An erosion detection method based on a proxy model

Abstract

The present application relates to a kind of based on agent model's scour detection method, obtain the target response data to be corrected of offshore wind power pile foundation;Based on the target response data to be corrected obtained, construct the finite element numerical model consistent with the size of physical model test, using improved PSO algorithm is corrected to finite element numerical model, obtain the finite element numerical model after correction;Wherein, using Latin hypercube sampling method selects several particles with good fitness as initial particle, while reducing the search space, optimization of PSO algorithm is carried out;The foundation response data under different scour depths is calculated by the finite element numerical model after correction;Foundation response data is used as input training set, and scour depth is used as output training set, constructs BP neural network and carries out training, obtains the agent model that can be used to detect scour depth, each working condition test data is used as test set to retrieve scour depth;The agent model constructed after model correction can be used to detect the scour of pile foundation in real time.

Description

Technical Field

[0001] This invention relates to a scour detection method based on a surrogate model, belonging to the field of engineering testing technology. Background Technology

[0002] In recent years, the offshore wind power market has developed rapidly, with fixed pile foundations, mainly monopile and multipile foundations, being the primary foundation type. Scour is a significant factor causing reduced durability and foundation instability in wind turbines, and existing scour detection methods primarily rely on multibeam and sonar scanning.

[0003] For example, the invention with application number 202211000642.2 discloses a method and system for detecting scour of offshore wind power pile foundations based on an underwater robot. After determining the area to be detected, the method determines multiple scanning areas, including an outer ring area and an inner ring area. First, an outer ring area is established. When multiple outer ring areas cannot cover all the detection areas, an inner ring area is set to ensure that the entire area to be detected can be covered. Then, the robot is deployed. After the robot scans each scanning area through a mechanical scanning sonar module, the scour image is finally obtained. The invention with application number 202410665927.0 proposes an intelligent sensing and monitoring system for offshore wind turbines, comprising: an application layer, a real-time operating system, a communication connection layer, a logic control layer, and an environmental hardware layer. The communication connection layer communicates with sensors and a PLC controller in the logic control layer. Sensors are used to acquire measurement data from wind turbine components in the environmental hardware layer. The PLC controller receives control signals and transmits them to the wind turbine components. The real-time operating system implements the interface between the operating system kernel and the hardware based on a hardware abstraction layer. The application layer deploys one or more of the following: a C++ / Rust / Python compatible library, a real-time database, a graphics processing framework, a SCADA system, a structural safety monitoring system, a fault diagnosis and analysis system, and a real-time control system for the wind turbine. The sensors are used to detect operational data, meteorological data, seabed environmental data, and marine environmental data of the offshore wind turbine. The structural safety equipment uses a multibeam echo sounder and side-scan sonar for scour detection and continuous monitoring of key parameters.

[0004] The existing technologies mentioned above, such as multibeam sonar scanning, cannot achieve real-time detection. Therefore, a method is needed in engineering to detect the scour progress of pile foundations in real time. Summary of the Invention

[0005] This invention provides a scour detection method based on a proxy model. The proxy model constructed after model correction can detect the scour of pile foundations in real time.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A scour detection method based on a surrogate model specifically includes the following steps:

[0008] Step S1: Obtain the target response data to be corrected for the offshore wind turbine pile foundation;

[0009] Step S2: Based on the obtained target response data to be corrected, a finite element numerical model with the same experimental size as the physical model is constructed. The improved PSO algorithm is used to correct the finite element numerical model to obtain the corrected finite element numerical model.

[0010] Among them, the Latin hypercube sampling method is used to select several particles with good fitness as initial particles, and the PSO algorithm is optimized after the search space is reduced.

[0011] Step S3: Calculate the basic response data at different scour depths using the modified finite element numerical model;

[0012] Step S4: Using the basic response data obtained in step S3 as the input training set and the scour depth as the output training set, construct a BP neural network and train it to obtain a surrogate model that can be used to detect the scour depth. Use the test data of each working condition as the test set to invert the scour depth.

[0013] Furthermore, the target response data to be corrected includes pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle;

[0014] Furthermore, the specific steps for obtaining the corrected finite element numerical model in step S2 are as follows:

[0015] Step S21: Select the symmetrical half of the numerical model in the constructed finite element numerical model for calculation, and select the parameters to be corrected to export the time history response data at the same position as the physical model under cyclic loading.

[0016] Step S22: Calculate the root mean square error of the finite element simulation results and the actual measurement normalized by the physical model;

[0017] Step S23: Continue to normalize the root mean square error;

[0018] Step S24: Evaluate the similarity of the spatial distribution of the target response to be corrected using the Pearson correlation coefficient;

[0019] Step S25: Obtain the improved objective function of the PSO algorithm based on the Pearson correlation coefficient;

[0020] Furthermore, in step S22, the formula for calculating the normalized root mean square error between the finite element simulation results and the actual measurements is as follows:

[0021]

[0022] In formula (1), RMSE is the root mean square error, n is the number of responses at the measurement points where the target to be corrected is located, and y i x represents the response data of the measurement points obtained from numerical simulation. i These are the response data measured at the measurement points in the physical model;

[0023] In step S23, the calculation formula for further normalization is:

[0024]

[0025] In formula (2), RMSE′ is the normalized root mean square error. This represents the average value of the response data measured at the measurement points in the physical model. This represents the average value of the measurement point response data obtained from numerical simulation.

[0026] In step S24, the formula for calculating the Pearson correlation coefficient is as follows:

[0027]

[0028] In formula (3), r is the Pearson correlation coefficient;

[0029] In step S25, the objective function of the improved PSO algorithm is obtained as follows:

[0030]

[0031] In formula (4), f min (x) is the fitness function, α is the weight of the normalized root mean square error, β is the weight of the Pearson coefficient, and α+β=1;

[0032] Furthermore, in step S2, before performing the PSO algorithm optimization, several initial particle positions are randomly sampled and the fitness functions of several initial particles are calculated. Based on the fitness functions, several initial particle positions with good fitness are selected from several particles to narrow the search space. PSO optimization is then performed within the narrowed search space.

[0033] Furthermore, when modifying the finite element numerical model, the number of particles in each iteration should be at least 20.

[0034] Furthermore, in step S4, the basic response data used as the training set is obtained based on sensitivity analysis, and the specific steps are as follows:

[0035] Step S41: Construct a surrogate model with scour depth as the input sample set and pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle as the output sample set.

[0036] Step S42: Using the Latin hypercube sampling method, 1000 sets of scour depths are randomly selected and substituted into the surrogate model to obtain 1000 sets of input-output pairs;

[0037] Step S43: Calculate the sensitivity of the input-output pair and select a suitable input training set.

[0038] By employing the above technical solutions, the present invention has the following beneficial effects compared to the prior art:

[0039] 1. The scour detection method based on the proxy model provided by this invention can detect the scour of pile foundations in real time after preliminary data collection, model correction and proxy model construction.

[0040] 2. The scour detection method based on the surrogate model provided by this invention provides a reference for the health detection of offshore wind turbine foundations, reduces operation and maintenance costs, and has high detection accuracy. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a schematic diagram illustrating the reduction of the search space provided by the present invention;

[0043] Figure 2 This is a schematic diagram of the sample space obtained by sampling in a three-dimensional sampling space using the Latin hypercube sampling method provided by the present invention;

[0044] Figure 3 This is a schematic diagram of evaluating the quality of initial particle positions by calculating fitness, provided by the present invention.

[0045] Figure 4 This is a graph showing the fitness changes of the PSO algorithm in two sets of model correction experiments provided by this invention.

[0046] Figure 5 This is a schematic diagram of the pile bending moment sensitivity provided by the present invention;

[0047] Figure 6 This is a schematic diagram of the scour detection results provided by the present invention, using bending moment as input;

[0048] Figure 7 This is a schematic diagram of the scour detection results provided by the present invention, using displacement as input;

[0049] Figure 8 This is the physical model diagram provided by the present invention;

[0050] Figure 9 This is a schematic diagram of the finite element numerical model provided by the present invention;

[0051] Figure 10 This is a scour simulation diagram provided by the present invention, which is tested using a physical model. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. In the description of this application, it should be understood that the terms "left side," "right side," "upper part," "lower part," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. "First," "second," etc., do not indicate the importance of the components, and therefore should not be construed as a limitation of the present invention. The specific dimensions used in this embodiment are only for illustrating the technical solution and do not limit the scope of protection of the present invention.

[0053] As described in the background section, current methods for detecting the scour progress of pile foundations mainly rely on multibeam sonar scanning. These methods have two main drawbacks: high detection costs and the inability to achieve real-time detection.

[0054] To address the aforementioned issues, this application provides a scour detection method based on a surrogate model. The main technical approach involves collecting response data from offshore wind turbines during operation. After model correction and surrogate model construction, real-time pile foundation scour detection can be achieved. Specifically, the method includes the following steps:

[0055] Step S1: Obtain the target response data to be corrected for the offshore wind turbine pile foundation; the target response data to be corrected includes pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle;

[0056] Step S2: Based on the obtained target response data to be corrected, a finite element numerical model with the same experimental size as the physical model is constructed. The improved PSO algorithm is used to correct the finite element numerical model to obtain the corrected finite element numerical model.

[0057] Among them, the Latin hypercube sampling method is used to select several particles with good fitness as initial particles, and the PSO algorithm (Particle Swarm Optimization Algorithm) is used to optimize after the search space is reduced.

[0058] Step S3: Calculate the basic response data at different scour depths using the modified finite element numerical model;

[0059] Step S4: Using the basic response data obtained in step S3 as the input training set and the scour depth as the output training set, construct a BP neural network and train it to obtain a surrogate model that can be used to detect the scour depth. Use the test data of each working condition as the test set to invert the scour depth.

[0060] In the above steps, the most significant innovation of this application is that it provides an improved PSO algorithm to correct the finite element numerical model. By reducing the search range in advance, the amount of computation is reduced, thereby improving the computational accuracy.

[0061] The specific steps for obtaining the corrected finite element numerical model are as follows:

[0062] Step S21: Select the symmetrical half of the numerical model in the constructed finite element numerical model for calculation, and select the parameters to be corrected to export the time history response data at the same position as the physical model under cyclic loading.

[0063] Step S22: Calculate the root mean square error (RMSE) of the finite element simulation results compared to the actual measurements of the physical model, normalized. The formula for calculating the RMSE of the finite element simulation results compared to the actual measurements is as follows:

[0064]

[0065] In formula (1), RMSE is the root mean square error, n is the number of responses at the measurement points where the target to be corrected is located, and y i x represents the response data of the measurement points obtained from numerical simulation. i These are the response data measured at the measurement points in the physical model;

[0066] Step S23, continue to normalize the root mean square error; the calculation formula for further normalization is:

[0067]

[0068] In formula (2), RMSE′ is the normalized root mean square error. This represents the average value of the response data measured at the measurement points in the physical model. This represents the average value of the measurement point response data obtained from numerical simulation.

[0069] Step S24: Evaluate the similarity of the spatial distribution of the target response to be corrected using the Pearson correlation coefficient; the formula for calculating the Pearson correlation coefficient is:

[0070]

[0071] In formula (3), r is the Pearson correlation coefficient;

[0072] Step S25: Based on the Pearson correlation coefficient, the improved PSO algorithm objective function is obtained. The improved PSO algorithm objective function is as follows:

[0073]

[0074] In formula (4), f min (x) is the fitness function, α is the weight of the normalized root mean square error, β is the weight of the Pearson coefficient, and α+β=1.

[0075] After obtaining the improved PSO algorithm objective function through the above methods, it is necessary to perform optimization operations on it. Before optimization, this application first narrows the search space by random sampling pre-experimentation. Specifically, it selects several initial particle positions by random sampling, calculates the fitness function of the particle, selects several initial particle positions with better fitness from these particles, and narrows the search space while including as many of these better initial particle positions as possible. Alternatively, the search space can be divided into several small search subspaces as needed, and PSO optimization is performed on each search subspace.

[0076] This diagram illustrates the reduction of the search space. Figure 1 As shown in the figure, it should be noted that even if the position of the sample particle with low fitness may find the global optimum nearby, the probability of this happening in a search space with a relatively uniform fitness distribution is low. To improve efficiency, the search range can be narrowed first using this method. If the expected result cannot be achieved, the search range can be appropriately expanded until the structure meets the expectations.

[0077] Regarding the methods used for random sampling, the most common method is Monte Carlo sampling. This method, based on probability distribution, ensures that samples fall randomly at any location in the sampling space. This can lead to excessively high sampling density in some areas and excessively low sampling density in others. In engineering, some extreme cases occur with low probability, but these are often of primary concern. Monte Carlo sampling may fail to obtain samples from these areas. This application preferably uses Latin hypercube sampling, the process of which can be summarized as stratification, sampling, and randomization. First, the sampling space is stratified, and random sampling is performed between each stratum. Then, the order of the obtained samples is shuffled. This ensures that there is at least one sample in every area, resulting in more uniform sampling and stronger coverage.

[0078] When conducting experimental verification in this application, the elastic modulus E of the physical model pile was selected. p The elastic modulus E of the soil s The model is optimized using three parameters: the soil's internal friction angle φ, and the initial search intervals for these three parameters are: E...p ∈(1~40)GPa, E s ∈(5~60)MPa, φ∈(10~45)°. The sample space is obtained by performing 100 samplings in the three-dimensional sampling space using the Latin hypercube sampling method, as shown below. Figure 2 As shown, the parameter values ​​of 100 sample particles were input into the finite element numerical model to calculate the fitness and evaluate the quality of the initial particle positions. The results are as follows. Figure 3 As shown. From Figure 3 As can be seen, the fitness of the initial particles at the sampling points is between 0.47 and 0.96. A large number of the initial particle positions have very low fitness, indicating that the values ​​of these particles are not reasonable and it is difficult to find the global optimal solution in their vicinity. Therefore, it can be inferred that the method of reducing the search space is reasonable.

[0079] In summary, narrowing the search space should include as many optimal initial particles as possible, while ensuring sufficient search space around each particle, it is necessary to avoid placing the initial particle position on the boundary of the search space.

[0080] After initial particle position evaluation and search space reduction, the PSO optimization algorithm is implemented. Typically, to ensure search accuracy, the PSO optimization algorithm requires 20-100 particles per generation and more than 100 iterations. In this application, since the initial positions obtained after LHS sampling and screening already have a high probability of being near the optimal position, the number of particles and iterations in the PSO optimization process can be appropriately reduced.

[0081] To verify the advantages of initial particle position selection and search space scaling, this application provides two sets of model correction experiments for PSO optimization:

[0082] The first group did not undergo LHS initial particle position screening and search space scaling. Each generation of particles was set to 20, and the iteration termination condition was that the number of iterations reached 100.

[0083] The second group adopted the approach of narrowing the search space, reducing the initial search space to two better search subspaces, and carried out PSO optimization in each subspace. Each generation of particles was set to 20, and the iteration termination condition was that the number of iterations reached 20.

[0084] The computational cost, computation time, and model correction accuracy of the two sets of model correction experiments were compared respectively. The fitness changes of the PSO algorithm in the two sets of model correction experiments are as follows: Figure 4 As shown, Figure 4 In the diagram, 4a is the fitness change graph of the first group, 4b is the fitness change graph of the first search subspace of the second group, and 4c is the fitness change graph of the second search subspace of the second group.

[0085] Next, regarding step S4, which involves obtaining the basic response data for the training set, this application preferably uses sensitivity analysis for confirmation. However, conducting local sensitivity analysis based on statistical data requires a large dataset, and a surrogate model can significantly reduce the computational cost of finite element simulation. The specific steps are as follows: Construct a surrogate model with scour depth as the input sample set and pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle as the output sample set; use the Latin hypercube sampling method to randomly select 1000 sets of scour depths and substitute them into the surrogate model to obtain 1000 input-output pairs; finally, obtain the sensitivity of the input-output pairs.

[0086] Table 1 shows the local sensitivity of each structure to scouring, based on the aforementioned target response data to be corrected.

[0087] Table 1. Local sensitivity of each structural response to scour

[0088]

[0089] In Table 1, the sensitivity of pile top displacement is 0.93, and the sensitivity of tower top displacement is 0.71. This shows that because pile top displacement is close to the scour pit, it is a relatively local structural response, while tower top displacement is a global structural response, making the former more sensitive to local scour pits. The sensitivity of tower top acceleration is 1.06, indicating that acceleration response is more sensitive to scour pits than displacement response. However, overall, the sensitivity coefficients of pile top displacement, tower top displacement, tower top acceleration, and tower top rotation are all around 1, indicating that these structural responses of displacement, velocity, and acceleration have similar scour detection accuracy when used as input training samples in the surrogate model. The local sensitivity of jacket nodal stress is between 0.07 and 0.16, much lower than the previous structural responses; therefore, jacket nodal stress is not suitable as input training samples for constructing a scour detection surrogate model.

[0090] Regarding the bending moment of the pile body Figure 5 As shown, the bending moment sensitivity in the middle of the pile is between 4 and 10, indicating that these bending moments are more suitable as input training sample sets compared to the former. Figure 6 (Scrap detection results using bending moment as input) Figure 7 The scour detection results (with displacement as input) also verify the above conclusions.

[0091] Finally, this application also provides a physical model diagram of the detection method described above, such as... Figure 8 As shown, Figure 9 It is a constructed finite element numerical model. Figure 10 A simulation diagram of its scouring was shown.

[0092] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0093] The meaning of "and / or" as used in this application includes situations where each exists alone or both exist simultaneously.

[0094] The term "connection" as used in this application can mean a direct connection between components or an indirect connection between components through other components.

[0095] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A scour detection method based on a surrogate model, characterized in that: Specifically, the following steps are included: Step S1: Obtain the target response data to be corrected for the offshore wind turbine pile foundation; Step S2: Based on the obtained target response data to be corrected, a finite element numerical model with the same experimental size as the physical model is constructed. The improved PSO algorithm is used to correct the finite element numerical model to obtain the corrected finite element numerical model. Among them, the Latin hypercube sampling method is used to select several particles with good fitness as initial particles, and the PSO algorithm is optimized after the search space is reduced. The specific steps to obtain the corrected finite element numerical model are as follows: Step S21: Select the symmetrical half of the numerical model in the constructed finite element numerical model for calculation, and select the parameters to be corrected to export the time history response data at the same position as the physical model under cyclic loading. Step S22: Calculate the root mean square error of the finite element simulation results and the actual measurement normalized by the physical model; The formula for calculating the normalized root mean square error between finite element simulation results and actual measurements is as follows: 1） In formula (1), The root mean square error, The number of measurement points where the target to be corrected is located. The data represents the response data of the measurement points obtained from numerical simulation. These are the response data measured at the measurement points in the physical model; Step S23: Continue to normalize the root mean square error; The formula for further normalization is: （2） In formula (2), For normalized root mean square error, This represents the average value of the response data measured at the measurement points in the physical model. This represents the average value of the measurement point response data obtained from numerical simulation. Step S24: Evaluate the similarity of the spatial distribution of the target response to be corrected using the Pearson correlation coefficient; The formula for calculating the Pearson correlation coefficient is as follows: （3） In formula (3), The Pearson correlation coefficient; Step S25: Obtain the improved objective function of the PSO algorithm based on the Pearson correlation coefficient; The objective function of the improved PSO algorithm is: （4） In formula (4), For the fitness function, The weights for the normalized root mean square error, The weights of the Pearson coefficients are given, and ; Step S3: Calculate the basic response data at different scour depths using the modified finite element numerical model; Step S4: Using the basic response data obtained in step S3 as the input training set and the scour depth as the output training set, construct a BP neural network and train it to obtain a surrogate model that can be used to detect the scour depth. Use the test data of each working condition as the test set to invert the scour depth.

2. The scour detection method based on a surrogate model according to claim 1, characterized in that: The target response data to be corrected includes pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle.

3. The scour detection method based on a surrogate model according to claim 1, characterized in that: In step S2, before performing the PSO algorithm optimization, several initial particle positions are randomly sampled and the fitness functions of these initial particles are calculated. Based on the fitness functions, several initial particle positions with good fitness are selected from these particles to narrow the search space. PSO optimization is then performed within the narrowed search space.

4. The scour detection method based on a surrogate model according to claim 3, characterized in that: When refining the finite element numerical model, the number of particles in each iteration should be at least 20.

5. The scour detection method based on a surrogate model according to claim 1, characterized in that: In step S4, the basic response data used as the training set is obtained based on sensitivity analysis. The specific steps are as follows: Step S41: Construct a surrogate model with scour depth as the input sample set and pile bending moment, pile top displacement, jacket nodal stress, tower top displacement, tower top acceleration, and tower top rotation angle as the output sample set. Step S42: Using the Latin hypercube sampling method, 1000 sets of scour depths are randomly selected and substituted into the surrogate model to obtain 1000 sets of input-output pairs; Step S43: Calculate the sensitivity of the input-output pair and select a suitable input training set.

Images