CN108763718A - The method for quick predicting of Field Characteristics amount when streaming object and operating mode change - Google Patents
The method for quick predicting of Field Characteristics amount when streaming object and operating mode change Download PDFInfo
- Publication number
- CN108763718A CN108763718A CN201810499129.XA CN201810499129A CN108763718A CN 108763718 A CN108763718 A CN 108763718A CN 201810499129 A CN201810499129 A CN 201810499129A CN 108763718 A CN108763718 A CN 108763718A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- level
- flow field
- deep
- neural network
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of method for quick predicting streaming Field Characteristics amount when object and operating mode change, this method includes：Deep learning network inputs data are generated, which is suitable for any two dimension or 3-D geometric model；Based on the input data and flow field primary condition and working condition, deep learning network structure of the structure for flow field prediction；The deep learning network is trained, to obtain flow field prediction model；The method achieve deep learnings to carry out arbitrary complex geometry object the requirement of flow field prediction, and can consider the influence of different working conditions simultaneously, and extensive use of the deep learning in terms of flow field is made to become possibility；And it can be applied in the occasions such as the foundation of fluid-structure coupling system reduced-order model.
Description
Technical field
The invention belongs to Flow Field Calculation and deep learning field, more particularly to Field Characteristics when streaming object and operating mode change
The method for quick predicting of amount.
Background technology
Tradition is needed using the Flow Field Calculation of CFD (Computational Fluid Dynamics, Fluid Mechanics Computation)
It takes a substantial amount of time and computing resource.When streaming object change, calculating operating mode change, it is required to re-start multiple
Miscellaneous and time-consuming CFD is calculated, this not only consumes a large amount of computing resource, while also resulting in inefficiency.In fact, flow field
As a system, there is its own characteristic, the way for re-starting CFD calculating every time has ignored this point.Deep learning
Simultaneously learning system characteristic, the ability of fast prediction are excavated with autonomous.Therefore, carrying out flow field prediction using deep learning method is
It is a kind of not only feasible, and the method with wide application prospect.
The associated deep learning method developed at present is only capable of to a few simple geometry bodies such as circles in specific work
Under the conditions of condition, generated Flow Field feature is made a prediction.Its network inputs, which can not conclusively show out to study, streams object
The geological information of body, and have many redundant datas；The feature for streaming object can not be considered simultaneously with operating mode.Therefore, traditional
Deep learning flow field prediction technique have significant limitation, the scope of application is very limited, for beyond the pre- of training data range
It is very low to survey accuracy.
Therefore, development is a kind of having generally applicable range, can stream object for difference simultaneously and different operating modes are flowed
Field prediction, general deep learning flow field prediction technique with higher robustness are necessary and have widespread demand.
Invention content
In order to overcome the problems, such as existing deep learning method in the presence of the Predicting Technique of flow field, the present invention propose it is a kind of around
The method for quick predicting of Field Characteristics amount when flowing object and operating mode change, the method achieve deep learnings to complex geometry object
The generalization requirement of Field Characteristics amount prediction is carried out, and can consider the influence of different operating modes simultaneously, deep learning is made to flow
Extensive use in terms of field prediction becomes possibility.
In order to achieve the above object, the present invention adopts the following technical scheme that：
The method for quick predicting of Field Characteristics amount when streaming object and operating mode change, including generate deep learning network inputs
Data and the deep learning model structure predicted for flow field for input data foundation；The specific method is as follows：
Step 1：Generate deep learning network inputs data：
1) partial differential equation of Laplace forms is used to generate flow field grid
To streaming the region around object in preset range, flow field grid is generated using the Laplace equations without source item；
The determination principle of the preset range is：Object will be streamed completely to be included；And ensure in the grid lines generated, close to the range
The grid lines curvature on boundary is 0 or 10-6Below magnitude；In addition, the requirement during Fluid Mechanics Computation should be followed, make grid
Point is intensive as possible, it is therefore an objective to ensure the intranet ruling for streaming object boundary and being generated discrete without distortion；
2) using each mesh point curvature value as deep learning network inputs data
Single channel multi-dimensional matrix (two-dimensional problems are single channel two-dimensional matrix) corresponding with net region is established, wherein each
Element corresponds to a unique mesh point, and index value of the element in single channel multi-dimensional matrix is equal to corresponding mesh point in entire net
Index value in lattice region；Each element data obtain in the following way in single channel multi-dimensional matrix：It is corresponding to calculate this element
Each grid lines curvature at mesh point, and take the product of wherein maximal and minmal value (straight for two-dimensional problems as the element data
It connects and two grid lines curvature is multiplied)；Using gained single channel multi-dimensional matrix as the input data of deep learning network；
Step 2：For the input data that step 1 generates, the deep learning model structure for flow field prediction is established：
1) the complex neural network structure of multilayer convolutional neural networks and deep neural network composition is established
Entire deep learning network is divided to two levels：It is to pass through that first level, which uses multilayer convolutional neural networks, input,
The input data that step 1 obtains, the hierarchical network are not provided with full articulamentum, using the output of the last one convolutional layer as first
The output of a level；It is speed of incoming flow and viscosity and operating mode that second level, which uses deep neural network structure, input,
Condition, and by the output of the first level directly as the hidden layer of the second level；Second level uses complete depth nerve net
Network structure exports the output as entire deep learning network；
2) to the training of complex neural network
To being trained inside multilayer convolutional neural networks and deep neural network, realized by data transfer between two rank
Docking；Specifically, first to the second level, i.e. deep neural network updates each parameter value using a training method, then will
The renewal amount of hidden layer in second level corresponding to the output of the first level passes to the first level, then is used to the first level
Training method；Two rank is all made of after a training method as primary complete learning process, and above-mentioned is repeated
Habit process, the training end condition until meeting deep learning network.
The present invention is compared to the prior art compared with having following advantage：
1. deep learning network inputs data creation method proposed by the present invention can be adapted for having any complex appearance
Stream object, and the feature of institute's research object can be extracted.Therefore this method has wide applicability, and deep learning is made to exist
It is provided with real application value in terms of the prediction of flow field.
2. the present invention, which is applicable not only to difference, streams object, and can consider the influence of operating mode simultaneously.Therefore the invention makes depth
Extensive use of the degree study in terms of the prediction of flow field becomes possibility.
3. the present invention considers the association between different scale range successively by using multilayer convolutional neural networks, finally
The influence of operating mode is introduced by deep neural network.On this basis flow field system is may learn using deep learning method more
More characteristic informations, therefore with compared with the more accurate prediction effect of traditional neural network method.
Description of the drawings
The Establishing process figure of Fig. 1 deep learning models.
Fig. 2 complex neural network structure charts.
Specific implementation mode
Invention is further described in detail with reference to the accompanying drawings and detailed description.
As shown in Figure 1, the present invention streams the method for quick predicting of Field Characteristics amount when object and operating mode change, including generate
Deep learning network inputs data and the deep learning model structure predicted for flow field for input data foundation；Specifically
Method is as follows：
Step 1：Generate deep learning network inputs data：
1) partial differential equation of Laplace forms is used to generate flow field grid
To streaming the region around object in preset range, flow field grid is generated using the Laplace equations without source item.
The determination principle of the preset range is：Object will be streamed completely to be included；And ensure in the grid lines generated, close to the range
The grid lines curvature on boundary is 0 or 10-6Below magnitude.In addition, CFD (Computational Fluid should be followed
Dynamics, Fluid Mechanics Computation) during requirement, keep mesh point intensive as possible, it is therefore an objective to ensure substantially without distortion from
Dissipate the intranet ruling for streaming object boundary and being generated.
2) using each mesh point curvature value as deep learning network inputs data
Single channel multi-dimensional matrix (two-dimensional problems are single channel two-dimensional matrix) corresponding with net region is established, wherein each
Element corresponds to a unique mesh point, and index value of the element in single channel multi-dimensional matrix is equal to corresponding mesh point in entire net
Index value in lattice region.Each element data obtain in the following way in single channel multi-dimensional matrix：It is corresponding to calculate this element
Each grid lines curvature at mesh point, and take the product of wherein maximal and minmal value (straight for two-dimensional problems as the element data
It connects and two grid lines curvature is multiplied).Using gained single channel multi-dimensional matrix as the input data of deep learning network.
Step 2：For the input data that step 1 generates, the deep learning model structure for flow field prediction is established：
1) the complex neural network structure of multilayer convolutional neural networks and deep neural network composition is established
Entire deep learning network is divided to two levels：It is to pass through that first level, which uses multilayer convolutional neural networks, input,
The input data that step 1 obtains, the hierarchical network are not provided with full articulamentum, using the output of the last one convolutional layer as first
The output of a level；It is that the flow fields such as speed of incoming flow, viscosity are initial that second level, which uses deep neural network structure, input,
Condition and working condition, and by the output of first level directly as the hidden layer of the second level.Second level is using complete
Deep neural network structure, export output as entire deep learning network.Fig. 2 is the complex neural network structure.
2) complex neural network is trained
Multilayer convolutional neural networks directly use conventional exercises algorithm such as BP (Back with deep neural network inside
Propagation, error backpropagation algorithm) algorithm is trained, and pass through data transfer between two rank and realizes docking.It is specific real
Shi Zhong, first to the second level, i.e. deep neural network updates each parameter value using a training method, then by the second level
In the first level output corresponding to the renewal amount of hidden layer pass to the first level, then to the first level using primary training
Method.Two rank is all made of after a training method as primary complete learning process, above-mentioned learning process is repeated, directly
To the training end condition for meeting deep learning network.
When carrying out the prediction of Field Characteristics amount, step 1 the method is used to generate input data first, using institute in step 2
Method input duty parameter is stated, deep learning network is run, acquired results are Field Characteristics amount to be predicted.
Claims (1)
1. the method for quick predicting of Field Characteristics amount when streaming object and operating mode change, it is characterised in that：Including generating depth
It practises network inputs data and establishes the deep learning model structure for flow field prediction for the input data；Specific method is such as
Under：
Step 1：Generate deep learning network inputs data：
1) partial differential equation of Laplace forms is used to generate flow field grid
To streaming the region around object in preset range, flow field grid is generated using the Laplace equations without source item；This is pre-
If the determination principle of range is：Object will be streamed completely to be included；And ensure in the grid lines generated, close to the range boundary
Grid lines curvature be 0 or 10-6Below magnitude；In addition, the requirement during Fluid Mechanics Computation should be followed, mesh point is made to use up
It measures intensive, it is therefore an objective to ensure the intranet ruling for streaming object boundary and being generated discrete without distortion；
2) using each mesh point curvature value as deep learning network inputs data
Single channel multi-dimensional matrix corresponding with net region is established, wherein each element corresponds to a unique mesh point, element
Index value in single channel multi-dimensional matrix is equal to index value of the corresponding mesh point in entire net region；Single channel multidimensional square
Each element data obtain in the following way in battle array：Each grid lines curvature at the corresponding mesh point of this element is calculated, and is taken wherein
The product of maximal and minmal value is as the element data；Using gained single channel multi-dimensional matrix as the input number of deep learning network
According to；
Step 2：For the input data that step 1 generates, the deep learning model structure for flow field prediction is established：
1) the complex neural network structure of multilayer convolutional neural networks and deep neural network composition is established
Entire deep learning network is divided to two levels：It is to pass through step 1 that first level, which uses multilayer convolutional neural networks, input,
Obtained input data, the hierarchical network are not provided with full articulamentum, using the output of the last one convolutional layer as first layer
The output of grade；Second level uses deep neural network structure, and input is speed of incoming flow and viscosity and working condition,
And by the output of the first level directly as the hidden layer of the second level；Second level uses complete deep neural network knot
Structure exports the output as entire deep learning network；
2) to the training of complex neural network
To being trained inside multilayer convolutional neural networks and deep neural network, pass through data transfer realization pair between two rank
It connects；Specifically, first to the second level, i.e., deep neural network updates each parameter value using a training method, then by the
The renewal amount of hidden layer in two levels corresponding to the output of the first level passes to the first level, then uses one to the first level
Secondary training method；Two rank is all made of after a training method as primary complete learning process, and above-mentioned study is repeated
Process, the training end condition until meeting deep learning network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810499129.XA CN108763718B (en) | 2018-05-23 | 2018-05-23 | Method for quickly predicting characteristic quantity of flow field when flow-around object and working condition change |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810499129.XA CN108763718B (en) | 2018-05-23 | 2018-05-23 | Method for quickly predicting characteristic quantity of flow field when flow-around object and working condition change |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108763718A true CN108763718A (en) | 2018-11-06 |
| CN108763718B CN108763718B (en) | 2020-08-14 |
Family
ID=64004791
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810499129.XA Active CN108763718B (en) | 2018-05-23 | 2018-05-23 | Method for quickly predicting characteristic quantity of flow field when flow-around object and working condition change |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108763718B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222828A (en) * | 2019-06-12 | 2019-09-10 | 西安交通大学 | A kind of Unsteady Flow method for quick predicting based on interacting depth neural network |
| CN110348059A (en) * | 2019-06-12 | 2019-10-18 | 西安交通大学 | A kind of channel flow field reconstructing method based on structured grid |
| CN113238076A (en) * | 2021-05-10 | 2021-08-10 | 西北工业大学 | Complex flow field measuring method and system based on deep learning |
| CN113836828A (en) * | 2020-06-08 | 2021-12-24 | 罗伯特·博世有限公司 | System and method for a combined differentiable partial differential equation solver and graphical neural network for fluid flow prediction |
| US11295045B2 (en) * | 2019-04-03 | 2022-04-05 | GM Global Technology Operations LLC | Tools and methods for aerodynamically optimizing the geometry of vehicle bodies |
| CN115062560A (en) * | 2022-06-16 | 2022-09-16 | 毕节高新技术产业开发区国家能源大规模物理储能技术研发中心 | Turbine multi-scale flow reduced-order coupling method based on machine learning |
| CN113836828B (en) * | 2020-06-08 | 2024-06-28 | 罗伯特·博世有限公司 | System and method for combining a differentiable partial differential equation solver and a graphical neural network for fluid flow prediction |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103954429A (en) * | 2014-04-02 | 2014-07-30 | 西北工业大学 | Lotus leaf stem imitative cylinder around flow drag reduction test model |
| US20180096249A1 (en) * | 2016-10-04 | 2018-04-05 | Electronics And Telecommunications Research Institute | Convolutional neural network system using adaptive pruning and weight sharing and operation method thereof |
| CN107944137A (en) * | 2017-11-23 | 2018-04-20 | 南京航空航天大学 | The thermographic curve computing technique of hypersonic aircraft trajectory state multi- scenarios method |
-
2018
- 2018-05-23 CN CN201810499129.XA patent/CN108763718B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103954429A (en) * | 2014-04-02 | 2014-07-30 | 西北工业大学 | Lotus leaf stem imitative cylinder around flow drag reduction test model |
| US20180096249A1 (en) * | 2016-10-04 | 2018-04-05 | Electronics And Telecommunications Research Institute | Convolutional neural network system using adaptive pruning and weight sharing and operation method thereof |
| CN107944137A (en) * | 2017-11-23 | 2018-04-20 | 南京航空航天大学 | The thermographic curve computing technique of hypersonic aircraft trajectory state multi- scenarios method |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11295045B2 (en) * | 2019-04-03 | 2022-04-05 | GM Global Technology Operations LLC | Tools and methods for aerodynamically optimizing the geometry of vehicle bodies |
| CN110222828A (en) * | 2019-06-12 | 2019-09-10 | 西安交通大学 | A kind of Unsteady Flow method for quick predicting based on interacting depth neural network |
| CN110348059A (en) * | 2019-06-12 | 2019-10-18 | 西安交通大学 | A kind of channel flow field reconstructing method based on structured grid |
| CN110222828B (en) * | 2019-06-12 | 2021-01-15 | 西安交通大学 | Unsteady flow field prediction method based on hybrid deep neural network |
| CN110348059B (en) * | 2019-06-12 | 2021-03-12 | 西安交通大学 | Channel internal flow field reconstruction method based on structured grid |
| CN113836828A (en) * | 2020-06-08 | 2021-12-24 | 罗伯特·博世有限公司 | System and method for a combined differentiable partial differential equation solver and graphical neural network for fluid flow prediction |
| CN113836828B (en) * | 2020-06-08 | 2024-06-28 | 罗伯特·博世有限公司 | System and method for combining a differentiable partial differential equation solver and a graphical neural network for fluid flow prediction |
| CN113238076A (en) * | 2021-05-10 | 2021-08-10 | 西北工业大学 | Complex flow field measuring method and system based on deep learning |
| CN113238076B (en) * | 2021-05-10 | 2022-12-06 | 西北工业大学 | Complex flow field measuring method based on deep learning |
| CN115062560A (en) * | 2022-06-16 | 2022-09-16 | 毕节高新技术产业开发区国家能源大规模物理储能技术研发中心 | Turbine multi-scale flow reduced-order coupling method based on machine learning |
| CN115062560B (en) * | 2022-06-16 | 2024-06-21 | 毕节高新技术产业开发区国家能源大规模物理储能技术研发中心 | Turbine multi-scale flow reduced-order coupling method based on machine learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108763718B (en) | 2020-08-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108763718A (en) | The method for quick predicting of Field Characteristics amount when streaming object and operating mode change | |
| CN114503121A (en) | Resource constrained neural network architecture search | |
| CN104034794B (en) | A kind of pipe leakage defect inspection method based on extreme learning machine | |
| Zhang et al. | An effective LS-SVM-based approach for surface roughness prediction in machined surfaces | |
| Wu et al. | Reduced order model using convolutional auto-encoder with self-attention | |
| Song et al. | Robust neural network tracking controller using simultaneous perturbation stochastic approximation | |
| MX2013008812A (en) | System and method for using an artificial neural network to simulate pipe hydraulics in a reservoir simulator. | |
| Lu et al. | A hydrodynamic optimization design methodology for a ship bulbous bow under multiple operating conditions | |
| CN110222828A (en) | A kind of Unsteady Flow method for quick predicting based on interacting depth neural network | |
| CN102129259A (en) | Neural network proportion integration (PI)-based intelligent temperature control system and method for sand dust environment test wind tunnel | |
| CN106777402A (en) | A kind of image retrieval text method based on sparse neural network | |
| CN103020709B (en) | Based on the one-dimensional water quality model parameter calibration method of artificial bee colony and quanta particle swarm optimization | |
| CN108595803A (en) | Shale gas well liquid loading pressure prediction method based on recurrent neural network | |
| CN105160539A (en) | Probability matrix decomposition recommendation method | |
| Djavareshkian et al. | Neuro-fuzzy based approach for estimation of Hydrofoil performance | |
| CN104634265A (en) | Soft measurement method for thickness of mineral floating foam layer based on multivariate image feature fusion | |
| CN109858798A (en) | It is associated with the electric grid investment decision-making modeling method and device of modification measures and voltage indexes | |
| CN113821724B (en) | Time interval enhancement-based graph neural network recommendation method | |
| CN114912364A (en) | Natural gas well flow prediction method, device, equipment and computer readable medium | |
| Maucec et al. | GeoDIN-Geoscience-Based Deep Interaction Networks for Predicting Flow Dynamics in Reservoir Simulation Models | |
| CN117786286A (en) | Fluid mechanics equation solving method based on physical information neural network | |
| Ballı et al. | An application of artificial neural networks for prediction and comparison with statistical methods | |
| Song et al. | Reconstruction of RANS model and cross-validation of flow field based on tensor basis neural network | |
| CN117034808A (en) | Natural gas pipe network pressure estimation method based on graph attention network | |
| CN112016956B (en) | Ore grade estimation method and device based on BP neural network |
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 |