CN108959728B - Radio frequency device parameter optimization method based on deep learning - Google Patents
Radio frequency device parameter optimization method based on deep learning Download PDFInfo
- Publication number
- CN108959728B CN108959728B CN201810602897.3A CN201810602897A CN108959728B CN 108959728 B CN108959728 B CN 108959728B CN 201810602897 A CN201810602897 A CN 201810602897A CN 108959728 B CN108959728 B CN 108959728B
- Authority
- CN
- China
- Prior art keywords
- neural network
- parameter
- value
- radio frequency
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computer Hardware Design (AREA)
- Physiology (AREA)
- Evolutionary Biology (AREA)
- Geometry (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Monitoring And Testing Of Transmission In General (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a radio frequency device parameter optimization method based on deep learning. Firstly, tensorflow is used for training, geometric parameters and signal frequency of the radio frequency device are extracted from a data set and used as input items, and S parameter information corresponding to the geometric parameters and the signal frequency is used as a true value. Initializing and segmenting original data, constructing a multilayer neural network, sending the segmented data into the neural network, training by using an optimization algorithm, and training to finish the storage model. And then, optimizing by using a genetic algorithm, inputting the range of the parameters to be optimized, determining an optimization target value, and setting an allowable error. Initializing a population, calling a Tensorflow model to predict the population, calculating the fitness, and continuously selecting, crossing and varying until an optimal optimization result is selected.
Description
Technical Field
The invention belongs to the technical field of microwave radio frequency engineering, and relates to a radio frequency device parameter optimization method based on deep learning.
Background
With the rapid development of wireless communication technology, the information society enters the age of 5G communication and cloud computing, and the market demand of microwave radio frequency devices at present and in future will continue to grow rapidly. With the rapid development of the artificial intelligence algorithm, the deep neural network has the capability of simulating highly complex nonlinear mapping, and the deep learning framework can be conveniently used for designing and packaging interfaces, so that the performance and the efficiency are very high.
For the parameter optimization design of radio frequency devices in microwave radio frequency engineering, general engineering software carries out simulation calculation of parameter scanning. The electromagnetic simulation algorithm has the problems of large algorithm calculation amount and time consumption. And the parameter scanning can only match the parameters on the scanning interval, and the accuracy is low. Some studies use genetic optimization algorithms, but still rely on simulation algorithms or tools, which are very time consuming if the genetic population is too large. The radio frequency device parameter optimization method based on deep learning integrates deep learning rapid prediction and genetic algorithm global optimization, and solves the problems.
Disclosure of Invention
The invention aims to provide a radio frequency device parameter optimization method based on deep learning aiming at the defects of the prior art, and the method is fast in calculation and high in precision.
The technical scheme adopted by the invention is as follows:
step (1) obtaining original data and preprocessing the data
The geometric parameters of the preprocessed existing radio frequency device (for example, the device parameters of the microstrip line rectangular spiral inductor are length l) 1 Width l 2 ) And the signal frequency freq added to the above-mentioned device is regarded as the input item x of the training model, take the known electromagnetic parameter that the input item x of the above-mentioned radio frequency device corresponds to namely S parameter (take 2 port network as example, S parameter includes S11 real part, S11 imaginary part, S12 real part, S12 imaginary part, S21 real part, S21 imaginary part, S22 real part, S22 imaginary part) as the true value y of the training model;
the input item x preprocessing process of the various existing radio frequency devices comprises the following steps: the specific operation is normalization processing, and the value is mapped to the range [ -1,1], where dispersion normalization is used, see formula (1):
wherein x min Is the minimum value of x, x max Is the maximum value of x;
and (2) performing training set and test set segmentation on the input parameters of the existing radio frequency device processed in the step (1) and the corresponding S parameters.
Preferably, 70% is taken as the training set and 30% as the test set.
And (3) after the data are segmented, constructing a deep multilayer neural network by using a deep learning framework:
3.1, setting the geometric parameters of the existing radio frequency device preprocessed in the step (1) and the frequency freq of the signal added to the device as input items to be an input layer of a neural network, and setting electromagnetic parameters (namely S parameters) corresponding to the radio frequency device as output items to be an output layer of the neural network;
3.2 connecting more than two hidden layers between the input layer and the output layer, wherein K neurons (which can be 100) are arranged in each hidden layer; adding a Relu activation function behind the hidden layer to add a nonlinear factor to the neural network, adding a dropout layer behind the activation function, and using a full-connection framework for the whole network.
The Relu activation function enables the neural network to fit a nonlinear model, the fitting speed of the activation function is high, and the phenomenon of gradient disappearance does not occur in back propagation. A dropout layer is added after the activation function to prevent the overfitting phenomenon.
Step (4), training the deep multilayer neural network model, defining a loss function call optimization algorithm, and changing a neural network input item x through back propagation i Weight parameter W of i And a deviation parameter b, and further reducing the loss value until the termination condition is met.
The loss function uses a mean squared error function (MSE):
where i denotes the ith sample, y i Representing an input item x i Corresponding true value, y' i Representing an input item x i And (4) obtaining a predicted value through calculation of the neural network, wherein n represents the number of samples input to the neural network.
The optimization algorithm used in the present invention is the Adam algorithm, which is an adaptive algorithm that stores the previously decayed squared gradient, while it also stores the previously decayed gradient. The learning rate can be adjusted by self during optimization, so that the loss function is prevented from falling into a local minimum or a saddle point to prevent the training from being continued.
The termination condition is that test set data is transmitted into the neural network when the loss value f (theta) is less than a preset loss threshold value, and the calculated loss value is less than the square of the preset loss threshold value after the test set data is put into the neural network.
And (5) optimizing the deep multilayer neural network model by adopting a genetic algorithm, wherein the optimization target is to transmit a desired result, namely an optimization target value, into the optimization algorithm, and the global optimal solution of the result can be obtained by the genetic algorithm.
Inputting the geometric parameter data set of the radio frequency device to be optimized, the variation range of the frequency freq data set of the input signal, the S parameter optimization target value and the allowable error (default 1 multiplied by 10) in the trained deep multilayer neural network model -8 ) (indicating the error between the target value of the S-parameter and the predicted value of the S-parameter).
The number of individuals in the population (representing the input items) dataset is set. And carrying out random assignment (namely random value taking) on the individuals according to the set number of the individuals. Calling a previously trained deep multilayer neural network model, and predicting all individuals in the population to obtain a plurality of S parameter predicted values;
then, selecting, crossing and mutating the fitness of the genetic algorithm to obtain an individual with the highest fitness, namely an individual with the smallest error, comparing the individual with an S parameter optimization target value, and directly outputting the individual if an allowable error range is met to obtain the best S parameter; otherwise, the selection, the crossover and the mutation are circularly carried out until the allowable error is met.
The fitness is shown in formula (3):
FIT (s,s＇) equation (3) =1/| s-s' |
Wherein S is the predicted value of the S parameter, and S' is the target value of the S parameter.
The invention has the beneficial effects that: the geometric parameter information of the radio frequency device to be optimized is input based on the trained model, so that the calculation result (such as S parameter) of the originally time-consuming electromagnetic simulation algorithm can be accurately and quickly predicted. Meanwhile, when the value of the S parameter is known, the optimal size information of the radio frequency device can be quickly obtained by utilizing a genetic algorithm, and the method is very accurate.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a spiral inductor structure;
FIG. 3 is a deep multi-layer neural network;
FIG. 4 is a graph showing the variation of loss values;
FIG. 5 is a graph comparing S parameter results predicted by Tensorflow with ADS; (a) Amplitude contrast of S parameter (b)
Comparing the phases of the S parameters;
fig. 6 is a diagram of a spiral inductance parameter optimization process.
Detailed Description
The technology is further described in detail with reference to the drawings and the embodiments.
Fig. 1 is a flowchart of a method for optimizing parameters of a radio frequency device based on deep learning according to the present invention. Example an example of optimizing spiral inductance is explained. The structure of the spiral inductor is shown in fig. 2.
Google deep learning framework tensoroflow was used for training. In this case, a microstrip rectangular spiral inductor is used. Take its length l 1 Width l 2 The frequency freq of the input signal is used as an input item x of a training model, and known electromagnetic parameters (S11 real part, S11 imaginary part, S12 real part, S12 imaginary part, S21 real part, S21 imaginary part, S22 real part, S22 imaginary part) corresponding to the input item x of the radio frequency device are used as a real value y of the training model.
Continuing to perform data preprocessing operation, specifically performing normalization operation on the input item x, where dispersion normalization is used, and the formula is:
wherein x min Is the minimum value of x, x max For the maximum value of x, the input vector x is defined as [ -1,1]In the interval, the purpose is to prevent that the learning rate needs to be continuously adjusted when the optimization is carried out along different gradient directions in the subsequent optimization process, so that the process of solving the minimum value of the loss function becomes a zigzag decline with low efficiency. This normalization also enables the Adam optimization algorithm to be used later with the best results.
After data preprocessing, the original data need to be segmented, and here, a tarin _ test _ split function of a model _ selection module of a skleann library can be directly called to segment a real value y consisting of an input item x and an S parameter into a 70% x _ train and y _ train training data set and a 30% x _ test and y _ test testing data set.
After the data are well segmented, a deep multilayer neural network is constructed, and the multilayer neural network is represented by a formula:
y＝σ(W n …σ(W 2 σ(W 1 x+b 1 )+b 2 )…+b n )
wherein W is a weight parameter, b is a bias execution parameter, and σ () is an activation function, and the neural network has more nonlinear characteristics after passing through a plurality of layers of activation functions. Setting an input item x of a radio frequency device to be trained as an input layer of a neural network, determining a matrix of n rows and 3 columns, wherein n is the number of samples fed into the neural network at each time, and the parameter l representing the spiral inductance is 3 according to the batch _ size (the batch _ size represents the number of training samples) of the samples fed into the neural network 1 ,l 2 Freq, i.e., the input layer has 3 neurons, the scattering S parameter value is set to the output layer in the neural network, and it is determined that it is a matrix of 100 rows and 8 columns (100 represents the number of neurons connected to the hidden layer, 8 represents the real part of S11, the imaginary part of S11, the real part of S12, the imaginary part of S12, the real part of S21, the imaginary part of S21, the real part of S22, the imaginary part of S22), i.e., the number of neurons in the output layer is 8. Connecting hidden layers between the input layer and the output layer, wherein two hidden layers are adopted, each hidden layer is provided with 100 neurons, and Relu activation function is added after the neuronThe neural network adds a nonlinear factor, and the formula of the Relu activation function is as follows:
f(x)＝max(0,x)
in this way, all output values equal to 0 are suppressed for all parts of the activation function that are less than 0, so that the neural network can be fitted to the nonlinear model, and it can be seen that the derivative of the Relu activation function is very simple, and the derivative is 1 for parts greater than 0. The fitting speed of the activation function is high, and the phenomenon that the gradient disappears can not occur when the back propagation derivation is carried out. In order to prevent the overfitting phenomenon from enhancing the generalization capability of the model, a dropout layer is added after each hidden layer passes through an activation function, the dropout operation has the effect that some neurons are randomly discarded from a certain hidden layer during one-time neural network training, the dropout proportion is 50%, and the whole network uses a full-connection architecture neural network architecture, as shown in fig. 3.
After the neural network is built, a loss function is defined, and a mean square error function (MSE) is used:
where i denotes the ith sample, y i Representing training model input x i Corresponding true value, y' i Representing an input item x i And (4) calculating a predicted value through the neural network, wherein n represents the number of samples input into the neural network.
After a mean square error loss function is defined, an adaptive optimization algorithm Adam algorithm tf, train, adamaOptimizer () in a Tensorflow library is called, and the formula of the Adam algorithm is as follows:
estimating a gradient mean value: m is a unit of t ＝β 1 m t-1 +(1-β 1 )g t
Estimating the mean square of the gradient:
correcting the gradient mean value by considering the iteration number:
correcting the mean square value of the gradient by considering the iteration number:
updating the weight and the bias:
wherein g is t Is the gradient of the t-th iteration loss function, beta 1 ，β 2 Is a delay factor, m t ,v t Respectively first and second moment estimates of the gradient,
is to m t And v t So that it can be approximated as an unbiased estimate of the expectation. Theta t+1 Is θ t The updated parameters, η, ε, are two constants. In the descending optimization of the loss function, the Adam optimization algorithm corrects the gradient mean value and the gradient square mean value by using the iteration times and the delay factor, accelerates the learning speed and efficiency, and can adjust the learning rate by self, wherein the learning rate base number alpha =0.001 and the delay factor beta is set 1 ＝0.9，β 2 ＝0.999，ε＝10 -8 . The iteration number t =0 during initialization, and the bias and weight matrix theta 0 Is a random matrix with mean value of 0 and small variance, m 0 ＝0，v 0 =0. The Adam algorithm makes the loss function not easy to fall into a local minimum or saddle points when optimizing, so that the training can not be continued.
When the optimization algorithm is determined to be good, a Session (Session) is opened in the Tensorflow deep learning framework, and all parameters in the Session are initialized by using tf. Run () using feed _ dic = assign operation, pass training set data x _ train, y _ train into the deep neural network model, using batch descent (Mini-batches L)earning) method, taking 100 for the batch _ size, taking 100 random samples of all samples for one iteration each time, and automatically updating the weight parameter W and the bias term b by the back propagation algorithm integrated by the adammomizer () optimization algorithm after the iteration is completed for one time. Thus, when all samples (one epoch) are trained, the W parameter and the bias item b are updated for a plurality of times, and the epoch is continuously circulated during training until the error requirement f (theta) is met<1×10 -2 。
When the loss value f (theta) of the trained model is less than 1 multiplied by 10 -2 The test set data x _ test is transmitted into the neural network, if the loss value calculated after the test set data is put into the neural network is less than 1 multiplied by 10 of the termination condition -4 Then the training is finished, the variation of the loss value is shown in fig. 4, the oscillation amplitude of the loss value is larger at the beginning because each gradient descent runs on min-batch instead of the whole data set, the data difference may cause the effect, but the oscillation amplitude of the loss value is reduced with the increase of the training times, and the loss value converges and approaches to 1 × 10 continuously -4 . And after the training is finished, calling a tf.train.saver () function, and saving the Session in which the deep learning network is positioned as a ckpt format file.
After the training is finished, the optimization operation is started, the optimization algorithm uses a genetic algorithm, and geometric parameters (l) of the radio frequency device to be optimized are input firstly 1 ，l 2 ) And the variation range (l) of the frequency freq of the signal added to the device 1 Is [30,40]，l 2 Is [20,25]Freq is [1,20]) Then, the target value of the S parameter is determined (here, the real part of S11 in the S parameter is taken as an optimization target, and the target value is set to be 0.8), and an allowable error (default is 1X 10-8) is selected. Setting the size of the population to be 500 individuals, and setting the size of the population to be l of all the individuals in the population according to the parameter range 1 ,l 2 And carrying out random initialization assignment on the freq parameter. Then calling a trained tenserflow model for prediction, building a neural network completely consistent with the training process in the prediction process, calling a restore () function of train. Saver to call the stored model, predicting S parameters of 500 individuals, and taking out real parts of S11 parameters and an optimized target value of 0And 8, making a difference value, and taking the difference value as an error. The fitness (fitness) formula for all individuals is calculated as:
FIT (s,s′) ＝1/|s-s′|
wherein S is the predicted value of the S parameter, and S' is the optimized target value of the S parameter.
After the fitness is determined, the population needs to be selected (selection), and the specific operation is to adopt a roulette method: 1. population size 500, individual s i Has a fitness of
Then the individual s i Probability of being selected is
2. Calculating cumulative probability
3. A random number r of evenly distributed 0-1 is generated.
4. R and q are i Compare if q i-1 ≤r≤q i Then the ith individual is selected to enter the next generation new population.
5. Repeat 3 and 4 until the number of new populations equals 500.
Next, cross-matching (crossover) is performed on the individuals of the selected new population, and an arithmetic cross mode is adopted in the design, and the formula is as follows:
the new individual is composed of two individuals by linear combination. Wherein, the first and the second end of the pipe are connected with each other,
and &>
Is the individual after the crossing, is>
And &>
Is two individuals randomly selected from the population, alpha is a crossed constant and takes the value of (0,1)]The probability of crossover is taken to be 0.8.
And (3) carrying out mutation on a part of new individuals while carrying out cross pairing, wherein the probability of the mutation is 0.01, and the formula of the mutation is as follows:
And after selecting, cross-matching and mutating the new generation population, taking out the individual with the highest fitness value, namely the individual with the smallest error, comparing the allowable error of the optimized target value, outputting the optimal individual if the allowable error range is met, continuously calculating the fitness to select, cross-matching and mutating the new generation population if the allowable error range is not met, and outputting the result until the error value meets the error range. Then, the real part value of S11 is obtained according to the expected real part value of 0.8Is best 1 ，l 2 Freq parameter.
As shown in FIG. 5, the S parameter result of Tensorflow prediction is compared with ADS, including the amplitude comparison of S parameter and the phase comparison of S parameter, and the mean square error of the S parameter result of Tensorflow prediction of spiral inductor in this example, the simulation result of amplitude and phase of ADS is 7.34 × 10 -5 The effect is very good.
As shown in fig. 6, a process diagram for spiral inductance parameter optimization. The optimization target is the scattering S11 parameter of the spiral inductor, and the error is less than 1 multiplied by 10 after 7 times of iteration -8 The result is very accurate and fast.
The above embodiments are not intended to limit the present invention, and the present invention is not limited to the above embodiments, and all embodiments are within the scope of the present invention as long as they meet the requirements of the present invention.
Claims (1)
1. The radio frequency device parameter optimization method based on deep learning is characterized by comprising the following steps:
step (1), acquiring original data and preprocessing the data
Taking the preprocessed geometric parameters of the existing radio frequency device and the signal frequency freq added to the device as input items x of a training model, and taking known electromagnetic parameters corresponding to the input items x of the radio frequency device as real values y of the training model;
the input item x preprocessing process of the various existing radio frequency devices comprises the following steps: the specific operation is normalization processing, the value of the normalization processing is mapped to the range of [ -1,1], and the dispersion normalization is used as shown in formula (1):
wherein x min Is the minimum value of x, x max Is the maximum value of x;
step (2), the input parameters of the existing radio frequency device processed in the step (1) and the corresponding S parameters are subjected to training set and test set segmentation;
and (3) after the data are segmented, constructing a deep multilayer neural network by using a deep learning framework:
3.1, setting the geometric parameters of the existing radio frequency device preprocessed in the step (1) and the frequency freq of the signal added to the device as input items to be input layers of a neural network, and setting the electromagnetic parameters corresponding to the radio frequency device as output items to be output layers of the neural network;
3.2 connecting more than two hidden layers between the input layer and the output layer, wherein K neurons are arranged in each hidden layer; adding a Relu activation function behind the neuron to add a nonlinear factor to the neural network, adding a dropout layer behind the activation function, and using a full-connection framework for the whole network;
step (4) training the deep multi-layer neural network model, defining a loss function call optimization algorithm, and changing a neural network input item x through back propagation i Weight parameter W of i And a deviation parameter b, so as to reduce the loss value until the termination condition is met;
the loss function uses the mean square error function:
where i denotes the ith sample, y i Representing an entry x i Corresponding true value, y' i Representing an entry x i A predicted value is obtained through calculation of the neural network, and n represents the number of samples input to the neural network;
the termination condition is that when the loss value f (theta) is smaller than a preset loss threshold value, test set data are transmitted into the neural network, and if the test set data are put into the neural network, the calculated loss value is smaller than the square of the preset loss threshold value;
and (5) optimizing a deep multilayer neural network model by adopting a genetic algorithm:
inputting a geometric parameter data set of a radio frequency device to be optimized, a variation range of a frequency freq data set of an input signal, an S parameter optimization target value and an allowable error in the trained deep multilayer neural network model;
regarding the input items as a population, and setting the number of individuals in a population data set; randomly assigning values to the individuals according to the set number of the individuals; calling the trained deep multilayer neural network model to predict all individuals in the population to obtain a plurality of S parameter predicted values;
then, selecting, crossing and mutating the fitness of the genetic algorithm to obtain an individual with the highest fitness, namely an individual with the minimum error, comparing the individual with an S parameter optimization target value, and directly outputting the individual if the allowable error range is met to obtain the optimal S parameter; otherwise, circularly selecting, crossing and mutating until the allowable error is met;
the fitness is shown in formula (3):
FIT (s,s′) equation (3) =1/| s-s' |
Wherein S is the predicted value of the S parameter, and S' is the target value of the S parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810602897.3A CN108959728B (en) | 2018-06-12 | 2018-06-12 | Radio frequency device parameter optimization method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810602897.3A CN108959728B (en) | 2018-06-12 | 2018-06-12 | Radio frequency device parameter optimization method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108959728A CN108959728A (en) | 2018-12-07 |
| CN108959728B true CN108959728B (en) | 2023-04-07 |
Family
ID=64488522
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810602897.3A Active CN108959728B (en) | 2018-06-12 | 2018-06-12 | Radio frequency device parameter optimization method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108959728B (en) |
Families Citing this family (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635420B (en) * | 2018-12-10 | 2021-07-30 | 南开大学 | Simulation method and system of microwave microstrip hairpin type filter |
| CN109583575B (en) * | 2018-12-17 | 2023-06-23 | 东南大学 | Processing method for improving instrument vector signal analysis performance based on deep learning |
| CN109407007A (en) * | 2018-12-24 | 2019-03-01 | 广东省智能制造研究所 | A kind of battery charge state detection method and device |
| CN109767429B (en) * | 2018-12-28 | 2021-08-06 | 上海联影智能医疗科技有限公司 | Image screening method and device |
| EP3912094A1 (en) * | 2019-01-18 | 2021-11-24 | Nokia Technologies Oy | Training in communication systems |
| CN110058943B (en) * | 2019-04-12 | 2021-09-21 | 三星（中国）半导体有限公司 | Memory optimization method and device for electronic device |
| CN112308278A (en) * | 2019-08-02 | 2021-02-02 | 中移信息技术有限公司 | Method, device, equipment and medium for optimizing prediction model |
| WO2021112665A1 (en) * | 2019-12-06 | 2021-06-10 | Collaborative Research In Engineering, Science & Technology (Crest) Center | Radio frequency device design simulator system and method |
| CN111191361A (en) * | 2019-12-27 | 2020-05-22 | 湖南大学 | Random forest based spiral inductance modeling method |
| CN111159915A (en) * | 2020-01-03 | 2020-05-15 | 山东天岳电子科技有限公司 | Parameter optimization method and device for device design |
| CN112556682B (en) * | 2021-02-07 | 2023-06-23 | 天津蓝鳍海洋工程有限公司 | Automatic detection algorithm for underwater composite sensor target |
| CN112861991B (en) * | 2021-03-09 | 2023-04-14 | 中山大学 | Learning rate adjusting method for neural network asynchronous training |
| CN112883670B (en) * | 2021-03-17 | 2022-07-29 | 电子科技大学 | Inductance automatic design comprehensive model and method based on artificial neural network |
| CN113505564B (en) * | 2021-06-30 | 2023-04-18 | 南京邮电大学 | RESURF transverse power device optimal drift region concentration design method based on deep learning |
| CN114239330B (en) * | 2021-11-01 | 2022-06-10 | 河海大学 | Deep learning-based large-span latticed shell structure form creation method |
| CN115146531A (en) * | 2022-06-22 | 2022-10-04 | 中国电力科学研究院有限公司 | Electromagnetic structure optimization design method based on deep learning |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6269351B1 (en) * | 1999-03-31 | 2001-07-31 | Dryken Technologies, Inc. | Method and system for training an artificial neural network |
| CN101706335A (en) * | 2009-11-11 | 2010-05-12 | 华南理工大学 | Wind power forecasting method based on genetic algorithm optimization BP neural network |
| CN102609587A (en) * | 2012-02-13 | 2012-07-25 | 苏州芯禾电子科技有限公司 | Method for designing radio frequency integrated circuit by utilizing three-dimensional fast electromagnetic field simulation technology |
| WO2015172560A1 (en) * | 2014-05-16 | 2015-11-19 | 华南理工大学 | Central air conditioner cooling load prediction method based on bp neural network |
| CN105975687A (en) * | 2016-05-04 | 2016-09-28 | 西安电子科技大学 | Method for constructing lumped model of band-pass coplanar waveguide micro-strip through hole-free transition structure |
| CN108090658A (en) * | 2017-12-06 | 2018-05-29 | 河北工业大学 | Arc fault diagnostic method based on time domain charactreristic parameter fusion |
| CN108140141A (en) * | 2015-08-15 | 2018-06-08 | 易享信息技术有限公司 | Using 3D batches of normalized three-dimensional (3D) convolution |
-
2018
- 2018-06-12 CN CN201810602897.3A patent/CN108959728B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6269351B1 (en) * | 1999-03-31 | 2001-07-31 | Dryken Technologies, Inc. | Method and system for training an artificial neural network |
| CN101706335A (en) * | 2009-11-11 | 2010-05-12 | 华南理工大学 | Wind power forecasting method based on genetic algorithm optimization BP neural network |
| CN102609587A (en) * | 2012-02-13 | 2012-07-25 | 苏州芯禾电子科技有限公司 | Method for designing radio frequency integrated circuit by utilizing three-dimensional fast electromagnetic field simulation technology |
| WO2015172560A1 (en) * | 2014-05-16 | 2015-11-19 | 华南理工大学 | Central air conditioner cooling load prediction method based on bp neural network |
| CN108140141A (en) * | 2015-08-15 | 2018-06-08 | 易享信息技术有限公司 | Using 3D batches of normalized three-dimensional (3D) convolution |
| CN105975687A (en) * | 2016-05-04 | 2016-09-28 | 西安电子科技大学 | Method for constructing lumped model of band-pass coplanar waveguide micro-strip through hole-free transition structure |
| CN108090658A (en) * | 2017-12-06 | 2018-05-29 | 河北工业大学 | Arc fault diagnostic method based on time domain charactreristic parameter fusion |
Non-Patent Citations (1)
| Title |
|---|
| 马惠珠 ; 宋朝晖 ; 季飞 ; 侯嘉 ; 熊小芸 ; .项目计算机辅助受理的研究方向与关键词――2012年度受理情况与2013年度注意事项.电子与信息学报.2013,(01),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108959728A (en) | 2018-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108959728B (en) | Radio frequency device parameter optimization method based on deep learning | |
| CN110782663B (en) | Road network traffic flow short-time prediction method combining time-space characteristics | |
| CN109360604B (en) | Ovarian cancer molecular typing prediction system | |
| CN111563706A (en) | Multivariable logistics freight volume prediction method based on LSTM network | |
| CN102831269A (en) | Method for determining technological parameters in flow industrial process | |
| AU2021245165B2 (en) | Method and device for processing quantum data | |
| CN111985825A (en) | Crystal face quality evaluation method for roller mill orientation instrument | |
| CN109787821B (en) | Intelligent prediction method for large-scale mobile client traffic consumption | |
| CN112766603A (en) | Traffic flow prediction method, system, computer device and storage medium | |
| US11550274B2 (en) | Information processing apparatus and information processing method | |
| CN111058840A (en) | Organic carbon content (TOC) evaluation method based on high-order neural network | |
| CN113743474A (en) | Digital picture classification method and system based on cooperative semi-supervised convolutional neural network | |
| CN115237878A (en) | Process database construction method and medium based on additive manufacturing | |
| CN113095466A (en) | Algorithm of satisfiability model theoretical solver based on meta-learning model | |
| CN113536939A (en) | Video duplication removing method based on 3D convolutional neural network | |
| CN111524348A (en) | Long-short term traffic flow prediction model and method | |
| CN109523386A (en) | A kind of investment portfolio risk prediction technique of GMM in conjunction with LSTM | |
| CN116415177A (en) | Classifier parameter identification method based on extreme learning machine | |
| EP4109374A1 (en) | Data processing method and device | |
| KR20190129422A (en) | Method and device for variational interference using neural network | |
| US11783194B1 (en) | Evolutionary deep learning with extended Kalman filter for modeling and data assimilation | |
| CN114492174A (en) | Full life cycle shield tunneling parameter prediction method and device | |
| CN114139677A (en) | Unequal interval time sequence data prediction method based on improved GRU neural network | |
| CN113763710A (en) | Short-term traffic flow prediction method based on nonlinear adaptive system | |
| CN113656707A (en) | Financing product recommendation method, system, storage medium and equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |