CN113657020B - Deep cluster neural network model construction method for pneumatic data processing - Google Patents
Deep cluster neural network model construction method for pneumatic data processing Download PDFInfo
- Publication number
- CN113657020B CN113657020B CN202110788390.3A CN202110788390A CN113657020B CN 113657020 B CN113657020 B CN 113657020B CN 202110788390 A CN202110788390 A CN 202110788390A CN 113657020 B CN113657020 B CN 113657020B
- Authority
- CN
- China
- Prior art keywords
- neural network
- deep
- data
- cluster
- pneumatic
- 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
Links
- 238000003062 neural network model Methods 0.000 title claims abstract description 52
- 238000012545 processing Methods 0.000 title claims abstract description 28
- 238000010276 construction Methods 0.000 title claims abstract description 12
- 238000000034 method Methods 0.000 claims abstract description 67
- 238000012549 training Methods 0.000 claims abstract description 48
- 238000013528 artificial neural network Methods 0.000 claims abstract description 36
- 238000012360 testing method Methods 0.000 claims abstract description 35
- 230000004044 response Effects 0.000 claims abstract description 22
- 238000013461 design Methods 0.000 claims abstract description 21
- 238000012795 verification Methods 0.000 claims abstract description 18
- 238000009826 distribution Methods 0.000 claims abstract description 10
- 238000002360 preparation method Methods 0.000 claims abstract description 8
- 239000012530 fluid Substances 0.000 claims abstract description 7
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 238000002372 labelling Methods 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims description 12
- 238000002474 experimental method Methods 0.000 claims description 9
- 238000004140 cleaning Methods 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000003064 k means clustering Methods 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 230000007547 defect Effects 0.000 abstract description 5
- 238000005070 sampling Methods 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 32
- 230000000694 effects Effects 0.000 description 14
- 238000005457 optimization Methods 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012356 Product development Methods 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010845 search algorithm Methods 0.000 description 1
Images
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
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- 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
- G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Hardware Design (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Pure & Applied Mathematics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Probability & Statistics with Applications (AREA)
- Fluid Mechanics (AREA)
- Algebra (AREA)
- Medical Informatics (AREA)
- Automation & Control Theory (AREA)
- Aviation & Aerospace Engineering (AREA)
- Computational Mathematics (AREA)
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
- Feedback Control In General (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a deep cluster neural network model construction method for pneumatic data processing, and relates to the technical field of deep cluster neural network model construction in the field of aerodynamic data processing. a. Preparation and preprocessing of data sets: firstly, acquiring a pneumatic data set by a computational fluid dynamics method, and extracting main design parameters and response parameters in the pneumatic data set; classifying the data into a plurality of subsets; labeling each subset; finally, dividing the pneumatic data set into a training set, a verification set and a test set; b. constructing a deep cluster neural network model; c. training a deep cluster neural network model; d. and (5) verifying a deep cluster neural network model. When the clustered neural network trained by the method is used for processing aerodynamic data, the defect of a traditional neural network model in an environment with insufficient sample quality caused by uneven sample sampling or uneven sample distribution can be overcome, so that the prediction accuracy of the model is improved.
Description
Technical Field
The invention relates to the technical field of deep cluster neural network model construction in the field of aerodynamic data processing.
Background
With the continuous development of the economy and science and technology of the country, the aerospace field is also advancing continuously. Aerodynamic research directly affects the development and advancement of the aerospace industry. In current aerodynamic studies, there are three main sources of data: numerical calculation, wind tunnel experiments and flight tests.
The numerical calculation method is a computational fluid dynamics method, and can provide a flow field numerical simulation result with high precision. However, the solution of the partial differential equation of aerodynamics is susceptible to turbulence and flow field grid density, requiring significant time and cost, and some complex partial differential equations do not have a numerical solution.
Wind tunnel experiments refer to the placement of a model in a wind tunnel, and the study of gas flow and its interactions with the model. The inherent simulation deficiency of wind tunnel experiments mainly has the following three aspects: the effects of the hole wall effect or hole wall interference, stent interference and the like are not met.
The flight test refers to a physical test such as an airplane flight test, a missile live-action launching test and the like, the simulation distortion problems of a model, an environment and the like can not occur, the method is always a final means for identifying the aerodynamic performance of the aircraft and calibrating other experimental results, the test is high in cost, the test environment is difficult to control, the boundary condition is difficult to reach, and the method cannot be carried out in the initial stage of product development.
An artificial Neural Network (Artificial Neural Network, ANN) is a mathematical model for simulating a complex information processing mechanism of a Neural system of a human brain based on a Network topology knowledge as a theoretical basis after understanding and abstracting the human brain structure and an external stimulus response mechanism based on a basic principle of the Neural Network in biology. In recent years, the application of the neural network method in the aerodynamic field is in progress, for example, a Chinese patent document with publication number of CN103488847A and publication date of 2016, 2 and 10 discloses a pneumatic shape optimization method based on neural network integration. According to the method, a plurality of aerodynamic profiles are constructed according to different aerodynamic profile parameters to serve as samples, and a numerical analysis method is adopted to obtain an objective function of each sample. Based on the sample data, constructing an objective function approximation model by adopting a neural network integration method. The approximation model is less accurate but much less computationally intensive than the numerical analysis method. And combining the objective function approximation model with a direct search algorithm to perform optimized search, and calling a numerical analysis method or an approximation model to acquire the objective function based on a certain strategy in the search process until the optimal aerodynamic appearance is obtained. The method can effectively reduce the number of times of numerical analysis in the optimization process on the premise of ensuring the optimization effect, improves the optimization design efficiency and quality, and is very suitable for the pneumatic appearance optimization design of the aircraft and related engineering problems.
However, the conventional neural network methods represented by the above patent documents themselves require training based on a large amount of sample data, and although each of the above three data sample acquisition methods has its own advantages, the obtained data sets may be unevenly distributed throughout the sample space, and the quality of the sample data sets may be affected. When the samples are unevenly sampled or the samples are unevenly distributed to cause low quality of the samples, predicting data with few sample characteristics is more difficult than predicting data with most sample characteristics, modeling accuracy is further affected, and even modeling failure is caused, so that further application and development of a neural network method in the whole pneumatic field are affected.
Disclosure of Invention
Aiming at the defects and shortcomings of the prior art, the invention provides a deep clustered neural network model construction method for pneumatic data processing, and the clustered neural network trained by the method can make up the defects of a traditional neural network model in an environment with insufficient sample quality caused by uneven sample sampling or uneven sample distribution when aerodynamic data are processed, so that the prediction precision of the model is improved.
The invention is realized by adopting the following technical scheme:
a method for constructing a deep cluster neural network model for pneumatic data processing is characterized by comprising the following steps:
a. preparation and preprocessing of data sets: firstly, obtaining a pneumatic data set through a computational fluid dynamics method, a wind tunnel experiment or a flight test, and extracting main design parameters and response parameters in the pneumatic data set; classifying the data in the pneumatic data set into a plurality of subsets according to the distribution characteristics of the data; labeling each subset for identifying which cluster in the deep clustered neural network a subset plays in training, each cluster in the deep clustered neural network correspondingly processing data in the labeled subset; finally, all data in the pneumatic data set are normalized, and are divided into a training set, a verification set and a test set according to a set proportion;
b. deep cluster neural network model construction: the deep cluster neural network consists of n clusters, and each cluster consists of a functional network and a context network; the output of each cluster is the product of the output of the functional network and the context network, and the final output of the deep cluster neural network is the sum of the output of all clusters;
c, constructing a model according to the data processed in the step a and the basic structure of the clusters in the deep cluster network, and determining the number of input nodes, the number of output nodes, the number of clusters, and the hidden layer number and the node number of the hidden layer of the fully-connected neural network of the functional network and the context network in the clusters;
c. training a deep cluster neural network model;
d. and (3) verifying a deep cluster neural network model: and (3) verifying the depth cluster neural network model constructed in the step three by adopting a K-fold cross verification method, and calculating the test error of the model according to the data in the test set and the data predicted by the model. The objective evaluation of the model effect is achieved.
In the step c, the training of the deep cluster neural network model is more specifically as follows: the training process of the functional network and the context network adopts alternate training, the context network is trained firstly, and then the functional network is trained after the parameters of the context network are fixed; the loss functions of the functional network and the context network are different, the loss function of the functional network is a regression loss function, the regression problem is processed, the loss function of the context network is a classification loss function, and the classification problem is processed; the functional network and the context network are trained alternately until both converge.
In the step c, the regression loss function of the functional network adopts a mean square error function:
wherein N represents the number of samples in the training set, y i Representing the predicted values obtained from the training of the samples,
representing the true value of the sample.
In the step c, the loss function of the context network adopts a cross entropy loss function:
wherein N represents the number of samples, c ij Representing the probability, p, that the ith input sample is partitioned by the jth context network into the jth class ij Representing the true class to which the i-th input sample belongs.
In the step a, before classifying the data in the pneumatic data set, there is a data cleaning step: the removal process is performed for data in the pneumatic dataset, such as the presence of outliers and nulls.
In the step a, the data in the pneumatic data set are classified, specifically, a K-means clustering algorithm is adopted, and clustering is carried out according to distance between each sample.
In the step a, the set proportion is specifically as follows: the training set, the validation set and the test set are divided according to the proportion of 8:1:1.
In the step a, the design parameters include the air flow angle, mach number, height, reynolds number and rudder deflection angle.
In the step a, the response parameters include a lift coefficient, a lateral force coefficient, a resistance coefficient, a pitch moment coefficient, a yaw moment coefficient and a roll moment coefficient.
In the step b, the number of the input nodes is determined by the dimension of the design parameter, and the parameters of the output nodes are determined by the dimension of the response parameter.
In the step b, the dimension of the design parameter refers to the number of the design parameter, such as the airflow angle, the mach number, the altitude, the reynolds number and the rudder deflection angle, which are 5, that is, 5 dimensions.
In the step b, the dimension of the response parameter refers to the number of response parameters, such as lift coefficient, lateral force coefficient, drag coefficient, pitch moment coefficient, yaw moment coefficient and roll moment coefficient, which are 6, that is, 6 dimensions.
In the step b, the number of clusters of the deep cluster neural network, the hidden number of functional networks and context networks in each cluster and the number of hidden nodes of the hidden layers are determined by a designer of the neural network according to experience, specifically, the number of clusters is set to be 4-6, the hidden number of functional networks in each cluster is 3-5, the number of hidden nodes is 32-1024, the hidden number of context networks in each cluster is 1-3, and the number of hidden nodes is 4-16.
Compared with the prior art, the invention has the following beneficial effects:
1. compared with the prior art represented by the publication No. CN103488847A, the preparation and pretreatment of the data set are adopted; b. constructing a deep cluster neural network model; c. training a deep cluster neural network model; d. the method for constructing the whole deep clustered neural network model formed by the deep clustered neural network model verification is used for processing pneumatic data, is particularly suitable for steady-state nonlinear pneumatic response prediction modeling when the quality of a sample data set is low due to uneven sample sampling or uneven sample distribution, and improves the prediction accuracy of the model compared with the traditional aerodynamic model.
2. In the preparation and pretreatment of the data set, a method of extracting main design parameters and response parameters in the pneumatic data set by adopting a computational fluid dynamics method is adopted, in an ideal flight environment, the real flight state of an aircraft can be simulated, meanwhile, the pneumatic data set is divided into a plurality of subsets, each subset is labeled for identifying which cluster training in a cluster network a certain subset plays, each cluster in the deep cluster network correspondingly processes data in the labeled subset, and the processing mode can enable each cluster to concentrate on processing the independent data subset, shield the mutual interference of training among the data subsets and achieve the effect of improving the model learning precision.
3. In the invention, all data in the pneumatic data set are normalized and divided into the training set, the verification set and the test set according to the set proportion, and the data are normalized in the setting mode, so that the convergence speed of the model can be accelerated, the precision of the model can be improved, and gradient explosion can be prevented. The training set, the verification set and the test set are divided according to the set proportion, and no intersection exists among the three sets, so that the model can be trained on the training set, the model is verified on the verification set and the model is tested on the test set, the problem of data leakage does not exist, and the effect of accurately evaluating the model learning precision is achieved.
4. In the invention, a model is constructed according to the data processed in the step a and the basic structure of the clusters in the cluster network, and the specific modes of determining the number of input nodes, the number of output nodes, the number of clusters, the hidden number of functional networks and context networks in each cluster and the number of nodes of the hidden layers have the following special technical effects: the number of the input nodes and the number of the output nodes are determined, so that modeling of the model aiming at design parameters and response parameters can be guaranteed, and any parameter is not omitted; the determination of the cluster number is one of key factors for ensuring the model learning effect, and a proper cluster number determines the learning effect of shielding mutual interference among all data subsets in the deep clustered neural network; the number of hidden layers and the number of nodes of hidden layers of the functional network and the context network in each cluster determine the learning effect of each cluster.
5. In the step c, the training process of the functional network and the context network adopts alternate training, so that the specific processing mode of the alternate training can ensure that the model can process classification and regression problems simultaneously, and the convergence rate of the model is accelerated.
6. In the step c of the invention, the regression loss function of the functional network adopts the mean square error function, and the loss function of the context network adopts the cross entropy loss function, so that the specific setting mode is inquired that the cross entropy loss function is not adopted for training the context network in the pneumatic data, the technical prejudice that the cross entropy loss function cannot be used for training the context network in the pneumatic data in the prior art is overcome, and unexpected technical effects are obtained: in the classification of the context network, the cross entropy loss function obtains lower classification errors, namely the classification of the context network adopting the cross entropy loss function on pneumatic data is more accurate, and in the prediction of the response parameters by the final model, a model of the cross entropy loss function is adopted, and the prediction error is lower than that of the model adopting the mean square error loss function.
7. According to the invention, before classifying the data in the pneumatic data set, a data cleaning step is further provided, and abnormal values and null values in the data can be removed through the data cleaning step, so that the prediction precision of the constructed deep cluster neural network model is enhanced.
8. According to the invention, the training set, the verification set and the test set are divided according to the proportion of 8:1:1, so that the specific proportion is divided, and experiments prove that most of pneumatic data can be used for training, and errors (namely, the phenomenon of under fitting) during model training are reduced; a small amount of pneumatic data is used for verification and testing to determine if the model has over-learned the unnecessary relationship between design parameters and response parameters (i.e., avoid over-fitting).
Drawings
The invention will be described in further detail with reference to the drawings and detailed description, wherein:
FIG. 1 is a flow chart of a method for constructing a deep clustered neural network model;
fig. 2 is a diagram of a deep clustered neural network model constructed by the method.
Detailed Description
Example 1
As a preferred embodiment of the invention, a deep clustered neural network model construction method for pneumatic data processing is disclosed, which comprises the following steps:
a. preparation and preprocessing of data sets: firstly, obtaining a pneumatic data set through a computational fluid dynamics method, a wind tunnel experiment or a flight test, and extracting main design parameters and response parameters in the pneumatic data set; classifying the data in the pneumatic data set into a plurality of subsets according to the distribution characteristics of the data; labeling each subset for identifying which cluster in the deep clustered neural network a subset plays in training, each cluster in the deep clustered neural network correspondingly processing data in the labeled subset; finally, all data in the pneumatic data set are normalized, and are divided into a training set, a verification set and a test set according to a set proportion (for example, 7:2:1 or 6:2:2);
b. deep cluster neural network model construction: the deep cluster neural network consists of n clusters, and each cluster consists of a functional network and a context network; the output of each cluster is the product of the output of the functional network and the context network, and the final output of the deep cluster neural network is the sum of the output of all clusters;
c, constructing a model according to the data processed in the step a and the basic structure of the clusters in the deep cluster network, and determining the number of input nodes, the number of output nodes, the number of clusters, and the hidden layer number and the node number of the hidden layer of the fully-connected neural network of the functional network and the context network in the clusters;
c. training a deep cluster neural network model: the step adopts a conventional network model training method in the prior art.
d. And (3) verifying a deep cluster neural network model: and (3) verifying the depth cluster neural network model constructed in the step three by adopting a K-fold cross verification method, and calculating the test error of the model according to the data in the test set and the data predicted by the model. The objective evaluation of the model effect is achieved.
Example 2
Referring to fig. 1 and 2, another preferred embodiment of the present invention is: aiming at the defects and shortcomings of the existing data acquisition method in the pneumatic data modeling field, the invention provides an aerodynamic modeling method based on a clustered neural network for modeling pneumatic data distribution characteristics. The method is based on a network structure of a cluster network, and provides a modeling method for pneumatic data distribution characteristics so as to meet the requirements of pneumatic data modeling. When the clustered neural network trained by the method is used for processing aerodynamic data, the defect of a traditional neural network model in an environment with insufficient sample quality caused by uneven sample sampling or uneven sample distribution can be overcome, so that the accuracy of the model is improved.
The method comprises the following steps:
a. preparation and preprocessing of data sets: firstly, obtaining a pneumatic data set through a computational fluid dynamics method, a wind tunnel experiment or a flight test, and extracting main design parameters and response parameters in the pneumatic data set; classifying the data in the pneumatic data set into a plurality of subsets according to the distribution characteristics of the data; labeling each subset for identifying which cluster in the deep clustered neural network a subset plays in training, each cluster in the deep clustered neural network correspondingly processing data in the labeled subset; finally, all data in the pneumatic data set are normalized, and are divided into a training set, a verification set and a test set according to a set proportion;
in the step a, the data in the pneumatic data set are classified, specifically, a K-means clustering algorithm is adopted, and clustering is carried out according to distance between each sample.
In the step a, the set proportion is specifically as follows: the training set, the validation set and the test set are divided according to the proportion of 8:1:1.
In the step a, the design parameters include the air flow angle, mach number, height, reynolds number and rudder deflection angle.
In the step a, the response parameters include a lift coefficient, a lateral force coefficient, a resistance coefficient, a pitch moment coefficient, a yaw moment coefficient and a roll moment coefficient.
b. Deep cluster neural network model construction: the cluster network consists of n clusters, and each cluster consists of a functional network and a context network; the output of each cluster is the product of the output of the functional network and the context network, and the final output of the cluster network is the sum of the outputs of all clusters:
where u represents the output of the model, f represents the output of the functional network, c represents the output of the context network, i represents the sequence number of the cluster;
c, constructing a model according to the data processed in the step a and the basic structure of the clusters in the deep cluster network, and determining the number of input nodes, the number of output nodes, the number of clusters, and the hidden layer number and the node number of the hidden layer of the fully-connected neural network of the functional network and the context network in the clusters;
in the step b, the number of the input nodes is determined by the dimension of the design parameter, and the parameters of the output nodes are determined by the dimension of the response parameter.
In the step b, the dimension of the design parameter refers to the number of the design parameter, such as the airflow angle, the mach number, the altitude, the reynolds number and the rudder deflection angle, which are 5, that is, 5 dimensions.
In the step b, the dimension of the response parameter refers to the number of response parameters, such as lift coefficient, lateral force coefficient, drag coefficient, pitch moment coefficient, yaw moment coefficient and roll moment coefficient, which are 6, that is, 6 dimensions.
In the step b, the number of clusters of the deep cluster neural network, the hidden number of functional networks and context networks in each cluster and the number of hidden nodes of the hidden layers are determined by a designer of the neural network according to experience, specifically, the number of clusters is set to be 4-6, the hidden number of functional networks in each cluster is 3-5, the number of hidden nodes is 32-1024, the hidden number of context networks in each cluster is 1-3, and the number of hidden nodes is 4-16.
c. Training a deep cluster neural network model: the training process of the functional network and the context network adopts alternate training, the context network is trained firstly, and then the functional network is trained after the parameters of the context network are fixed; the loss functions of the functional network and the context network are different, the loss function of the functional network is a regression loss function, the regression problem is processed, the loss function of the context network is a classification loss function, and the classification problem is processed; the functional network and the context network are trained alternately until both converge.
The regression loss function of the functional network adopts a mean square error function:
wherein N represents the number of samples in the training set, y i Representing the predicted values obtained from the training of the samples,
representing the true value of the sample.
The loss function of the context network employs a cross entropy loss function:
wherein N represents a sampleNumber, c ij Representing the probability, p, that the ith input sample is partitioned by the jth context network into the jth class ij Representing the true class to which the i-th input sample belongs.
d. And (3) verifying a deep cluster neural network model: and (3) verifying the depth cluster neural network model constructed in the step three by adopting a K-fold cross verification method, and calculating the test error of the model according to the data in the test set and the data predicted by the model. The objective evaluation of the model effect is achieved.
Example 3
Referring to fig. 1 and 2, as the best mode of the present invention, it is: in the step a, before classifying the data in the pneumatic data set, there is a data cleaning step: the removal process is performed for data in the pneumatic dataset, such as the presence of outliers and nulls. Thereby enhancing the prediction accuracy of the constructed deep clustered neural network model.
Example 4
Specific examples of applications of the present invention are as follows:
preparing a hardware environment: firstly, one PC is needed, and simultaneously, an Nvidia Tesla K80 deep learning display card is provided.
Preparing a software environment: the PC is provided with a Windows system or a Linux system. Installing more than Python3.0 version, tensorFlow1.4 version and Keras2.0 version.
Preparation and preprocessing of data sets: the data set may be acquired by CFD calculation or wind tunnel experiments. Extracting main parameters in pneumatic data, including: design parameters (airflow angle, mach number, altitude and reynolds number) and response parameters (rudder deflection angle, lift coefficient, lateral force coefficient, drag coefficient, pitch moment coefficient, yaw moment coefficient and roll moment coefficient). The second to-be-preprocessed data includes data cleaning: outliers and nulls present in the data need to be screened out. Each data then needs to be tagged to indicate in which cluster the data is trained. Finally, the data set is also required to be divided into a training set, a verification set and a test set.
And (3) constructing a model: and obtaining an input-output dimension through a preprocessing process of the data in the earlier stage, and determining the number of input nodes and output nodes of the model through the dimension. Thus, a neural network model is preliminarily constructed, and the complexity of the model is roughly matched with the size of the data set.
And (3) verifying a model: and (3) verifying the depth cluster neural network model constructed in the step three by adopting a K-fold cross verification method, and calculating the test error of the model according to the data in the test set and the data predicted by the model. The objective evaluation of the model effect is achieved.
And (3) construction: after the model passes the verification, a complete deep clustered neural network based on the clustered network is established.
Claims (10)
1. A method for constructing a deep cluster neural network model for pneumatic data processing is characterized by comprising the following steps:
a. preparation and preprocessing of data sets: firstly, obtaining a pneumatic data set through a computational fluid dynamics method, a wind tunnel experiment or a flight test, and extracting main design parameters and response parameters in the pneumatic data set; classifying the data in the pneumatic data set into a plurality of subsets according to the distribution characteristics of the data; labeling each subset for identifying which cluster in the deep clustered neural network a subset plays in training, each cluster in the deep clustered neural network correspondingly processing data in the labeled subset; finally, all data in the pneumatic data set are normalized, and are divided into a training set, a verification set and a test set according to a set proportion;
b. deep cluster neural network model construction: the deep cluster neural network consists of n clusters, and each cluster consists of a functional network and a context network; the output of each cluster is the product of the output of the functional network and the context network, and the final output of the deep cluster neural network is the sum of the output of all clusters; c, constructing a model according to the data processed in the step a and the basic structure of the clusters in the deep cluster network, and determining the number of input nodes, the number of output nodes, the number of clusters, and the hidden layer number and the node number of the hidden layer of the fully-connected neural network of the functional network and the context network in the clusters;
c. training a deep cluster neural network model;
d. and (3) verifying a deep cluster neural network model: and (3) verifying the depth cluster neural network model constructed in the step three by adopting a K-fold cross verification method, and calculating the test error of the model according to the data in the test set and the data predicted by the model.
2. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step c, the training of the deep cluster neural network model is more specifically as follows: the training process of the functional network and the context network adopts alternate training, the context network is trained firstly, and then the functional network is trained after the parameters of the context network are fixed; the loss function of the functional network is a regression loss function, the regression problem is processed, the loss function of the context network is a classification loss function, and the classification problem is processed; the functional network and the context network are trained alternately until both converge.
3. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 2, wherein the method comprises the following steps: in the step c, the regression loss function of the functional network adopts a mean square error function:
wherein ,
representing the number of samples in the training set, +.>
Representing the predicted value obtained from the sample training, +.>
Representing the true value of the sample.
4. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 2, wherein the method comprises the following steps: in the step c, the loss function of the context network adopts a cross entropy loss function:
wherein N represents the number of samples,
representing the probability of dividing the ith input sample into the jth class by the jth context network, for example>
Representing the true class to which the i-th input sample belongs.
5. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step a, before classifying the data in the pneumatic data set, there is a data cleaning step: the removal process is performed for data in the pneumatic dataset, such as the presence of outliers and nulls.
6. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step a, the data in the pneumatic data set are classified, specifically, a K-means clustering algorithm is adopted, and clustering is carried out according to distance between each sample.
7. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step a, the set proportion is specifically as follows: the training set, the validation set and the test set are divided according to the proportion of 8:1:1.
8. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step a, the design parameters include the air flow angle, mach number, height, reynolds number and rudder deflection angle.
9. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step a, the response parameters include a lift coefficient, a lateral force coefficient, a resistance coefficient, a pitch moment coefficient, a yaw moment coefficient and a roll moment coefficient.
10. The method for constructing the deep clustered neural network model for pneumatic data processing according to claim 1, wherein the method comprises the following steps: in the step b, the number of the input nodes is determined by the dimension of the design parameter, and the parameters of the output nodes are determined by the dimension of the response parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110788390.3A CN113657020B (en) | 2021-07-13 | 2021-07-13 | Deep cluster neural network model construction method for pneumatic data processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110788390.3A CN113657020B (en) | 2021-07-13 | 2021-07-13 | Deep cluster neural network model construction method for pneumatic data processing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113657020A CN113657020A (en) | 2021-11-16 |
| CN113657020B true CN113657020B (en) | 2023-05-05 |
Family
ID=78489315
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110788390.3A Active CN113657020B (en) | 2021-07-13 | 2021-07-13 | Deep cluster neural network model construction method for pneumatic data processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113657020B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115099275B (en) * | 2022-06-29 | 2023-07-07 | 西南医科大学 | Training method of arrhythmia diagnosis model based on artificial neural network |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108549934A (en) * | 2018-04-25 | 2018-09-18 | 福州瑞芯微电子股份有限公司 | A kind of operation method and device based on automated cluster neural network chip group |
| CN109193649A (en) * | 2018-10-23 | 2019-01-11 | 东南大学 | Distributed photovoltaic cluster dynamic equivalent modeling method based on deepness belief network |
| CN109492516A (en) * | 2018-09-01 | 2019-03-19 | 哈尔滨工程大学 | A kind of UUV Aggregation behaviour recognition methods based on DGRU neural network |
| CN109871976A (en) * | 2018-12-20 | 2019-06-11 | 浙江工业大学 | A kind of prediction technique of power quality containing distributed power distribution network based on cluster and neural network |
| CN112019381A (en) * | 2020-08-12 | 2020-12-01 | 苏州浪潮智能科技有限公司 | Cluster link detection method and system based on deep learning |
| CN112597702A (en) * | 2020-12-21 | 2021-04-02 | 电子科技大学 | Pneumatic modeling generation type confrontation network model training method based on radial basis function |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11023441B2 (en) * | 2017-05-11 | 2021-06-01 | Oracle International Corporation | Distributed storage and processing of hierarchical data structures |
| CN110390387B (en) * | 2018-04-20 | 2023-07-18 | 伊姆西Ip控股有限责任公司 | Assessment of resources used by deep learning applications |
-
2021
- 2021-07-13 CN CN202110788390.3A patent/CN113657020B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108549934A (en) * | 2018-04-25 | 2018-09-18 | 福州瑞芯微电子股份有限公司 | A kind of operation method and device based on automated cluster neural network chip group |
| CN109492516A (en) * | 2018-09-01 | 2019-03-19 | 哈尔滨工程大学 | A kind of UUV Aggregation behaviour recognition methods based on DGRU neural network |
| CN109193649A (en) * | 2018-10-23 | 2019-01-11 | 东南大学 | Distributed photovoltaic cluster dynamic equivalent modeling method based on deepness belief network |
| CN109871976A (en) * | 2018-12-20 | 2019-06-11 | 浙江工业大学 | A kind of prediction technique of power quality containing distributed power distribution network based on cluster and neural network |
| CN112019381A (en) * | 2020-08-12 | 2020-12-01 | 苏州浪潮智能科技有限公司 | Cluster link detection method and system based on deep learning |
| CN112597702A (en) * | 2020-12-21 | 2021-04-02 | 电子科技大学 | Pneumatic modeling generation type confrontation network model training method based on radial basis function |
Non-Patent Citations (2)
| Title |
|---|
| Lisha CUI ; Rui MA ; Pei LV ; Xiaoheng JIANG ; Zhimin GAO ; Bing ZHOU ; Mingliang XU ; .MDSSD:multi-scale deconvolutional single shot detector for small objects.Science China(Information Sciences).2020,(第02期),98-100. * |
| 钱光武 ; 张蕾 ; 王炎 ; .基于概念器的深度神经网络模型.中国科学:信息科学.2018,(第05期),31-40. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113657020A (en) | 2021-11-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113609596B (en) | Aircraft aerodynamic characteristic prediction method based on neural network | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN107766668B (en) | Complex simulation model verification method based on neural network | |
| CN110082738B (en) | Radar target identification method based on Gaussian mixture and tensor recurrent neural network | |
| CN108549866B (en) | Remote sensing airplane identification method based on dense convolutional neural network | |
| CN110443448B (en) | Bidirectional LSTM-based airplane position classification prediction method and system | |
| CN112597702B (en) | Pneumatic modeling generation type confrontation network model training method based on radial basis function | |
| CN106203483A (en) | A kind of zero sample image sorting technique of multi-modal mapping method of being correlated with based on semanteme | |
| CN104850867A (en) | Object identification method based on intuitive fuzzy c-means clustering | |
| CN113777931B (en) | Icing wing type pneumatic model construction method, device, equipment and medium | |
| CN111539302B (en) | Glass insulator self-explosion identification method based on multi-scale deep disturbance neural network | |
| CN113657020B (en) | Deep cluster neural network model construction method for pneumatic data processing | |
| CN115880497A (en) | Wing-type icing ice shape prediction method based on combination of self-encoder and multi-layer perceptron | |
| CN114973019A (en) | Deep learning-based geospatial information change detection classification method and system | |
| CN112947080B (en) | Scene parameter transformation-based intelligent decision model performance evaluation system | |
| CN108984864A (en) | Higher-dimension complex aircraft model characterizing method based on Gaussian process variance evaluation | |
| CN115083511A (en) | Peripheral gene regulation and control feature extraction method based on graph representation learning and attention | |
| CN113722951B (en) | Scatterer three-dimensional finite element grid optimization method based on neural network | |
| CN115206455A (en) | Deep neural network-based rare earth element component content prediction method and system | |
| CN113108949A (en) | Model fusion-based sonde temperature sensor error prediction method | |
| CN117634365B (en) | Airplane aerodynamic force prediction method, electronic equipment and storage medium | |
| Gao et al. | High precision detection technology of infrared wall cracks based on improved single shot multibox detector | |
| Sha et al. | Ensemble 1D DenseNet Damage Identification Method Based on Vibration Acceleration. | |
| Asher et al. | Developing a probabilistic load spectrum for fatigue modeling | |
| CN116776230B (en) | Method and system for identifying signal based on feature imprinting and feature migration |
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 |