CN114692347A - Temperature field proxy model construction method based on neural network architecture search technology - Google Patents
Temperature field proxy model construction method based on neural network architecture search technology Download PDFInfo
- Publication number
- CN114692347A CN114692347A CN202210258055.7A CN202210258055A CN114692347A CN 114692347 A CN114692347 A CN 114692347A CN 202210258055 A CN202210258055 A CN 202210258055A CN 114692347 A CN114692347 A CN 114692347A
- Authority
- CN
- China
- Prior art keywords
- model
- neural network
- temperature field
- sub
- training
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/18—Network design, e.g. design based on topological or interconnect aspects of utility systems, piping, heating ventilation air conditioning [HVAC] or cabling
-
- 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
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/08—Thermal analysis or thermal optimisation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computer Networks & Wireless Communication (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a temperature field proxy model construction method based on a neural network architecture search technology, which comprises the following steps: generating training data according to the component layout design requirement; preprocessing training data; determining a neural network reference model, setting search parameters according to the neural network reference model, and constructing a super network model containing all possible paths; based on the super network model, randomly selecting search parameters to obtain a sub-model, training the sub-model by using training data to update the model parameters of the sub-model, and repeating the randomly selecting sub-model and the training process until the training times reach a preset value; searching an approximate optimal sub-model by utilizing a multi-objective optimization algorithm based on the trained super network model; and retraining the approximate optimal submodel by using the training data to obtain a temperature field agent model. The method can realize automatic design of the neural network structure according to the layout design requirements of the components, and the obtained model has the advantages of less parameters, high prediction precision and short prediction time.
Description
Technical Field
The invention relates to the technical field of component layout optimization design, in particular to a temperature field proxy model construction method based on a neural network architecture search technology.
Background
The satellite component layout design is an important component in the overall satellite design, and needs to consider a plurality of index factor constraints, wherein the temperature field constraint index is one of key indexes and is directly related to the performance and the service life of the satellite. In a high-temperature environment, the reliability of the electronic element is reduced by 10% every 2 degrees of temperature rise, and when the temperature rises to 50 degrees, the service life is 1/6 of 25 degrees. Therefore, the temperature field of the satellite component layout needs to be constructed and analyzed to realize the satellite component layout optimization design considering the temperature field constraint. The traditional component layout temperature field construction depends on high-fidelity simulation software calculation, the calculation time is long, and the component layout design scheme needs simulation calculation once every time the optimization adjustment is carried out, so that the time cost and the calculation cost of the whole design flow are extremely high. Therefore, how to improve the temperature field construction efficiency is an important problem to be solved urgently in the field of current satellite layout design.
Constructing a low-cost temperature field proxy model to predict the temperature field of the component layout instead of high-fidelity thermal simulation analysis is a main solution for improving the construction efficiency of the temperature field at present. For example, a temperature field proxy model is constructed based on a response surface method of a polynomial, a spline function interpolation method, a Kriging function interpolation approximation method and the like. However, the above method can generally process only variables of several tens of dimensions, and it is difficult to process a high-dimensional proxy model. For example, 20 components with the same heating power need to be placed in one two-dimensional layout area, 40 design variables are needed to determine two-dimensional coordinates of the 20 components under the condition of a fixed placing angle, and when a temperature field is calculated, 40000-dimensional variables are needed to describe the whole temperature field if the layout area is divided into 200 × 200 grids, and at this time, the construction of a temperature field proxy model cannot be realized by using the method.
In order to implement the construction of a temperature field proxy model with high dimensional variables, some researches propose to use a deep neural Network, such as a Feature Pyramid Network (FPN), to learn the mapping relationship between a heat source layout and a corresponding temperature field, so as to construct a corresponding temperature field proxy model and improve the efficiency of component layout optimization design.
However, although the efficiency of calculating the satellite component layout temperature field can be improved by using the deep neural network to construct the proxy model, it is still a difficult point for how to select and design a proper neural network model, especially for the problem of component layout optimization design, once the setting of the problem of component layout optimization design is changed, such as different boundary conditions, different heat source strengths, and different shapes of component layouts, the neural network model needs to be designed and adjusted elaborately and repeatedly during training to obtain higher training precision, and the time cost and the calculation cost are greatly increased.
Disclosure of Invention
In order to solve part or all of the technical problems in the prior art, the invention provides a temperature field proxy model construction method based on a neural network architecture search technology.
The technical scheme of the invention is as follows:
a temperature field proxy model building method based on neural network architecture search technology is provided, and the method comprises the following steps:
generating training data according to the component layout design requirement, wherein the training data comprises a component layout and a temperature field corresponding to the component layout;
preprocessing the training data;
determining a neural network reference model, setting search parameters according to the neural network reference model, and constructing a super network model containing all possible paths;
randomly selecting search parameters to obtain a sub-model based on the super network model, training the sub-model by using training data to update model parameters, and repeating the random selection of the sub-model and the training process until the training times reach a preset value to obtain the trained super network model comprising multiple sub-models;
searching an approximate optimal sub-model by utilizing a multi-objective optimization algorithm based on the trained super network model;
and retraining the approximate optimal submodel by using training data to fit the mapping relation between the component layout and the temperature field, and acquiring a temperature field proxy model.
In some possible implementations, the generating training data may include:
grid division is carried out on the component layout area, and a component layout X is obtained by randomly selecting a corresponding number of grid placement components according to the number of the componentsmSimulating and calculating the temperature field T corresponding to the component layout by using a finite element analysis methodmObtaining a training data (X) comprising the component layout and the temperature field corresponding to the component layoutm,Tm) Repeating the random generation process for multiple times until a preset number M of training data is obtained, and obtaining a training data set { (X) comprising M training datam,Tm)|m＝1,2,…,M}。
In some possible implementations, the preprocessing the training data includes:
partitioning a layout area of a satellite component layout into M1×M2A grid for laying out the satellite components by M1×M2In matrix representation, the matrix element corresponding to the grid position with the component is the component power, and the matrix element corresponding to the grid position without the component is 0.
In some possible implementation manners, a MobileNetV2 model with a preset number of layers is adopted as a neural network model reference model.
In some possible implementations, the search parameter includes: the spreading factor set for each layer of the MobileNetV2 model, the number of convolution paths set for each layer of the MobileNetV2 model, and the convolution kernel size for each convolution in each layer of the MobileNetV2 model.
In some possible implementations, the expansion rate of each layer of the neural network of the MobileNetV2 model is selected from [3,6], the number of convolution paths is selected from [1,2,3,4], and after the number of convolution paths is determined, the convolution kernel of each convolution is selected from [3,5,7,9] without repetition.
In some possible implementations, the building the super network model including all possible paths includes:
determining a network structure of a MobileNet V2 model containing all possible paths according to set search parameters, dividing the network structure of the MobileNet V2 model into a plurality of parts according to the number of layers of the MobileNet V2 model, replacing a main structure ResNet model of a feature pyramid network with the network structure of the divided MobileNet V2 model to obtain an improved feature pyramid network as a super network model containing all possible paths, wherein the plurality of parts correspond to a left side structure of the feature pyramid network taking ResNet50 as the main structure, the resolution of each part is sequentially reduced by half, and different parts extract image features with different resolutions.
In some possible implementation manners, the randomly selecting a search parameter based on the super network model to obtain a sub-model, training the sub-model by using training data to update the model parameter, and repeating the randomly selecting the sub-model and the training process until the number of training times reaches a preset value includes:
step S41, setting a total training frequency E, a total layer number L of the neural network reference model, and setting the current training frequency E as 1;
step S42, setting the current neural network structure layer number l to 1, and randomly selecting a numerical value from the set expansion rate range as the expansion rate of the first layer of the neural network reference model;
step S43, randomly selecting a numerical value from the set convolution path number range as the convolution path number of the I layer of the neural network reference model;
step S44, selecting a corresponding number of convolution kernels of different sizes from the set range of convolution kernel sizes according to the number of convolution paths determined in step S43;
step S45, determining whether L is less than L, if so, making L equal to L +1, and returning to step S43, otherwise, performing step S46;
step S46, configuring a neural network reference model according to the obtained expansion rate, the number of convolution paths and the size of a convolution kernel to obtain a sub-model of the super network model;
step S47, training the sub-model by using the training data to update the model parameters of the sub-model;
in step S48, it is determined whether E is smaller than E, if so, E is made to be E +1, and the process returns to step S42.
In some possible implementations, the searching for the approximately optimal sub-model by using the multi-objective optimization algorithm based on the trained super network model includes:
step S51, generating a parent population P with n individuals, wherein the individuals are sub models obtained by randomly sampling the expansion rate, the number of convolution paths and the size of a convolution kernel on each layer of the neural network reference model;
step S52, setting the current iteration number T to 0 and the maximum iteration number T;
step S53, setting the current individual count j equal to 0, and establishing a child population set Q equal to Φ, where Φ represents an empty set;
step S54, two individuals P in P are randomly selectedmAnd pnPerforming crossover operation to generate new individual qj+1Wherein the crossover operation is represented at pmRandomly selecting a layer of neural network structure in the network to replace pnA layer of neural network structure is randomly selected;
step S55, randomly selecting an individual P in PkPerforming mutation operation to generate new individual qj+2Q is mixing qj+1And q isj+2Adding a set Q of offspring populations, wherein the crossover operation represents a random choice of pkIn the first layer of neural network structure, randomly selecting the expansion rate, the convolution path number and the convolution kernel size again for configuration;
step S56, changing j to j +2, determining whether j is smaller than n-1, if yes, returning to step S54, if no, performing step S57;
step S57, merging the parent population and the child population to obtain a population R, where R ═ P ═ u Q;
step S58, calculating the prediction precision and parameter quantity of each individual in the population R;
step S59, calculating a pareto front F by using a rapid non-dominated sorting algorithm;
step S510, calculating the crowding degrees of all individuals in the population R, and selecting n individuals from the population R according to the crowding degree distance sorting to obtain a new parent population P;
and step S511, making T equal to T +1, determining whether T is less than T, if so, returning to step S53, and if not, taking all individuals in the currently obtained pareto frontier F as the searched approximately optimal submodel.
In some possible implementations, when the sub-model is trained using the training data, an average absolute error between the predicted value of the temperature field and the true value of the temperature field is used as a loss function.
The technical scheme of the invention has the following main advantages:
according to the temperature field agent model construction method based on the neural network architecture search technology, the super network model containing all possible paths is constructed and trained, the near-optimal sub-model is searched by utilizing a multi-objective algorithm based on the trained super network model, the automatic design of the neural network structure can be realized according to different component layout design requirements, the obtained temperature field agent model has less parameter quantity, the model has higher prediction precision, and the time required by model prediction is shorter.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flowchart of a temperature field proxy model construction method based on neural network architecture search technology according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of the MobileNet V2 model according to an embodiment of the present invention;
FIG. 3 is a schematic structural diagram of a feature pyramid network according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a component layout according to an embodiment of the present invention;
FIG. 5 is a diagram illustrating 10 iterative optimization results of the multi-objective optimization algorithm according to an embodiment of the present invention;
fig. 6 is a schematic diagram of a backbone structure of a selected super network model according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the specific embodiments of the present invention and the accompanying drawings. It is to be understood that the described embodiments are merely a few embodiments of the invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
The technical scheme provided by the embodiment of the invention is explained in detail in the following with the accompanying drawings.
Referring to fig. 1, an embodiment of the present invention provides a temperature field proxy model building method based on a neural network architecture search technology, including the following steps:
step S1, generating training data according to the design requirements of the component layout, wherein the training data comprises the component layout and a temperature field corresponding to the component layout;
step S2, preprocessing the training data;
step S3, determining a neural network reference model, setting search parameters according to the neural network reference model, and constructing a super network model containing all possible paths;
step S4, based on the super network model, randomly selecting search parameters to obtain a sub-model, training the sub-model by using training data to update the model parameters, and repeating the randomly selecting sub-model and the training process until the training times reach a preset value to obtain the trained super network model comprising multiple sub-models;
step S5, searching an approximate optimal sub model by using a multi-objective optimization algorithm based on the trained super network model;
and step S6, retraining the approximate optimal sub-model by using the training data to fit the mapping relation between the component layout and the temperature field, and acquiring the temperature field proxy model.
According to the temperature field proxy model construction method based on the neural network architecture search technology, the super network model containing all possible paths is constructed and trained, the near-optimal sub-model is searched by using the multi-objective algorithm based on the trained super network model, automatic design of the neural network structure can be achieved according to different component layout design requirements, the obtained temperature field proxy model has fewer parameters, the model is high in prediction precision, and the time required by model prediction is shorter.
In the case of the component layout optimization design problem considering the temperature field index, each component can be simplified to be a heat source. The component layout optimization design may be regarded as a heat source layout optimization design. When a certain number of components need to be placed in a certain layout area, the purpose of the component layout optimization design is to enhance the heat conduction efficiency of the layout area by optimizing the positions of the components to minimize the maximum temperature of the layout area. Considering the heat source layout optimization design problem in a two-dimensional plane, the mathematical model can be expressed as:
wherein X represents the layout scheme of the heat source, (X)i,yi) Position coordinates of the ith heat source, NsDenotes the number of heat sources, T denotes the temperature field of the layout area, k denotes the thermal conductivity of the layout area, (x, y) denotes the position coordinates of any point of the layout area, [ phi ] (x, y) denotes the intensity distribution function of the heat sources, T0Representing the temperature value at the isothermal boundary, n representing the edgeThe normal direction at the boundary, h, represents the surface heat transfer coefficient between the object and the surrounding fluid at the boundary.
The intensity distribution function phi (x, y) of the heat source is determined by the location of the heat source, and is specifically expressed as:
wherein phiiRepresenting the intensity of the ith heat source, ΓiIndicating the layout area covered by the heat source.
The following specifically describes the steps and principles of the method for constructing a temperature field proxy model based on a neural network architecture search technique according to an embodiment of the present invention, taking optimization design of component layout in a two-dimensional plane as an example:
and step S1, generating training data according to the component layout design requirement.
Setting component layout design requirements includes: in a square layout area, N is required to be placedsAnd (3) a component.
Generating training data based on the set component layout design requirements may include:
grid division is carried out on the component layout area, and a component layout X is obtained by randomly selecting a corresponding number of grid placement components according to the number of the componentsmSimulating and calculating the temperature field T corresponding to the component layout by using a finite element analysis methodmObtaining a training data (X) comprising the component layout and the temperature field corresponding to the component layoutm,Tm) Repeating the random generation process for multiple times until a preset number M of training data is obtained, and obtaining a training data set { (X) comprising M training datam,Tm)|m＝1,2,…,M}。
The specific amount of training data can be determined according to the actually required model precision and the component layout optimization time. For example, the number of training data may be 20000, 30000, 40000, 50000, or the like.
Step S2, training data is preprocessed.
Specifically, the preprocessing the training data based on the acquired training data includes:
partitioning a layout area of a satellite component layout into M1×M2A grid for laying out the satellite components by M1×M2In matrix representation, the matrix element corresponding to the grid position with the component is the component power, and the matrix element corresponding to the grid position without the component is 0.
And step S3, determining a neural network reference model, setting search parameters according to the neural network reference model, and constructing a super network model containing all possible paths.
In an embodiment of the invention, a MobileNetV2 model with a preset number of layers is used as a neural network model reference model. The MobileNetV2 model is a small and efficient deep learning model, and is formed by stacking a plurality of inverted residual blocks, and due to the design of the inverted residual blocks, the whole network is small, but has strong ability to extract features of input images.
Referring to fig. 2, each layer of the MobileNetV2 model includes: an extension layer (extension layer), a convolution layer (convolution layer), and a Projection layer (Projection layer). The extension layer extends the channel dimension of the feature map by a set extension rate (Expansion rate), the convolution layer extracts the feature by a set convolution kernel (conv) of different sizes, and the projection layer reduces the extended feature map channel dimension to a predefined good channel dimension by using a 1x1 convolution kernel (1x1 conv).
Further, based on the determined neural network reference model, the searching for parameters may include: the spreading factor set for each layer of the MobileNetV2 model, the number of convolution paths set for each layer of the MobileNetV2 model, and the convolution kernel size for each convolution in each layer of the MobileNetV2 model.
In an embodiment of the present invention, when constructing the MobileNetV2 model, the expansion rate of each layer of the neural network of the MobileNetV2 model may be selected from [3,6], the number of convolution paths may be selected from [1,2,3,4], and after the number of convolution paths is determined, the size of convolution kernel of each convolution may be selected from [3,5,7,9] without repetition, that is, according to the determined number of convolution paths, the size of convolution kernel of a corresponding number is selected from 3x3conv, 5x5conv, 7x7conv, and 9x9conv without repetition to configure each convolution path.
Based on the search parameters set above, each layer of the MobileNetV2 model includes two possible extension paths and four possible convolution paths.
Further, constructing a super network model including all possible paths based on the specifically set neural network reference model and the search parameters includes:
the method comprises the steps of determining a network structure of a MobileNet V2 model containing all possible paths according to set search parameters, dividing the network structure of the MobileNet V2 model into a plurality of parts according to the number of layers of the MobileNet V2 model, replacing a main structure ResNet model of a feature pyramid network with the divided network structure of the MobileNet V2 model to obtain an improved feature pyramid network serving as a super network model containing all the possible paths, wherein the plurality of parts correspond to the left side structure of the feature pyramid network with ResNet50 as the main structure, the resolution of each part is reduced by half in sequence, and image features with different resolutions are extracted by different parts.
Specifically, referring to fig. 3, C2, C3, C4 and C5 in the feature pyramid network are replaced with the determined MobileNetV2 model containing all possible paths, so as to obtain a super network model containing all possible paths. In order to ensure that the MobileNetV2 model can be divided into 4 parts, the number of layers of the MobileNetV2 model can be set to be an integral multiple of 4, for example, 12 layers.
By combining a network structure containing the MobileNetV2 model of all possible paths with a feature pyramid network architecture, the network structure can be adapted for mapping learning between component layout and temperature field.
And step S4, based on the super network model, randomly selecting search parameters to obtain a sub-model, training the sub-model by using training data to update the model parameters, and repeating the randomly selecting sub-model and the training process until the training times reach a preset value to obtain the trained super network model comprising a plurality of sub-models.
Specifically, on the basis of the set search parameters, based on the super network model, randomly selecting the search parameters to obtain a sub-model, training the sub-model by using training data to update the model parameters, and repeating the randomly selecting the sub-model and the training process until the training times reach a preset value, the method comprises the following steps:
step S41, setting a total training frequency E, a total layer number L of the neural network reference model, and setting the current training frequency E as 1;
step S42, setting the current neural network structure layer number l to 1, and randomly selecting a numerical value from the set expansion rate range as the expansion rate of the first layer of the neural network reference model;
step S43, randomly selecting a numerical value from the set convolution path number range as the convolution path number of the I layer of the neural network reference model;
step S44, selecting a corresponding number of convolution kernels of different sizes from the set range of convolution kernel sizes according to the number of convolution paths determined in step S43;
step S45, determining whether L is less than L, if so, making L equal to L +1, and returning to step S43, otherwise, performing step S46;
step S46, configuring a neural network reference model according to the obtained expansion rate, the number of convolution paths and the size of a convolution kernel to obtain a sub-model of the super network model;
step S47, training the sub-model by using the training data to update the model parameters of the sub-model;
in step S48, it is determined whether E is smaller than E, if so, E is made equal to E +1, and the process returns to step S42.
Based on the set steps, each path of the super network model is randomly activated in each training period, and when training is terminated, even if the obtained super network model cannot achieve the precision directly used for predicting the temperature field, the super network model has certain model sequencing capability, namely, a proxy model capable of evaluating the precision of different models is equivalently provided.
And step S5, searching an approximate optimal sub-model by using a multi-objective optimization algorithm based on the trained super network model.
Based on the trained super network model, an approximately optimal model architecture can be searched by utilizing a multi-objective optimization algorithm to obtain an approximately optimal submodel, and the approximately optimal submodel can meet the required temperature field prediction precision requirement by retraining the approximately optimal submodel obtained by searching for a sufficient period.
Specifically, according to the setting, based on the trained super network model, the multi-objective optimization algorithm is utilized to search for the approximate optimal sub-model, and the method comprises the following steps:
step S51, generating a parent population P with n individuals, wherein the individuals are sub models obtained by randomly sampling the expansion rate, the number of convolution paths and the size of a convolution kernel on each layer of the neural network reference model;
step S52, setting the current iteration count T equal to 0 and the maximum iteration count T;
step S53, setting the current individual count j equal to 0, and establishing a child population set Q equal to Φ, where Φ represents an empty set;
step S54, two individuals P in P are randomly selectedmAnd pnPerforming crossover operation to generate new individual qj+1Wherein the crossover operation is represented at pmRandomly selecting a layer of neural network structure in the network to replace pnA layer of neural network structure is randomly selected;
step S55, randomly selecting an individual P in PkPerforming mutation operation to generate new individual qj+2Q is prepared byj+1And q isj+2Adding a set Q of offspring populations, wherein the crossover operation represents a random choice of pkIn the first layer of neural network structure, randomly selecting the expansion rate, the number of convolution paths and the size of a convolution kernel again for configuration;
step S56, let j equal j +2, determine whether j is less than n-1, if yes, return to step S54, if no, go to step S57;
step S57, merging the parent population and the child population to obtain a population R, where R ═ P ═ u Q;
step S58, calculating the prediction precision and parameter quantity of each individual in the population R;
step S59, calculating a pareto front F by using a rapid non-dominated sorting algorithm;
step S510, calculating the crowding degrees of all individuals in the population R, and selecting n individuals from the population R according to the crowding degree distance sorting to obtain a new parent population P;
and step S511, making T equal to T +1, determining whether T is less than T, if so, returning to step S53, and if not, taking all individuals in the currently obtained pareto frontier F as the searched approximately optimal submodel.
In an embodiment of the present invention, the obtained individuals in the pareto frontier can be regarded as an approximately optimal model architecture, that is, an approximately optimal submodel, and can meet the requirements of rapid deployment or high prediction accuracy.
In step S510, n individuals are selected from R in descending order of the congestion degree distance to obtain a new parent population P including the selected n individuals.
And step S6, retraining the approximate optimal sub-model by using the training data to fit the mapping relation between the component layout and the temperature field, and acquiring the temperature field proxy model.
Specifically, after the approximately optimal submodel is determined, namely the model architecture of the super network model is determined, the approximately optimal submodel is retrained by using the preprocessed training data to fit the mapping relation between the component layout and the temperature field, and the temperature field proxy model is obtained.
In an embodiment of the present invention, when the model is trained and updated by using the training data, an average Absolute Error (MAE) between a predicted value of the temperature field and a true value of the temperature field may be used as a loss function.
In particular, to divide the component layout into M1×M2For example, the average absolute error between the predicted temperature field value and the true temperature field value can be expressed as:
wherein, MAE (Y, Y)*) Showing the predicted value Y and the true value Y of the temperature field*The average absolute error between the two is calculated,
denotes the m-th1Line m2The predicted value of the temperature field at the column grid,
denotes the m-th1Line m2The true value of the temperature field at the column grid,
indicating the predicted value of the temperature field
True value of temperature field
Absolute error between.
And carrying out iterative updating on the model parameters in a back propagation mode according to the determined loss function until the set iteration times are reached, and then stopping to obtain the final temperature field proxy model.
The iteration number during training can be set according to the prediction precision and the training time which are actually required, and generally, the more the iteration number is, the higher the prediction precision of the finally obtained model is, and the longer the required training time is. For example, the number of iterations may be set to 50, 60, etc.
The following explains beneficial effects of the temperature field proxy model construction method based on the neural network architecture search technology according to an embodiment of the present invention with reference to specific examples.
Referring to FIG. 4, the volume-to-point (VP) thermal conduction problem in a two-dimensional square area is taken as an example, i.e. in a square layout areaArrangement NsEach having the same power phi0The module of (2), the border department of this square overall arrangement region is provided with a louvre, and louvre regional temperature is fixed, and other borders of outside except louvre region are adiabatic.
Generating 50000 training data and 5000 test data according to the set component layout design requirement; in the structure searching process of the super network model, the training times of the super network model are set to be 600, the learning rate is set to be 0.001, and the batch processing size of input data is set to be 32; in the process of searching the approximate optimal sub-model by using the multi-objective optimization algorithm, the size of the population is set to be 40; in the process of training the approximate optimal submodel, retraining the neural network from the beginning, and setting the training times to be 50; the number of channels of each layer of the whole super network model after passing through the projection layer is set to be [32,48,48,96,96,96,192, 256,320 ]. And constructing a temperature field proxy model based on the specifically set parameters.
Referring to fig. 5 and 6, fig. 5 shows an optimization result obtained after the multi-objective optimization algorithm iterates 10 times according to an embodiment of the present invention, and fig. 6 shows a backbone structure of the selected super network model. In an embodiment of the invention, one submodel is selected in a compromise mode to test based on a non-dominated solution obtained by iterating for 10 times.
In order to evaluate the effect of the temperature field proxy model constructed by the method provided by the embodiment of the invention and the existing manual design model, 6 main flow models in image segmentation are selected for comparison, and the performance of the network model searched by the example and other models is tested on four performance indexes by using test data.
Wherein, 6 mainstream models include: legacy FPN, ResNet18_ FPN, DetNet _ FPN, MobileNet _ FPN, legacy Unet, and ResNet34_ Unet. Since the encoder-to-decoder structure model is suitable for the image-to-image regression task, FPN, Unet and its variants were chosen for comparison. The MobileNet is a series of small models on the mobile terminal, can effectively classify images, and is combined into an FPN frame as a backbone structure to obtain MobileNet _ FPN. Other models (such as ResNet18, DetNet, ResNet34) are also incorporated into the FPN or Unet framework as a backbone to obtain a corresponding neural network model. The number of channels of Unet is [64,128,256,512,1024 ].
Wherein, four performance indicators include: precision, parameter number, number of Floating point operations (FLOPs), and inference time. The accuracy is expressed by MAE, and the lower the MAE is, the higher the model prediction accuracy is; the parameter quantity represents the total quantity of weight parameters of the neural network model, and the smaller the total parameter quantity is, the lower the training cost of the neural network model is; the number of floating point operations may evaluate the temporal complexity of the neural network model; the inference time represents the average time to perform the forward computation of the neural network model on a single input layout image.
In an embodiment of the present invention, the performance of each model is tested on the same test data under the same computing environment to obtain the performance results of different neural network models as shown in table 1, where the computing environment is NVIDIA GTX3060Ti GPU.
TABLE 1
| Model (model) | Accuracy of measurement | Amount of ginseng | FLOPs | Inferring time |
| FPN | 0.108K | 26M | 4.97G | 0.0228s |
| ResNet18_FPN | 0.126K | 13.04M | 2.77G | 0.0203s |
| DetNet_FPN | 0.208K | 20.68M | 1.67G | 0.0658s |
| MobileNet_FPN | 3.801K | 3.11M | 1.10G | 0.0807s |
| Unet | 0.123K | 22.93M | 14.56G | 0.0173s |
| ResNet34_Unet | 0.104K | 103.08M | 21.05G | 0.0828s |
| Model of the invention | 0.105K | 6.57M | 1.81G | 0.0124s |
It can be seen that with similar prediction accuracy, the model obtained by an embodiment of the present invention is only 1/4 times the total parameter size of the FPN, the conventional FPN has 26M parameters, and the model obtained by an embodiment of the present invention is only 6.57M. While the parameter of the MobileNet _ FPN is only 3.11M, the MAE reaches 3.801K, which is much lower than that of other models. In addition, compared with the traditional FPN, the model obtained by the embodiment of the invention reduces FLOPs from 4.97G to 1.81G, has lower training cost, reduces inference time from 0.0228s to 0.0124s, and has lower calculation cost and higher calculation efficiency.
Further, in order to evaluate the generality of the model searched in the embodiment of the present invention and the conventional FPN, in the embodiment of the present invention, a test sample is randomly extracted from the test data, the model searched in the embodiment of the present invention and the conventional FPN are predicted separately, and the average absolute error and the maximum absolute error shown in table 2 are calculated according to the predicted temperature field.
TABLE 2
| Model (model) | Mean absolute error | Maximum absolute error |
| FPN | 0.1754K | 1.82K |
| Model of the invention | 0.0998K | 0.86K |
It can be seen that, compared with the conventional FPN, the model obtained by searching according to an embodiment of the present invention has a smaller average absolute error and a smaller maximum absolute error, and has a better prediction effect.
It is noted that, in this document, relational terms such as "first" and "second," and the like, may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, "front", "rear", "left", "right", "upper" and "lower" in this document are all referred to the placement state shown in the drawings.
Finally, it should be noted that: the above examples are only for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
Claims (10)
1. A temperature field proxy model construction method based on a neural network architecture search technology is characterized by comprising the following steps:
generating training data according to the component layout design requirement, wherein the training data comprises a component layout and a temperature field corresponding to the component layout;
preprocessing the training data;
determining a neural network reference model, setting search parameters according to the neural network reference model, and constructing a super network model containing all possible paths;
randomly selecting search parameters to obtain a sub-model based on the super network model, training the sub-model by using training data to update model parameters, and repeating the random selection of the sub-model and the training process until the training times reach a preset value to obtain the trained super network model comprising multiple sub-models;
searching an approximate optimal sub-model by utilizing a multi-objective optimization algorithm based on the trained super network model;
and retraining the approximate optimal submodel by using training data to fit the mapping relation between the component layout and the temperature field, and acquiring a temperature field proxy model.
2. The method for constructing the temperature field proxy model based on the neural network architecture search technology according to claim 1, wherein the generating training data includes:
grid division is carried out on the component layout area, and a component layout X is obtained by randomly selecting a corresponding number of grid placement components according to the number of the componentsmSimulating and calculating the temperature field T corresponding to the component layout by using a finite element analysis methodmObtaining a training data (X) comprising the component layout and the temperature field corresponding to the component layoutm,Tm) Repeating the random generation process for multiple times until a preset number M of training data is obtained, and obtaining a training data set { (X) comprising M training datam,Tm)|m＝1,2,…,M}。
3. The method for constructing the temperature field proxy model based on the neural network architecture search technology according to claim 2, wherein the preprocessing the training data comprises:
partitioning a layout area of a satellite component layout into M1×M2A grid for laying out the satellite components by M1×M2Matrix representation, with componentsThe matrix element corresponding to the grid position is the component power, and the matrix element corresponding to the grid position without the component is 0.
4. The temperature field proxy model construction method based on neural network architecture search technology according to any one of claims 1 to 3, characterized in that a MobileNet V2 model with a preset number of layers is adopted as a neural network model reference model.
5. The method for constructing the temperature field proxy model based on the neural network architecture search technology according to claim 4, wherein the search parameters comprise: the spreading factor set for each layer of the MobileNetV2 model, the number of convolution paths set for each layer of the MobileNetV2 model, and the convolution kernel size for each convolution in each layer of the MobileNetV2 model.
6. The temperature field proxy model construction method based on the neural network architecture search technology as claimed in claim 5, wherein the expansion rate of each layer of the neural network of the MobileNet V2 model is selected from [3,6], the number of convolution paths is selected from [1,2,3,4], and after the number of convolution paths is determined, the convolution kernel of each convolution is selected from [3,5,7,9] without repeated selection.
7. The method for constructing the temperature field proxy model based on the neural network architecture search technology according to claim 5, wherein the constructing the super network model containing all possible paths comprises:
determining a network structure of a MobileNet V2 model containing all possible paths according to set search parameters, dividing the network structure of the MobileNet V2 model into a plurality of parts according to the number of layers of the MobileNet V2 model, replacing a main structure ResNet model of a feature pyramid network with the network structure of the divided MobileNet V2 model to obtain an improved feature pyramid network as a super network model containing all possible paths, wherein the plurality of parts correspond to a left side structure of the feature pyramid network taking ResNet50 as the main structure, the resolution of each part is sequentially reduced by half, and different parts extract image features with different resolutions.
8. The temperature field agent model construction method based on neural network architecture search technology as claimed in any one of claims 5 to 7, wherein the step of randomly selecting search parameters to obtain a sub-model based on the super network model, training the sub-model with training data to update model parameters thereof, and repeating the randomly selecting sub-model and training process until the number of training times reaches a preset value comprises:
step S41, setting a total training frequency E, a total layer number L of the neural network reference model, and setting the current training frequency E as 1;
step S42, setting the current neural network structure layer number l to 1, and randomly selecting a numerical value from the set expansion rate range as the expansion rate of the first layer of the neural network reference model;
step S43, randomly selecting a numerical value from the range of the number of the set convolution paths as the number of the convolution paths of the I layer of the neural network reference model;
step S44, selecting a corresponding number of convolution kernels of different sizes from the set range of convolution kernel sizes according to the number of convolution paths determined in step S43;
step S45, determining whether L is less than L, if so, making L equal to L +1, and returning to step S43, otherwise, performing step S46;
step S46, configuring a neural network reference model according to the obtained expansion rate, the number of convolution paths and the size of a convolution kernel to obtain a sub-model of the super network model;
step S47, training the sub-model by using the training data to update the model parameters of the sub-model;
in step S48, it is determined whether E is smaller than E, if so, E is made to be E +1, and the process returns to step S42.
9. The temperature field agent model construction method based on neural network architecture search technology as claimed in claim 8, wherein the searching for the approximately optimal sub-model using the multi-objective optimization algorithm based on the trained super network model comprises:
step S51, generating a parent population P with n individuals, wherein the individuals are sub models obtained by randomly sampling the expansion rate, the number of convolution paths and the size of a convolution kernel on each layer of the neural network reference model;
step S52, setting the current iteration number T to 0 and the maximum iteration number T;
step S53, setting the current individual count j equal to 0, and establishing a child population set Q equal to Φ, where Φ represents an empty set;
step S54, two individuals P in P are randomly selectedmAnd pnPerforming crossover operation to generate new individual qj+1Wherein the crossover operation is represented at pmRandomly selecting a layer of neural network structure in the network to replace pnA layer of neural network structure is randomly selected;
step S55, randomly selecting an individual P in PkPerforming mutation operation to generate new individual qj+2Q is prepared byj+1And q isj+2Adding a set Q of offspring populations, wherein the crossover operation represents a random choice of pkIn the first layer of neural network structure, randomly selecting the expansion rate, the convolution path number and the convolution kernel size again for configuration;
step S56, let j equal j +2, determine whether j is less than n-1, if yes, return to step S54, if no, go to step S57;
step S57, merging the parent population and the child population to obtain a population R, where R ═ P ═ u Q;
step S58, calculating the prediction precision and parameter quantity of each individual in the population R;
step S59, calculating a pareto front F by using a rapid non-dominated sorting algorithm;
step S510, calculating the crowding degrees of all individuals in the population R, and selecting n individuals from the population R according to the crowding degree distance sorting to obtain a new parent population P;
and step S511, making T equal to T +1, determining whether T is less than T, if so, returning to step S53, and if not, taking all individuals in the currently obtained pareto frontier F as the searched approximately optimal submodel.
10. The temperature field proxy model construction method based on neural network architecture search technology according to any one of claims 1 to 9, characterized in that, when training the sub-model with training data, the average absolute error between the predicted value of the temperature field and the true value of the temperature field is used as a loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210258055.7A CN114692347A (en) | 2022-03-16 | 2022-03-16 | Temperature field proxy model construction method based on neural network architecture search technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210258055.7A CN114692347A (en) | 2022-03-16 | 2022-03-16 | Temperature field proxy model construction method based on neural network architecture search technology |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114692347A true CN114692347A (en) | 2022-07-01 |
Family
ID=82139612
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210258055.7A Pending CN114692347A (en) | 2022-03-16 | 2022-03-16 | Temperature field proxy model construction method based on neural network architecture search technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114692347A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115329710A (en) * | 2022-08-26 | 2022-11-11 | 苏州大学 | Heat pipe constraint component layout optimization method based on genetic algorithm |
| CN115730509A (en) * | 2022-09-20 | 2023-03-03 | 中国人民解放军军事科学院国防科技创新研究院 | Spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning |
-
2022
- 2022-03-16 CN CN202210258055.7A patent/CN114692347A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115329710A (en) * | 2022-08-26 | 2022-11-11 | 苏州大学 | Heat pipe constraint component layout optimization method based on genetic algorithm |
| CN115329710B (en) * | 2022-08-26 | 2023-08-29 | 苏州大学 | Heat pipe constraint assembly layout optimization method based on genetic algorithm |
| CN115730509A (en) * | 2022-09-20 | 2023-03-03 | 中国人民解放军军事科学院国防科技创新研究院 | Spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114692347A (en) | Temperature field proxy model construction method based on neural network architecture search technology | |
| Kollat et al. | A new epsilon-dominance hierarchical Bayesian optimization algorithm for large multiobjective monitoring network design problems | |
| WO2020117991A1 (en) | Generating integrated circuit floorplans using neural networks | |
| CN115315703A (en) | Generating an integrated circuit layout using a neural network | |
| Tencaliec et al. | Flexible semiparametric generalized Pareto modeling of the entire range of rainfall amount | |
| Ryberg et al. | Metamodel-based multidisciplinary design optimization for automotive applications | |
| CN111639787A (en) | Spatio-temporal data prediction method based on graph convolution network | |
| CN111507032A (en) | Deep learning technology-based component layout optimization design method for predicting temperature distribution | |
| CN112257321B (en) | Unsupervised learning-based satellite component temperature field determination method | |
| CN108540988B (en) | Scene division method and device | |
| CN105893669A (en) | Global simulation performance predication method based on data digging | |
| Ling et al. | Learning-based superresolution land cover mapping | |
| Veerbeek et al. | Enhancing the calibration of an urban growth model using a memetic algorithm | |
| CN113177290A (en) | Satellite component temperature field prediction method based on depth proxy model normalization | |
| JP2023510922A (en) | Optimizing High-Cost Functions for Complex Multidimensional Constraints | |
| CN110991621A (en) | Method for searching convolutional neural network based on channel number | |
| Xie et al. | Efficient mesh motion using radial basis functions with volume grid points reduction algorithm | |
| CN117077771B (en) | Causal association driven aviation traffic network sweep effect prediction method | |
| CN114611350A (en) | FCN-based phase change heat storage device flow channel topology optimization design method | |
| CN114781207B (en) | Heat source layout temperature field prediction method based on uncertainty and semi-supervised learning | |
| Brierley et al. | Bayesian estimation of migration flows | |
| Zareifard et al. | A heterogeneous Bayesian regression model for skewed spatial data | |
| CN116415482A (en) | Wing flow field analysis method based on graph neural network MeshGraphNet | |
| CN114548526A (en) | Satellite component layout temperature field prediction method based on physical prior neural network | |
| CN113505929A (en) | Topological optimal structure prediction method based on embedded physical constraint deep learning technology |
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 |