CN109214116A - A kind of auxiliary design method of high-speed EMUs head dummy - Google Patents
A kind of auxiliary design method of high-speed EMUs head dummy Download PDFInfo
- Publication number
- CN109214116A CN109214116A CN201811184702.4A CN201811184702A CN109214116A CN 109214116 A CN109214116 A CN 109214116A CN 201811184702 A CN201811184702 A CN 201811184702A CN 109214116 A CN109214116 A CN 109214116A
- Authority
- CN
- China
- Prior art keywords
- head dummy
- data
- speed emus
- training
- sample
- 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.)
- Withdrawn
Links
Classifications
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Hardware Design (AREA)
- Mathematical Analysis (AREA)
- Aviation & Aerospace Engineering (AREA)
- Automation & Control Theory (AREA)
- Biomedical Technology (AREA)
- Image Analysis (AREA)
- Mathematical Optimization (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Mathematics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
Abstract
The present invention relates to a kind of auxiliary design methods of high-speed EMUs head dummy, the steps include: S1, establish high-speed EMUs head dummy image sample data collection；S2, it is established according to high-speed EMUs head dummy sample data set for polytypic training data；S3, training depth convolutional neural networks model；The test sample that test data is concentrated is input in trained depth convolutional neural networks model by S4, a sample data for choosing high-speed EMUs head dummy image pattern collection as test data set, carries out convolution, the pondization operation of multilayer；S5, " marking " is carried out to test sample, obtains the head dummy score of test sample, auxiliary head dummy design is completed.The present invention combines ordinal regression and deep learning method, the data of existing mature vehicle are made full use of to assist the design selection of the following new model head dummy, head dummy type selecting quantity can be effectively reduced, it is obviously improved the efficiency of high-speed EMUs head dummy type selecting, reduces the consuming time in high-speed EMUs head dummy design process and cost.
Description
Technical field
The invention belongs to rail vehicle technical fields, are related to high-speed EMUs, specifically, relating to a kind of high speed motor car
The auxiliary design method of group head dummy.
Background technique
The design of high-speed EMUs car body head dummy highly important key technology when being design high-speed EMUs.High speed motor car
The head dummy of group mainly influences the speed of service and stress condition of train.Head dummy design be one comprising air drag, lift, cross
The dynamic load system of multiple physical parameters such as pressure wave, air pressure wave in tunnel, aesthetics, each physical parameter are not independent
Linear change, but the interaction of multiple physical parameters, mutually restriction, the stress and beauty of joint effect head dummy.Therefore, head portrait
Design the prominent features with complication systems such as complicated composition parameter, calculating complexity and correlations.
Existing high-speed EMUs head dummy designing technique, mostly carries out the drag reduction based on bionic theory from aerodynamics angle
Design.High-speed EMUs aerodynamic studies method mainly has real train test, model test, numerical value calculating and theory analysis.
In terms of numerical value calculating, propose to utilize hydrodynamics (referred to as: CFD)；Research means are mainly wind-tunnel examination in terms of model test
It tests, moving model experiment, sink or water tunnel test.For high-speed EMUs head dummy design problem, no matter using which kind of mode, progress
Which kind of optimization is difficult to make head dummy to keep good performance by a secondary design, and has that consuming time is long, at high cost asks
Topic.
Summary of the invention
Time-consuming, the above problems such as at high cost for of the existing technology by the present invention, provides a kind of high-speed EMUs head
The auxiliary design method of type, for reducing the time expended in head dummy design process and cost.
In order to achieve the above object, the present invention provides a kind of auxiliary design methods of high-speed EMUs head dummy, specific
Step are as follows:
S1, high-speed EMUs head dummy image sample data collection is established
S11, the head dummy image data for collecting the various models of high-speed EMUs；
S12, collection do the label data of label to the head dummy image data in step S11；
S13, noise filtering is carried out to the head dummy image data and label data being collected into, filters out non-genuine high-speed EMUs
Headstock noise data obtains high-speed EMUs head dummy image sample data collection；
S2, it is established according to high-speed EMUs head dummy sample data set for polytypic training data
S21, by the resonable degree scoring of headstock design for 1,2 ..., M grade, 1 be it is worst, M be it is best, form M and divide
Class device；
S22, " marking " is carried out to the head dummy image sample data collection being collected into, obtains marking data set, each sample number
According to including importation and mark part；Importation is the head dummy image data of certain model high-speed EMUs, and mark part is
Experiment entirety score after model high-speed EMUs test；
S23, input head dummy image sample data collection and marking data set generate M points a series of according to head dummy and marking
Class category obtains the training dataset comprising head dummy image, M classification category and weight；
S3, training depth convolutional neural networks model
A sample data of high-speed EMUs head dummy image pattern collection is randomly selected as validation data set, by step S2
Obtained training dataset and verify data while depth convolutional neural networks model is acted on, training data is utilized to be fitted depth
Convolutional neural networks model prevents depth convolutional Neural net by image rotation, pattern reversal, random cut using verify data
Network model transition fitting, training output depth convolutional neural networks model, so that each output is a M of M classifier
Classification category；Depth convolutional neural networks model is finally optimal after gradient optimization algorithm, obtains by constantly fitting
To trained depth convolutional neural networks model；
S4, a sample data for choosing high-speed EMUs head dummy image pattern collection will test number as test data set
It is input in trained depth convolutional neural networks model according to the test sample of concentration, carries out convolution, the pondization operation of multilayer；
S5, " marking " is carried out to test sample, obtains the head dummy score of test sample, auxiliary head dummy design is completed.
Preferably, in step S2, a series of M is generated according to head dummy and marking and is classified category, obtain comprising head dummy image,
The specific steps of the training dataset of M classification category and weight are as follows: marking mechanism transformation is sorted for score with a series of M points
The processing of class device constructs corresponding training data to each M classifier of high-speed EMUs head dummy image, gives ordered sequence
Training dataWherein, XiFor input data, the input data includes image data and numeric data, Yi∈
It [1, k] is marking grade, k=M is total number of grades, and N is the sum of training data；For k-th of M classifier, M classification categoryIndicate the ordered sequence label y of i-th of image patterniIt is whether bigger than k, definition:
Work as yiWhen > k, otherwise it is 0 that M classification class, which is designated as 1,；The training data of k-th of M classifier is configured to
Wherein, xiIt is the input data of i-th of image pattern for input data,For the weight of i-th of image pattern, using absolute generation
Valence matrix value is
Preferably, in step S3, depth convolutional neural networks model training uses intersection entropy loss, cross entropy loss function
Are as follows:
In formula, ljiIndicate whether the marking rank of j-th of training sample belongs to i grades, ojiIt indicates for j-th of trained sample
The output marking rank of this depth convolutional neural networks model is the probability of i rank.
Preferably, in step S3, the numeric data includes anti-air drag data, lift data, the pressure that crosses wave number
According to, air pressure wave in tunnel data and aesthetics data.
Preferably, in step S1, the method for the collection various model head dummy image datas of high-speed EMUs are as follows:
The head dummy image data of the various models of high-speed EMUs is collected by social networks using web crawlers；
Based on existing engineering experiment room high-speed EMUs head dummy model, various model high-speed EMUs head dummy models are collected
Random angles screenshot image data.
Preferably, in step S4, when input test sample, head dummy image of the test sample selection without marking label.
Preferably, the specific steps that the label data of label is done to the head dummy image data in step S11 are collected are as follows:
Air drag, the lift, the pressure wave that crosses, tunnel pressure of each headstock head dummy design are collected based on existing developmental achievement
Five kinds of wave, aesthetics significant in value type indexs；
High-speed EMUs head dummy image based on collection is being asked in such a way that social user carries out questionnaire ballot
Existing vehicle head dummy photo is enumerated in volume, is allowed social user to vote in and is thought personally most U.S. head dummy, is collected simultaneously social user
Age, gender, occupational information, in this, as head dummy aesthetics label；The aesthetics label is divided into 10 class, the 1st class
Aesthetics is worst, and numerical value is described as 0.1；10th class aesthetics highest, numerical value are described as 1.0；
The corresponding headstock model of high-speed EMUs head dummy image has been collected in determination, and collects the resistance of air known to the headstock
Power, lift, the pressure wave that crosses, four physical parameters of air pressure wave in tunnel experimental data, four physical parameters are normalized
Processing, as label data.
Compared with prior art, the beneficial effects of the present invention are:
The present invention utilizes data-driven Modeling Theory, establishes high-speed EMUs head dummy sample data set, and establish for more
The training data of classification, training depth convolutional neural networks model, is input to trained convolutional neural networks for test sample
Model obtains the marking estimation of test sample, scores the high-speed EMUs head dummy scheme of design, by ordinal regression and depth
Degree learning method combines, and the data of existing mature vehicle is made full use of to assist the design selection of the following new model head dummy, can
Head dummy type selecting quantity is effectively reduced, the efficiency of high-speed EMUs head dummy type selecting has been obviously improved, reduces high-speed EMUs head dummy
Consuming time and cost in design process propose new solution route for the design of conventional high rate EMU head dummy, enhance
The diversity of high-speed EMUs vehicle design method.
Detailed description of the invention
Fig. 1 is the flow chart of the auxiliary design method of high-speed EMUs of embodiment of the present invention head dummy；
Fig. 2 is the flow chart that the embodiment of the present invention establishes high-speed EMUs head dummy image sample data collection；
Fig. 3 is the flow chart for the training data that the embodiment of the present invention is established for 8 classification；
Fig. 4 is the frame construction drawing of depth of embodiment of the present invention convolutional neural networks model；
Fig. 5 is the structural schematic diagram of classics of embodiment of the present invention depth convolutional neural networks model.
Specific embodiment
In the following, the present invention is specifically described by illustrative embodiment.It should be appreciated, however, that not into one
In the case where step narration, element, structure and features in an embodiment can also be advantageously incorporated into other embodiments
In.
Present invention discloses a kind of auxiliary design methods of high-speed EMUs head dummy, the specific steps are that:
S1, high-speed EMUs head dummy image sample data collection is established
S11, the head dummy image data for collecting the various models of high-speed EMUs；Its method particularly includes:
A, the head dummy image data of the various models of high-speed EMUs is collected by social networks using web crawlers；
B, it is based on existing engineering experiment room high-speed EMUs head dummy model, collects various model high-speed EMUs head dummy models
Random angles screenshot image data；
In above-mentioned steps S11, the method and step A of head dummy image data collection, the sequence of step B be can be interchanged, step
A, step B can also be carried out simultaneously；
S12, collection do the label data of label to the head dummy image data in step S11；The specific steps are that:
A, air drag, the lift, the pressure wave that crosses, tunnel pressure of each headstock head dummy design are collected based on existing developmental achievement
Five kinds of Reeb, aesthetics significant in value type indexs；
B, the high-speed EMUs head dummy image based on collection, in such a way that social user carries out questionnaire ballot,
Existing vehicle head dummy photo is enumerated in questionnaire, is allowed social user to vote in and is thought personally most U.S. head dummy, is collected simultaneously social use
The age at family, gender, occupational information, in this, as head dummy aesthetics label；The aesthetics label is divided into 10 class, and the 1st grade
Secondary aesthetics is worst, and numerical value is described as 0.1；10th class aesthetics highest, numerical value are described as 1.0；
C, the corresponding headstock model of high-speed EMUs head dummy image has been collected in determination, and collects the resistance of air known to the headstock
Power, lift, the pressure wave that crosses, four physical parameters of air pressure wave in tunnel experimental data, four physical parameters are normalized
Processing, as label data；Such as: CH380A head dummy windage is 0.45, input data of the label data as later period sample；
S13, noise filtering is carried out to the head dummy image data and label data being collected into, filters out non-genuine high-speed EMUs
Headstock noise data obtains high-speed EMUs head dummy image sample data collection.
S2, it is established according to high-speed EMUs head dummy sample data set for polytypic training data
S21, by the resonable degree scoring of headstock design for 1,2 ..., M grade, 1 be it is worst, M be it is best, form M and divide
Class device；
S22, " marking " is carried out to the head dummy image sample data collection being collected into, obtains marking data set, each sample number
According to including importation and mark part；Importation is the head dummy image data of certain model high-speed EMUs, and mark part is
Experiment entirety score after model high-speed EMUs test；
S23, input head dummy image sample data collection and marking data set generate M points a series of according to head dummy and marking
Class category obtains the training dataset comprising head dummy image, M classification category and weight；The specific steps are that:
It is that a series of M classifier of score sequence is handled marking mechanism transformation, to high-speed EMUs head dummy image
Each M classifier constructs corresponding training data, gives the training data of ordered sequenceWherein, XiIt is defeated
Enter data, the input data includes image data and numeric data, Yi∈ [1, k] is marking grade, and k=M is total number of grades,
N is the sum of training data；For k-th of M classifier, M classification categoryIndicate the orderly of i-th of image pattern
Sequential labeling yiIt is whether bigger than k, definition:Work as yiWhen > k, otherwise it is 0 that M classification class, which is designated as 1,；
The training data of k-th of M classifier is configured toWherein, xiIt is i-th of image sample for input data
This input data,For the weight of i-th of image pattern, use absolute cost matrix value for
S3, training depth convolutional neural networks model
A sample data of high-speed EMUs head dummy image pattern collection is randomly selected as validation data set, by step S2
Obtained training dataset and verify data while depth convolutional neural networks model is acted on, training data is utilized to be fitted depth
Convolutional neural networks model prevents depth convolutional Neural net by image rotation, pattern reversal, random cut using verify data
Network model transition fitting, training output depth convolutional neural networks model, so that each output is a M of M classifier
Classification category；Depth convolutional neural networks model is finally optimal after gradient optimization algorithm, obtains by constantly fitting
To trained depth convolutional neural networks model；
When training output depth convolutional neural networks model, intersection entropy loss, cross entropy loss function are used are as follows:
In formula, ljiIndicate whether the marking rank of j-th of training sample belongs to i grades, ojiIt indicates for j-th of trained sample
The output marking rank of this depth convolutional neural networks model is the probability of i rank.
S4, a sample data for choosing high-speed EMUs head dummy image pattern collection will test number as test data set
It is input in trained depth convolutional neural networks model according to the test sample of concentration, carries out convolution, the pondization operation of multilayer.
When input test sample, head dummy image of the test sample selection without marking label.
S5, " marking " is carried out to test sample, obtains the head dummy score of test sample, auxiliary head dummy design is completed.
The auxiliary design method of above-mentioned high-speed EMUs head dummy makes full use of the data auxiliary of existing mature vehicle following new
The design selection of vehicle head dummy can effectively reduce head dummy type selecting quantity, be obviously improved the effect of high-speed EMUs head dummy type selecting
Rate saves human and material resources, and saves time and cost, and auxiliary solves high-speed EMUs head dummy design calculation process complexity, modeling
Problem cumbersome, that capital investment is big, the period is long.
Below with auxiliary design method of the specific embodiment to the above-mentioned high-speed EMUs head dummy of the present invention make into
One step explanation.
Referring to Fig. 1, Fig. 2, Fig. 3, a kind of auxiliary design method of high-speed EMUs head dummy, the specific steps are that:
S1, high-speed EMUs head dummy image sample data collection is established
S11, the head dummy image data for collecting the various models of high-speed EMUs；Its method particularly includes:
A, the head dummy image data of the various models of high-speed EMUs is collected by social networks using web crawlers；
B, it is based on existing engineering experiment room high-speed EMUs head dummy model, collects various model high-speed EMUs head dummy models
Random angles screenshot image data；
S12, collection do the label data of label to the head dummy image data in step S11；The specific steps are that:
A, air drag, the lift, the pressure wave that crosses, tunnel pressure of each headstock head dummy design are collected based on existing developmental achievement
Five kinds of Reeb, aesthetics significant in value type indexs；
B, the high-speed EMUs head dummy image based on collection, in such a way that social user carries out questionnaire ballot,
Existing vehicle head dummy photo is enumerated in questionnaire, is allowed social user to vote in and is thought personally most U.S. head dummy, is collected simultaneously social use
The age at family, gender, occupational information, in this, as head dummy aesthetics label；The aesthetics label is divided into 10 class, and the 1st grade
Secondary aesthetics is worst, and numerical value is described as 0.1；10th class aesthetics highest, numerical value are described as 1.0；
C, the corresponding headstock model of high-speed EMUs head dummy image has been collected in determination, and collects the resistance of air known to the headstock
Power, lift, the pressure wave that crosses, four physical parameters of air pressure wave in tunnel experimental data, four physical parameters are normalized
Processing, as label data；Such as: CH380A head dummy windage is 0.45, input data of the label data as later period sample；
S13, noise filtering is carried out to the head dummy image data and label data being collected into, filters out non-genuine high-speed EMUs
Headstock noise data obtains high-speed EMUs head dummy image sample data collection.
S2, the training data for 8 classification is established according to high-speed EMUs head dummy sample data set
S21, by the resonable degree scoring of headstock design be 1,2,3,4,5,6,7,8 grade, 1 is worst, 8 be it is best,
Form 8 classifiers；
S22, " marking " is carried out to the head dummy image sample data collection being collected into, obtains marking data set, each sample number
According to including importation and mark part；Importation is the head dummy image data of certain model high-speed EMUs, and mark part is
Experiment entirety score after model high-speed EMUs test；
S23, input head dummy image sample data collection and marking data set, generate a series of 8 points according to head dummy and marking
Class category obtains the training dataset comprising head dummy image, 8 classification categories and weight；The specific steps are that:
It is that a series of 8 classifier of score sequence is handled marking mechanism transformation, to high-speed EMUs head dummy image
Each 8 classifier constructs corresponding training data, gives the training data of ordered sequenceWherein, XiIt is defeated
Enter data, the input data includes image data and numeric data, Yi∈ [1, k] is marking grade, and k=8 is total number of grades,
N is the sum of training data；For k-th of 8 classifiers, 8 classification categoriesIndicate the orderly of i-th of image pattern
Sequential labeling yiIt is whether bigger than k, definition:Work as yiWhen > k, otherwise it is 0 that 8 classification classes, which are designated as 1,；
The training data of k-th of 8 classifiers is configured toWherein, xiIt is i-th of image sample for input data
This input data,For the weight of i-th of image pattern, use absolute cost matrix value for
S3, training depth convolutional neural networks model
A sample data of high-speed EMUs head dummy image pattern collection is randomly selected as validation data set, by step S2
Obtained training dataset and verify data while depth convolutional neural networks model is acted on, the depth convolutional neural networks mould
Type is existing classical depth convolutional neural networks model, and specific structure is fitted depth convolution referring to Fig. 5, using training data
Neural network model prevents depth convolutional neural networks mould by image rotation, pattern reversal, random cut using verify data
Type transition fitting, training output depth convolutional neural networks model, so that each output is one 8 classification of 8 classifiers
Category；Depth convolutional neural networks model is finally optimal after gradient optimization algorithm, is instructed by constantly fitting
The depth convolutional neural networks model perfected；
When training output depth convolutional neural networks model, intersection entropy loss, cross entropy loss function are used are as follows:
In formula, ljiIndicate whether the marking rank of j-th of training sample belongs to i grades, ojiIt indicates for j-th of trained sample
The output marking rank of this depth convolutional neural networks model is the probability of i rank.
S4, a sample data for choosing high-speed EMUs head dummy image pattern collection will test number as test data set
It is input in trained depth convolutional neural networks model according to the test sample of concentration, carries out convolution, the pondization operation of multilayer.
When input test sample, head dummy image of the test sample selection without marking label.
S5, " marking " is carried out to test sample, obtains the head dummy score of test sample, auxiliary head dummy design is completed.
In the present embodiment above method, learn the process of segmentation data, the high speed motor car that will be collected into according to standard depth
Group head dummy image sample data collection is divided into 10 parts, randomly selects 1 part as validation data set, 1 part is used as test data set, remaining
8 parts are used as training dataset.For training dataset for training depth convolutional neural networks model, data volume accounts for entire high quick-action
The 80% of vehicle group head dummy image sample data collection.Validation data set is in the training process, verifying depth convolutional neural networks
The training of model judges that current depth convolutional neural networks model prediction accuracy rate, data volume account for entire high speed motor car
The 10% of group head dummy image sample data collection.The data volume of test data set accounts for entire high-speed EMUs head dummy image sample data
The 10% of collection, trained depth convolutional neural networks model is predicted in test data set, to depth convolutional Neural net
Network model carries out total evaluation.High-speed EMUs head dummy image sample data collection is made of several samples, and each sample includes
The input data and output data of depth convolutional neural networks model.Wherein input data includes image data and numeric data,
Image data includes that the image 20 of the pixel 512*512 size of train headstock different angle shooting is opened, and numeric data includes air
5 resistance, lift, the pressure wave that crosses, air pressure wave in tunnel, aesthetics numeric datas.Output data refers to the label manually marked
Value.The input data of each sample can correspond to the label value manually marked, and headstock is obtained through test when the label value
Whole grade, grade classification is 8 ranks, and 1 rank indicates that the headstock test result of design is least ideal, 8 grades of expression correspondences
It is optimal to design headstock.
Referring to fig. 4, the input data of depth convolutional neural networks model includes 20 image datas and 5 kinds of numeric datas,
Predicted value, predicted value o are exported after depth convolutional neural networks modeli, indicate that the final marking of test sample is i-stage
Probability, i ∈ [1,8], wherein8 classification classes are designated as li,YiThe category mark of 8 classification categories
Number,
When there is the demand for designing new high-speed EMUs head dummy, by the three of the threedimensional model of the new head dummy of preliminary type selecting
View investment utilize the trained depth convolutional neural networks model of the above method, obtain the head dummy based on air drag, rise
Power, the pressure wave that crosses, air pressure wave in tunnel four physical parameters and aesthetics score, complete auxiliary head dummy design.
Above-described embodiment is can also be using other more classification methods such as 2 classification based on 8 classification methods, with specific reference to
Depending on practical head dummy design requirement.8 classification methods are based on using cylinder is above-mentioned the step of other more classification methods, it is no longer superfluous herein
It states.
Embodiment provided above only with illustrating the present invention for convenience, and it is not intended to limit the protection scope of the present invention,
Technical solution scope of the present invention, person of ordinary skill in the field make various simple deformations and modification, should all include
In the above claim.
Claims (7)
1. a kind of auxiliary design method of high-speed EMUs head dummy, which is characterized in that the specific steps are that:
S1, high-speed EMUs head dummy image sample data collection is established
S11, the head dummy image data for collecting the various models of high-speed EMUs；
S12, collection do the label data of label to the head dummy image data in step S11；
S13, noise filtering is carried out to the head dummy image data and label data being collected into, filters out non-genuine high-speed EMUs headstock
Noise data obtains high-speed EMUs head dummy image sample data collection；
S2, it is established according to high-speed EMUs head dummy sample data set for polytypic training data
S21, by the resonable degree scoring of headstock design for 1,2 ..., M grade, 1 be it is worst, M is best, formation M classifier；
S22, " marking " is carried out to the head dummy image sample data collection being collected into, obtains marking data set, each sample data packet
Include importation and mark part；Importation is the head dummy image data of certain model high-speed EMUs, and mark part is the type
Experiment entirety score after the test of number high-speed EMUs；
S23, head dummy image sample data collection and marking data set are inputted, a series of M classification class is generated according to head dummy and marking
Mark obtains the training dataset comprising head dummy image, M classification category and weight；
S3, training depth convolutional neural networks model
A sample data of high-speed EMUs head dummy image pattern collection is randomly selected as validation data set, step S2 is obtained
Training dataset and verify data simultaneously act on depth convolutional neural networks model, utilize training data be fitted depth convolution
Neural network model prevents depth convolutional neural networks mould by image rotation, pattern reversal, random cut using verify data
Type transition fitting, training output depth convolutional neural networks model, so that each output is the M classification of M classifier
Category；Depth convolutional neural networks model is finally optimal after gradient optimization algorithm, is instructed by constantly fitting
The depth convolutional neural networks model perfected；
S4, choose high-speed EMUs head dummy image pattern collection a sample data as test data set, by test data set
In test sample be input in trained depth convolutional neural networks model, carry out multilayer convolution, pondization operation；
S5, " marking " is carried out to test sample, obtains the head dummy score of test sample, auxiliary head dummy design is completed.
2. the auxiliary design method of high-speed EMUs head dummy as described in claim 1, which is characterized in that in step S2, according to
Head dummy and marking generate a series of M classification category, obtain the training dataset comprising head dummy image, M classification category and weight
Specific steps are as follows: marking mechanism transformation be score sequence with a series of M classifier processing, to high-speed EMUs head dummy figure
Each M classifier of picture constructs corresponding training data, gives the training data of ordered sequenceWherein, Xi
For input data, the input data includes image data and numeric data, Yi∈ [1, k] is marking grade, and k=M is grade
Sum, N are the sum of training data；For k-th of M classifier, M classification categoryIndicate i-th of image pattern
Ordered sequence label yiIt is whether bigger than k, definition:Work as yiWhen > k, M classification class is designated as 1, otherwise
It is 0；The training data of k-th of M classifier is configured toWherein, xiIt is i-th of figure for input data
Decent input data,For the weight of i-th of image pattern, use absolute cost matrix value for
3. the auxiliary design method of high-speed EMUs head dummy as claimed in claim 2, which is characterized in that in step S3, depth
Convolutional neural networks model training uses intersection entropy loss, cross entropy loss function are as follows:
In formula, ljkIndicate whether the marking rank of j-th of training sample belongs to k grades, ojkIt indicates for j-th of training sample depth
The output marking rank of convolutional neural networks model is the probability of k rank.
4. the auxiliary design method of high-speed EMUs head dummy as claimed in claim 2, which is characterized in that described in step S3
Numeric data includes anti-air drag data, lift data, the pressure wave data that cross, air pressure wave in tunnel data and aesthetics number
According to.
5. the auxiliary design method of high-speed EMUs head dummy as claimed in claim 1 or 2 or 3 or 4, which is characterized in that step
In S11, the method for the collection various model head dummy image datas of high-speed EMUs are as follows:
The head dummy image data of the various models of high-speed EMUs is collected by social networks using web crawlers；
Based on existing engineering experiment room high-speed EMUs head dummy model, the random of various model high-speed EMUs head dummy models is collected
Angle screenshot image data.
6. the auxiliary design method of high-speed EMUs head dummy as claimed in claim 5, which is characterized in that in step S4, input
When test sample, head dummy image of the test sample selection without marking label.
7. the auxiliary design method of high-speed EMUs head dummy as described in claim 1, which is characterized in that in step S12, collect
The specific steps of the label data of label are done to the head dummy image data in step S11 are as follows:
Based on existing developmental achievement collect the air drag of each headstock head dummy design, lift, the pressure wave that crosses, air pressure wave in tunnel,
Five kinds of significant in value type indexs of aesthetics；
High-speed EMUs head dummy image based on collection, in such a way that social user carries out questionnaire ballot, in questionnaire
Existing vehicle head dummy photo is enumerated, allows social user to vote in and thinks personally most U.S. head dummy, be collected simultaneously the year of social user
Age, gender, occupational information, in this, as head dummy aesthetics label；The aesthetics label is divided into 10 class, and the 1st class is beautiful
Property is worst, and numerical value is described as 0.1；10th class aesthetics highest, numerical value are described as 1.0；
The corresponding headstock model of high-speed EMUs head dummy image has been collected in determination, and is collected air drag known to the headstock, risen
Power, the pressure wave that crosses, four physical parameters of air pressure wave in tunnel experimental data, four physical parameters are normalized,
As label data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811184702.4A CN109214116A (en) | 2018-10-11 | 2018-10-11 | A kind of auxiliary design method of high-speed EMUs head dummy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811184702.4A CN109214116A (en) | 2018-10-11 | 2018-10-11 | A kind of auxiliary design method of high-speed EMUs head dummy |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109214116A true CN109214116A (en) | 2019-01-15 |
Family
ID=64979620
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811184702.4A Withdrawn CN109214116A (en) | 2018-10-11 | 2018-10-11 | A kind of auxiliary design method of high-speed EMUs head dummy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109214116A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109948278A (en) * | 2019-03-29 | 2019-06-28 | 山东建筑大学 | Building Design aesthetic measure appraisal procedure and system based on width study |
| CN117994447A (en) * | 2024-04-07 | 2024-05-07 | 大强信息技术(深圳)有限公司 | Auxiliary generation method and system for 3D image of vehicle model design oriented to sheet |
-
2018
- 2018-10-11 CN CN201811184702.4A patent/CN109214116A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109948278A (en) * | 2019-03-29 | 2019-06-28 | 山东建筑大学 | Building Design aesthetic measure appraisal procedure and system based on width study |
| CN117994447A (en) * | 2024-04-07 | 2024-05-07 | 大强信息技术(深圳)有限公司 | Auxiliary generation method and system for 3D image of vehicle model design oriented to sheet |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107341506A (en) | A kind of Image emotional semantic classification method based on the expression of many-sided deep learning | |
| CN106156003B (en) | A kind of question sentence understanding method in question answering system | |
| CN107273490B (en) | Combined wrong question recommendation method based on knowledge graph | |
| CN106203395B (en) | Face attribute recognition method based on multitask deep learning | |
| CN103824054B (en) | A kind of face character recognition methods based on cascade deep neural network | |
| CN103258204B (en) | A kind of automatic micro-expression recognition method based on Gabor and EOH feature | |
| CN106650806A (en) | Cooperative type deep network model method for pedestrian detection | |
| CN103413050B (en) | Mental imagery EEG signals temporal voting strategy sorting technique based on very fast learning machine | |
| CN101447020B (en) | Pornographic image recognizing method based on intuitionistic fuzzy | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN107330446A (en) | A kind of optimization method of depth convolutional neural networks towards image classification | |
| CN105005765A (en) | Facial expression identification method based on Gabor wavelet and gray-level co-occurrence matrix | |
| CN110046248A (en) | Model training method, file classification method and device for text analyzing | |
| CN112446591A (en) | Evaluation system for student comprehensive capacity evaluation and zero sample evaluation method | |
| CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
| CN106203296B (en) | The video actions recognition methods of one attribute auxiliary | |
| CN106529503A (en) | Method for recognizing face emotion by using integrated convolutional neural network | |
| CN102156871A (en) | Image classification method based on category correlated codebook and classifier voting strategy | |
| CN101853250A (en) | Method and device for classifying documents | |
| CN107203752A (en) | A kind of combined depth study and the face identification method of the norm constraint of feature two | |
| CN106022363A (en) | Method for recognizing Chinese characters in natural scene | |
| CN101593273A (en) | A kind of video feeling content identification method based on fuzzy overall evaluation | |
| CN107480141A (en) | It is a kind of that allocating method is aided in based on the software defect of text and developer's liveness | |
| Huang et al. | Research on urban modern architectural art based on artificial intelligence and GIS image recognition system | |
| CN106503616A (en) | A kind of Mental imagery Method of EEG signals classification of the learning machine that transfinited based on layering |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190115 |
|
| WW01 | Invention patent application withdrawn after publication |