+

CN110705029B - A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning - Google Patents

A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning Download PDF

Info

Publication number
CN110705029B
CN110705029B CN201910838392.1A CN201910838392A CN110705029B CN 110705029 B CN110705029 B CN 110705029B CN 201910838392 A CN201910838392 A CN 201910838392A CN 110705029 B CN110705029 B CN 110705029B
Authority
CN
China
Prior art keywords
domain
information
fluid
net
flow field
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910838392.1A
Other languages
Chinese (zh)
Other versions
CN110705029A (en
Inventor
谢永慧
陈悦
刘天源
张荻
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN201910838392.1A priority Critical patent/CN110705029B/en
Publication of CN110705029A publication Critical patent/CN110705029A/en
Application granted granted Critical
Publication of CN110705029B publication Critical patent/CN110705029B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于迁移学习的振荡扑翼能量采集系统流场预测方法，该方法从连续几个经时间离散后的时间点的已知流体域信息，预测下一时间点的未知流体域信息。本发明通过获取和预处理已知流体域信息，建立生成式对抗神经网络迁移学习的生成和判别网络，经训练实现不同参数数值下流体域信息预测，将预测流体域信息作为已知流体域信息，可实现流体域信息随着时间连续预测；由预测流体域信息还原流场，实现流场预测，可计算出在预测时间点上的流体动力学参数及性能参数。本发明在大幅度降低时间和经济成本的同时，能够准确地对振荡扑翼能量采集系统进行流场预测并可得到流体动力学参数及性能参数等，对振荡扑翼能量采集系统的研发具有重要意义。

Figure 201910838392

The invention discloses a method for predicting the flow field of an oscillating flapping wing energy acquisition system based on migration learning. The method predicts the unknown fluid domain at the next time point from the known fluid domain information at several time discrete time points. information. The invention establishes the generation and discrimination network of generative confrontation neural network migration learning by acquiring and preprocessing known fluid domain information, realizes the prediction of fluid domain information under different parameter values after training, and uses the predicted fluid domain information as known fluid domain information , which can realize the continuous prediction of fluid domain information over time; restore the flow field from the predicted fluid domain information, realize the flow field prediction, and calculate the hydrodynamic parameters and performance parameters at the predicted time point. While greatly reducing time and economic cost, the invention can accurately predict the flow field of the oscillating flapping-wing energy collection system, and can obtain fluid dynamic parameters and performance parameters, etc., which is important for the research and development of the oscillating flapping-wing energy collection system. significance.

Figure 201910838392

Description

Flow field prediction method of oscillating flapping wing energy acquisition system based on transfer learning
Technical Field
The invention belongs to the technical field of fluid dynamics, and particularly relates to a flow field prediction method of an oscillating flapping wing energy acquisition system based on transfer learning.
Background
With the rapid development of various social industries, the world energy situation is increasingly severe, and the exploration and application of renewable energy sources gradually become the development theme of various countries all over the world. The oscillating flapping wing energy acquisition system has a huge development prospect due to the fact that the oscillating flapping wing energy acquisition system is simple in structure and can effectively utilize low-quality renewable energy sources such as wind energy and ocean energy, and the oscillating flapping wing energy acquisition system becomes a research hotspot in recent years.
In the process of researching, designing and developing the oscillating flapping wing energy collecting system, two basic methods of numerical simulation and experimental research are mainly relied on at the present stage. The actual working condition of the oscillating flapping wing energy collecting system is determined by a complex external environment, so that the working conditions are numerous, and a large amount of numerical simulation or experimental research is bound to be carried out in the process of researching, designing and developing the oscillating flapping wing energy collecting system. For the numerical simulation method, because the calculation of the oscillating flapping wing energy collecting system belongs to unsteady flow field calculation, a large amount of calculation by using the traditional commercial CFD software brings huge time cost. For the experimental research method, because the actual working conditions of the oscillating flapping wing energy collecting system are complicated, different experimental conditions are needed for the simulation of different working conditions, and the experimental research method brings huge economic cost.
In recent years, the field of artificial intelligence rises rapidly, and simultaneously along with the rapid development of the field of computers, the computing capability is greatly improved, so that new solutions are brought to many traditional fields, and a new revolution of revolution is brought to many traditional industries. For the field of fluid dynamics, the application of artificial intelligence is increasingly wide, and great progress is made from single-parameter prediction to multi-parameter prediction of flow performance and from steady-state flow field prediction to unsteady-state flow field prediction of a flow field state.
Disclosure of Invention
The invention aims to provide a flow field prediction method of an oscillating flapping wing energy acquisition system based on transfer learning, which can greatly reduce time cost and economic cost while ensuring the accuracy of a predicted flow field.
The invention is realized by adopting the following technical scheme:
a method for predicting a flow field of an oscillating flapping wing energy acquisition system based on transfer learning is disclosed, wherein unknown fluid domain information of a next time point is predicted from known fluid domain information of a plurality of continuous time points after time dispersion; specifically, by acquiring node position information and node flow field parameter information of a fluid domain after spatial dispersion, performing matrixing processing and standardization processing, distributing a source domain and a target domain according to a set proportion, further extracting input data sets and corresponding theoretical output data sets of the source domain and the target domain, then respectively establishing a generation network and a discrimination network of generation type antagonistic neural network transfer learning, after training, realizing fluid domain information prediction under different parameter values, and taking unknown fluid domain information obtained by prediction as new known fluid domain information, namely realizing continuous prediction of the fluid domain information along with time under different parameter values; the whole flow field is restored through node position information contained in the fluid domain information obtained through prediction and flow field parameter information on the nodes, flow field prediction is achieved, fluid dynamics parameters and performance parameters of the oscillating flapping wing energy collecting system at the prediction time point are calculated, and therefore researchers can analyze the oscillating flapping wing energy collecting system conveniently.
The invention is further improved in that the method specifically comprises the following steps:
1) obtaining model data of oscillating flapping wing energy collecting system
Respectively collecting known fluid domain information of a plurality of continuous time-dispersed time points of the model aiming at different parameter values, wherein the known fluid domain information comprises node position information and node flow field parameter information of the fluid domain after space dispersion; carrying out space discretization on the model to obtain grid information Hj,qPerforming unsteady state numerical simulation calculation, calculating an application layer flow model, spatially adopting a second-order difference format, temporally adopting a first-order difference format, solving by using a dynamic grid technology, and obtaining node position information at each time point according to a solving result
Figure BDA0002192911010000021
And flow field parameter information on the node
Figure BDA0002192911010000022
J is 1, 2.. the J, J is the number of grid units, Q is 1, 2.. the Q, Q is the number of nodes of each grid unit, R is 1, 2.. the R, R is the total number of different selected parameter values, T is 1, 2.. the T, T is the number of time points after time dispersion, I is 1, 2.. the I, I is the total number of nodes, N is 1, 2.. the N, N is the number of coordinate dimensions, M is 1, 2.. the M, M is the number of flow field parameters to be considered;
2) model data preprocessing of oscillating flapping wing energy collecting system
Regular matrixing is carried out on the model data of the oscillating flapping wing energy acquisition system, and then node position information is rearranged
Figure BDA0002192911010000031
And flow field parameter information on the node
Figure BDA0002192911010000032
Corresponding elements in the node are transformed to obtain the position information of the nodes after matrixing
Figure BDA0002192911010000033
And flow field parameter information on the node
Figure BDA0002192911010000034
Wherein a is 1,2, a, B is 1,2, a, B, C is 1,2, C, A, B, C are the number of nodes in three mutually orthogonal directions, respectively, and a × B × C is satisfied;
for the node position information after matrixing
Figure BDA0002192911010000035
And flow field parameter information on the node
Figure BDA0002192911010000036
Carrying out standardization processing to obtain node position information G after standardization processingr，t，a，b，c，nAnd flow field parameter information F on the noder，t，a，b，c，m
3) Model data splitting source domain and target domain of oscillating flapping wing energy acquisition system
The normalized node position information G corresponding to different values selected by the parametersr,t,a,b,c,nAnd flow field parameter information F on the noder,t,a,b,c,mSplitting at random or according to the requirement according to the set proportion, using one part as the source domain and the node position information of the source domain
Figure BDA0002192911010000039
And flow field parameter information on the node
Figure BDA00021929110100000310
Another part is used as target domain, target domain node position information
Figure BDA00021929110100000311
And flow field parameter information on the node
Figure BDA00021929110100000312
RS, RS are different total numbers of values selected as parameters of a source domain, RT is 1,2,., RT are different total numbers of values selected as parameters of a target domain, and RS + RT is satisfied;
4) extracting respective input data set and theoretical output data set in source domain and target domain
The number N of continuous time points required according to actual input dataTBCarrying out further data extraction on the source domain and the target domain to obtain respective input data set and theoretical output data set;
for the source domain, the node position information of the source domain is firstly
Figure BDA00021929110100000313
And flow field parameter information on the node
Figure BDA00021929110100000314
Splicing to obtain the total data information of the source domain
Figure BDA00021929110100000315
Where K1, 2,., K N + M, a source domain Input dataset Input is obtainedSDAnd a corresponding source domain theoretical output data set LableSD；
For the target domain, firstly, the node position information of the target domain
Figure BDA00021929110100000316
And flow field parameter information on the node
Figure BDA00021929110100000317
Splicing to obtain the total data information of the target domain
Figure BDA00021929110100000318
Where K is 1,2,., K is N + M, the target domain Input dataset Input is obtainedTDAnd a corresponding target domain theoretical output data set LableTD；
5) Fluid domain information prediction GAN migration learning network for constructing oscillating flapping wing energy acquisition system
The fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system comprises a generation network and a judgment network, wherein the generation network is called G Net for short, and the judgment network is called D Net for short;
firstly, a convolution residual error network is adopted to establish a characteristic extraction part of G Net, and Input data Input from a source domain is realizedSDOr target field Input data InputTDExtracting corresponding source domain fluid domain Feature information FeatureSDOr target domain fluid domain Feature information FeatureTDThen, a deconvolution residual error network is adopted to establish a predicted fluid domain information part of G Net, and the Feature information Feature of the source domain fluid domain is realizedSDOr target domain fluid domain Feature information FeatureTDGenerating corresponding source domain actual Output data OutputSDOr target domain actual Output data OutputTD(ii) a Wherein each source domain fluid domain characteristic information
Figure BDA0002192911010000041
ES, each source domain actually outputs data as
Figure BDA0002192911010000042
ES, characteristic information of each target domain fluid domain
Figure BDA0002192911010000043
ET 1,2, ET, each target field actually outputting data as
Figure BDA0002192911010000044
ET 1, 2.. ET, G Net total 21 layers of network; secondly, a D Net network is established by adopting a convolution residual error network to judge whether the fluid domain information is true or false, and the actual output data of each source domain is realized
Figure BDA0002192911010000045
And each source domain theoretical output data corresponding to the source domain theoretical output data
Figure BDA0002192911010000046
To obtain the corresponding discrimination result
Figure BDA0002192911010000047
Is 0 and
Figure BDA0002192911010000048
1, the data sets of the discrimination result are respectively expressed as D _ OutputSDAnd D _ LableSDSimultaneously undertake the actual output of data from each target domain
Figure BDA0002192911010000049
Obtaining a discrimination result
Figure BDA00021929110100000410
The corresponding data set of the discrimination result is D _ OutputTDThe two-classification loss function part is used for calculating the network training of G Net, and D Net has 19 layers of networks in total;
with the quantity SSDA batch of source domain input data and quantity STDA batch of target domain input data for calculating maximum mean difference loss;
g Net in GAN network adopts two-norm loss function L2Loss of class, two-class loss function LbceGLoss function of loss and maximum mean difference LmmdA weighted average of _ loss as a total loss function;
d Net in GAN network adopts two-classification loss of source domain actual output data
Figure BDA00021929110100000411
Two-classification loss L of source domain theoretical output databceLbLoss of class two for the _lossand actual output data of the target domain
Figure BDA00021929110100000412
The sum as a function of total loss;
6) fluid domain information prediction GAN migration learning network for training oscillating flapping wing energy acquisition system
Firstly, the number of the oscillating flapping wing energy acquisition systems is SSDThe fluid domain information of a batch of source domain input data is predicted through G Net in a GAN transfer learning network, and the obtained quantity is SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, and then the number of the oscillating flapping wing energy acquisition systems is STDThe fluid domain information prediction is carried out on a batch of target domain input data for calculating the maximum mean difference loss through G Net in a GAN migration learning network, and the obtained quantities are STDA batch of target domain fluid domain characteristic information and a batch of target domain actual output data;
then the obtained quantity is SSDA batch of source domain actual output data, the corresponding quantity is SSDA batch of source domain theoretical output data and the quantity is STDThe batch of target domain actual output data are respectively judged through D Net in the GAN transfer learning network, and the D Net is subjected to parameter updating once according to the obtained judgment result, so that the D Net can better distinguish a fluid domain information prediction result predicted by G Net and a fluid domain information theoretical result;
then the obtained quantity is SSDA batch of source domain actual output data and the number STDThe actual output data of a batch of target domains are respectively judged through D Net which is subjected to parameter updating once in the GAN migration learning network, and according to the obtained judgment result and the obtained quantity are SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, the quantity is STDIs known to correspond to the actual output of the source domain, and has a quantity SSDThe G Net is subjected to parameter updating once by a batch of source domain theoretical output data, so that the fluid domain information prediction result of the G Net is closer to the fluid domain information theoretical result, and the judgment of the fluid domain information prediction result and the fluid domain information theoretical result of the D Net is confused as much as possible;
through the continuous and repeated confrontation of G Net and D Net, the judgment of the fluid domain information prediction result and the fluid domain information theoretical result by the D Net is more accurate, and the fluid domain information prediction result of the G Net is forced to be closer to the fluid domain information theoretical result;
7) flow field prediction result processing and analysis of oscillating flapping wing energy acquisition system
The trained fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system is used for predicting fluid domain information under other parameter values, and the predicted unknown fluid domain information at the next time point is used as new known fluid domain information, so that the fluid domain information prediction at the next time point can be continuously carried out, and the continuous prediction of the fluid domain information along with time is realized under different parameter values; the whole flow field can be restored through node position information and flow field parameter information on the nodes contained in the fluid domain information obtained through prediction; furthermore, fluid dynamics parameters and performance parameters such as torque, lift coefficient and resistance coefficient of the oscillating flapping wing energy collecting system at a prediction time point are calculated according to a flow field prediction result so as to provide convenience for researchers to analyze the oscillating flapping wing energy collecting system.
The invention is further improved in that in step 2), the position information of the nodes after matrixing is processed
Figure BDA0002192911010000061
And flow field parameter information on the node
Figure BDA0002192911010000062
The normalization process was performed as follows:
for the node position information after matrixing
Figure BDA0002192911010000063
The standardization processing method comprises the following steps:
Figure BDA0002192911010000064
Figure BDA0002192911010000065
Figure BDA0002192911010000066
wherein, MaxGnRepresents the maximum value of the nth coordinate dimension of the node in all data, wherein N is 1,2nRepresenting the minimum value of the nth coordinate dimension of the node in all data of the nth coordinate dimension;
to the parameter information of the flow field on the nodes after matrixing
Figure BDA0002192911010000067
The standardization processing method comprises the following steps:
Figure BDA0002192911010000068
Figure BDA0002192911010000069
Figure BDA00021929110100000610
wherein, MaxFmRepresents the maximum value of the mth parameter of the node in all data, wherein M is 1,2mRepresents the minimum value of the mth parameter of the node in all data thereof.
In a further development of the invention, in step 4), the source domain Input dataset Input is obtainedSDAnd a corresponding source domain theoretical output data set LableSDThe method comprises the following steps:
the source domain input dataset is:
Figure BDA0002192911010000071
wherein each source domain input data is
Figure BDA0002192911010000072
ES, which is the total amount of source domain data and satisfies
ES＝RS×(T-NTB)
The corresponding source domain theoretical output data set is:
Figure BDA0002192911010000073
wherein each source domain theoretical output data is
Figure BDA0002192911010000074
ES, which is the total amount of source domain data and satisfies ES RS × (T-N)TB)。
The invention further improves that in the step 4), a target domain Input data set Input is obtainedTDAnd a corresponding target domain theoretical output data set LableTDThe method comprises the following steps:
the target domain input dataset is:
Figure BDA0002192911010000075
wherein each target domain inputs data as
Figure BDA0002192911010000076
ET 1,2, ET is the total amount of target domain data, and there are multiple time period data missing, i.e. ET RT × (T-N) may not be satisfiedTB)；
The corresponding target domain theoretical output data set is:
Figure BDA0002192911010000077
wherein each target domain theoretical output data is
Figure BDA0002192911010000078
ET 1, 2.. ET, ET is the total amount of target domain data, and there may be a plurality of time segment data deletions corresponding to the target domain input data set, i.e., ET RT × (T-N) may not be satisfiedTB)。
The invention further improves that in the step 5), G Net in the GAN network adopts a two-norm loss function L2Loss of class, two-class loss function LbceGLoss function of loss and maximum mean difference LmmdThe weighted average of _ loss is used as the total loss function, and the specific loss function formula is as follows:
two-norm loss function L2The formula of _ loss is:
Figure BDA0002192911010000081
binary loss function LbceGThe formula of _ loss is:
Figure BDA0002192911010000082
maximum mean difference loss function LmmdThe formula of _ loss is:
Figure BDA0002192911010000083
wherein
Figure BDA0002192911010000089
Representing the Regenerated Kernel Hilbert Space (RKHS), phi (-) represents the mapping function of the original feature space to RKHS, further developed to obtain:
Figure BDA0002192911010000084
for ease of calculation, a simplified equation is introduced:
Figure BDA0002192911010000085
wherein delta is a certain value parameter;
the G Net total loss function is as follows:
LGNet_loss＝L2_loss+WbceG×LbceG_loss+Wmmd×Lmmd_loss
wherein, WbceGIs a binary loss function LbceGWeight of _ loss, WmmdIs a maximum mean difference loss function LmmdWeight of _ loss, according to experience WbceG∈[0.04,0.2]，Wmmd∈[0.04,0.2]。
The invention further improves that in the step 5), D Net in the GAN network adopts the two-classification loss of the actual output data of the source domain
Figure BDA0002192911010000086
Two-classification loss L of source domain theoretical output databceLbLoss of class two for the _lossand actual output data of the target domain
Figure BDA0002192911010000087
The sum is taken as a total loss function, and the specific loss function formula is as follows:
binary loss of source domain actual output data
Figure BDA0002192911010000088
The formula is as follows:
Figure BDA0002192911010000091
two-classification loss L of source domain theoretical output databceLbThe formula of _ loss is:
Figure BDA0002192911010000092
two-classification loss of target domain actual output data
Figure BDA0002192911010000093
The formula is as follows:
Figure BDA0002192911010000094
the D Net total loss function is as follows:
Figure BDA0002192911010000095
the invention further improves the method that in the step 6), during the training of the GAN transfer learning network, the optimizer selection and learning rate setting of G Net and the optimizer selection and learning rate setting of D Net are as follows:
in the training process, an optimizer of G Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data cycle training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after each 100 times of data cycle training;
the optimizer of D Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data cycle training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after every 100 times of data cycle training.
The invention has the following beneficial technical effects:
the invention provides a flow field prediction method of an oscillating flapping wing energy acquisition system based on transfer learning, which is characterized in that unknown fluid domain information of the next time point is predicted by transfer learning of a generative antagonistic neural network (GAN) from known fluid domain information of a plurality of time points after time dispersion, including node position information of a fluid domain after space dispersion and flow field parameter information on a node, and the unknown fluid domain information obtained by prediction is used as new known fluid domain information, so that the continuous prediction of the fluid domain information along with time can be realized under different parameter values. The whole flow field can be restored through node position information contained in the fluid domain information obtained through prediction and flow field parameter information on the nodes, flow field prediction is achieved, and fluid dynamics parameters, performance parameters and the like of the oscillating flapping wing energy collecting system at the prediction time point can be calculated. The method can greatly reduce time cost and economic cost, can efficiently and accurately predict the flow field of the oscillating flapping wing energy acquisition system, further can obtain fluid dynamics parameters, performance parameters and the like of the oscillating flapping wing energy acquisition system, and has important significance for research, design and development of the oscillating flapping wing energy acquisition system.
The method has the advantages that on one hand, compared with the traditional commercial CFD software, the time cost can be greatly saved, for similar flow fields or similar flow conditions, the method can directly predict the corresponding flow field change conditions, and can further calculate the fluid dynamics parameters, the performance parameters and the like of the oscillating flapping wing energy collecting system under the corresponding conditions. On the other hand, compared with experimental research, the method can greatly save economic cost and does not need expensive experimental instruments. In addition, compared with other artificial intelligence technologies, the method has the advantages that direct conversion from the fluid domain information at the known time point to the fluid domain information at the next unknown time point can be realized, so that the data structure is clear, and the further processing and analysis of the data are very convenient.
Drawings
FIG. 1 is a flow chart of the present invention for flow field prediction for an oscillating flapping wing energy harvesting system based on transfer learning;
FIG. 2 is a schematic diagram of a physical model of the oscillating flapping wing energy harvesting system of the present invention;
FIG. 3 is a spatial discrete model and a partial enlarged view of a grid of the oscillating flapping wing energy harvesting system of the present invention;
FIG. 4 is an illustration of a GAN migration learning flow field prediction network of the oscillating flapping wing energy harvesting system of the present invention;
FIG. 5 is an illustration of the oscillating flapping wing energy harvesting system generation network (G Net) architecture of the present invention;
FIG. 6 is an illustration of the oscillating flapping wing energy harvesting system discrimination network (D Net) architecture of the present invention;
FIG. 7 is a schematic diagram of loss functions and influence ranges of G Net and D Net of the oscillating flapping wing energy collecting system of the present invention;
FIG. 8 is a graph comparing the theoretical results and the predicted results of the pressure field of the oscillating flapping wing energy acquisition system of the present invention; wherein FIG. 8(a) is theoretical fluency and FIG. 8(b) is predicted fluency;
FIG. 9 is a comparison graph of the theoretical result and the predicted result of the X-direction velocity field of the oscillating flapping wing energy collecting system of the present invention; wherein FIG. 9(a) is theoretical fluency and FIG. 9(b) is predicted fluency;
FIG. 10 is a comparison graph of the theoretical results and the predicted results of the velocity field in the Y direction of the oscillating flapping wing energy collecting system of the present invention; where FIG. 10(a) is theoretical fluency and FIG. 10(b) is predicted fluency.
Detailed Description
The invention is described in more detail below with reference to a specific example according to the summary of the invention. The following is an application of the present invention, but not limited to this, and the implementer may change the parameters according to the specific problem and the actual application.
The invention provides a method for predicting a flow field of an oscillating flapping wing energy acquisition system based on transfer learning. The method comprises the steps of acquiring node position information and flow field parameter information on nodes after a fluid domain is spatially dispersed, performing matrixing processing and standardization processing, distributing a source domain and a target domain according to a set proportion, further extracting input data sets and corresponding theoretical output data sets of the source domain and the target domain, then respectively establishing a generation network (G Net) and a discrimination network (D Net) of generative antagonistic neural network (GAN) transfer learning, training, realizing fluid domain information prediction under different parameter values, and taking predicted unknown fluid domain information as new known fluid domain information, namely realizing continuous prediction of the fluid domain information along with time under different parameter values; the whole flow field is restored through node position information contained in the fluid domain information obtained through prediction and flow field parameter information on the nodes, flow field prediction is achieved, fluid dynamics parameters and performance parameters of the oscillating flapping wing energy collecting system at the prediction time point are calculated, and therefore researchers can analyze the oscillating flapping wing energy collecting system conveniently.
According to the flow chart of the flow field prediction of the oscillation flapping wing energy acquisition system based on the transfer learning in the figure 1, the flow field prediction of different Re values is carried out on the oscillation flapping wing energy acquisition system of a specific wing section, the physical model schematic diagram of the system is shown in the figure 2, and the system comprises the wing section 1, an inner area 2 which does rigid motion along with the wing section, an inner area interface 3, an outer area 4 which adopts a moving grid technology, an inlet boundary 5, an outlet boundary 6 and a symmetrical plane 7. The spatially discrete model and the grid partial magnification of the system are shown in fig. 3.
Firstly, obtaining model data of an oscillating flapping wing energy acquisition system;
and respectively collecting known fluid domain information of a plurality of continuous time-dispersed time points of the model aiming at different Re values, wherein the known fluid domain information comprises node position information and node flow field parameter information of the fluid domain after space dispersion. The node position information comprises an x coordinate and a y coordinate of a node, and the flow field parameter information comprises a speed u of a fluid on the node in the x direction, a speed v of the fluid on the node in the y direction and a pressure value p. Carrying out space discretization on the model to obtain grid information Hj,qPerforming unsteady state numerical simulation calculation, calculating an application layer flow model, spatially adopting a second-order difference format, temporally adopting a first-order difference format, solving by using a dynamic grid technology, and obtaining node position information at each time point according to a solving result
Figure BDA0002192911010000121
And flow field parameter information on the node
Figure BDA0002192911010000122
Wherein J is 1,2,.. gth, J is the number of grid units, Q is 1,2,.. gth, Q is the number of nodes of each grid unit, here, a two-dimensional four-node structured grid, Q is 1,2,3,4, which respectively represents the number of four nodes of each grid unit, R is 1,2,. gth, R is the total number of different selected Re values, T is 1,2,. gth, T is the number of time points after time dispersion, I is 1,2,. gth, I is 1,2The total number of nodes, N, 1,2, M, 3, N, M, N, M, p, M, N, 1,2, M, p, M, p, M, p, M, p, M, y, v, w, p, w, y, p, w, y, w, y, w, p, w, y, w, p, w, y, w, p, w.
Secondly, preprocessing the model data of the oscillating flapping wing energy collecting system;
regular matrixing is carried out on the model data of the oscillating flapping wing energy acquisition system, and then node position information is rearranged
Figure BDA0002192911010000123
And flow field parameter information on the node
Figure BDA0002192911010000124
Corresponding elements in the node are transformed to obtain the position information of the nodes after matrixing
Figure BDA0002192911010000125
And flow field parameter information on the node
Figure BDA0002192911010000126
Where a 1,2, the., a, B1, 2, the., B, C1, 2, the., C, A, B, C are the number of nodes in three mutually orthogonal directions, respectively, satisfying a × B × C ═ I, where in the problem, a is the number of nodes in the circumferential direction along the airfoil surface, B is the number of nodes in the radial direction along the airfoil, and where in the problem, C is the two-dimensional plane problem, C1.
For the node position information after matrixing
Figure BDA0002192911010000127
And flow field parameter information on the node
Figure BDA0002192911010000128
Carrying out standardization processing to obtain node position information G after standardization processingr,t,a,b,c,nAnd flow field parameter information F on the noder,t,a,b,c,m。
For the node position information after matrixing
Figure BDA0002192911010000129
The standardization processing method comprises the following steps:
Figure BDA00021929110100001210
Figure BDA00021929110100001211
Figure BDA0002192911010000131
wherein, MaxGnDenotes the maximum value of the nth coordinate dimension of a node among all its data, N is 1,2nRespectively representing the maximum value of the x-coordinate and the y-coordinate of the node, MinG, respectively, in all its datanDenotes the minimum value of the nth coordinate dimension of the node among all its data, N is 1,2nRespectively, the x-coordinate and the y-coordinate of the node each represent the minimum of all of its data.
To the parameter information of the flow field on the nodes after matrixing
Figure BDA0002192911010000132
The standardization processing method comprises the following steps:
Figure BDA0002192911010000133
Figure BDA0002192911010000139
Figure BDA0002192911010000134
wherein, MaxFmThe M-th parameter of a node is represented as the maximum value of all data, M is 1,2,.. the M is the number of flow field parameters to be considered, three parameters of the speed u of the fluid on the node in the x direction, the speed v of the fluid on the node in the y direction and the pressure value p are considered, and M is 1,2,3, MaxFmThe maximum value, MinF, of each of the three parameters, i.e., the velocity u in the x-direction, the velocity v in the y-direction, and the pressure value p, of the fluid at the node pointmThe M-th parameter of a node is represented as the minimum value of all data, M is 1,2,.. the M is the number of flow field parameters to be considered, three parameters of the speed u of the fluid on the node in the x direction, the speed v of the fluid on the node in the y direction and the pressure value p are considered, and M is 1,2,3, MinFmThe three parameters, i.e. the velocity u in the x-direction of the fluid at the node, the velocity v in the y-direction and the pressure value p, respectively, represent the minimum of all their data.
Thirdly, splitting a source domain and a target domain of the model data of the oscillating flapping wing energy collecting system;
the normalized node position information G corresponding to different values selected by the parametersr,t,a,b,c,nAnd flow field parameter information F on the noder,t,a,b,c,mSplitting at random or according to requirement according to a certain proportion, using one part as source domain and source domain node position information
Figure BDA0002192911010000135
And flow field parameter information on the node
Figure BDA0002192911010000136
Another part is used as target domain, target domain node position information
Figure BDA0002192911010000137
And flow field parameter information on the node
Figure BDA0002192911010000138
Wherein RS is the total number of different values selected as the parameters of the source domain, RT is 1,2, and RT is the total number of different values selected as the parameters of the target domain, and RS + RT is satisfied. Here, a 1:1 relationship is selected to select the source domain and the target domain, and RS ═ RT ═ R/2.
Fourthly, respectively extracting respective input data sets and theoretical output data sets in the source domain and the target domain;
the number N of continuous time points required according to actual input dataTBFor example, where N is selectedTBAnd 2, further extracting data of the source domain and the target domain to obtain respective input data sets and theoretical output data sets.
For the source domain, the node position information of the source domain is firstly
Figure BDA0002192911010000141
And flow field parameter information on the node
Figure BDA0002192911010000142
Splicing to obtain the total data information of the source domain
Figure BDA0002192911010000143
Where K is 1,2, the., K is N + M, where N is 2, where three parameters of the velocity u in the x direction, the velocity v in the y direction, and the pressure value p of the fluid on the node are considered, and M is 3, where K is 5, K is 1,2,3,4,5, which respectively represent the x coordinate and the y coordinate of the node and the velocity u in the x direction, the velocity v in the y direction, and the pressure value p of the fluid on the node.
The source domain input dataset is:
Figure BDA0002192911010000144
wherein each source domain input data is
Figure BDA0002192911010000145
ES, which is the total amount of source domain data and satisfies ES RS × (T-N)TB)。
The corresponding source domain theoretical output data set is:
Figure BDA0002192911010000146
wherein each source domain theoretical output data is
Figure BDA0002192911010000147
ES, which is the total amount of source domain data and satisfies ES RS × (T-N)TB)。
For the target domain, firstly, the node position information of the target domain
Figure BDA0002192911010000148
And flow field parameter information on the node
Figure BDA0002192911010000149
Splicing to obtain the total data information of the target domain
Figure BDA00021929110100001410
Where K is 1,2, the., K is N + M, where N is 2, where three parameters of the velocity u in the x direction, the velocity v in the y direction, and the pressure value p of the fluid on the node are considered, and M is 3, where K is 5, K is 1,2,3,4,5, which respectively represent the x coordinate and the y coordinate of the node and the velocity u in the x direction, the velocity v in the y direction, and the pressure value p of the fluid on the node.
The target domain input dataset is:
Figure BDA0002192911010000151
wherein each target domain inputs data as
Figure BDA0002192911010000152
ET 1, 2.. ET, ET is the total amount of target domain data, and there may be a plurality of time period data missing, that is, ET RT × (T-N) may not be satisfiedTB)。
The corresponding target domain theoretical output data set is:
Figure BDA0002192911010000153
wherein each target domain theoretical output data is
Figure BDA0002192911010000154
ET 1, 2.. ET, ET is the total amount of target domain data, and there may be a plurality of time segment data deletions corresponding to the target domain input data set, i.e., ET RT × (T-N) may not be satisfiedTB)。
Fifthly, constructing a fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system;
the fluid domain information prediction GAN transfer learning network of the oscillating flapping wing energy acquisition system comprises a generation network (G Net) and a discrimination network (D Net), and the specific structure of the GAN transfer learning network is shown in FIG. 4.
Firstly, a convolution residual error network is adopted to establish a characteristic extraction part of G Net, and Input data Input from a source domain is realizedSDOr target field Input data InputTDExtracting corresponding source domain fluid domain Feature information FeatureSDOr target domain fluid domain Feature information FeatureTDThen, a deconvolution residual error network is adopted to establish a predicted fluid domain information part of G Net, and the Feature information Feature of the source domain fluid domain is realizedSDOr target domain fluid domain Feature information FeatureTDGenerating corresponding source domain actual Output data OutputSDOr target domain actual Output data OutputTD. Wherein each source domain fluid domain characteristic information
Figure BDA0002192911010000155
ES, each source domain actually outputs data as
Figure BDA0002192911010000156
ES, characteristic information of each target domain fluid domain
Figure BDA0002192911010000157
ET 1,2, ET, each target field actually outputting data as
Figure BDA0002192911010000158
The ET 1,2, the ET and the G Net have 21 layers of networks, and a specific G Net framework is shown in fig. 5. Secondly, a D Net network is established by adopting a convolution residual error network to judge whether the fluid domain information is true or false, and the actual output data of each source domain is realized
Figure BDA0002192911010000159
And each source domain theoretical output data corresponding to the source domain theoretical output data
Figure BDA00021929110100001510
To obtain the corresponding discrimination result
Figure BDA00021929110100001511
Is 0 and
Figure BDA00021929110100001512
1, the data sets of the discrimination result are respectively expressed as D _ OutputSDAnd D _ LableSDSimultaneously undertake the actual output of data from each target domain
Figure BDA0002192911010000161
Obtaining a discrimination result
Figure BDA0002192911010000162
The corresponding data set of the discrimination result is D _ OutputTDAnd a two-class loss function part for calculating network training on G Net, wherein D Net has 19 layers of networks, and a specific D Net framework is shown in FIG. 6.
With the quantity SSDA batch of source domain input data and quantity STDIs used to calculate the loss of the maximum mean difference, where S is takenSD＝16，STD＝16。
In a GAN networkG Net of (1) adopts a two-norm loss function L2Loss of class, two-class loss function LbceGLoss function of loss and maximum mean difference LmmdA weighted average of loss as a function of total loss.
Two-norm loss function L2The formula of _ loss is:
Figure BDA0002192911010000163
binary loss function LbceGThe formula of _ loss is:
Figure BDA0002192911010000164
maximum mean difference loss function LmmdThe formula of _ loss is:
Figure BDA0002192911010000165
wherein
Figure BDA0002192911010000168
Representing the Regenerated Kernel Hilbert Space (RKHS), phi (-) represents the mapping function of the original feature space to RKHS, further developed to obtain:
Figure BDA0002192911010000166
for ease of calculation, a simplified equation is introduced:
Figure BDA0002192911010000167
where δ is a parameter of value.
The G Net total loss function is as follows:
LGNet_loss＝L2_loss+WbceG×LbceG_loss+Wmmd×Lmmd_loss
wherein, WbceGIs a binary loss function LbceGWeight of _ loss, WmmdIs a maximum mean difference loss function LmmdWeight of _ loss, according to experience WbceG∈[0.04,0.2]，Wmmd∈[0.04,0.2]At which W is takenbceG＝0.1，WmmdThe loss functions and their range of influence are shown in fig. 7, which is 0.1.
D Net in GAN network adopts two-classification loss of source domain actual output data
Figure BDA0002192911010000171
Two-classification loss L of source domain theoretical output databceLbLoss of class two for the _lossand actual output data of the target domain
Figure BDA0002192911010000172
The sum as a function of total loss.
Binary loss of source domain actual output data
Figure BDA0002192911010000173
The formula is as follows:
Figure BDA0002192911010000174
two-classification loss L of source domain theoretical output databceLbThe formula of _ loss is:
Figure BDA0002192911010000175
two-classification loss of target domain actual output data
Figure BDA0002192911010000176
The formula is as follows:
Figure BDA0002192911010000177
the D Net total loss function is as follows:
Figure BDA0002192911010000178
the loss functions and their range of influence are shown in figure 7.
Sixthly, training a fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system;
firstly, the number of the oscillating flapping wing energy acquisition systems is SSDThe fluid domain information of a batch of source domain input data is predicted through G Net in a GAN transfer learning network, and the obtained quantity is SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, and then the number of the oscillating flapping wing energy acquisition systems is STDThe fluid domain information prediction is carried out on a batch of target domain input data for calculating the maximum mean difference loss through G Net in a GAN migration learning network, and the obtained quantities are STDA batch of target domain fluid domain characteristic information and a batch of target domain actual output data.
Then the obtained quantity is SSDA batch of source domain actual output data, the corresponding quantity is SSDA batch of source domain theoretical output data and the quantity is STDAnd (3) distinguishing the batch of target domain actual output data through D Net in the GAN transfer learning network, and updating parameters of the D Net once according to an obtained distinguishing result, so that the D Net can better distinguish a fluid domain information prediction result and a fluid domain information theoretical result predicted by G Net.
Then the obtained quantity is SSDA batch of source domain actual output data and the number STDThe actual output data of a batch of target domains are respectively judged through D Net which is subjected to parameter updating once in the GAN migration learning network, and according to the obtained judgment result and the obtained quantity are SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, the quantity is STDA batch of target domain fluid domain characteristic informationThe known number corresponding to the actual output of the source domain is SSDThe G Net is subjected to parameter updating once by a batch of source domain theoretical output data, so that the fluid domain information prediction result of the G Net is closer to the fluid domain information theoretical result, and the judgment of the fluid domain information prediction result and the fluid domain information theoretical result of the D Net is confused as much as possible.
Through the continuous and repeated confrontation of G Net and D Net, the judgment of the fluid domain information prediction result and the fluid domain information theoretical result by the D Net is more accurate, and the fluid domain information prediction result of the G Net is forced to be closer to the fluid domain information theoretical result.
In the training process, the optimizer of G Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data loop training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after every 100 times of data loop training. The optimizer of D Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data cycle training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after every 100 times of data cycle training.
Seventhly, processing and analyzing flow field prediction results of the oscillating flapping wing energy acquisition system
The trained fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system is used for predicting fluid domain information under other Re, and the predicted unknown fluid domain information at the next time point is used as new known fluid domain information, so that the fluid domain information prediction at the next time point can be continuously carried out, and the continuous prediction of the fluid domain information along with time is realized under different parameter values. The whole flow field can be restored through node position information and flow field parameter information on the nodes contained in the fluid domain information obtained through prediction, the flow field prediction results are shown in fig. 8-10, theoretical flow field conditions and predicted flow field results of a pressure field, an X-direction speed field and a Y-direction speed field at a certain time point are respectively given, the predicted flow field results are well matched with the theoretical flow field conditions, and errors are within a reasonable range. Furthermore, fluid dynamics parameters and performance parameters such as torque, lift coefficient and resistance coefficient of the oscillating flapping wing energy collecting system at a prediction time point can be calculated through a flow field prediction result, so that researchers can analyze the oscillating flapping wing energy collecting system.

Claims (7)

1. A method for predicting a flow field of an oscillating flapping wing energy acquisition system based on transfer learning is characterized in that unknown fluid domain information of a next time point is predicted from known fluid domain information of a plurality of continuous time points after time dispersion; specifically, by acquiring node position information and node flow field parameter information of a fluid domain after spatial dispersion, performing matrixing processing and standardization processing, distributing a source domain and a target domain according to a set proportion, further extracting input data sets and corresponding theoretical output data sets of the source domain and the target domain, then respectively establishing a generation network and a discrimination network of generation type antagonistic neural network transfer learning, after training, realizing fluid domain information prediction under different parameter values, and taking unknown fluid domain information obtained by prediction as new known fluid domain information, namely realizing continuous prediction of the fluid domain information along with time under different parameter values; restoring the whole flow field through node position information contained in the fluid domain information obtained through prediction and flow field parameter information on the nodes, realizing flow field prediction, and calculating fluid dynamics parameters and performance parameters of the oscillating flapping wing energy acquisition system at a prediction time point so as to provide researchers with analysis on the oscillating flapping wing energy acquisition system; the method specifically comprises the following steps:
1) obtaining model data of oscillating flapping wing energy collecting system
Respectively collecting known fluid domain information of a plurality of continuous time-dispersed time points of the model aiming at different parameter values, wherein the known fluid domain information comprises node position information and node flow field parameter information of the fluid domain after space dispersion; carrying out space discretization on the model to obtain grid information Hj,qPerforming unsteady state numerical simulation calculation, calculating an application layer flow model, spatially adopting a second-order difference format, temporally adopting a first-order difference format, solving by using a dynamic grid technology, and obtaining node position information at each time point according to a solving result
Figure FDA0003033230050000011
And flow field parameter information on the node
Figure FDA0003033230050000012
J is 1, 2., J is the number of grid units, Q is 1, 2., Q is the number of nodes of each grid unit, R is 1, 2., R is R, R is the total number of different selected parameter values, T is 1,2, …, T is the number of time points after time dispersion, I is 1, 2., I is the total number of nodes, N is 1, 2., N is the number of coordinate dimensions, M is 1, 2., M is the number of flow field parameters to be considered;
2) model data preprocessing of oscillating flapping wing energy collecting system
Regular matrixing is carried out on the model data of the oscillating flapping wing energy acquisition system, and then node position information is rearranged
Figure FDA0003033230050000013
And flow field parameter information on the node
Figure FDA0003033230050000014
Corresponding elements in the node are transformed to obtain the position information of the nodes after matrixing
Figure FDA0003033230050000015
And flow field parameter information on the node
Figure FDA0003033230050000016
Wherein a is 1,2, a, B is 1,2, a, B, C is 1,2, C, A, B, C are the number of nodes in three mutually orthogonal directions, respectively, and a × B × C is satisfied;
for the node position information after matrixing
Figure FDA0003033230050000021
And flow field parameter information on the node
Figure FDA0003033230050000022
Carrying out standardization processing to obtain node position information G after standardization processingr,t,a,b,c,nAnd flow field parameter information F on the noder,t,a,b,c,m；
3) Model data splitting source domain and target domain of oscillating flapping wing energy acquisition system
The normalized node position information G corresponding to different values selected by the parametersr,t,a,b,c,nAnd flow field parameter information F on the noder,t,a,b,c,mSplitting at random or according to the requirement according to the set proportion, using one part as the source domain and the node position information of the source domain
Figure FDA0003033230050000023
And flow field parameter information on the node
Figure FDA0003033230050000024
Another part is used as target domain, target domain node position information
Figure FDA0003033230050000025
And flow field parameter information on the node
Figure FDA0003033230050000026
RS, RS are different total numbers of values selected as parameters of a source domain, RT is 1,2,., RT are different total numbers of values selected as parameters of a target domain, and RS + RT is satisfied;
4) extracting respective input data set and theoretical output data set in source domain and target domain
The number N of continuous time points required according to actual input dataTBCarrying out further data extraction on the source domain and the target domain to obtain respective input data set and theoretical output data set;
for the source domain, the node position information of the source domain is firstly
Figure FDA0003033230050000027
And node upflow fieldParameter information
Figure FDA0003033230050000028
Splicing to obtain the total data information of the source domain
Figure FDA0003033230050000029
Where K1, 2,., K N + M, a source domain Input dataset Input is obtainedSDAnd a corresponding source domain theoretical output data set LableSD；
For the target domain, firstly, the node position information of the target domain
Figure FDA00030332300500000210
And flow field parameter information on the node
Figure FDA00030332300500000211
Splicing to obtain the total data information of the target domain
Figure FDA00030332300500000212
Where K is 1,2, …, K is N + M, the target domain Input dataset Input is obtainedTDAnd a corresponding target domain theoretical output data set LableTD；
5) Fluid domain information prediction GAN migration learning network for constructing oscillating flapping wing energy acquisition system
The fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system comprises a generation network and a judgment network, wherein the generation network is called G Net for short, and the judgment network is called D Net for short;
firstly, a convolution residual error network is adopted to establish a characteristic extraction part of G Net, and Input data Input from a source domain is realizedSDOr target field Input data InputTDExtracting corresponding source domain fluid domain Feature information FeatureSDOr target domain fluid domain Feature information FeatureTDThen, a deconvolution residual error network is adopted to establish a predicted fluid domain information part of G Net, and the Feature information Feature of the source domain fluid domain is realizedSDOr target domain fluid domain Feature information FeatureTDGenerating correspondencesSource domain actual Output data OutputSDOr target domain actual Output data OutputTD(ii) a Wherein each source domain fluid domain characteristic information
Figure FDA0003033230050000031
Figure FDA0003033230050000032
Each source domain actually outputs data of
Figure FDA0003033230050000033
Fluid domain characteristic information of each target domain
Figure FDA0003033230050000034
Each target domain actually outputs data of
Figure FDA0003033230050000035
G Net comprises 21 layers of networks; secondly, a D Net network is established by adopting a convolution residual error network to judge whether the fluid domain information is true or false, and the actual output data of each source domain is realized
Figure FDA0003033230050000036
And each source domain theoretical output data corresponding to the source domain theoretical output data
Figure FDA0003033230050000037
To obtain the corresponding discrimination result
Figure FDA0003033230050000038
Is 0 and
Figure FDA0003033230050000039
1, the data sets of the discrimination result are respectively expressed as D _ OutputSDAnd D _ LableSDSimultaneously undertake the actual output of data from each target domain
Figure FDA00030332300500000310
Obtaining a discrimination result
Figure FDA00030332300500000311
The corresponding data set of the discrimination result is D _ OutputTDThe two-classification loss function part is used for calculating the network training of G Net, and D Net has 19 layers of networks in total;
with the quantity SSDA batch of source domain input data and quantity STDA batch of target domain input data for calculating maximum mean difference loss;
g Net in GAN network adopts two-norm loss function L2Loss of class, two-class loss function LbceGLoss function of loss and maximum mean difference LmmdA weighted average of _ loss as a total loss function;
d Net in GAN network adopts two-classification loss of source domain actual output data
Figure FDA00030332300500000312
Two-classification loss L of source domain theoretical output databceLbLoss of class two for the _lossand actual output data of the target domain
Figure FDA00030332300500000313
The sum as a function of total loss;
6) fluid domain information prediction GAN migration learning network for training oscillating flapping wing energy acquisition system
Firstly, the number of the oscillating flapping wing energy acquisition systems is SSDThe fluid domain information of a batch of source domain input data is predicted through G Net in a GAN transfer learning network, and the obtained quantity is SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, and then the number of the oscillating flapping wing energy acquisition systems is STDThe fluid domain information prediction is carried out on a batch of target domain input data for calculating the maximum mean difference loss through G Net in a GAN migration learning network, and the obtained quantities are STDA batch of target domain fluid domain characteristic information and a batch of target domain actual output data;
then the obtained quantity is SSDA batch of source domain actual output data, the corresponding quantity is SSDA batch of source domain theoretical output data and the quantity is STDThe batch of target domain actual output data are respectively judged through D Net in the GAN transfer learning network, and the D Net is subjected to parameter updating once according to the obtained judgment result, so that the D Net can better distinguish a fluid domain information prediction result predicted by G Net and a fluid domain information theoretical result;
then the obtained quantity is SSDA batch of source domain actual output data and the number STDThe actual output data of a batch of target domains are respectively judged through D Net which is subjected to parameter updating once in the GAN migration learning network, and according to the obtained judgment result and the obtained quantity are SSDA batch of source domain fluid domain characteristic information and a batch of source domain actual output data, the quantity is STDIs known to correspond to the actual output of the source domain, and has a quantity SSDThe G Net is subjected to parameter updating once by a batch of source domain theoretical output data, so that the fluid domain information prediction result of the G Net is closer to the fluid domain information theoretical result, and the judgment of the fluid domain information prediction result and the fluid domain information theoretical result of the D Net is confused as much as possible;
through the continuous and repeated confrontation of G Net and D Net, the judgment of the fluid domain information prediction result and the fluid domain information theoretical result by the D Net is more accurate, and the fluid domain information prediction result of the G Net is forced to be closer to the fluid domain information theoretical result;
7) flow field prediction result processing and analysis of oscillating flapping wing energy acquisition system
The trained fluid domain information prediction GAN migration learning network of the oscillating flapping wing energy acquisition system is used for predicting fluid domain information under other parameter values, and the predicted unknown fluid domain information at the next time point is used as new known fluid domain information, so that the fluid domain information prediction at the next time point can be continuously carried out, and the continuous prediction of the fluid domain information along with time is realized under different parameter values; the whole flow field can be restored through node position information and flow field parameter information on the nodes contained in the fluid domain information obtained through prediction; further, the torque, the lift coefficient and the resistance coefficient of the oscillating flapping wing energy collecting system at the prediction time point are calculated according to the flow field prediction result, so that a researcher can analyze the oscillating flapping wing energy collecting system.
2. The method for predicting the flow field of the oscillating flapping wing energy collecting system based on the transfer learning of claim 1, wherein in the step 2), the position information of the nodes after the matrixing is carried out
Figure FDA0003033230050000051
And flow field parameter information on the node
Figure FDA0003033230050000052
The normalization process was performed as follows:
for the node position information after matrixing
Figure FDA0003033230050000053
The standardization processing method comprises the following steps:
Figure FDA0003033230050000054
Figure FDA0003033230050000055
Figure FDA0003033230050000056
wherein, MaxGnRepresents the maximum value of the nth coordinate dimension of the node in all data, wherein N is 1,2nRepresents the minimum value of the nth coordinate dimension of the node in all data thereof；
To the parameter information of the flow field on the nodes after matrixing
Figure FDA0003033230050000057
The standardization processing method comprises the following steps:
Figure FDA0003033230050000058
Figure FDA0003033230050000059
Figure FDA00030332300500000510
wherein, MaxFmRepresents the maximum value of the mth parameter of the node in all data, wherein M is 1,2mRepresents the minimum value of the mth parameter of the node in all data thereof.
3. The method for predicting the flow field of the oscillatory flapping wing energy collecting system based on the transfer learning of claim 2, wherein in the step 4), a source domain Input data set Input is obtainedSDAnd a corresponding source domain theoretical output data set LableSDThe method comprises the following steps:
the source domain input dataset is:
Figure FDA00030332300500000511
wherein each source domain input data is
Figure FDA00030332300500000512
ES is the total amount of source domain data and satisfies
ES＝RS×(T-NTB)
The corresponding source domain theoretical output data set is:
Figure FDA0003033230050000061
wherein each source domain theoretical output data is
Figure FDA0003033230050000062
ES is the total amount of source domain data, and satisfies ES RS x (T-N)TB)。
4. The method for predicting the flow field of the oscillatory flapping wing energy collecting system based on the transfer learning of claim 3, wherein in the step 4), the target domain Input data set Input is obtainedTDAnd a corresponding target domain theoretical output data set LableTDThe method comprises the following steps:
the target domain input dataset is:
Figure FDA0003033230050000063
wherein each target domain inputs data as
Figure FDA0003033230050000064
ET is the total amount of target domain data, and a plurality of time period data are missing, namely ET ═ RT × (T-N) is not satisfiedTB)；
The corresponding target domain theoretical output data set is:
Figure FDA0003033230050000065
wherein each target domain theoretical output data is
Figure FDA0003033230050000066
ET is the total amount of target domain data, and there are a plurality of time segment data losses corresponding to the target domain input data set, that is, ET ═ RT × (T-N) is not satisfiedTB)。
5. The method for predicting the flow field of the oscillatory flapping wing energy collecting system based on the transfer learning of claim 4, wherein in the step 5), G Net in a GAN network adopts a two-norm loss function L2Loss of class, two-class loss function LbceGLoss function of loss and maximum mean difference LmmdThe weighted average of _ loss is used as the total loss function, and the specific loss function formula is as follows:
two-norm loss function L2The formula of _ loss is:
Figure FDA0003033230050000067
binary loss function LbceGThe formula of _ loss is:
Figure FDA0003033230050000071
maximum mean difference loss function LmmdThe formula of _ loss is:
Figure FDA0003033230050000072
wherein
Figure FDA0003033230050000073
Representing the regenerated kernel hilbert space RKHS, phi (-) represents the mapping function from the original feature space to the RKHS, and further develops to obtain:
Figure FDA0003033230050000074
for ease of calculation, a simplified equation is introduced:
Figure FDA0003033230050000075
wherein delta is a certain value parameter;
the G Net total loss function is as follows:
LGNet_loss＝L2_loss+WbceG×LbceG_loss+Wmmd×Lmmd_loss
wherein, WbceGIs a binary loss function LbceGWeight of _ loss, WmmdIs a maximum mean difference loss function LmmdWeight of _ loss, according to experience WbceG∈[0.04,0.2]，Wmmd∈[0.04,0.2]。
6. The method for predicting the flow field of the oscillatory flapping wing energy collecting system based on the transfer learning of claim 5, wherein in the step 5), D Net in the GAN network adopts the binary classification loss of the actual output data of the source domain
Figure FDA0003033230050000076
Two-classification loss L of source domain theoretical output databceLbLoss of class two for the _lossand actual output data of the target domain
Figure FDA0003033230050000077
The sum is taken as a total loss function, and the specific loss function formula is as follows:
binary loss of source domain actual output data
Figure FDA0003033230050000078
The formula is as follows:
Figure FDA0003033230050000081
two-classification loss L of source domain theoretical output databceLbThe formula of _ loss is:
Figure FDA0003033230050000082
two-classification loss of target domain actual output data
Figure FDA0003033230050000083
The formula is as follows:
Figure FDA0003033230050000084
the D Net total loss function is as follows:
Figure FDA0003033230050000085
7. the method for predicting the flow field of the oscillatory flapping wing energy collecting system based on the transfer learning of claim 6, wherein in the step 6), in the process of training the GAN transfer learning network, the method for selecting the optimizer and setting the learning rate of G Net and the method for selecting the optimizer and setting the learning rate of D Net are as follows:
in the training process, an optimizer of G Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data cycle training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after each 100 times of data cycle training;
the optimizer of D Net selects an Adam optimization algorithm, the initial learning rate is set to be 0.002, after 200 times of data cycle training, the learning rate is set to be 0.0002, and then the learning rate is reduced by 90% after every 100 times of data cycle training.
CN201910838392.1A 2019-09-05 2019-09-05 A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning Active CN110705029B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910838392.1A CN110705029B (en) 2019-09-05 2019-09-05 A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910838392.1A CN110705029B (en) 2019-09-05 2019-09-05 A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning

Publications (2)

Publication Number Publication Date
CN110705029A CN110705029A (en) 2020-01-17
CN110705029B true CN110705029B (en) 2021-09-07

Family

ID=69194721

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910838392.1A Active CN110705029B (en) 2019-09-05 2019-09-05 A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning

Country Status (1)

Country Link
CN (1) CN110705029B (en)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111444614B (en) * 2020-03-26 2021-12-28 西安交通大学 Flow field reconstruction method based on graph convolution
CN111581784B (en) * 2020-04-17 2021-12-21 浙江大学 A data-driven adaptive quasi-steady-state model for the optimization of flapping wing motion parameters
CN111724487B (en) * 2020-06-19 2023-05-16 广东浪潮大数据研究有限公司 Flow field data visualization method, device, equipment and storage medium
CN111859746B (en) * 2020-07-10 2022-12-09 西安交通大学 A Performance Prediction Method of Turbomachinery under Variable Conditions Based on Flow Field Reconstruction
CN112784508A (en) * 2021-02-12 2021-05-11 西北工业大学 Deep learning-based airfoil flow field rapid prediction method
CN114004170B (en) * 2021-10-19 2025-01-28 苏州湍流数字科技有限公司 Reconstruction method of cascade flow field based on limited measurement points
CN114996658B (en) * 2022-07-20 2022-11-08 中国空气动力研究与发展中心计算空气动力研究所 Projection-based hypersonic aircraft aerodynamic heat prediction method
CN115758911B (en) * 2022-12-07 2023-05-26 中国石油大学(华东) Fusion point cloud residual error network and flow field and pressure field prediction method considering slippage
CN118348789B (en) * 2024-04-15 2025-03-25 西安交通大学 A method, system, device and medium for controlling flapping-wing energy harvesting performance

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102135760A (en) * 2010-12-16 2011-07-27 天津工业大学 Neural network energy coordinated controller for microgrid
CN106251401A (en) * 2016-07-06 2016-12-21 西安交通大学 A kind of generation system and method for hexahedron 20 node unit grid
CN108932554B (en) * 2017-05-26 2021-03-23 西安交通大学 Method and device for configuration optimization of flow field measurement points in wind farms
CN107039975B (en) * 2017-05-27 2020-03-24 上海电气分布式能源科技有限公司 Energy management method for distributed energy system
US20190243933A1 (en) * 2018-02-07 2019-08-08 Incucomm, Inc. System and method that characterizes an object employing virtual representations thereof
CN108563906B (en) * 2018-05-02 2022-03-22 北京航空航天大学 A deep learning-based prediction method for macroscopic properties of short fiber reinforced composites
CN109060001B (en) * 2018-05-29 2021-02-26 浙江工业大学 Multi-working-condition process soft measurement modeling method based on feature transfer learning
CN109102126B (en) * 2018-08-30 2021-12-10 燕山大学 Theoretical line loss rate prediction model based on deep migration learning
CN109918752B (en) * 2019-02-26 2022-12-16 华南理工大学 Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network

Also Published As

Publication number Publication date
CN110705029A (en) 2020-01-17

Similar Documents

Publication Publication Date Title
CN110705029B (en) A flow field prediction method for oscillating flapping-wing energy harvesting system based on transfer learning
CN110288048B (en) Submarine pipeline risk assessment method of SVM directed acyclic graph
Obayashi et al. Multi-objective design exploration for aerodynamic configurations
CN107133181B (en) A kind of construction method of difference wavelet neural network software fault prediction technology
Deng et al. Prediction of transonic flow over supercritical airfoils using geometric-encoding and deep-learning strategies
CN112541572A (en) Residual oil distribution prediction method based on convolutional encoder-decoder network
CN112182306B (en) Uncertain graph-based community discovery method
Chen et al. MVE-Net: An automatic 3-D structured mesh validity evaluation framework using deep neural networks
CN112016253A (en) High-fidelity chaotic polynomial correction method suitable for CFD uncertainty quantification
CN116628854B (en) Wing section aerodynamic characteristic prediction method, system, electronic equipment and storage medium
CN106886657A (en) A kind of FEM model method for building up based on kriging functions
Zhang et al. Convolutional neural networks-based surrogate model for fast computational fluid dynamics simulations of indoor airflow distribution
CN116629352A (en) Hundred million-level parameter optimizing platform
Stein et al. Operator-and template-based modeling of solid geometry for Isogeometric Analysis with application to Vertical Axis Wind Turbine simulation
CN103455612A (en) Method for detecting non-overlapping network communities and overlapping network communities based on two-stage strategy
CN115119242A (en) Cellular network fault diagnosis method based on knowledge and data fusion
CN104732067A (en) Industrial process modeling forecasting method oriented at flow object
CN117540664A (en) Two-dimensional flow field prediction and correction method based on graph neural network
CN113516019A (en) Hyperspectral image unmixing method, device and electronic device
CN119337510A (en) Intelligent reconstruction method of aircraft aerodynamic characteristics based on multi-source data fusion
Hu et al. Better neural PDE solvers through data-free mesh movers
Zhao et al. Review of empowering computer-aided engineering with artificial intelligence
CN118395886B (en) A flow field prediction method based on token selection Transformer
Obayashi et al. Multi-objective design exploration and its application to regional-jet wing design
CN118709560A (en) Machine tool structure optimization method based on self-guided online learning and efficient sampling

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载