present invention are:

[0061] 1. The present invention establishes a GAN-BP machine learning algorithm taking into account a small sample database, and establishes a suction bucket pull-out soil failure mode prediction model based on the algorithm. The measured data is obtained through a suction anchor pull-out visualization experimental device and a small sample expansion is performed to obtain the influence of the suction bucket length, diameter and pull-out rate on the soil rupture surface size.

[0062] 2. Taking into account the limitations of experimental equipment and conditions, the present invention establishes a hybrid machine learning method for small sample databases by integrating the advantages of the BP neural network algorithm for the good mapping relationship between multiple inputs and outputs and the GAN algorithm for expanding the small sample problem.

[0063] 3. The present invention adopts a machine learning method to comprehensively consider various factors to predict the foundation bearing capacity, and controls a single variable to explain the foundation bearing capacity, and explains the suction bucket pulling foundation bearing capacity from the perspective of a single factor and multiple factors coupling.

[0064] 4. The present invention sets up a suction anchor pull-up visualization experimental device and uses experimental data as a machine learning sample library, thereby avoiding the disadvantage that the prediction results are inconsistent with the actual results caused by using numerical simulation results as a sample library. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A schematic diagram of a flow chart of a method for predicting failure modes of soil bodies lifted by suction buckets based on a GAN-BP algorithm proposed in an embodiment of the present invention;

[0066] Figure 2 Schematic diagram of a suction anchor pull-out visualization test device in an embodiment of the present invention;

[0067] Figure 3A cross-sectional view of a drum suction anchor structure of a visualization test device in an embodiment of the present invention;

[0068] Figure 4 It is a contour map of the land displacement around the suction bucket corresponding to the suction anchor at different loading speeds in the embodiment method of the present invention;

[0069] Figure 5 (a) and (b) are schematic diagrams of the local failure mode and the overall failure mode of the drum suction anchor in the method of the embodiment of the present invention, respectively;

[0070] Figure 6 (a) and (b) are schematic diagrams showing comparison results of predicted values ​​and actual values ​​of fracture surface width B and fracture surface depth H by GAN-BP algorithm in the method of the embodiment of the present invention respectively;

[0071] Figure 7 (a) and (b) are schematic diagrams showing the comparison between the predicted value and the actual value of the fracture surface width B and the fracture surface depth H using the BP algorithm in the comparative method. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0073] In the following embodiments, BP refers to BP neural network (Back Propagation Artificial Network), and GAN refers to Generative Adversarial Network (Generative Adversarial Network).

[0074] This embodiment proposes a method for predicting the failure mode of soil body lifted by suction bucket based on GAN-BP algorithm. Figure 1 As shown, the method comprises the following steps:

[0075] S1: Establish a visual experimental device for suction anchor pull-out;

[0076] In this embodiment, Figure 2 As shown, the suction anchor pull-up visualization experimental device includes a suction anchor system, an image acquisition system and a microcomputer 11:

[0077] The suction anchor system includes a hollow thin-walled barrel suction anchor 2 with a closed top and an open bottom. The hollow thin-walled barrel suction anchor 2 is made of a series of acrylic organic glasses, such as Figure 3As shown, the top anchor cover of the hollow barrel suction anchor 2 is provided with a level gauge 23 and a drainage and exhaust hole 21 with a valve, which are used to discharge the water and air trapped in the anchor during the sinking process. Before pulling up, the valve is closed to seal the drainage and exhaust hole 21. A threaded hole 22 with a diameter of 8 mm is provided at the center of the anchor cover, through which a pull rod is connected, and the top of the pull rod is connected to the linear servo electric cylinder 6 through a flange 4; a tension sensor 5 is also installed on the pull rod, and the tension sensor 5 is connected to the tension sensor data acquisition instrument 8 through a cable; the suction anchor 2 is fixed in an acrylic organic glass box 1 of 18 cm×18 cm×40 cm, the acrylic organic glass box 1 has a wall thickness of 1 cm, and transparent sand 3 is filled inside. The glass box 1 is fixed in a metal reaction frame 9, and a guide rod 7 is fixed at the top of the reaction frame 9 to control the relative position of the suction barrel in the soil.

[0078] The linear servo electric cylinder 6, the tension sensor 5 and the metal reaction force frame 9 constitute a loading system, and the microcomputer 11 is connected with the loading system and the image acquisition system by signals.

[0079] In this embodiment, the reaction frame 9 is welded from square steel pipes and plays a role in stabilizing and transmitting the reaction force; the linear servo electric cylinder 6 is connected to the pull rod drive to stably provide different speeds to drive the pull rod to move. The pull rod lifting range is 0 to 300 mm, the tension range is 0 to 1000 N, and the moving speed range is 1 to 500 mm / s; the moving speed and lifting position can be controlled and adjusted by the microcomputer 9; the tension sensor has a range of 500 N, an accuracy of 0.01 N, and the sampling frequency of the tension data acquisition card is 10 Hz.

[0080] like Figure 2 As shown, the image acquisition system includes a Canon camera 10 and a light source mechanism, the light source mechanism includes a laser emitter 12, a straight line lens 13 and a laser power supply group 14, wherein the laser power supply group 14 is connected to the laser emitter 12 through a cable, and the laser emitter 12 is facing the suction anchor 2 in the glass box 1 through the straight line lens 13 to generate a fan-shaped laser speckle image; the camera 10 is arranged outside the side of the reaction frame 9 and facing the side of the glass box 1, and is used to shoot the laser speckle image generated by the light source mechanism. The image acquisition system is connected to the microcomputer 11 for signal control, and the control system of the microcomputer 11 includes an image processing system. The image processing system conducts a soil visualization study of a 1g suction anchor high-speed pull-out model test based on transparent soil and PIV technology, and uses PIVlab software to process images in an incremental sequence. After processing the images, all incremental image sequence result cloud maps before the suction anchor 2 reaches the maximum pull-out force are superimposed, so as to obtain the displacement field of the suction anchor 2 at the maximum pull-out moment.

[0081] S2: Using the above experimental device, suction anchor pull-up experiments are carried out under different working conditions to obtain the original video of the laser speckle image of the suction anchor during different loading rates.

[0082] The specific implementation of step S2 is as follows:

[0083] S2.1: Design working condition, the parameters that can be changed in this working condition include the inner diameter of the suction bucket, the outer diameter of the suction bucket, the length of the suction bucket and the upward pull force loading rate;

[0084] S2.2: Fill the prepared transparent sand 3 into the organic acrylic glass box 1;

[0085] S2.3: Use the linear servo electric cylinder 6 to penetrate the suction anchor 2 into the transparent sand 3 in the glass box 1. The specific method is as follows: first, open the suction anchor drainage valve 21, adjust the position of the bottom glass box 1, so that the suction anchor 2 is facing the center of the bottom plane of the glass box 1; second, use the servo electric cylinder 6 to penetrate the suction anchor 2 into the specified position in the transparent sand 3.

[0086] In this embodiment, five suction anchors are selected and numbered M-1, M-2, M-3, M-4, and M-5, respectively. The average diameter L and the side wall thickness t 1、 The top wall thickness t2 parameters are shown in Table 1.

[0087] Table 1 Suction anchor size parameter table

[0088]

[0089] S2.4: After the suction anchor 2 penetrates the transparent sand 3, it is left to stand for a period of time;

[0090] S2.5: Arrange the land visualization system, start and adjust the image acquisition system, and complete the acquisition preparation work.

[0091] In this embodiment, the land visualization system is arranged as follows: the position of the transparent glass box 1 is adjusted so that the suction anchor 2 is directly opposite to the center of the bottom plane of the glass box 1; the distance between the laser emitter 12 and the glass box 1 is arranged and adjusted so that the laser can completely illuminate the inside of the transparent sand 3 after passing through the straight optical prism 13, thereby generating a stable fan-shaped laser speckle image. The lens aperture size and ISO sensitivity value of the camera 10 are adjusted so that the image obtained is clear and visible;

[0092] The collection preparation work is specifically as follows: start the camera 10 to record video, prepare to record the laser speckle image of the suction anchor 2 during different loading rates; connect the tension sensor data acquisition instrument 8 to the computer 10 via a cable, start the computer 10, start the data acquisition system, and prepare to record the top pulling force of the suction anchor 2 during the pulling process in real time.

[0093] S2.6: Pull the suction anchor 2 upward, and record the pulling process of the suction anchor 2 in real time, and record the original laser speckle image video and tension sensor reading of the suction anchor 2 in different loading rates. The method of pulling the suction anchor 2 is displacement control, and the pulling speed is controlled by the linear servo electric cylinder 6, and the suction anchor 2 is pulled up at different loading rates, so that the suction anchor 2 is pulled up at different loading rates respectively, and the microcomputer 10 records the reading of the tension sensor 5, as shown in Table 2 below:

[0094] Table 2 Record of pull-up loading rate of different suction anchors

[0095]

[0096] S3: Process and calculate the original video of the laser speckle image to obtain the displacement contour map of the soil around the suction anchor barrel when the maximum pull-out force is reached;

[0097] S3.1: The suction anchor pulling process is recorded by the camera 10, and the original video of the laser speckle image is obtained. The original video of the laser speckle image is converted to black and white by using the Pr editing software, and then each frame of the video is batch extracted by using the FreeVideoToJPGConverter software;

[0098] S3.2: Import the frames of the laser speckle slices extracted at different times and working conditions into the image processing and analysis software PIVlab for image processing and calculation;

[0099] S3.3: After exporting the calculated data, use Origin software to draw the soil velocity field, displacement field, and displacement vector diagram during the pull-up process, and then use MATLAB programming to calculate and obtain the maximum pull-up force F max Contour map of land displacement around the suction anchor barrel.

[0100] like Figure 4 As shown in the figure, the displacement contour map of the land around the suction bucket corresponding to the suction anchor with diameter D = 50mm and length L = 25mm at different loading speeds, where: Figure 4 (a) is the contour map of the land displacement around the suction bucket corresponding to the suction anchor when the loading speed v = 1 mm / s; Figure 4 (b) is the contour map of the land displacement around the suction bucket corresponding to the suction anchor when the loading speed v = 12.5 mm / s; Figure 4 (c) is the contour map of the land displacement around the suction bucket corresponding to the suction anchor when the loading speed v = 25 mm / s; Figure 4 (d) is the contour map of the land displacement around the suction bucket corresponding to the suction anchor when the loading speed v = 50 mm / s; Figure 4(e) is the contour map of the land displacement around the suction bucket corresponding to the suction anchor when the loading speed v = 100 mm / s.

[0101] S4: Through the land displacement contour map, the width and depth of the land rupture surface are measured, and a "factor-failure size" database is established. The "factors" of the "factor-failure size" database include the suction bucket length, suction bucket diameter and pull-up rate; the "failure size" includes the rupture surface width B and the rupture surface depth H.

[0102] Table 3 below records the “factor-failure size” data of the suction anchor M-3 (diameter D=50mm, length L=25mm) and the suction anchor M-1 (average diameter D=50mm, length L=50mm) at pull-up rates of V=12.5mm, V=25mm, V=50mm, and V=100mm, respectively.

[0103] Table 3 “Factor-Failure Size” Database (Part)

[0104]

[0105] S5: Establish a GAN-BP machine learning algorithm to predict small sample data;

[0106] In this embodiment, the GAN-BP machine learning algorithm includes a GAN module and a BP module. The method for establishing the GAN-BP machine learning algorithm model is:

[0107] S5.1: Generate high-quality data samples through the GAN module so that the generated data is close to the real data distribution. The method is as follows:

[0108] S5.11: Define the generator G(z; θ G ) and the discriminator D(x;θ D );

[0109] Generator G(z;θ G ): Input random noise z~p z (z), generate sample x fake =G(z);

[0110] Discriminator D(x;θ D ): Input data x, the probability that the output data is true D(x)

[0111] S5.12: Train the discriminator D and the generator G by optimizing the loss function until the samples generated by the generator G are difficult to distinguish;

[0112] The goal of the GAN module is to maximize the ability of the discriminator D to distinguish between real data and generated data, while minimizing the probability of the generator G being recognized by the discriminator; therefore, its objective function is:

[0113]

[0114] By optimizing the above loss function, train D and G until the samples generated by the generator are difficult to distinguish;

[0115] S5.13: After training, use the generator G to input random noise z to generate augmented samples x augmented , the expanded sample x augmented This is the expanded data sample obtained in step S4, and some of the data are shown in Table 4:

[0116] Table 4 GAN expansion samples (part)

[0117] Diameter D(mm) Length L(mm) Loading rate v(mm / s) Fracture surface width B(mm) Fracture surface depth H(mm) 44.31518173 68.41151428 28.2653923 84.93772888 90.00748444 54.74298096 57.27037048 22.36504173 83.10439301 84.56375885 45.3184166 61.41879654 11.4386301 66.49040985 81.75670624 40.28317261 60.88147736 27.43096733 83.2386322 76.64796448 52.46657562 73.30093384 80.52485657 78.88573456 112.2959061 51.99222183 50.20989227 21.87613678 79.08688354 79.97119904 43.28781128 52.85329819 21.70853233 79.03794861 81.4047699 47.32524109 47.0251503 52.17623138 88.63056183 86.86373138 44.13137436 61.69446182 20.72089386 81.01996613 80.69803619 57.11587906 49.82167053 22.88331604 83.21014404 82.46383667 57.58080292 54.06962585 12.35421562 73.38467407 80.63221741 43.01916122 62.54060745 21.55706215 82.27147675 77.31987 58.78965378 49.02116776 39.20677185 89.68482971 90.51895142 47.80777359 35.50246429 95.53018951 81.99420929 69.41500092 43.06548309 51.82941055 28.85625458 85.93470764 75.40726471 55.61429977 60.53258133 10.71063805 73.97647858 83.76602173 54.35240173 19.05724716 99.8663559 65.19730377 54.34417725 49.18434906 56.70740509 24.64125061 83.2882843 79.88928223 47.5256691 71.38578796 40.03570557 70.89803314 103.9092407 44.07012558 35.29263306 56.65210724 67.41448975 57.59306717

[0118] S5.2: Combine the high-quality data samples generated by the GAN module with the original data to form a new combined training set X train =[X real ,X augmented ],Y train =[Y real ,Y augmented ], X train is the input data, including the suction bucket length, diameter and pull-up rate; Y train It is the output data, including the fracture surface width and fracture surface depth;

[0119] S5.3: Use the above combined training set to train the BP module and establish a GAN-BP hybrid model. The method is as follows:

[0120] 1) Define a BP neural network, wherein the BP neural network comprises an input layer, a hidden layer and an output layer;

[0121] enter:

[0122] Output:

[0123] Parameters: weight matrix W and bias b;

[0124] 2) Forward propagation, through the activation function f, calculate the values ​​of the hidden layer and the output layer:

[0125] h=f(W hidden x+b hidden )

[0126]

[0127] Where, f is the activation function, which is used to introduce nonlinearity, such as Sigmoid and ReLU; h is the output value of the hidden layer; W ishidden : The weight matrix from the input layer to the hidden layer; b hidden : The bias vector of the hidden layer; W output : The weight matrix from the hidden layer to the output layer; b output : The bias vector of the output layer;

[0128] 3) The mean squared error MSE is used as the loss function:

[0129]

[0130] In the formula, The loss function; N: The total number of training data samples; y i : The true value of the i-th sample; The predicted value of the i-th sample;

[0131] 4) Backpropagation is performed to optimize the network parameters through gradient descent.

[0132]

[0133] In the formula, η: The learning rate, which controls the step size of parameter update; The partial derivative of the loss function with respect to the weight matrix W; The partial derivative of the loss function with respect to the bias vector b.

[0134] S6: According to the prediction results, analyze the influence trends of different factors on the foundation failure surface, and comprehensively judge the development law of the foundation failure envelope surface by using image analysis and machine learning methods.

[0135] From the displacement contour maps under different working conditions as shown above Figure 4 It can be seen that when the suction anchor is pulled out at different loading rates, the soil failure mode can be roughly divided into the overall failure mode and the local failure mode. When the pulling-out rate v = 1 mm / s, there is basically no vertical displacement of the soil around the suction bucket, and the failure mode of the suction bucket is the local failure mode. A clear shear rupture zone can be seen along the bucket wall, as shown in Figure 5 (a). When the pulling-out rate of the suction bucket v = 100 mm / s, the soil plug in the suction bucket and part of the foundation soil at the bottom of the bucket will be pulled out together with the anchor, as shown in Figure 5 (b). The failure mode of the anchor is the overall failure mode. At this time, the suction bucket is in the overall failure mode during pulling out, and its peak pulling-out force reaches the undrained ultimate bearing capacity. When 1 mm / s < v < 100 mm / s, the suction bucket is in a partial drainage state. As the loading rate increases, the soil flow range caused by the bucket pulling out also continuously increases.

[0136] The prediction results of the GAN-BP machine learning algorithm model in this embodiment are compared with the measured data. The sample number obtained by sampling any sample in the database is used as the horizontal axis, and the output value of the GAN-BP algorithm is and the true value y i is the vertical axis, such as Figure 6 The comparison results are shown in Figure 1, where Figure 6 (a) is the comparison result of the predicted value and measured value of the fracture surface width B by GAN-BP algorithm. Figure 6 (b) is the comparison result of the predicted value and measured value of the fracture surface depth H by GAN-BP algorithm. The mean square error of the test set is 0.0189 and the mean absolute error is 0.1161.

[0137] Comparative Example

[0138] In the comparative example, the original samples before GAN data expansion are directly used for BP neural network fitting, and the results of comparing the predicted values ​​with the measured values ​​are as follows: Figure 7 As shown, Figure 7 (a) is the comparison result of the predicted value and measured value of the fracture surface width B by BP algorithm. Figure 7 (b) is the comparison result of the predicted value and measured value of the fracture surface depth H by BP algorithm. The mean square error of the calculated test set is 0.0910 and the mean absolute error is 0.2617.

[0139] Therefore, it can be seen that the use of the GAN-BP hybrid algorithm greatly improves the robustness and accuracy of the model, and can better explain the mapping relationship between the suction bucket diameter, length and loading rate and the size of the foundation damage envelope surface, which can serve as a reference for the design of foundation bearing capacity.

[0140] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A method for predicting the failure mode of soil body lifted by suction bucket based on GAN-BP algorithm, characterized in that: The steps include: S1: Establish a visual experimental device for suction anchor pull-out; S2: Using the experimental device, a suction anchor pull-up experiment is performed under different working conditions to obtain the original video of the laser speckle image of the suction anchor during different loading rates; S3: Processing and calculating the original video of the laser speckle image to obtain a displacement contour map of the soil around the suction bucket when the maximum upward pulling force is applied; S4: measuring the width and depth of the land rupture surface through the land displacement contour map, and establishing a land "factor-damage size" database; S5: Establish a GAN-BP machine learning algorithm to predict small sample data; S6: Based on the prediction results, analyze the impact trend of different factors on the foundation failure surface.

2. The method for predicting the failure mode of soil body lifted by suction bucket according to claim 1 is characterized in that: In step S1, the suction anchor pull-up visualization experimental device includes a suction anchor system, an image acquisition system and a control mechanism: The suction anchor system comprises a hollow barrel suction anchor with a closed top and an open bottom. A drainage and exhaust hole with a valve is provided on the top anchor cover of the hollow barrel. A pull rod is connected to the center of the anchor cover, and the pull rod is connected to the driving mechanism. A tension sensor is installed on the pull rod. The hollow barrel is fixed in a glass box, and the glass box is filled with transparent sand. The glass box is arranged in the reaction frame. The driving mechanism, the tension sensor and the reaction force frame constitute a loading system; The control mechanism is signal-connected to the loading system and the image acquisition system.

3. The method for predicting failure modes of soil bodies lifted by suction buckets according to claim 2 is characterized in that: The image acquisition system includes a camera mechanism and a light source mechanism, the light source mechanism includes a laser emitter, a straight line lens and a laser power supply group, the laser power supply group is connected to the laser emitter cable, the laser emitter is facing the suction anchor through the straight line lens, and generates a fan-shaped laser speckle image; the camera mechanism is arranged outside the reaction force frame and facing the side of the glass box, and is used to shoot the laser speckle image; the image acquisition system is connected to the control mechanism, and the control mechanism includes an image processing system, and the image processing system uses PIVlab software to process images in an incremental sequence to obtain the displacement field of the suction anchor at the moment of maximum upward pulling force.

4. The method for predicting failure modes of soil bodies lifted by suction buckets according to claim 2 is characterized in that: The reaction frame is welded from square steel pipes; the driving mechanism is a linear servo electric cylinder, which is used to provide different speeds to drive the pull rod to move. The pull rod lifting range is 0 to 300 mm, the pulling force range is 0 to 1000 N, and the moving speed range is 1 to 500 mm / s; the control mechanism is a computer.

5. The method for predicting failure modes of soil bodies lifted by suction buckets according to claim 2 is characterized in that: Step S2 is specifically as follows: S2.1: Design working condition, the parameters that can be changed in the working condition include the inner diameter of the suction bucket, the outer diameter of the suction bucket, the length of the suction bucket and the upward pull force loading rate; S2.2: Fill the prepared transparent sand into the plexiglass box; S2.3: Use the driving mechanism to penetrate the suction anchor into the transparent sand in the glass box; S2.4: After the suction anchor penetrates the transparent sand, let it stand for a while; S2.5: Arrange the land visualization system, start and adjust the image acquisition system, and complete the acquisition preparation work; S2.6: Pull the suction anchor upwards and record the pulling-up process of the suction anchor in real time, and record the original laser speckle image video and tension sensor readings of the suction anchor at different loading rates.

6. The method for predicting failure modes of soil bodies lifted by suction buckets according to claim 5 is characterized in that: In step S2.5, the method of pulling the suction anchor is displacement control, and the pulling speed is controlled by the driving mechanism, and the suction anchor is pulled up at different loading rates.

7. The method for predicting failure modes of soil lifted by suction bucket according to claim 1, characterized in that: Step S3 is specifically as follows: S3.1: After converting the original laser speckle image video to black and white, batch extract each frame of the video; S3.2: Import the frames of the laser speckle slices extracted at different times and working conditions into the image processing and analysis software for image processing and calculation; S3.3: After exporting the calculated data, the velocity field, displacement field and displacement vector diagram of the soil during the pull-out process are plotted, and then the contour map of the soil displacement at the maximum pull-out force is obtained through programming calculation.

8. The method for predicting failure modes of soil lifted by suction bucket according to claim 1, characterized in that: In step S4, the factors of the "factor-failure size" database include the suction bucket length, the suction bucket diameter and the pull-up rate; the failure size of the "factor-failure size" database includes the fracture surface width and the fracture surface depth.

9. The method for predicting failure modes of soil lifted by suction bucket according to claim 1, characterized in that: In step S5, the GAN-BP machine learning algorithm includes a GAN module and a BP module; the method for establishing the GAN-BP machine learning algorithm model is: S5.1: Generate high-quality data samples through the GAN module, so that the generated data is close to the real data distribution; S5.2: Combine the high-quality data samples generated by the GAN module with the original data to form a new combined training set; S5.3: Use the combined training set to train the BP module and establish a GAN-BP hybrid model.

10. The method for predicting failure modes of soil bodies lifted by suction buckets according to claim 9, characterized in that: In step S5.1, the method of generating high-quality data samples through the GAN module is: S5.11: Define the generator G(z; θ G ) and the discriminator D(x;θ D ); S5.12: Train the discriminator D and the generator G by optimizing the loss function until the samples generated by the generator G are difficult to distinguish. S5.13: After training, use the generator G to input random noise z to generate augmented samples x augmented , the expanded sample x augmented This is the expanded data sample of the "factor-damage size" database in step S4. The training method of the BP module is: S5.31: Prepare the dataset, X train =[X real ,X augmented ],Y train =[Y real ,Y augmented ]; S5.32: define a BP neural network, wherein the BP neural network comprises an input layer, a hidden layer and an output layer; S5.33: Forward propagation, through the activation function f, calculate the values ​​of the hidden layer and the output layer; S5.34: Use mean square error MSE as the loss function; S5.35: Optimize network parameters by gradient descent.

Images