WO2021164250A1 - 一种湍流场更新方法、装置及其相关设备 - Google Patents

一种湍流场更新方法、装置及其相关设备 Download PDF

Info

Publication number
WO2021164250A1
WO2021164250A1 PCT/CN2020/117025 CN2020117025W WO2021164250A1 WO 2021164250 A1 WO2021164250 A1 WO 2021164250A1 CN 2020117025 W CN2020117025 W CN 2020117025W WO 2021164250 A1 WO2021164250 A1 WO 2021164250A1
Authority
WO
WIPO (PCT)
Prior art keywords
turbulence
data
model
sample
reinforcement learning
Prior art date
Application number
PCT/CN2020/117025
Other languages
English (en)
French (fr)
Inventor
李茹杨
赵雅倩
李仁刚
Original Assignee
苏州浪潮智能科技有限公司
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 苏州浪潮智能科技有限公司 filed Critical 苏州浪潮智能科技有限公司
Priority to US17/801,283 priority Critical patent/US11741373B2/en
Publication of WO2021164250A1 publication Critical patent/WO2021164250A1/zh

Links

Images

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/092—Reinforcement learning
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Definitions

  • This application relates to the field of reinforcement learning technology, in particular to a turbulence field update method, and also to a turbulence field update device, equipment, and computer-readable storage medium.
  • Turbulence phenomena are common in nature and industry. For example, extreme weather such as sandstorms, typhoons and tsunamis in nature, the complex flow environment of large civil aviation aircraft and passenger ships, and the internal flow of engines involved in independent research and development of aero engines are all Typical turbulence phenomenon. Turbulence is a complex flow phenomenon with irregular spatial and temporal distribution, which is characterized by strong nonlinearity, randomness, and multi-scale characteristics.
  • the computer-readable storage medium also has the above-mentioned beneficial effects.
  • this application provides a turbulence field update method, including:
  • the predicted Reynolds stress is calculated by the RANS equation to obtain updated turbulence data.
  • the method before using the sample turbulence data for model training and obtaining a reinforcement learning turbulence model, the method further includes:
  • Preprocessing the sample turbulence data to obtain standard sample turbulence data Preprocessing the sample turbulence data to obtain standard sample turbulence data.
  • the preprocessing the sample turbulence data to obtain standard sample turbulence data includes:
  • using the sample turbulence data to perform model training to obtain a reinforcement learning turbulence model includes:
  • Model training is performed using the sample turbulence feature to obtain the reinforcement learning turbulence model.
  • said using the sample turbulence feature to perform model training to obtain the reinforcement learning turbulence model includes:
  • a preset objective function is used as an iterative condition, and the DQN neural network is iteratively trained using the sample turbulence feature to obtain the reinforcement learning turbulence model.
  • the calculation of the initial turbulence data of the turbulence field by the RANS equation includes:
  • the initial Reynolds stress is used to close the Reynolds stress term of the RANS equation, and the initial turbulence data is obtained by calculation.
  • the processing the initial turbulence data through the reinforcement learning turbulence model to obtain the predicted Reynolds stress includes:
  • the initial turbulence data is processed by the optimized reinforcement learning turbulence model to obtain the predicted Reynolds stress.
  • this application also provides a turbulence field updating device, including:
  • the sample acquisition module is used to acquire sample turbulence data
  • the model training module is used for model training using the sample turbulence data to obtain a reinforced learning turbulence model
  • the initial data calculation module is used to calculate the initial turbulence data of the turbulence field through the RANS equation calculation;
  • An initial data processing module configured to process the initial turbulence data through the reinforcement learning turbulence model to obtain predicted Reynolds stress
  • the turbulence field update module is used to calculate the predicted Reynolds stress through the RANS equation to obtain updated turbulence data.
  • this application also discloses a turbulence field updating device, including:
  • Memory used to store computer programs
  • the processor is configured to execute the computer program to realize the steps of any of the turbulence field updating methods described above.
  • the present application also discloses a computer-readable storage medium in which a computer program is stored, and when the computer program is executed by a processor, it is used to realize any of the above-mentioned turbulences. Steps of the field update method.
  • the turbulence field update method includes obtaining sample turbulence data; using the sample turbulence data for model training to obtain a reinforcement learning turbulence model; obtaining initial turbulence data of the turbulence field through RANS equation calculation; The turbulence model is learned to process the initial turbulence data to obtain predicted Reynolds stress; the predicted Reynolds stress is calculated through the RANS equation to obtain updated turbulence data.
  • the turbulence field update method is based on the generalization, label-free, serialized decision-making, and closed-loop update capabilities of reinforcement learning.
  • the reinforcement learning technology is used to realize the construction of turbulence model, and the RANS equation solver is combined with Reinforcement learning turbulence model is coupled, training based on current turbulence field data, to obtain a model with stronger generalization ability, so as to realize the turbulence field update.
  • the model construction based on reinforcement learning technology effectively reduces the target high Reynolds number turbulence and
  • the impact of low Reynolds number training data difference improves the generalization ability of the model; by alternately solving the RANS equation to calculate the turbulence field and using the reinforcement learning turbulence model to predict the Reynolds stress, it effectively reduces the computational difficulty of the high Reynolds number turbulence field and achieves high The accuracy is solved quickly, so as to achieve a more accurate turbulence field update.
  • the turbulence field updating device, equipment, and computer-readable storage medium provided in the present application all have the above-mentioned beneficial effects, and will not be repeated here.
  • Fig. 1 is a schematic flow chart of a turbulent field update method provided by this application
  • Fig. 2 is a schematic flowchart of another turbulence field update method provided by this application.
  • FIG. 3 is a schematic structural diagram of a turbulence field updating device provided by this application.
  • Fig. 4 is a schematic structural diagram of a turbulence field updating device provided by this application.
  • the computer-readable storage medium also has the above-mentioned beneficial effects.
  • Figure 1 is a schematic flow chart of a turbulent field update method provided by this application, including:
  • This step aims to obtain sample turbulence data, which is the high-precision and high-resolution flow field data of the published DNS (Direct Numerical Simulation), which can be used to implement subsequent model training.
  • This step aims to implement model training to obtain a reinforcement learning turbulence model.
  • reinforcement learning technology has strong generalization, unlabeled, serialized decision-making, and closed-loop update capabilities.
  • the sample turbulence data is processed to realize the construction of reinforcement learning turbulence model, which can effectively reduce the target height.
  • the impact of the difference between Reynolds number turbulence and low Reynolds number training data improves model accuracy.
  • the above-mentioned model training using sample turbulence data, and before obtaining the reinforcement learning turbulence model may further include: preprocessing the sample turbulence data to obtain standard sample turbulence data.
  • the sample turbulence data can also be preprocessed to obtain the standard sample turbulence data.
  • the implementation method of the foregoing preprocessing operation is not unique, for example, it may be standardized processing, normalization processing, sampling processing, etc., which is not limited in this application.
  • the foregoing preprocessing of the sample turbulence data to obtain standard sample turbulence data may include: normalizing the sample turbulence data to obtain normalized sample turbulence data; and performing normalized sample turbulence data Perform equal interval extraction processing to obtain standard sample turbulence data.
  • the preferred embodiment provides a more specific preprocessing method of sample turbulence data, that is, data normalization processing and data extraction processing.
  • data normalization processing and data extraction processing First, normalize the sample turbulence data with the mainstream velocity and density of the flat turbulence; further, extract the normalized sample turbulence data at equal intervals in the three directions of the sample turbulence data space to obtain the standard sample turbulence. data.
  • the foregoing model training using sample turbulence data to obtain a reinforced learning turbulence model may include: performing feature extraction on the sample turbulence data to obtain sample turbulence features; using sample turbulence features for model training to obtain reinforcement learning turbulence Model.
  • the preferred embodiment provides a more specific model training method, that is, model construction based on sample characteristics.
  • model construction based on sample characteristics.
  • the feature extraction operation can use any of the existing technologies, which is not limited in this application.
  • the foregoing model training using sample turbulence features to obtain a reinforced learning turbulence model may include: establishing a DQN neural network (Deep-Q-Network, deep value function neural network); using the preset objective function as an iteration Condition, the DQN neural network is iteratively trained using the turbulence characteristics of the sample to obtain a reinforcement learning turbulence model.
  • a DQN neural network (Deep-Q-Network, deep value function neural network)
  • the preset objective function as an iteration Condition
  • the preferred embodiment provides a specific type of reinforcement learning turbulence model, that is, a training model based on a DQN neural network.
  • a DQN neural network is established, and the above-mentioned sample turbulence characteristics are input for iterative training based on reinforcement learning.
  • an objective function is pre-established as an iterative condition to ensure model convergence, so as to obtain an accurate reinforcement learning turbulence model.
  • This step aims to realize the calculation of initial turbulence data, which is based on the RANS equation, where the RANS equation is an ensemble average NS equation (Navior-Stokes, Navier-Stokes equation) describing the evolution of turbulence statistics.
  • the turbulence data is the turbulence field data before the turbulence field is updated.
  • the foregoing calculation of the initial turbulence data of the turbulence field through the RANS equation calculation may include: obtaining the initial Reynolds stress of the turbulence field; using the initial Reynolds stress to close the Reynolds stress term of the RANS equation to obtain the initial turbulence data by calculation .
  • This preferred embodiment provides a more specific calculation method for initial turbulence data.
  • the initial Reynolds stress is the preset Reynolds stress value in the current turbulence field, and then use this value to compare the RANS equation
  • the above-mentioned initial turbulence data can be obtained by closed solution for the Reynolds stress term of.
  • S104 Process the initial turbulence data through the reinforcement learning turbulence model to obtain the predicted Reynolds stress
  • the initial turbulence data can be input into the reinforcement learning model for processing, so as to predict and obtain the Reynolds stress corresponding to the updated turbulence field, that is, the above-mentioned predicted Reynolds stress. Therefore, the turbulence field can be achieved based on the predicted Reynolds stress. renew. Further, in the process of updating the turbulence field, it can also be realized based on the RANS equation. Specifically, the above-mentioned predicted Reynolds stress can be used to solve the Reynolds stress term of the RANS equation in a closed solution to obtain the above-mentioned updated turbulence data. The updated turbulence data corresponds to the updated turbulence data. The data information of the turbulence field, so far, the update of the turbulence field is completed.
  • processing the initial turbulence data through the reinforcement learning turbulence model to obtain the predicted Reynolds stress may include: acquiring the learning experience and network parameters during the training process of the reinforcement learning turbulence model; using the learning experience and the network parameter pair The reinforcement learning turbulence model is optimized to obtain the optimized reinforcement learning turbulence model; the initial turbulence data is processed by the optimized reinforcement learning turbulence model to obtain the predicted Reynolds stress.
  • the data information recorded during the training process of the reinforcement learning turbulence model can be used to optimize the model.
  • This data information can be used to optimize the model.
  • the network parameters can be weights, biases, etc.
  • the optimized reinforcement learning turbulence model can be used to process the initial turbulence data to obtain the predicted Reynolds stress.
  • S101 to S102 are the training process of the reinforcement learning turbulence model. After the model training is completed, it can be stored in the pre-established storage space. Furthermore, in the actual turbulence field update process, the model training process only needs to be executed once, and when it needs to be used multiple times in the future, it can be directly retrieved from the storage space without repeated training.
  • the turbulence field update method is based on the generalization, label-free, serialized decision-making, and closed-loop update capabilities of reinforcement learning.
  • the reinforcement learning technology is used to realize the construction of turbulence model, and the RANS equation solver is combined with Reinforcement learning turbulence model is coupled, training based on current turbulence field data, to obtain a model with stronger generalization ability, so as to realize the turbulence field update.
  • the model construction based on reinforcement learning technology effectively reduces the target high Reynolds number turbulence and
  • the impact of low Reynolds number training data difference improves the generalization ability of the model; by alternately solving the RANS equation to calculate the turbulence field and using the reinforcement learning turbulence model to predict the Reynolds stress, it effectively reduces the computational difficulty of the high Reynolds number turbulence field and achieves high The accuracy is solved quickly, so as to achieve a more accurate turbulence field update.
  • FIG. 2 is a schematic flowchart of another turbulence field update method provided by this application.
  • the turbulence field update method couples the deep reinforcement learning algorithm DQN with the Reynolds average RANS equation solver, and uses turbulence field data (including the x, y, and z direction velocity components u, v, w and pressure p predict the 6 components (3 normal stress components and 3 shear stress components) of the Reynolds stress ⁇ tensor in the RANS equation, and then use the Reynolds stress to close the RANS equation to solve the next flow field to achieve turbulence Field update.
  • the specific implementation process of the turbulent field update method is as follows:
  • the direct numerical simulation DNS calculation has normalized the flow field velocity and pressure data using the characteristic velocity U and the fluid density ⁇ . If not, the mainstream velocity U and density ⁇ of the flat turbulence can be used for normalization. . Furthermore, considering that the DNS method uses a computational grid with a resolution much higher than the average Reynolds RANS equation, in order to match the two computational grids and reduce training costs, the original DNS data can be analyzed in three spatial directions. Extract data at equal intervals to form new sparse 3D flow field data sorted by time and used for training.
  • the velocity components u, v, w and pressure p on each computing grid point are selected as features, which are also the state s (State) in the reinforcement learning algorithm. Furthermore, the state s on all grid points constitutes the environment E (Environment ).
  • a typical reinforcement learning parameter space can form a four-tuple ⁇ A, S, R, T>, namely action space A, state space S, reward space R and transfer function space T.
  • the agent observes the environment and its current state s in the environment, and makes an action a (Action) according to a certain rule or strategy ⁇ , so that it obtains the current reward r and the long-term cumulative reward R (Reward), which affects changes in the environment Or transfer T (Transition).
  • a new action a1 is made, a new reward r1 and a cumulative reward R1 are obtained, and then this process is repeated.
  • the classic DQN method can be used to construct two neural networks with the same structure but different parameters, which are the target network (target_net) that updates the parameters at a certain interval and the prediction network (pred_net) that updates the parameters every step.
  • the prediction network that updates the parameters at each step is the Reynolds stress prediction neural network, which is used to predict the 6 components of the Reynolds stress ⁇ .
  • the existing high-precision and high-resolution DNS data can be used to train the Renault response prediction neural network, and the learning experience in this process (s t , a t , q t , r t , s t + 1 ) Stored in the memory bank, so as to be randomly extracted and replayed during the coupling solution of the reinforcement learning turbulence model RLTM and RANS equation to assist in predicting the Reynolds stress ⁇ pred ; at the same time, the parameters of the Reynolds stress prediction neural network , That is, the weight w and the bias b are stored to provide a better set of initial parameters of the neural network for subsequent coupling calculations.
  • the implementation process is as follows:
  • Input state S velocity components u, v, w and pressure p
  • the first layer (input layer) of the Reynolds stress prediction neural network uses the ReLu function as the activation function
  • the second layer does not use the activation function to obtain
  • the total predicted value Q pred (related to the reward r).
  • the target network also has the same network structure, outputs all the intermediate target values Q', and is used to calculate and update the target value Q target .
  • the discount factor ⁇ reflects the deeper into the future, the smaller the impact on the current return, and the value is a constant between 0 and 1.
  • the specific form of the network is as follows:
  • w1, w2 and b1, b2 are the network parameters of the first and second layers, namely the weight and bias;
  • L1 is the output of the first layer network, and the value Q is the final output;
  • Q target and Q pred are:
  • s t+1 , a t+1 , and r t+1 are the state, action, and reward at the next moment, respectively.
  • the action a corresponding to the largest Q pred is selected, that is, the current Reynolds stress prediction value ⁇ pred and the obtained reward r, which is defined as:
  • ⁇ DNS is the known Reynolds stress of high-precision, high-resolution DNS data
  • ⁇ pred is the current Reynolds stress prediction value
  • Q target and Q pred can be used to calculate the loss, and the network parameters can be updated through the back propagation operation.
  • the objective function of the Reynolds stress prediction neural network's back propagation and update parameters is defined as minimize: ( ⁇ DNS - ⁇ pred 2 ), the training network model is corrected by ⁇ DNS , and the weight w and bias are updated by the RMSProp optimizer.
  • Set b wherein, in each round of the calculation process, the network parameters and the learning experience (s, ⁇ pred, Q ' t, r, s') records and saves them to memory.
  • the first formula is a continuum equation, which reflects the incompressibility of the fluid;
  • the second formula is a momentum equation, which is essentially Newton’s second law, which reflects the force of the fluid microclusters.
  • u and p are fluid velocity and pressure respectively, ⁇ is density;
  • the subscripts i and j represent the physical quantity components in different directions of x, y, and z;
  • ⁇ ij is the stress tensor, Average velocity component
  • the composition of the spatial partial derivatives of x, y, and z reflects the pressure (or tension) and shear conditions of the fluid; It is the additional Reynolds stress tensor that appears due to the Reynolds averaging operation, that is, the above-mentioned ⁇ , which needs to be closed.
  • the initial flow field data is obtained by solving the RANS equation, it is input as the state s into the reinforcement learning turbulence model RLTM.
  • the predicted Reynolds is calculated Stress ⁇ pred .
  • the main steps of the process are similar to the training link of the reinforcement learning turbulence model, but because the high-precision and high-resolution ⁇ DNS data is not obtained in advance in the calculation of specific turbulence problems, the Reynolds stress prediction neural network is The objective function of backpropagation and parameter update needs to be changed to minimize:(rr pred 2 ).
  • r is the actual reward obtained by predicting ⁇ pred at each step
  • r pred is the predicted reward obtained through the additional two-layer reward feedback estimation neural network (reward_net) calculation.
  • the first layer uses relu as the incentive function.
  • the specific form is as follows:
  • the network structure parameters stored in the model training link will be assigned as initial values to the predictive neural network to improve the performance of the initial network, and with the help of the unique experience replay in the DQN algorithm and the parameters of the staged fixed target neural network.
  • DQN algorithm updates each time can be subjected to random (s, ⁇ pred, Q ' t, r, s') from the training phase of some memory in learning, for example, during use of the target neural net Q target of ,
  • the Q'used in is the parameters extracted from the memory bank.
  • This random extraction method can effectively break the correlation between the turbulent physical quantities developed in the time direction, and can effectively ensure the efficiency of the neural network.
  • the structural parameters of the target neural network remain unchanged for a period of time, and then are updated by the Reynolds stress prediction neural network with the latest parameters at any time to effectively break the correlation between the two networks.
  • the Reynolds stress term of the RANS equation is closed, and the updated flow field information u, v, w, p is obtained by solving, and as the next state s1, the reinforcement learning turbulence model is used to predict the next step
  • the Reynolds stress is predicted, and the cycle continues to realize the update of the turbulence field.
  • the turbulence field update method is based on the generalization, label-free, serialized decision-making, and closed-loop update capabilities of reinforcement learning.
  • the reinforcement learning technology is used to realize the construction of the turbulence model and solve the RANS equation.
  • the turbulence model is coupled with the reinforcement learning turbulence model, and the current turbulence field data is used for training to obtain a model with stronger generalization ability, thereby realizing the turbulence field update.
  • the model construction based on reinforcement learning technology effectively reduces the target high Reynolds number
  • the influence of the difference between turbulence and low Reynolds number training data improves the generalization ability of the model; the turbulence field is calculated by alternately solving the RANS equation and the reinforcement learning turbulence model is used to predict the Reynolds stress, which effectively reduces the computational difficulty of the high Reynolds number turbulence field. In order to achieve a more accurate turbulence field update.
  • FIG. 3 is a schematic structural diagram of a turbulent flow field updating device provided by this application, including:
  • the sample acquisition module 1 is used to acquire sample turbulence data
  • Model training module 2 is used for model training using sample turbulence data to obtain a reinforcement learning turbulence model
  • the initial data calculation module 3 is used to obtain the initial turbulence data of the turbulence field through the RANS equation calculation;
  • the initial data processing module 4 is used to process the initial turbulence data through the reinforcement learning turbulence model to obtain the predicted Reynolds stress;
  • the turbulence field update module 5 is used to calculate the predicted Reynolds stress through the RANS equation to obtain updated turbulence data.
  • the turbulence field update device based on the generalization, labelless, serialized decision-making, and closed-loop update capabilities of reinforcement learning, realizes the construction of a turbulence model by using reinforcement learning technology, and solves the RANS equation
  • the turbulence model is coupled with the reinforcement learning turbulence model, and the current turbulence field data is used for training to obtain a model with stronger generalization ability, thereby realizing the turbulence field update.
  • the model construction based on reinforcement learning technology effectively reduces the target high Reynolds number
  • the effect of the difference between turbulence and low Reynolds number training data improves the generalization ability of the model; by alternately solving the RANS equation to calculate the turbulence field and using the reinforcement learning turbulence model to predict the Reynolds stress, it effectively reduces the computational difficulty of the high Reynolds number turbulence field. In order to achieve a more accurate and fast update of the turbulence field.
  • the turbulence field update device may further include a data preprocessing module for preprocessing the sample turbulence data to obtain standard sample turbulence data.
  • the aforementioned data preprocessing module may include:
  • the normalization unit is used to normalize the sample turbulence data to obtain the normalized sample turbulence data
  • the data extraction unit is used to extract the normalized sample turbulence data at equal intervals to obtain the standard sample turbulence data.
  • model training module 2 may include:
  • the feature extraction unit is used to perform feature extraction on sample turbulence data to obtain sample turbulence features
  • the model training unit is used for model training using sample turbulence features to obtain a reinforcement learning turbulence model.
  • the above-mentioned model training unit can be specifically used to establish a DQN neural network; using a preset objective function as an iterative condition, the DQN neural network is iteratively trained using sample turbulence characteristics to obtain a reinforced learning turbulence model.
  • the initial data calculation module 3 can be specifically used to obtain the initial Reynolds stress of the turbulence field; the initial Reynolds stress is used to close the Reynolds stress term of the RANS equation, and the initial turbulence data is obtained by calculation.
  • the aforementioned turbulence field update module 5 may be specifically used to obtain learning experience and network parameters during the training process of the reinforcement learning turbulence model; use the learning experience and network parameters to optimize the reinforcement learning turbulence model to obtain optimized reinforcement learning Turbulence model: The initial turbulence data is processed by optimizing the reinforcement learning turbulence model to obtain the predicted Reynolds stress.
  • FIG. 4 is a schematic structural diagram of a turbulence field updating device provided by this application.
  • the turbulence field updating device may include:
  • the memory 10 is used to store computer programs
  • the processor 20 is configured to implement the steps of any one of the above-mentioned turbulence field updating methods when executing a computer program.
  • the present application also provides a computer-readable storage medium with a computer program stored on the computer-readable storage medium.
  • the computer program can be executed by a processor to implement the steps of any of the above-mentioned turbulence field update methods. .
  • the computer-readable storage medium may include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk, etc., which can store program codes Medium.
  • the steps of the method or algorithm described in the embodiments disclosed in this document can be directly implemented by hardware, a software module executed by a processor, or a combination of the two.
  • the software module can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or all areas in the technical field. Any other known storage media.

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

一种湍流场更新方法、装置、设备及计算机可读存储介质，方法包括获取样本湍流数据；利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；通过RANS方程计算获得湍流场的初始湍流数据；通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力；通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据；该湍流场更新方法可以有效提高计算精度，实现更为准确的湍流场更新。

Description

一种湍流场更新方法、装置及其相关设备
本申请要求于2020年2月21日提交中国专利局、申请号为202010110908.3、发明名称为“一种湍流场更新方法、装置及其相关设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及强化学习技术领域，特别涉及一种湍流场更新方法，还涉及一种湍流场更新装置、设备和计算机可读存储介质。
背景技术
湍流现象普遍存在于自然界与工业界中，例如，自然界中的沙尘暴、台风海啸等极端天气，大型民航客机、客船所处的复杂流动环境，以及航空发动机自主研发涉及的发动机内部流动等，都是典型的湍流现象。湍流是空间和时间分布不规则的复杂流动现象，表现为强非线性、随机性、多尺度性等特点。
在湍流场中，实际湍流环境的雷诺数往往达到Re～O(10 10)，针对这种强非线性的超高雷诺数的复杂流动问题，一般采用风洞、水洞试验观测和雷诺平均RANS方程数值模拟求解，但由于试验观测技术的限制和试验成本的约束，雷诺平均的数值模拟方法是现在主要的湍流研究手段。
然而，由于湍流模型多来源于简单流动，在面对高雷诺数分离流等复杂问题时，RANS方程的计算结果往往与实际情况存在较大偏差，无法实现更为精确的湍流场更新。
因此，如何有效提高计算精度，实现更为准确的湍流场更新是本领域技术人员亟待解决的问题。
发明内容
本申请的目的是提供一种湍流场更新方法，该湍流场更新方法可以有效提高计算精度，实现更为准确的湍流场更新；本申请的另一目的是提供一种湍流场更新装置、设备和计算机可读存储介质，也具有上述有益效果。
为解决上述技术问题，第一方面，本申请提供了一种湍流场更新方法，包括：
获取样本湍流数据；
利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；
通过RANS方程计算获得湍流场的初始湍流数据；
通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力；
通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据。
优选的，所述利用所述样本湍流数据进行模型训练，获得强化学习湍流模型之前，还包括：
对所述样本湍流数据进行预处理，获得标准样本湍流数据。
优选的，所述对所述样本湍流数据进行预处理，获得标准样本湍流数据，包括：
对所述样本湍流数据进行归一化处理，获得归一化样本湍流数据；
对所述归一化样本湍流数据进行等间隔抽取处理，获得所述标准样本湍流数据。
优选的，所述利用所述样本湍流数据进行模型训练，获得强化学习湍流模型，包括：
对所述样本湍流数据进行特征提取，获得样本湍流特征；
利用所述样本湍流特征进行模型训练，获得所述强化学习湍流模型。
优选的，所述利用所述样本湍流特征进行模型训练，获得所述强化学习湍流模型，包括：
建立DQN神经网络；
将预设目标函数作为迭代条件，利用所述样本湍流特征对所述DQN神经网络进行迭代训练，获得所述强化学习湍流模型。
优选的，所述通过RANS方程计算获得湍流场的初始湍流数据，包括：
获取所述湍流场的初始雷诺应力；
利用所述初始雷诺应力对所述RANS方程的雷诺应力项进行封闭，计算获得所述初始湍流数据。
优选的，所述通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力，包括：
获取所述强化学习湍流模型训练过程中的学习经历和网络参数；
利用所述学习经历和网络参数对所述强化学习湍流模型进行优化，获得优化强化学习湍流模型；
通过所述优化强化学习湍流模型对所述初始湍流数据进行处理，获得所述预测雷诺应力。
第二方面，本申请还提供了一种湍流场更新装置，包括：
样本获取模块，用于获取样本湍流数据；
模型训练模块，用于利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；
初始数据计算模块，用于通过RANS方程计算获得湍流场的初始湍流数据；
初始数据处理模块，用于通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力；
湍流场更新模块，用于通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据。
第三方面，本申请还公开了一种湍流场更新设备，包括：
存储器，用于存储计算机程序；
处理器，用于执行所述计算机程序以实现如上所述的任一种湍流场更新方法的步骤。
第四方面，本申请还公开了一种计算机可读存储介质，所述计算机可读存储介质中存储有计算机程序，所述计算机程序被处理器执行时用以实现如上所述的任一种湍流场更新方法的步骤。
本申请所提供的一种湍流场更新方法，包括获取样本湍流数据；利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；通过RANS方程计算获得湍流场的初始湍流数据；通过所述强化学习湍流模型对所述初 始湍流数据进行处理，获得预测雷诺应力；通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据。
可见，本申请所提供的湍流场更新方法，基于强化学习具备的通用化、无标签、序列化决策、闭环更新的能力，利用强化学习技术实现了湍流模型的构建，并将RANS方程求解器与强化学习湍流模型进行耦合，依靠当前湍流场数据进行训练，获得泛化能力更强的模型，从而实现了湍流场更新，可见，基于强化学习技术的模型构建，有效降低了目标高雷诺数湍流与低雷诺数训练数据差异的影响，提高了模型的泛化能力；通过交替求解RANS方程计算湍流场和使用强化学习湍流模型预测雷诺应力，有效降低了高雷诺数湍流场的计算难度，实现了高精度快速求解，从而实现了更为准确的湍流场更新。
本申请所提供的一种湍流场更新装置、设备和计算机可读存储介质，均具有上述有益效果，在此不再赘述。
附图说明
为了更清楚地说明现有技术和本申请实施例中的技术方案，下面将对现有技术和本申请实施例描述中需要使用的附图作简要的介绍。当然，下面有关本申请实施例的附图描述的仅仅是本申请中的一部分实施例，对于本领域普通技术人员来说，在不付出创造性劳动的前提下，还可以根据提供的附图获得其他的附图，所获得的其他附图也属于本申请的保护范围。
图1为本申请所提供的一种湍流场更新方法的流程示意图；
图2为本申请所提供的另一种湍流场更新方法的流程示意图；
图3为本申请所提供的一种湍流场更新装置的结构示意图；
图4为本申请所提供的一种湍流场更新设备的结构示意图。
具体实施方式
本申请的核心是提供一种湍流场更新方法，该湍流场更新方法可以有效提高计算精度，实现更为准确的湍流场更新；本申请的另一核心是提供 一种湍流场更新装置、设备和计算机可读存储介质，也具有上述有益效果。
为了对本申请实施例中的技术方案进行更加清楚、完整地描述，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行介绍。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
请参考图1，图1为本申请所提供的一种湍流场更新方法的流程示意图，包括：
S101：获取样本湍流数据；
本步骤旨在实现样本湍流数据的获取，该样本湍流数据即为已经公开的DNS(Direct Numerical Simulation，直接数值模拟)的高精度、高分辨率的流场数据，可用于实现后续模型训练。
S102：利用样本湍流数据进行模型训练，获得强化学习湍流模型；
本步骤旨在实现模型训练，以获得强化学习湍流模型。具体而言，强化学习技术具有较强的通用化、无标签、序列化决策、闭环更新的能力，基于强化学习技术对样本湍流数据进行处理，实现强化学习湍流模型的构建，可以有效降低目标高雷诺数湍流与低雷诺数训练数据差异的影响，提高模型精度。
作为一种优选实施例，上述利用样本湍流数据进行模型训练，获得强化学习湍流模型之前，还可以包括：对样本湍流数据进行预处理，获得标准样本湍流数据。
为有效提高模型精度，在进行模型训练之前，还可以对样本湍流数据进行预处理，以获得标准样本湍流数据。其中，上述预处理操作的实现方法并不唯一，例如，可以为标准化处理、归一化处理、采样处理等，本申请对此不做限定。
作为一种优选实施例，上述对样本湍流数据进行预处理，获得标准样本湍流数据，可以包括：对样本湍流数据进行归一化处理，获得归一化样本湍流数据；对归一化样本湍流数据进行等间隔抽取处理，获得标准样本 湍流数据。
本优选实施例提供了一种较为具体的样本湍流数据的预处理方法，即数据归一化处理和数据抽取处理。首先，采用平板湍流的主流速度和密度对样本湍流数据进行归一化；进一步，在样本湍流数据空间的三个方向上对归一化样本湍流数据进行等间隔抽取，即可获得上述标准样本湍流数据。
作为一种优选实施例，上述利用样本湍流数据进行模型训练，获得强化学习湍流模型，可以包括：对样本湍流数据进行特征提取，获得样本湍流特征；利用样本湍流特征进行模型训练，获得强化学习湍流模型。
本优选实施例提供了一种较为具体的模型训练方法，即基于样本特征的模型构建。首先对预处理后的样本湍流数据进行湍流场特征提取，进而利用该样本湍流特征进行模型构建，即可获得上述强化学习湍流模型。其中，特征提取操作可采用已有技术中的任意一种，本申请对此不做限定。
作为一种优选实施例，上述利用样本湍流特征进行模型训练，获得强化学习湍流模型，可以包括：建立DQN神经网络(Deep-Q-Network，深度价值函数神经网络)；将预设目标函数作为迭代条件，利用样本湍流特征对DQN神经网络进行迭代训练，获得强化学习湍流模型。
本优选实施例提供了一种具体类型的强化学习湍流模型，即基于DQN神经网络的训练模型。具体而言，建立DQN神经网络，并输入上述样本湍流特征进行基于强化学习的迭代训练，在迭代过程中，预先建立目标函数作为迭代条件，以保证模型收敛，从而获得准确的强化学习湍流模型。
S103：通过RANS方程(Reynolds Averaged Navior-Stokes，雷诺平均纳维尔-斯托克斯方程/模型)计算获得湍流场的初始湍流数据；
本步骤旨在实现初始湍流数据的计算，即基于RANS方程实现，其中，RANS方程是描述湍流统计量的演化的系综平均N-S方程(Navior-Stokes，纳维尔-斯托克斯方程)，初始湍流数据则为湍流场未更新之前的湍流场数据。
作为一种优选实施例，上述通过RANS方程计算获得湍流场的初始湍流数据，可以包括：获取湍流场的初始雷诺应力；利用初始雷诺应力对RANS方程的雷诺应力项进行封闭，计算获得初始湍流数据。
本优选实施例提供了较为具体的初始湍流数据的计算方法，首先，获取湍流场的初始雷诺应力，该初始雷诺应力为当前湍流场中预先给定的雷诺应力值，进而利用该值对RANS方程的雷诺应力项进行封闭求解，即可获得上述初始湍流数据。
S104：通过强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力；
S105：通过RANS方程对预测雷诺应力进行计算，获得更新湍流数据。
具体的，可将初始湍流数据输入至强化学习模型中进行处理，从而预测获得对应于更新后湍流场的雷诺应力，即上述预测雷诺应力，由此，即可根据该预测雷诺应力实现湍流场的更新。进一步，在湍流场更新过程中，同样可基于RANS方程实现，具体可利用上述预测雷诺应力对RANS方程的雷诺应力项进行封闭求解，获得上述更新湍流数据，该更新湍流数据即为对应于更新后湍流场的数据信息，至此，完成湍流场更新。
作为一种优选实施例，上述通过强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力，可以包括：获取强化学习湍流模型训练过程中的学习经历和网络参数；利用学习经历和网络参数对强化学习湍流模型进行优化，获得优化强化学习湍流模型；通过优化强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力。
为进一步保证预测雷诺应力取值的准确性，在利用强化学习湍流模型对初始湍流数据进行计算之前，可先利用强化学习湍流模型训练过程中所记录的数据信息对模型进行优化，该数据信息可包括学习经历和网络参数，在模型训练过程中产生并记录于记忆库中，可直接调用，其中，网络参数可以为权重、偏置等。由此，即可利用优化强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力。
需要说明的是，上述S101至S102为强化学习湍流模型的训练过程，在模型训练完成后，将其存储至预先建立的存储空间中即可。进一步，在实际的湍流场更新过程中，模型训练过程仅需执行一次，在后续需要多次使用时，直接从存储空间中调取即可，无需重复训练。
可见，本申请所提供的湍流场更新方法，基于强化学习具备的通用化、 无标签、序列化决策、闭环更新的能力，利用强化学习技术实现了湍流模型的构建，并将RANS方程求解器与强化学习湍流模型进行耦合，依靠当前湍流场数据进行训练，获得泛化能力更强的模型，从而实现了湍流场更新，可见，基于强化学习技术的模型构建，有效降低了目标高雷诺数湍流与低雷诺数训练数据差异的影响，提高了模型的泛化能力；通过交替求解RANS方程计算湍流场和使用强化学习湍流模型预测雷诺应力，有效降低了高雷诺数湍流场的计算难度，实现了高精度快速求解，从而实现了更为准确的湍流场更新。
在上述各个实施例的基础上，本申请实施例提供了一种更为具体的湍流场更新方法，请参考图2，图2为本申请所提供的另一种湍流场更新方法的流程示意图。
具体而言，本申请所提供的湍流场更新方法，将深度强化学习算法DQN与雷诺平均的RANS方程求解器进行耦合，使用湍流场数据(包括x、y、z方向的速度分量u、v、w和压力p对RANS方程中的雷诺应力τ张量的6个分量(3个正应力分量和3个切应力分量)进行预测，再用雷诺应力封闭RANS方程进行下一步流场求解，实现湍流场更新。基于此，该湍流场更新方法的具体实现流程如下：
1、训练强化学习湍流模型RLTM(Regularized Lifelong Topic Model，正则化终身主题模型)
目前已有大量公开的直接数值模拟DNS的高精度、高分辨率流场数据供研究人员使用，这些数据通常以一定的时间间隔存储，包含三维流场在x、y、z方向的速度数据u、v、w和压力数据p，因此，可基于这些数据来训练可用于预测流场雷诺应力的深度神经网络模型。
(1)已有DNS流场数据准备与预处理：
通常，直接数值模拟DNS计算当中已经将流场速度和压力数据使用特征速度U和流体密度ρ进行了归一化，若没有，则可以使用平板湍流的主流速度U和密度ρ进行归一化处理。进一步，考虑到DNS方法使用分辨率远高于平均雷诺RANS方程的计算网格，为使两套计算网格相匹配，同 时减小训练成本，可以对原始的DNS数据在空间三个方向上进行等间隔抽取数据，组成新的按时间排序的、训练使用的稀疏三维流场数据。
(2)湍流场特征(强化学习状态)选取：
选用每个计算网格点上的速度分量u、v、w和压力p作为特征，同时也是强化学习算法中的状态s(State)，进一步，所有网格点上的状态s构成环境E(Environment)。
(3)强化学习湍流模型RLTM训练与记忆库积累：
典型的强化学习参数空间可组成一个四元组<A，S，R，T>，即动作空间A、状态空间S、奖励空间R和转移函数空间T。智能体在环境当中观测环境和自身当前的状态s，根据一定的规则或策略π，做出一个动作a(Action)，因此获得当前的奖励r和长期累计奖励R(Reward)，影响环境发生变化或转移T(Transition)。在新的环境中，根据观测到新的状态s1，而作出新的动作a1，获得新的奖励r1和累计奖励R1，之后重复这个过程。
具体的，可以使用经典的DQN方法，构建2个结构相同但参数不同的神经网络，分别是间隔一定时间更新参数的目标网络(target_net)和每步更新参数的预测网络(pred_net)。其中，每步更新参数的预测网络即为雷诺应力预测神经网络，用于预测雷诺应力τ的6个分量。由此，即可使用已有的高精度、高分辨率的DNS数据对雷诺应预测神经网络进行训练，将这个过程中的学习经历(s t,a t,q t,r t,s t+1)存储到记忆库当中，以便在强化学习湍流模型RLTM与RANS方程进行耦合求解的过程中被随机抽取出来进行重放，以辅助预测雷诺应力τ pred；同时，将雷诺应力预测神经网络的参数，即权重w与偏置b进行存储，为之后的耦合计算提供一组较好的神经网络初始参数。基于此，实现过程如下：
输入状态S(速度分量u、v、w和压力p)信息，雷诺应力预测神经网络的第1层(输入层)使用ReLu函数作为激活函数，第2层(输出层)不使用激活函数，获得全部的预测价值Q pred(与奖励r相关)。目标网络也具有同样的网络结构，输出全部的中间目标价值Q'，并用于计算和更新目标价值Q target。计算中，折扣因子γ体现越是深入未来，对当前回报的影 响越小，取值为介于0与1之间的常数。网络的具体形式如下：
L1＝relu(w1*S+b1)
Q＝w2*L1+b2
其中，w1、w2和b1、b2分别为第1层和第2层的网络参数，即权重与偏置；L1为第一层网络的输出，价值Q为最终输出；
Q target和Q pred的具体形式为：
Figure PCTCN2020117025-appb-000001
Q pred＝Q(s t+1,a t+1)
其中，s t+1,a t+1,r t+1分别为下一时刻的状态、动作和奖励。
进一步，获得全部Q pred之后，选择最大的Q pred对应的动作a，即当前的雷诺应力预测值τ pred，以及获得的奖励r，奖励r定义为：
Figure PCTCN2020117025-appb-000002
其中，τ DNS为已知的高精度、高分辨率DNS数据的雷诺应力；τ pred为当前的雷诺应力预测值。
至此，强化学习湍流模型RLTM的一轮训练结束。
(4)更新雷诺应力预测神经网络参数：
对于雷诺应力预测神经网络，可以使用Q target和Q pred计算损失，并通过反向传播运算更新网络参数。具体来说，雷诺应力预测神经网络的反向传播和更新参数的目标函数定义为minimize:(τ DNS-τ pred 2)，通过τ DNS来修正训练网络模型，通过RMSProp优化器更新权重w与偏置b。其中，在每轮计算过程中，对网络参数和将学习经历(s,τ pred,Q′ t,r,s')进行记录并存储至记忆库当中。
2、强化学习湍流模型RLTM与RANS方程的耦合计算
(1)基于ANS方程计算初始流场：
针对具体求解的湍流问题，利用预先给定的初始雷诺应力τ 0，如全零分布，对RANS方程的雷诺应力项进行封闭，求解出初始流场u,v,w,p。
在笛卡尔坐标系下，不可压缩的RANS方程如下：
Figure PCTCN2020117025-appb-000003
Figure PCTCN2020117025-appb-000004
其中，第一式为连续方程，体现流体的不可压缩性；第二式为动量方程，本质为牛顿第二定律，体现流体微团的受力情况。u和p分别为流体速度与压强，ρ为密度；
Figure PCTCN2020117025-appb-000005
表示雷诺平均的物理量；下标i和j表示x、y、z不同方向上的物理量分量；上标'表示去除平均值后的脉动量，体现湍流的高脉动性；σ ij为应力张量，由平均速度分量
Figure PCTCN2020117025-appb-000006
对x、y、z的空间偏导数组成，体现流体的受压(或受拉)和剪切情况；
Figure PCTCN2020117025-appb-000007
是由于雷诺平均操作而额外出现的雷诺应力张量，即上述τ，需要进行封闭。
(2)使用强化学习湍流模型RLTM预测雷诺应力：
在通过求解RANS方程获得初始流场数据后，将其作为状态s输入到强化学习湍流模型RLTM当中，通过交互使用与训练模型时同样结构的雷诺应力预测神经网络和目标神经网络，计算获得预测雷诺应力τ pred。
具体来说，该过程的主要步骤与强化学习湍流模型的训练环节类似，但由于针对具体湍流问题的计算中没有预先获知高精度、高分辨率的τ DNS数据，因此，雷诺应力预测神经网络的反向传播与参数更新的目标函数需要更改为minimize:(r-r pred 2)。其中，r为每一步预测τ pred获得的实际奖励，而r pred为通过额外补充的2层奖励反馈估计神经网络(reward_net)计算获得的预测奖励，其中第一层使用relu作为激励函数，其网络的具体形式如下：
L1＝relu(w2*τ pred+b2)
r pred＝w1*L1+b1
特别指出的是，模型训练环节存储的网络结构参数会作为初始值赋予预测神经网络，以提高初始网络的性能，并借助DQN算法中特有的经验重播和阶段性固定目标神经网络的参数。每次DQN算法更新时，都可以 从记忆库中随机抽取一些训练阶段的经历(s,τ pred,Q′ t,r,s')进行学习，例如在使用目标神经网络计算Q target的过程中，
Figure PCTCN2020117025-appb-000008
中使用到的Q'就是从记忆库中抽取出的参数，这种随机抽取的方法可以有效打破时间方向上发展的湍流物理量之间的相关性，能够有效保证神经网络的效率。同时，在计算过程当中，目标神经网络的结构参数在一段时间内保持不变，再由随时具备最新参数的雷诺应力预测神经网络进行更新，以有效打破两个网络之间的相关性。
(3)RANS方程计算流场更新：
基于上述预测雷诺应力τ pred对RANS方程的雷诺应力项进行封闭，求解获得更新后的流场信息u,v,w,p，并作为下一步的状态s1，使用强化学习湍流模型预测下一步的预测雷诺应力，以此循环下去，实现湍流场的更新。
可见，本申请实施例所提供的湍流场更新方法，基于强化学习具备的通用化、无标签、序列化决策、闭环更新的能力，利用强化学习技术实现了湍流模型的构建，并将RANS方程求解器与强化学习湍流模型进行耦合，依靠当前湍流场数据进行训练，获得泛化能力更强的模型，从而实现了湍流场更新，可见，基于强化学习技术的模型构建，有效降低了目标高雷诺数湍流与低雷诺数训练数据差异的影响，提高了模型的泛化能力；通过交替求解RANS方程计算湍流场和使用强化学习湍流模型预测雷诺应力，有效降低了高雷诺数湍流场的计算难度，实现了高精度快速求解，从而实现了更为准确的湍流场更新。
为解决上述技术问题，本申请还提供了一种湍流场更新装置，请参考图3，图3为本申请所提供的一种湍流场更新装置的结构示意图，包括：
样本获取模块1，用于获取样本湍流数据；
模型训练模块2，用于利用样本湍流数据进行模型训练，获得强化学习湍流模型；
初始数据计算模块3，用于通过RANS方程计算获得湍流场的初始湍流数据；
初始数据处理模块4，用于通过强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力；
湍流场更新模块5，用于通过RANS方程对预测雷诺应力进行计算，获得更新湍流数据。
可见，本申请实施例所提供的湍流场更新装置，基于强化学习具备的通用化、无标签、序列化决策、闭环更新的能力，利用强化学习技术实现了湍流模型的构建，并将RANS方程求解器与强化学习湍流模型进行耦合，依靠当前湍流场数据进行训练，获得泛化能力更强的模型，从而实现了湍流场更新，可见，基于强化学习技术的模型构建，有效降低了目标高雷诺数湍流与低雷诺数训练数据差异的影响，提高了模型的泛化能力；通过交替求解RANS方程计算湍流场和使用强化学习湍流模型预测雷诺应力，有效降低了高雷诺数湍流场的计算难度，实现了高精度快速求解，从而实现了更为准确的湍流场更新。
作为一种优选实施例，该湍流场更新装置还可以包括数据预处理模块，用于对样本湍流数据进行预处理，获得标准样本湍流数据。
作为一种优选实施例，上述数据预处理模块可以包括：
归一化单元，用于对样本湍流数据进行归一化处理，获得归一化样本湍流数据；
数据抽取单元，用于对归一化样本湍流数据进行等间隔抽取处理，获得标准样本湍流数据。
作为一种优选实施例，上述模型训练模块2可以包括：
特征提取单元，用于对样本湍流数据进行特征提取，获得样本湍流特征；
模型训练单元，用于利用样本湍流特征进行模型训练，获得强化学习湍流模型。
作为一种优选实施例，上述模型训练单元可具体用于建立DQN神经网络；将预设目标函数作为迭代条件，利用样本湍流特征对DQN神经网络进行迭代训练，获得强化学习湍流模型。
作为一种优选实施例，初始数据计算模块3可具体用于获取湍流场的 初始雷诺应力；利用初始雷诺应力对RANS方程的雷诺应力项进行封闭，计算获得初始湍流数据。
作为一种优选实施例，上述湍流场更新模块5可具体用于获取强化学习湍流模型训练过程中的学习经历和网络参数；利用学习经历和网络参数对强化学习湍流模型进行优化，获得优化强化学习湍流模型；通过优化强化学习湍流模型对初始湍流数据进行处理，获得预测雷诺应力。
对于本申请提供的装置的介绍请参照上述方法实施例，本申请在此不做赘述。
为解决上述技术问题，本申请还提供了一种服务器管理系统，请参考图4，图4为本申请所提供的一种湍流场更新设备的结构示意图，该湍流场更新设备可包括：
存储器10，用于存储计算机程序；
处理器20，用于执行计算机程序时可实现如上述任意一种湍流场更新方法的步骤。
对于本申请提供的系统的介绍请参照上述方法实施例，本申请在此不做赘述。
为解决上述问题，本申请还提供了一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，计算机程序被处理器执行时可实现如上述任意一种湍流场更新方法的步骤。
该计算机可读存储介质可以包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
对于本申请提供的计算机可读存储介质的介绍请参照上述方法实施例，本申请在此不做赘述。
说明书中各个实施例采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似部分互相参见即可。 对于实施例公开的装置而言，由于其与实施例公开的方法相对应，所以描述的比较简单，相关之处参见方法部分说明即可。
专业人员还可以进一步意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
结合本文中所公开的实施例描述的方法或算法的步骤可以直接用硬件、处理器执行的软件模块，或者二者的结合来实施。软件模块可以置于随机存储器(RAM)、内存、只读存储器(ROM)、电可编程ROM、电可擦除可编程ROM、寄存器、硬盘、可移动磁盘、CD-ROM、或技术领域内所公知的任意其它形式的存储介质中。
以上对本申请所提供的技术方案进行了详细介绍。本文中应用了具体个例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想。应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以对本申请进行若干改进和修饰，这些改进和修饰也落入本申请的保护范围内。

Claims (10)

  1. 一种湍流场更新方法，其特征在于，包括：
    获取样本湍流数据；
    利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；
    通过RANS方程计算获得湍流场的初始湍流数据；
    通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力；
    通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据。
  2. 根据权利要求1所述的湍流场更新方法，其特征在于，所述利用所述样本湍流数据进行模型训练，获得强化学习湍流模型之前，还包括：
    对所述样本湍流数据进行预处理，获得标准样本湍流数据。
  3. 根据权利要求2所述的湍流场更新方法，其特征在于，所述对所述样本湍流数据进行预处理，获得标准样本湍流数据，包括：
    对所述样本湍流数据进行归一化处理，获得归一化样本湍流数据；
    对所述归一化样本湍流数据进行等间隔抽取处理，获得所述标准样本湍流数据。
  4. 根据权利要求1所述的湍流场更新方法，其特征在于，所述利用所述样本湍流数据进行模型训练，获得强化学习湍流模型，包括：
    对所述样本湍流数据进行特征提取，获得样本湍流特征；
    利用所述样本湍流特征进行模型训练，获得所述强化学习湍流模型。
  5. 根据权利要求4所述的湍流场更新方法，其特征在于，所述利用所述样本湍流特征进行模型训练，获得所述强化学习湍流模型，包括：
    建立DQN神经网络；
    将预设目标函数作为迭代条件，利用所述样本湍流特征对所述DQN神经网络进行迭代训练，获得所述强化学习湍流模型。
  6. 根据权利要求1所述的湍流场更新方法，其特征在于，所述通过RANS方程计算获得湍流场的初始湍流数据，包括：
    获取所述湍流场的初始雷诺应力；
    利用所述初始雷诺应力对所述RANS方程的雷诺应力项进行封闭，计算获得所述初始湍流数据。
  7. 根据权利要求1所述的湍流场更新方法，其特征在于，所述通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力，包括：
    获取所述强化学习湍流模型训练过程中的学习经历和网络参数；
    利用所述学习经历和网络参数对所述强化学习湍流模型进行优化，获得优化强化学习湍流模型；
    通过所述优化强化学习湍流模型对所述初始湍流数据进行处理，获得所述预测雷诺应力。
  8. 一种湍流场更新装置，其特征在于，包括：
    样本获取模块，用于获取样本湍流数据；
    模型训练模块，用于利用所述样本湍流数据进行模型训练，获得强化学习湍流模型；
    初始数据计算模块，用于通过RANS方程计算获得湍流场的初始湍流数据；
    初始数据处理模块，用于通过所述强化学习湍流模型对所述初始湍流数据进行处理，获得预测雷诺应力；
    湍流场更新模块，用于通过所述RANS方程对所述预测雷诺应力进行计算，获得更新湍流数据。
  9. 一种湍流场更新设备，其特征在于，包括：
    存储器，用于存储计算机程序；
    处理器，用于执行所述计算机程序以实现如权利要求1至7任一项所述的湍流场更新方法的步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有计算机程序，所述计算机程序被处理器执行时用以实现如权利要求1至7任一项所述的湍流场更新方法的步骤。
PCT/CN2020/117025 2020-02-21 2020-09-23 一种湍流场更新方法、装置及其相关设备 WO2021164250A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/801,283 US11741373B2 (en) 2020-02-21 2020-09-23 Turbulence field update method and apparatus, and related device thereof

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010110908.3 2020-02-21
CN202010110908.3A CN111324993B (zh) 2020-02-21 2020-02-21 一种湍流场更新方法、装置及其相关设备

Publications (1)

Publication Number Publication Date
WO2021164250A1 true WO2021164250A1 (zh) 2021-08-26

Family

ID=71172898

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/117025 WO2021164250A1 (zh) 2020-02-21 2020-09-23 一种湍流场更新方法、装置及其相关设备

Country Status (3)

Country Link
US (1) US11741373B2 (zh)
CN (1) CN111324993B (zh)
WO (1) WO2021164250A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111324993B (zh) * 2020-02-21 2022-06-07 苏州浪潮智能科技有限公司 一种湍流场更新方法、装置及其相关设备
CN113420507B (zh) * 2021-07-08 2023-11-21 刘维烁 流体流动模拟方法、装置以及介质
CN113536709A (zh) * 2021-07-26 2021-10-22 厦门大学 一种基于湍流建模机器学习的湍流模型修正方法
CN114282448A (zh) * 2021-11-16 2022-04-05 北京百度网讯科技有限公司 流场信息获取方法、模型训练方法、装置及电子设备
CN114417748B (zh) * 2022-01-19 2022-09-23 中山大学 一种水上飞机的水气交混状态湍流场的模拟方法及装置
CN116070123A (zh) * 2023-04-07 2023-05-05 中国空气动力研究与发展中心计算空气动力研究所 一种壁湍流冗余数据剔除方法、装置、设备及介质
CN116341397B (zh) * 2023-05-30 2023-09-05 湖南光华防务科技集团有限公司 一种基于深度神经网络的灭火弹温度场模拟方法
CN116361927B (zh) * 2023-06-02 2023-08-29 浙江大学 面向高速飞行器复杂内流数值模拟的高精度湍流建模方法
CN117574029B (zh) * 2024-01-19 2024-04-26 中国空气动力研究与发展中心计算空气动力研究所 高分辨率雷诺应力与雷诺平均Navier-Stokes方程求解器的兼容方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106777542A (zh) * 2016-11-23 2017-05-31 大连理工大学 弹性叶片螺旋桨流噪声预测方法
US20180262291A1 (en) * 2017-03-07 2018-09-13 Government Of The United States Of America, As Represented By The Secretary Of The Navy Method for free space optical communication utilizing patterned light and convolutional neural networks
CN108717722A (zh) * 2018-04-10 2018-10-30 天津大学 基于深度学习和sph框架的流体动画生成方法及装置
CN109165440A (zh) * 2018-08-22 2019-01-08 西北工业大学 一种轴流压气机全三维级间气动匹配优化方法
CN110083977A (zh) * 2019-05-14 2019-08-02 南京大学 一种基于深度学习的大气湍流监测方法
CN111324993A (zh) * 2020-02-21 2020-06-23 苏州浪潮智能科技有限公司 一种湍流场更新方法、装置及其相关设备

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10255807B1 (en) * 2017-09-28 2019-04-09 Here Global B.V. Method and apparatus for providing a map data update based on region-specific data turbulence
CN107742050B (zh) * 2017-11-14 2020-09-22 中国空气动力研究与发展中心计算空气动力研究所 一种面向混合网格高精度阻力预测的熵修正方法
CN108073547B (zh) * 2017-12-06 2021-05-18 苏州大学 一种基于能量耗散的溶血经验预测方法及装置
CN108985929B (zh) * 2018-06-11 2022-04-08 创新先进技术有限公司 训练方法、业务数据分类处理方法及装置、电子设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106777542A (zh) * 2016-11-23 2017-05-31 大连理工大学 弹性叶片螺旋桨流噪声预测方法
US20180262291A1 (en) * 2017-03-07 2018-09-13 Government Of The United States Of America, As Represented By The Secretary Of The Navy Method for free space optical communication utilizing patterned light and convolutional neural networks
CN108717722A (zh) * 2018-04-10 2018-10-30 天津大学 基于深度学习和sph框架的流体动画生成方法及装置
CN109165440A (zh) * 2018-08-22 2019-01-08 西北工业大学 一种轴流压气机全三维级间气动匹配优化方法
CN110083977A (zh) * 2019-05-14 2019-08-02 南京大学 一种基于深度学习的大气湍流监测方法
CN111324993A (zh) * 2020-02-21 2020-06-23 苏州浪潮智能科技有限公司 一种湍流场更新方法、装置及其相关设备

Also Published As

Publication number Publication date
CN111324993A (zh) 2020-06-23
CN111324993B (zh) 2022-06-07
US20230102815A1 (en) 2023-03-30
US11741373B2 (en) 2023-08-29

Similar Documents

Publication Publication Date Title
WO2021164250A1 (zh) 一种湍流场更新方法、装置及其相关设备
JP7234370B2 (ja) ニューラルネットワークを使用した集積回路フロアプランの生成
US20220391771A1 (en) Method, apparatus, and computer device and storage medium for distributed training of machine learning model
CN111970163B (zh) 一种基于注意力机制的lstm模型的网络流量预测方法
CN111130839A (zh) 一种流量需求矩阵预测方法及其系统
CN114862656B (zh) 基于多gpu的分布式深度学习模型训练代价的获取方法
Yang et al. An integer-valued threshold autoregressive process based on negative binomial thinning
US20220027774A1 (en) Quantum control pulse generation method, device, and storage medium
CN110832509A (zh) 使用神经网络的黑盒优化
Shu et al. An on-line variable fidelity metamodel assisted multi-objective genetic algorithm for engineering design optimization
CN112700326A (zh) 一种基于灰狼算法优化bp神经网络的信贷违约预测方法
CN115150471B (zh) 数据处理方法、装置、设备、存储介质及程序产品
CN114490065A (zh) 一种负载预测方法、装置及设备
CN111667189A (zh) 一种基于一维卷积神经网络的建筑工程项目风险预测方法
CN110147284B (zh) 基于二维长短期记忆神经网络的超级计算机工作负载预测方法
Chen et al. Towards high-accuracy deep learning inference of compressible turbulent flows over aerofoils
CN117057258B (zh) 基于权重分配相关系数的黑启动过电压预测方法及系统
Wang et al. Adaptive echo state network with a recursive inverse-free weight update algorithm
Luo et al. Active Kriging-based conjugate first-order reliability method for highly efficient structural reliability analysis using resample strategy
CN116011071A (zh) 基于主动学习的空中造楼机结构可靠性分析方法及系统
CN112241811B (zh) “互联网+”环境下定制产品的分层混合性能预测方法
WO2023065309A1 (zh) 一种电路设计方法及装置
CN117435308B (zh) 一种基于并行计算算法的Modelica模型仿真方法及系统
Jin et al. Application Load prediction model based on SA-IPSO-BiLSTM
WO2020252634A1 (zh) 一种基于现场可编程门阵列的点击率预估的系统及方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20919952

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20919952

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20919952

Country of ref document: EP

Kind code of ref document: A1