A method for predicting scour of offshore pile foundation under multi-element coupling conditions

Abstract

The application discloses a method for predicting scouring of a marine pile foundation under a multi-element coupling condition, and comprises the following steps: acquiring environmental data and constructing a deep learning model, wherein the deep learning model adopts a DNN network model; acquiring training samples; training the deep learning model based on the training samples through a transfer learning mode to obtain a prediction model; and predicting the environmental data through the prediction model to obtain a prediction result of a scouring depth of the marine pile foundation. Through the technical solution, the prediction model can be trained and generated according to prior knowledge and related simulation and measured data as a data basis for transfer learning, and then the scouring depth can be effectively predicted through the prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pile scour prediction, in particular to a method for predicting offshore pile scour under multi-element coupling conditions. BACKGROUND

[0002] Due to the rich wind energy resources on the sea and the feasibility of technology and other related factors, offshore wind farm technology has been developed. Offshore wind farm technology is mainly based on offshore wind turbines, supported by different structures and equipment, and the bottom fixed support is the main fixed mode. The single pile fixed mode in the bottom fixed support is fixed by a single pile foundation, which is composed of a steel pile installed below the seabed. The installation depth is determined by the type of seabed surface. The single pile foundation effectively extends the wind tower underwater and into the seabed. One of the advantages of this foundation is that it does not need to arrange the seabed. However, the pile foundation needs to prevent the scouring of the seabed by the sea current. During the scouring process, a scour hole will appear around the pile foundation, which will weaken the bearing capacity of the pile foundation, reduce the natural frequency of the foundation, and affect the fatigue life of the structure. Therefore, the prediction of offshore pile scouring is of great significance to the early pile installation design and the later maintenance. In the prior art, different empirical formulas are used to calculate and predict the scouring depth. Due to the complexity and uncertainty of the scouring problem, most of the empirical formulas are obtained by simplifying the problem. There are certain differences between different empirical formulas and the solution of the actual scouring problem, and the coupling between the parameters is not considered. At the same time, most of the experiments of the empirical formula are indoor simulation experiments and flume experiments, which have certain deviations from the actual field observation. In summary, the scouring calculation in the prior art has certain errors in the actual field calculation, and cannot effectively predict the scouring. SUMMARY

[0003] To solve the above-mentioned problems existing in the prior art, the present application provides a method for predicting offshore pile scouring under multi-element coupling conditions, which can use prior knowledge and related simulation and measured data as the data basis for training the prediction model through transfer learning, and then effectively predict the scouring through the prediction model, greatly reducing the prediction error.

[0004] In order to achieve the above technical purpose, the present application provides the following technical scheme:

[0005] A method for predicting offshore pile scouring under multi-element coupling conditions, comprising:

[0006] Obtain environmental data and build a deep learning model, wherein the deep learning model adopts a DNN network model;

[0007] Obtaining training samples; training the deep learning model based on the training samples by a transfer learning method to obtain a prediction model, and predicting the environmental data by the prediction model to obtain a prediction result of the scour depth of the offshore pile foundation.

[0008] Optionally, the environmental data comprises pile foundation seabed shape data and wave current data, the pile foundation seabed data comprises pile foundation shape data, seabed shape data, sediment size data, sand particle size data, water depth data and Froude number, and the wave current data comprises wave velocity, current velocity, impact angle data, wave height, wave number, wave period and wave KC factor.

[0009] Optionally, the training samples comprise simulation data obtained by prior knowledge, indoor simulation data and measured historical data, wherein the prior knowledge comprises different algorithms for offshore scour prediction.

[0010] Optionally, the process of training the deep learning model comprises:

[0011] training the deep learning model based on the simulation data to generate a theoretical model;

[0012] training the theoretical model based on indoor simulation data by a feature transfer learning method to obtain a simulation model;

[0013] training the simulation model based on measured historical data by a feature transfer learning method and a structure transfer learning method to obtain a prediction model.

[0014] Optionally, after obtaining the prediction model, the process further comprises:

[0015] testing the prediction model, obtaining a performance index of the prediction model, judging the performance index, if the performance index is qualified, using the qualified prediction model to predict the environmental data, and if the performance index is not qualified, increasing the sample quantity of the measured historical data to train the prediction model until the performance index of the prediction model is qualified.

[0016] Optionally, after obtaining the training samples, the training samples are preprocessed to obtain preprocessed training samples, and the deep learning model is trained based on the preprocessed training samples; wherein the preprocessing comprises missing value filling and standardization processing.

[0017] Optionally, in the process of training the deep learning model, the gradient descent method is used to update the network parameters of the deep learning model.

[0018] Optionally, the process of generating the theoretical model comprises:

[0019] The training samples corresponding to different algorithms in simulation data are integrated; the training samples are sorted according to the number of variables in the different algorithms, the deep learning model is trained in turn through the sorted training samples, and a theoretical model is obtained.

[0020] The present application has the following technical effects:

[0021] The present application uses the above technical solution, can learn the interaction relationship between parameters between different algorithms through the data simulated by prior knowledge, and then migrate to the indoor simulation aspect, adjust the deficiencies of the theoretical model through the indoor simulation aspect, and through the actual training data again migrate to the field aspect, realize the prediction of the actual offshore scour depth in the field, in the training process, can improve the prediction accuracy of the model under the condition that the training data in the field is less. The model adopts a fully connected neural network (DNN), the hidden layer nodes in the model adopt a fully connected mode, can effectively reflect the coupling relationship between the input parameters in the network, and effectively predict the scour depth according to the coupling relationship of multiple input elements. BRIEF DESCRIPTION OF DRAWINGS

[0022] 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 will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Fig. 1 The method flowchart provided by the embodiments of the present application is shown in the figure;

[0024] Fig. 2 The task flowchart provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] As Figs. 1-2As shown, the present application provides a method for predicting scouring of offshore pile foundation under multi-element coupling condition, in which, by collecting relevant prior knowledge in advance, the prior knowledge of the present application includes an algorithm for pile scouring, using the data related to the above algorithm to train a theoretical model, and collecting training samples with similar characteristics, i.e. indoor simulation data, through related indoor experiments, the model is transferred from the theoretical model to the simulation aspect through the sample feature transfer learning mode, thereby training a model close to the actual field prediction, and the model is transferred from the simulation aspect to the field aspect by retraining the model using the actual collected historical sample data, generating a model that can be actually applied to the coastal prediction.

[0027] The model is used to predict the actual collected data, and the final scouring depth is generated. The present application uses the above technical solution, which can learn the interaction relationship between parameters of different algorithms through the data simulated by prior knowledge, and then migrate to the indoor simulation aspect, adjust the shortcomings of the theoretical model through the indoor simulation aspect, and migrate to the field aspect again through the actual training data, realize the prediction of the actual offshore scouring depth in the field, and improve the prediction accuracy of the model under the condition that the training data in the field is less. The model of the present application adopts a fully connected neural network (DNN), the hidden layer nodes in the model adopt a fully connected mode, which can effectively reflect the coupling relationship between the input parameters in the network, and effectively predict the scouring depth according to the multi-input element coupling relationship.

[0028] The above content is described through specific steps:

[0029] The application is used for training collected relevant data, that is, training samples, and comprises simulation data, indoor simulation data and measured historical data based on prior knowledge, wherein the prior knowledge comprises relevant algorithms for offshore pile scouring prediction calculation, such as CSU equation, Han Haiqian formula, Sumer method and the like, after the relevant algorithms are collected, an offshore pile scouring environment simulation model is constructed, wherein the same data as the input data of the deep learning model are set as environment parameters in the model, the environment parameters comprise pile foundation seabed data and wave flow data, the above data are obtained by direct setting collection or relevant calculation, the above environment simulation model is solved by the algorithms of the above prior knowledge, and then the scouring depth under different algorithms for the same environment data is generated, the environment simulation data and the scouring depth under the same algorithm are integrated to form the training sample of a certain algorithm, the training samples under multiple algorithms are obtained by simulating different algorithms, before training, the training samples under multiple algorithms are sorted from few to many based on the number of variable parameters in the algorithm, the training sample corresponding to the algorithm with few variable parameters is preferentially trained, the training sample under the next algorithm is used for training after the training sample under one algorithm is trained, and the training is completed until all the samples are trained, and then a theoretical model is generated. The indoor simulation data is obtained by relevant indoor simulation tests and water tank tests and the like, the measured historical data is obtained by sensing monitoring on the offshore pile foundation site, the indoor simulation data and the historical data all comprise the data category in the environment parameters and the corresponding scouring depth, after the different training samples are obtained, the training samples are filled with missing values and standardized pretreated, and the model is trained by the training samples after pretreatment.

[0030] The deep learning model used in the application adopts a full connection neural network (DNN), the DNN can learn the strong correlation between input-output and input-input parameters, does not need any special adaptation, and can realize high-precision convergence only by adjusting the internal parameters of the network, changes the problem from fitting on a nonlinear function to letting the DNN search nonlinear multi-functions. Although it is not clear whether this is advantageous in a strict mathematical sense, this is exactly the task that the DNN is good at in various practical applications. The application hopes to introduce the deep learning neural network into the scouring prediction field and provide a possibility for parameter solving by using the significant progress made in the deep learning field.

[0031] The present application is aimed at the structure design of a full connection neural network, including an input layer, a hidden layer and an output layer, the input data of the input layer being offshore pile foundation environment data, including pile foundation seabed data and wave flow data, the pile foundation seabed data including pile foundation shape data, offshore shape data, sediment size data, sand particle size data, water depth data and Froude number; the wave flow data including wave velocity, flow velocity, impact angle data, wave height, wave number, wave period and wave KC coefficient data, the number of hidden layers in the hidden layer being set to 7, the number of hidden layer nodes being set to 500, and the output layer being set to scour depth data. After the construction is completed, the weights in the nodes are randomly initialized, and the gradient descent method is used to update the network parameters in the training process.

[0032] First, the training samples under different algorithms in prior knowledge are trained in turn through the above sequence, after the training is completed, the theoretical model containing different algorithm characteristics is generated, that is, the optimization model under the source domain (theoretical aspect), then the training set is replaced and the network parameters of the theoretical model are retained, the training samples of indoor simulation data are used to continue optimizing the optimization model under the source domain through the way of transfer learning, the simulation model is generated, that is, the optimization model under the simulation aspect (intermediate domain), then the network parameters of the simulation model are retained, and the optimization model under the intermediate domain is trained again through the way of transfer learning using historical data, the prediction model capable of predicting the offshore pile foundation field data is generated, that is, the model under the actual aspect (target domain). The above transfer learning all adopts the feature transfer learning method, different scour algorithms can reflect part of the rules of actual scour, that is, there is a certain fitting relationship between the algorithm simulation data and the actual prediction data, but the above different scour algorithms are idealized algorithms for actual scour environment, and there is a certain difference between the scour phenomenon and the actual multi-element coupling, while the indoor simulation is the main test means of the algorithm fitting result, the algorithm and the indoor simulation data have a certain similarity, and there is a small deviation between the test and the actual data, so the intermediate domain is set as the transfer learning buffer to reduce the use of the training samples of the target domain, and through the step-by-step transfer of the source domain, the intermediate domain and the target domain, the model can be effectively trained using less actual scour data for the actual scour method, and higher precision is obtained, at the same time, the DNN model can fit the correlation between different parameters, such as the coupling between the pile foundation seabed and the wave flow, the coupling between the wave flow, and ensure the coupling between the fitting parameters, in the training process, the less coupling relationship in the algorithm, the coupling relationship between the indoor data and the coupling relationship between the actual data are optimized and fitted step by step, finally, the fitting relationship in the actual data can be effectively reflected, and the scour prediction of multi-element coupling is realized.

[0033] In the above, in order to better reflect the coupling relationship of actual data, prevent the absence of coupling relationship between actual data, in the process of migrating the intermediate domain to the target domain, the model needs to be adjusted, the application inserts a hidden layer between the output layer and the last hidden layer, and the number of hidden layer nodes is 500, and the inserted hidden layer is randomly initialized, at the same time, the pre-sequence network structure and parameters are not modified, and the model parameters before the inserted hidden layer are frozen during training, a small part of the actual training set is used to train the frozen model, and then the remaining actual training set is used to train the unfrozen model. The training sample of the historical data, i.e. the actual data, is used to train the adjusted model to realize feature and structure migration, and generate the final prediction model. After generating the prediction model, the model is tested by collecting actual data, and the test is performed by using accurate performance indicators, such as accuracy, recall rate, specificity or Matthew correlation coefficient. When the performance index is greater than a certain pre-set threshold, the model is considered to be qualified and can be used for prediction, if not, the actual collected data is obtained to train the prediction model until it is qualified.

[0034] After generating the model, the relevant environmental data of the offshore pile foundation is directly obtained, the relevant environmental data is preprocessed and input into the prediction model, and the scour depth is obtained by outputting the prediction model, so as to realize the scour prediction under the condition of considering the relevant coupling between the environmental parameters and the scour depth and the environmental parameters.

[0035] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting scour of marine pile foundations under multi-factor coupling conditions, characterized in that, include: Acquire environmental data and construct a deep learning model, wherein the deep learning model adopts a DNN network model; Obtain training samples; The deep learning model is trained using transfer learning based on training samples to obtain a prediction model. The prediction model is then used to predict the environmental data to obtain the prediction result of the scour depth of the marine pile foundation. The environmental data includes pile foundation seabed shape data and wave and current data. The pile foundation seabed data includes pile foundation shape data, seabed shape data, sediment size data, sand particle size data, water depth data, and Froude number. The wave and current data includes wave velocity, current velocity, impact angle data, wave height, wave number, wave period, and wave KC coefficient. The training samples include simulation data obtained through prior knowledge, indoor simulation data, and measured historical data, wherein the prior knowledge includes different algorithms for predicting marine scour. The process of training the deep learning model includes: training the deep learning model based on the simulation data to generate a theoretical model; The theoretical model is trained using feature transfer learning based on indoor simulation data to obtain the simulation model; The simulation model is trained using feature transfer learning and structure transfer learning methods based on measured historical data to obtain a prediction model.

2. The method according to claim 1, characterized in that: After obtaining the prediction model, the process also includes: testing the prediction model, wherein the performance index of the prediction model is obtained, the performance index is judged, and if the performance index is qualified, the qualified prediction model is used to predict the environmental data; if the performance index is unqualified, the number of samples of measured historical data is increased to train the prediction model until the performance index of the prediction model is qualified.

3. The method according to claim 1, characterized in that: After obtaining the training samples, the training samples are preprocessed to obtain preprocessed training samples, and the deep learning model is trained based on the preprocessed training samples. The preprocessing mentioned above includes missing value imputation and normalization.

4. The method according to claim 1, characterized in that: During the training of the deep learning model, the network parameters of the deep learning model are updated using the gradient descent method.

5. The method according to claim 1, characterized in that: The process of generating a theoretical model includes: integrating training samples corresponding to different algorithms in the simulation data; sorting the training samples according to the number of variables in the different algorithms; and training the deep learning model sequentially using the sorted training samples to obtain the theoretical model.

Images