CN105787558A - Knowledge-based neural network micro-strip filter design method based on ADS - Google Patents

Abstract

The present invention discloses a knowledge-based neural network micro-strip filter design method based on ADS. The knowledge-based neural network micro-strip filter design method based on ADS is a method taking ADS as prior knowledge on the basis of a known knowledge-based neural network to overcome the problems that the neural network is complex in structure and difficult in prior knowledge acquisition. Through combination of a neural network and simulation software, the method provided by the invention employs ADS and HFSS simulation results to respectively take as prior knowledge and teacher signals and performs training through a particle swarm optimization to construct a corresponding neural network model so as to effectively reduce the complexity of a neural network structure. Through adoption of the designed network, the method provided by the invention perform optimization design of a high-low impedance low pass filter and a micro-strip cassette band-pass filter, wherein the designed filter satisfies the design index, so that lots of complex formula derivation are avoided, the structure of a neural network is simplified, the hidden layers and the numbers of the hidden layer are reduced, and the cost is effectively reduced.

Description

Knowledge neutral net microstrip filter method for designing based on ADS

Technical field

The present invention relates to a kind of microstrip filter method for designing, particularly to a kind of knowledge neutral net microstrip filter method for designing based on ADS, belong to microwave electromagnetic technical field.

Background technology

It is little and be easy to the advantages such as integrated that microstrip filter has volume, is widely used in microwave circuit.Neutral net, as one quickly and flexibly instrument, was widely used in the modeling of microstrip filter in recent years, and yielded good result.However, to ensure that the accuracy of neural network model, generally require large batch of training sample and go to describe the change of different geometry and structural parameters in microwave circuit structure in operating frequency range, training sample is often by Numerical Calculation of Electromagnetic Fields or what measurement obtained, when the training sample quantity needed is too much, the amount of calculation of electromagnetic field can increase, the consumption of human and material resources can become much larger so that the foundation of neural network model becomes very difficult.

The complexity reducing input-output mappings relation is to reduce a kind of effective method of number of training purpose.According to this idea, based on acquainted neural network model and Knowledge-Based Neural Network Model (knowledge-basedneuralnetwork, KBNN) it is suggested, this priori is existing empirical equation mostly, they include the essential information of microwave circuit structure, but are unable to reach required precision in working range.KBNN model is repeatedly proved while ensureing model accuracy, moreover it is possible to be effectively reduced the quantity of training sample.Although KBNN can effectively reduce sample size, but in existing method, the acquisition of priori relies on empirical equation or neutral net.And the derivation of correlation formula is excessively complicated loaded down with trivial details in electromagnetic problem, the training of neutral net needs again great amount of samples, so both approaches all existing defects.The present invention is on the basis of existing KBNN model, it is proposed that the building method of new priori, and ADS simulation result constitutes the knowledge neuron of hidden layer as priori.

Summary of the invention

It is an object of the invention to provide a kind of based on ADS (AdvancedDesignsystem, Advanced Design System) knowledge neutral net microstrip filter method for designing, use it in microstrip filter, with the problem overcoming existing neural network structure complicated and priori obtains difficulty.

The purpose of the present invention is achieved by the following technical programs:

A kind of knowledge neutral net microstrip filter method for designing based on ADS, comprises the following steps:

The first step: the modeling of knowledge neutral net

1) acquisition of training sample

Using the simulation result of ADS as priori, constructing knowledge neutral net, the teacher signal of knowledge neutral net is emulated (HighFrequencyStructureSimulator, HFSS) by high-frequency structure and obtains；

2) structure of knowledge neutral net

Knowledge neutral net is made up of input layer, hidden layer, output layer, the neuron of described hidden layer is divided into traditional neural unit and priori neuron, described input layer is connected with traditional neural unit, described traditional neural unit is connected with output layer, described priori neuron is connected with output layer, described input layer with priori neuron without being connected；

Knowledge neutral net is the structure of n × m × 1, and its input is x_i(i=1,2 ..., n)；The neuronic number of priori of hidden layer is p, is output as hk_j(j=1,2 ..., p)；The number of the traditional neural unit of hidden layer is q (p+q=m), is output as h_k(k=1,2 ..., q)；Knowledge neutral net is set to an output network, is output as y；Its input of traditional neural unit is defined as:

xh_k=ω_ikx_i+b_k, i=1,2 ..., n, k=1,2 ..., q

Wherein ω_ikFor the connection weights of input layer and traditional neural unit, b_kFor threshold value；

The excitation function of traditional neural unit selects Sigmoid function, is output as:

$h_k=\frac{1}{1+e^{-{\mathrm{\λxh}}_k}},k=1,2,...,q$

Gain λ=1 of Sigmoid function；

Knowledge neutral net is output as each hidden layer weighting sum, it may be assumed that

Y=ω_jhk_j+ω_kh_k+ b, j=1,2 ..., p, k=1,2 ..., q

Wherein ω_jFor the connection weights of priori neuron and the output layer of hidden layer, ω_kFor the first connection weights with output layer of traditional neural of hidden layer, b is threshold value；

Knowledge neutral net is output as forward transmission coefficient S₂₁, input reflection coefficient S₁₁With forward transmission coefficient S₂₁There is following relation:

S₁₁ ²+S₂₁ ²=1

The neuronic number of priori of hidden layer is 2, takes input reflection coefficient S respectively₁₁With forward transmission coefficient S₂₁；

3) study of knowledge neutral net

After knowledge neural network configuration is good, adopts particle cluster algorithm that weights and threshold value are updated, stop updating when iteration reaches maximum times or mean square error reaches preset value, obtain the neutral net trained；

4) reliability of network is detected

By test sample to 3) in neutral net test, whether checking network output consistent with the result of high-frequency structure emulation HFSS, and the correlation coefficient of calculation knowledge neutral net judges the reliability of neutral net；

Second step: optimize design

1) stochastic generation initial population, as the input of knowledge neutral net, is exported accordingly by knowledge neutral net；

2), in optimization process, corresponding fitness function is set according to design objective；Assume that design objective is at f_i(i=1,2 ...) place require S₂₁Amplitude is s respectively_i(i=1,2 ...), then fitness function is

Fit=min (Σ β_i|y_i-s_i|)

Wherein, y_iFor f_i(i=1,2 ...) output of place's knowledge neutral net, β_iFor the weight selected by respective frequencies point place；

3) update initial population by particle cluster algorithm, find the optimal solution meeting fitness function；

4) by 3) in the optimal solution that obtains bring into and HFSS verify whether meet design objective.

The purpose of the present invention can also be realized further by techniques below measure:

The aforementioned knowledge neutral net microstrip filter method for designing based on ADS, when micro-strip high low-impedance filter is optimized design, constructed knowledge neutral net hidden neuron is 7.

The aforementioned knowledge neutral net microstrip filter method for designing based on ADS, when micro-strip hair fastener band filter is optimized design, constructed knowledge neutral net hidden neuron is 5.

The aforementioned knowledge neutral net microstrip filter method for designing based on ADS, wherein in particle cluster algorithm, the more new formula of speed and position is:

$v_{i,d}^{k+1}=v_{i,d}^k+c_{1}rand\left(\right)(p_{i,d}^k-x_{i,d}^k)+c_{2}rand\left(\right)(p_{g,d}^k-x_{i,d}^k)$

$x_{i,d}^{k+1}=x_{i,d}^k+v_{i,d}^{k+1}$

In formula, c₁And c₂For Studying factors；Rand () is the random number between (0,1)；WithThe respectively particle i speed that d ties up in k iteration and position；For the particle i position in the d individual extreme value tieed up；For the colony position at the d global extremum tieed up.

The aforementioned knowledge neutral net microstrip filter method for designing based on ADS, wherein c₁=2.8, c₂=1.3.

Compared with prior art, the invention has the beneficial effects as follows: the present invention using the simulation result of ADS as priori, structure knowledge neutral net, the derivation not only avoiding a large amount of numerous and diverse formula also simplify the structure of neutral net, decrease the number of hidden layer and hidden neuron, significantly reduce cost.The present invention can quickly obtain building the priori needed for knowledge neutral net, with the problem overcoming existing neural network structure complicated and priori obtains difficulty.Neutral net is trained as priori and teacher signal by simulation result, after network training completes, it is possible to use microstrip filter is optimized design by this model, is met the microstrip filter size of index.No matter being high Low ESR low pass filter or micro-strip hair fastener band filter, network test exports all better HFSS exact numerical of must having fitted, it is seen that the knowledge neutral net of this method structure has higher accuracy really.

Accompanying drawing explanation

Fig. 1 is knowledge neural network structure figure；

Fig. 2 is high Low ESR low pass filter schematic diagram；

Fig. 3 is high Low ESR low pass filter model, and wherein Fig. 3 (a) is ADS phantom, and Fig. 3 (b) is HFSS phantom；

Fig. 4 is that the knowledge neutral net test that low pass filter is set up exports and HFSS simulation result comparison diagram；

Fig. 5 is the HFSS analogous diagram of low pass filter optimized dimensions；

Fig. 6 is micro-strip hair fastener band filter schematic diagram；

Fig. 7 is micro-strip hair fastener band filter model, wherein Fig. 7 (a) ADS phantom, and Fig. 7 (b) is HFSS phantom；

Fig. 8 is that the knowledge neutral net test that band filter is set up exports and HFSS simulation result comparison diagram；

Fig. 9 is the HFSS analogous diagram of band filter optimized dimensions.

Detailed description of the invention

In existing knowledge neutral net, the acquisition of priori is typically all empirical equation, and minority adopts neural metwork training to generate, and these methods are all more numerous and diverse.The simulation result of ADS as priori, is constructed knowledge neutral net by the present invention, will effectively prevent the calculating of complexity, it is possible to quickly obtain building the priori needed for knowledge neutral net.

The method that neutral net is combined by the present invention with simulation software, utilize ADS and high-frequency structure emulation (HighFrequencyStructureSimulator, HFSS) obtain the scattering parameter (S parameter) of microstrip filter, neutral net is trained as priori and teacher signal by simulation result.After network training completes, it is possible to use microstrip filter is optimized design by this model, it is met the microstrip filter size of index.Therefore present invention is broadly divided into the modeling of knowledge neutral net and is used for model optimizing two parts of design.

One, the modeling of knowledge neutral net

1) acquisition of training sample

Training data is broadly divided into the teacher signal of priori and training network, and the former is obtained by ADS, and the latter is obtained by HFSS.Although knowledge neutral net can reduce sample size, but this minimizing is relative, if data adopt the form being manually entered and manually deriving, not only does not reduce workload, more makes this process complicate on the contrary；And the follow-up process optimizing design, each iteration dimensional parameters is unforeseen, and trains required priori to rely on dimensional parameters just can obtain, and is manually entered and seems unrealistic.Based on above-mentioned two reason, in the present invention, training data all adopts and writes script and call each simulation software, generates dimensional parameters, pass to each model computer sim-ulation result, then pass Matlab process back in Matlab.The acquisition of HFSS teacher signal has only to by calling vbs program file it is achieved that for same model different parameters, it is possible to only in amendment file to have related parameter to perform just passable again, this has just been provided in association with interface for HFSS software and Matlab；Similarly, the method obtaining employing of ADS priori is similar, by ADS distinctive AEL (ApplicationExtensionLanguage) language itself, model is operated.

2) structure of knowledge neutral net

Knowledge-Based Neural Network Model of the present invention is as it is shown in figure 1, a portion hidden neuron is made up of priori, and another part is then identical with traditional hidden neuron.Assuming the structure that knowledge neutral net is n × m × 1, its input is x_i(i=1,2 ..., n)；The neuronic number of hidden layer knowledge is p, and each knowledge is neuronic is output as hk_j(j=1,2 ..., p)；The number of hidden layer traditional neural unit is q (p+q=m), each neuronic is output as h_k(k=1,2 ..., q)；For convenience of describing, network is set to an output network, is output as y.For knowledge neuron, the present invention does not adopt empirical equation to calculate, therefore the input of network is not connected with knowledge neuron, simplifies neural network structure further.Its input of traditional neuron is defined as:

xh_k=ω_ikx_i+b_k, i=1,2 ..., n, k=1,2 ..., q (1)

Wherein ω_ikFor the connection weights of input layer and hidden layer traditional neural unit, b_kFor threshold value.

The excitation function of traditional neural unit generally selects Sigmoid function, is output as:

$h_k=\frac{1}{1+e^{-{\mathrm{\λxh}}_k}},k=1,2,...,q---\left(2\right)$

In the present invention, gain λ=1 of Sigmoid function.

Knowledge neutral net is output as each hidden layer weighting sum, it may be assumed that

Y=ω_jhk_j+ω_kh_k+ b, j=1,2 ..., p, k=1,2 ..., q (3)

Wherein ω_jFor the connection weights of hidden layer knowledge neuron Yu output layer, ω_kFor the connection weights of hidden layer traditional neural unit with output layer, b is threshold value.Set up knowledge neutral net by introducing ADS priori neuron, simplify the structure of neutral net, decrease the number of hidden neuron and the advantage still ensuring that knowledge neutral net.

In the present invention, knowledge neutral net is output as forward transmission coefficient S₂₁, due to input reflection coefficient S₁₁With forward transmission coefficient S₂₁There is following relation:

S₁₁ ²+S₂₁ ²=1 (4)

Therefore, the neuronic number of hidden layer knowledge is 2, takes input reflection coefficient S11 and forward transmission coefficient S21 respectively.

In general, choosing of the optimum number m of the hidden neuron of neutral net follows following formula:

$m=\sqrt{n+t}+\mathrm{\α}---\left(5\right)$

Wherein t is the number of neutral net output, and in the present invention, neutral net output number t=1, α are the constant between [1,10].In the present invention, 7 are taken for high Low ESR low pass filter hidden neuron, and micro-strip hair fastener band filter hidden neuron takes 5, the best results obtained.

3) study of knowledge neutral net

After knowledge neural network framework is put up, just neutral net is trained.In the present invention, choosing particle cluster algorithm and update weights and the threshold value of network, the method can be prevented effectively from the problem being absorbed in local optimum.In particle cluster algorithm, the more new formula of speed and position is:

$v_{i,d}^{k+1}=v_{i,d}^k+c_{1}rand\left(\right)(p_{i,d}^k-x_{i,d}^k)+c_{2}rand\left(\right)(p_{g,d}^k-x_{i,d}^k)---\left(6\right)$

$x_{i,d}^{k+1}=x_{i,d}^k+v_{i,d}^{k+1}---\left(7\right)$

In formula, c₁And c₂It is referred to as Studying factors or aceleration pulse, takes c in the present invention₁=2.8, c₂=1.3；Rand () is the random number between (0,1)；WithThe respectively particle i speed that d ties up in k iteration and position；For the particle i position in the d individual extreme value tieed up；For the colony position at the d global extremum tieed up.In training process, position x is the weights and threshold value to update, network exports the y mean square error with teacher signal as the fitness function of particle cluster algorithm, updates their value in each iteration.When error condition or the maximum iteration time of satisfied setting, algorithm stops updating, and global optimum now is weights and the threshold value of network.

4) reliability of network is detected

Randomly selecting sample in parameter area, the knowledge neutral net constructed is tested, whether the output of verification test sample is consistent with the simulation result of HFSS.The correlation coefficient of computing network judges the accuracy of network simultaneously, and computing formula is as shown in (8).

$r=\frac{\Σ(z_i-\stackrel{\‾}{z})(y_i-\stackrel{\‾}{y})}{\sqrt{\Σ{(z_i-\stackrel{\‾}{z})}^2{(y_i-\stackrel{\‾}{y})}^2}}---\left(8\right)$

Wherein, z_iIt is Electromagnetic Simulation value of calculation, y_iIt is neural network model value of calculation,It is the average of Electromagnetic Simulation value of calculation,It it is the average of neural network model value of calculation.Correlation coefficient, closer to 1, illustrates that neural network model value of calculation is closer to sample value, and model is set up more reasonable.If correlation coefficient is relatively low, then re-training network.

Two, design microstrip filter is optimized

1) stochastic generation initial population (dimensional parameters in corresponding microstrip filter), as the input of knowledge neutral net, is exported accordingly by knowledge neutral net；

2), in optimization process, corresponding fitness function is set according to design objective.Assume that design objective is at f_i(i=1,2 ...) place require S₂₁Amplitude is s respectively_i(i=1,2 ...), then fitness function is

Fit=min (Σ β_i|y_i-s_i|)(9)

Wherein, y_iFor f_i(i=1,2 ...) output of place's knowledge neutral net, β_iFor the weight selected by respective frequencies point place.

3) update initial population by particle cluster algorithm, find the optimal solution meeting fitness function；

4) by 3) in the optimal solution that obtains bring into and HFSS verify whether meet design objective, if meeting, illustrate that the invention of the present invention has practical feasibility.

It is used for optimizing design microstrip filter based on the knowledge neutral net of ADS by the present invention, adopts this neural net method respectively high Low ESR low pass filter and micro-strip hair fastener band filter to be designed.Embodiment is further illustrating the present invention in detail below, rather than restriction scope of invention:

Embodiment 1:

Design a high low-impedance filter as in figure 2 it is shown, optimization design objective is: 1. cut-off frequency is about 3GHz；2. in bandpass ripple more than-1dB；3. it is not more than-20dB in the decay of 4GHz place.This example selects optimize W_i(i=1,2 ..., 5) reach design objective, each parameter value is in Table 1.

The each dimensional parameters table of table 1 high Low ESR low pass filter

Specifically comprise the following steps that

(1) setting up the model of ADS, HFSS, as it is shown on figure 3, medium substrate is 1mm, and relative dielectric constant is 2.7.Called the simulation result obtained corresponding to each size by Matlab, wherein choose the simulation result S of ADS₁₁And S₂₁As priori；Choose the simulation result S of HFSS₂₁As teacher signal.

(2) network input sample is W_i(i=1,2 ..., 5) and operating frequency f, wherein, operating frequency f ranges for 0.1～5GHz, and step-length is 0.2GHz.Training sample adopts part combination orthogonal to obtain, totally 450 groups of samples.The hidden layer number of knowledge neutral net is 7, and wherein 2 is knowledge neuron, and all the other 5 is traditional neural unit.

(3) adopt particle cluster algorithm training neutral net, and by the reliability of test sample knowledge on testing neutral net, calculate the correlation coefficient of this network.

(4) neutral net trained is substituted HFSS model, adopt particle cluster algorithm to be optimized design, obtain dimensional parameters and verified with HFSS.

By particle cluster algorithm, network being trained, the correlation coefficient of the knowledge neutral net of final gained is 0.9936, and with test sample, network is tested, and as shown in Figure 4, its vertical coordinate is S to result₂₁Range value, mean absolute error is 1.1%.Being used for optimizing design by this network, the dimensional parameters finally given is [110.252211.45310.24811] respectively.HFSS model inputs this size and obtains S₂₁Parameter, as it is shown in figure 5, this size of empirical tests meets design objective, it was demonstrated that the present invention is practical.

Embodiment 2:

Designing a micro-strip hair fastener band filter, as shown in Figure 6, optimizing design objective is: 1. bandpass range is 2.3GHz～2.8GHz；2. in bandpass ripple more than-2dB；3. in the decay of 1.95GHz and 3.1GHz place less than-40dB；This example select optimization carry out L₀、L₁、L₂、L₅、S₁、S₂Reaching design objective, each parameter value is in Table 2.

Specifically comprise the following steps that

(1) setting up the model of ADS, HFSS, as it is shown in fig. 7, medium substrate is 1mm, and relative dielectric constant is 4.4.Called the simulation result obtained corresponding to each size by Matlab, wherein choose the simulation result S of ADS₁₁And S₂₁As priori；Choose the simulation result S of HFSS₂₁As teacher signal.

(2) network input sample is L₀,L₁,L₂,L₅,S₁,S₂And operating frequency f is wherein, operating frequency f ranges for 1.85～3.20GHz, and step-length is 50MHz.Training sample adopts part combination orthogonal to obtain, totally 1400 groups of samples.The hidden layer number of knowledge neutral net is 5, and wherein 2 is knowledge neuron, and all the other 3 is traditional neural unit.

The each dimensional parameters table of table 2 micro-strip hair fastener band filter

(3) adopt particle cluster algorithm training neutral net, and by the reliability of test sample knowledge on testing neutral net, calculate the correlation coefficient of this network.

(4) neutral net trained is substituted HFSS model, adopt particle cluster algorithm to be optimized design, obtain dimensional parameters and verified with HFSS.

By particle cluster algorithm, network being trained, the correlation coefficient of the knowledge neutral net of final gained is 0.9962, and with test sample, network is tested, and as shown in Figure 8, its vertical coordinate is S to result₂₁The decibel value of amplitude, average relative error is 0.7807.Being used for optimizing design by this network, the dimensional parameters finally given is [5.99.97991.22.65260.28620.4838] respectively.HFSS model inputs this size and obtains S₂₁Parameter, as it is shown in figure 9, this size of empirical tests meets design objective, it was demonstrated that the present invention is practical.

As shown in figures 4 and 8, for the comparison of the network test output of this method structure with HFSS exact numerical.It can be seen that no matter be high Low ESR low pass filter or micro-strip hair fastener band filter, network test exports all better HFSS exact numerical of must having fitted, it is seen that the knowledge neutral net of this method structure has higher accuracy really.The employing ADS that chooses of priori obtains, and not only avoids the derivation of a large amount of numerous and diverse formula to also simplify the structure of neutral net, decreases the number of hidden layer and hidden neuron, significantly reduce cost.And in optimizing design, Fig. 5 is the value that after optimizing, the size HFSS of high Low ESR low pass filter calculates gained.From figure 5 it can be seen that minimum is-0.8729dB at 0.9GHz place in ripple, cut-off frequency corresponding to three dB bandwidth is 2.9458GHz, decays to-21.2754dB at 4GHz place, meets design objective.From fig. 9, it can be seen that 3dB cut-off frequency bandwidth corresponding to the micro-strip hairpin filter of this size to be 2.3084GHz～2.8176GHz basically identical with the 2.3GHz～2.8GHz of setting.The minimum place m of ripple in passband₁Vertical coordinate is-1.7802dB, more than-2dB.At 1.95GHz and 3.1GHz place decay respectively-41.1832dB and-40.7268dB, it is both less than-40dB, meets design objective.

Claims (5)

1. the knowledge neutral net microstrip filter method for designing based on ADS, it is characterised in that comprise the following steps:

The first step: the modeling of knowledge neutral net

1) acquisition of training sample

Using the simulation result of ADS as priori, constructing knowledge neutral net, the teacher signal of knowledge neutral net is emulated HFSS by high-frequency structure and obtains；

2) structure of knowledge neutral net

Knowledge neutral net is made up of input layer, hidden layer, output layer, the neuron of described hidden layer is divided into traditional neural unit and priori neuron, described input layer is connected with traditional neural unit, described traditional neural unit is connected with output layer, described priori neuron is connected with output layer, described input layer with priori neuron without being connected；

Knowledge neutral net is the structure of n × m × 1, and its input is x_i(i=1,2 ..., n)；The neuronic number of priori of hidden layer is p, is output as hk_j(j=1,2 ..., p)；The number of the traditional neural unit of hidden layer is q (p+q=m), is output as h_k(k=1,2 ..., q)；Knowledge neutral net is set to an output network, is output as y；Its input of traditional neural unit is defined as:

xh_k=ω_ikx_i+b_k, i=1,2 ..., n, k=1,2 ..., q

Wherein ω_ikFor the connection weights of input layer and traditional neural unit, b_kFor threshold value；

The excitation function of traditional neural unit selects Sigmoid function, is output as:

$h_k=\frac{1}{1+e^{-{\mathrm{\λxh}}_k}},k=1,2,...,q$

Gain λ=1 of Sigmoid function；

Knowledge neutral net is output as each hidden layer weighting sum, it may be assumed that

Y=ω_jhk_j+ω_kh_k+ b, j=1,2 ..., p, k=1,2 ..., q

Wherein ω_jFor the connection weights of priori neuron and the output layer of hidden layer, ω_kFor the first connection weights with output layer of traditional neural of hidden layer, b is threshold value；

Knowledge neutral net is output as forward transmission coefficient S₂₁, input reflection coefficient S₁₁With forward transmission coefficient S₂₁There is following relation:

S₁₁ ²+S₂₁ ²=1

The neuronic number of priori of hidden layer is 2, takes input reflection coefficient S respectively₁₁With forward transmission coefficient S₂₁；

3) study of knowledge neutral net

After knowledge neural network configuration is good, adopts particle cluster algorithm that weights and threshold value are updated, stop updating when iteration reaches maximum times or mean square error reaches preset value, obtain the neutral net trained；

4) reliability of network is detected

By test sample to 3) in neutral net test, whether checking network output consistent with the result of high-frequency structure emulation HFSS, and the correlation coefficient of calculation knowledge neutral net judges the reliability of neutral net；

Second step: optimize design

1) stochastic generation initial population, as the input of knowledge neutral net, is exported accordingly by knowledge neutral net；

2), in optimization process, corresponding fitness function is set according to design objective；Assume that design objective is at f_i(i=1,2 ...) place require S₂₁Amplitude is s respectively_i(i=1,2 ...), then fitness function is

Fit=min (Σ β_i|y_i-s_i|)

Wherein, y_iFor f_i(i=1,2 ...) output of place's knowledge neutral net, β_iFor the weight selected by respective frequencies point place；

3) update initial population by particle cluster algorithm, find the optimal solution meeting fitness function；

4) by 3) in the optimal solution that obtains bring into and HFSS verify whether meet design objective.

2. the knowledge neutral net microstrip filter method for designing based on ADS as claimed in claim 1, it is characterised in that when micro-strip high low-impedance filter is optimized design, constructed knowledge neutral net hidden neuron is 7.

3. the knowledge neutral net microstrip filter method for designing based on ADS as claimed in claim 1, it is characterised in that when micro-strip hair fastener band filter is optimized design, constructed knowledge neutral net hidden neuron is 5.

4. the knowledge neutral net microstrip filter method for designing based on ADS as described in claim 1 or 2 or 3, it is characterised in that the step 3 of the described first step) in particle cluster algorithm, the more new formula of speed and position is:

$v_{i,d}^{k+1}=v_{i,d}^k+c_{1}rand\left(\right)(p_{i,d}^k-x_{i,d}^k)+c_{2}rand\left(\right)(p_{R,d}^k-x_{i,d}^k)$

$x_{i,d}^{k+1}=x_{i,d}^k+v_{i,d}^{k+1}$

In formula, c₁And c₂For Studying factors；Rand () is the random number between (0,1)；WithThe respectively particle i speed that d ties up in k iteration and position；For the particle i position in the d individual extreme value tieed up；For the colony position at the d global extremum tieed up.

5. the knowledge neutral net microstrip filter method for designing based on ADS as claimed in claim 4, it is characterised in that wherein c₁=2.8, c₂=1.3.