CN110427669B

CN110427669B - Neural network model calculation method for phased array scanning radiation beams

Info

Publication number: CN110427669B
Application number: CN201910657700.0A
Authority: CN
Inventors: 温亚庆; 周雷
Original assignee: 724th Research Institute of CSIC
Current assignee: 724th Research Institute of CSIC
Priority date: 2019-07-20
Filing date: 2019-07-20
Publication date: 2023-04-07
Anticipated expiration: 2039-07-20
Also published as: CN110427669A

Abstract

The invention relates to a neural network model calculation method for phased array scanning radiation beams, which mainly comprises the following steps: and constructing a feedback type neural network model, taking the obtained array beam calculation result as a fixed input value to participate in the training of the neural network model, taking the trained neural network model as a phased array radiation beam calculation mode, and verifying the constructed model by using a similarity function. The method has accuracy and large-range coverage on the calculation of the phased array radiation wave beam, and improves the calculation efficiency and accuracy in the phased array wave beam analysis.

Description

Neural network model calculation method for phased array scanning radiation beams

Technical Field

The invention relates to phased array beam calculation, in particular to a calculation method of spatial domain scanning radiation beams.

Background

The beam quality of the planar phased array antenna gradually decreases with the increase of the scanning angle, which is mainly reflected in the stability of the beam gain and the size of the side lobe. Accurate calculation of the phased array scanning beam pattern can be used for exposing the problems in the early stage of design, and the problems can be effectively avoided and solved in engineering practice. The main difficulty in the calculation of phased array scanning beam patterns is the consideration of the coupling strength between array elements and the different scattering environments of each array element, and these conditions are difficult to be reflected in an analytic mode in the calculation of the array factor product theorem, so that the accuracy of calculation is influenced. Especially for planar phased array antennas with large angular scanning range, the above factors result in higher accuracy of the array radiation beam pattern in the edge-fire region, and a substantial reduction in computational accuracy in the low elevation range.

The invention mainly aims at the planar phased array antenna with large-angle scanning characteristic, and accurately calculates the prediction method of the neural network model introduced into the scanning beam in the whole scanning area.

Disclosure of Invention

The invention aims to provide a method for calculating a radiation beam of a planar large-angle scanning phased array.

The technical solution for realizing the invention is as follows: and selecting K groups of input values of random feeding amplitude to be substituted into the array full-wave model, selecting stepping and increasing the phase difference according to the actual phase shifter, and calculating the corresponding accurate radiation output under the condition of fixed phase difference by using full-wave simulation. An Elman type feedback neural network model is established, and an initial input value is added into an input layer and mainly comprises port feed phase and amplitude information. The network model comprises k hidden layers and a final output layer, wherein a neuron function is established in each layer, usually a Sigmoid function is established in each hidden layer, and a linear function is established in each output layer. And adding feedback information and a feedback operator between the input information and the output information in the hidden layer to construct a single-loop feedback system. And substituting the feed amplitude input value and the radiation beam output value which are subjected to full-wave calculation into the neural network model, training by adopting a conjugate gradient algorithm, and terminating the training when the prediction precision is reached or the preset times are reached. And after the neural network model is established, re-selecting random input I (N), and respectively calculating an array radiation directional diagram output value matrix by adopting a full-wave model simulation method and a neural network model method. And calculating the similarity of the two, setting a precision cut-off value, and verifying the calculation correctness of the neural network model.

The invention is characterized in that:

(1) The neural network model is adopted to calculate the scanning wave beams of the phased array, the coupling influence of an array element does not need to be additionally added in the calculation, and the calculation of the scanning wave beams of the large-angle planar phased array is more convenient and accurate;

(2) After an input and output database of the full-wave model is established, a general calculation model is generated through a neural network model, the calculation precision is similar to that of full-wave simulation, and meanwhile, the calculation time is greatly reduced;

(3) And quantized input amplitude distribution is adopted in the calculation process, and the phase value is set aiming at the stepping of the actual phase shifter, so that the practical engineering application is facilitated.

Drawings

Fig. 1 is a schematic diagram of the Elman neural network model.

FIG. 2 is a flow chart of neural network model optimization.

Fig. 3 is a normal radiation beam pattern of a phased array obtained by full-wave simulation under the same amplitude distribution.

FIG. 4 is a diagram of a neural network model calculating a phased array normal radiation beam pattern under the same amplitude distribution.

Fig. 5 is a phased array low elevation directional radiation beam pattern obtained from full wave simulation under the same amplitude distribution.

FIG. 6 is a diagram of a neural network model calculating a phased array low elevation angle directional radiation beam pattern under the same amplitude distribution.

Detailed Description

The present invention is described with an eight-element phased array antenna as an embodiment. The array elements adopt dipole units, the number of the array elements is 8, and the spacing between the array elements is 0.5 lambda. The specific process is as follows:

the method comprises the following steps: establishing a neuron function, and processing the neuron with information x from other neurons _i The connection between them and the neuron being processed is weighted by w _i The inputs to the processed neuron are then:

the output of the processed neuron is:

where f is a neuron function, also referred to as an activation function, which determines the output of the neuron. θ represents the threshold of the neural node within the hidden layer. Sigmoid function (hyperbolic tangent continuous increasing function) is adopted in the hidden layer:

a linear function is adopted in the output layer:

f(x)＝x；

step two: establishing connection form between neurons, adopting single-ring feedback system, inputting information x _i (n), feedback information x' _i (n) output information y _i (n) of (a). Assuming that the system is linear and consists of a forward operator f and a feedback operator g, the output relationship is as follows:

y _i (n)＝f(x′ _i (n))；

x′ _i (n)＝x _i (n)+g(y _i (n))；

the output information is thus:

wherein, f is a fixed weight w, g is a time delay operator z ^-1 Thus, this single loop operator can be written as:

the last term is developed with the maculing equation to obtain:

/>

wherein the content of the first and second substances,

z ^-k (x _i (n))＝x _i (n-k)；

step three: in the neural network model of the eight-element microstrip dipole phased array, the feeding amplitude ratio of each port is counted as I (N) = [ I1, I2, \8230;, I8]. Before the neural network model building process, K groups of randomly selected feed amplitude ratios are used as fixed input values to participate in the training of the neural network. In each group of feed, different fixed values of the six digital phase shifters are selected for the phase difference, so that the array radiation beam achieves the scanning effect. The fixed phase difference is 0 degrees, + -11.25 degrees, + -22.5 degrees, + -45 degrees, + -90 degrees, + -123.75 degrees, + -151.875 degrees and + -174.375 degrees in this order. And carrying out full-wave simulation calculation on the array radiation directional diagrams of the K groups of different port feed amplitude ratios under different phase differences to obtain output values corresponding to each group of inputs.

In the Elman neural network model, the input layer is eight feed portsAnd the network model comprises three hidden layers and a final output layer. The activation function in the hidden layer is a Sigmoid function, and the activation function in the output layer is a linear function. A conjugate gradient algorithm (conjugate gradient) is adopted in the training mode, the K groups of known input and output are substituted into the neural network training, and the training rate is e _l . The prediction precision of the neural network model reaches e after M times of cyclic training _r ；

Step four: regenerating input values I of Kn groups of new port feeding amplitude ratios _n (N), calculating corresponding output values Ep and Eq of the Kn groups of phased array beam radiation pattern by respectively adopting a full-wave simulation method and a neural network model:

calculating the similarity of the two output value matrixes, and setting a threshold value r of the similarity _i This value is set to 0.85 in this example.

If the similarity calculation results of each group are greater than a preset threshold value, constructing a result of the neural network model;

and if at least one group of similarity calculation results are smaller than the preset threshold value, returning to the step three, and retraining the model.

And calculating the normal radiation beam pattern and the low-elevation radiation beam pattern of the eight-element phased array by using the constructed neural network model, wherein the result is basically consistent with the comparison of full-wave simulation.

Therefore, the invention can accurately calculate the radiation beam pattern of the phased array in a large-angle scanning range.

Claims

1. A phased array scanning radiation beam neural network model calculation method is characterized by comprising the following steps:

the method comprises the following steps: establishing a neural network model neuron function, wherein the information from other neurons is x _i And the connection weight with the processed neuron is w _i The inputs to the neuron being processed are then:

the output of the processed neuron is:

step two: establishing connection form between neurons, adopting single-ring feedback system, inputting information x _i (n), feedback information x' _i (n) output information y _i (n); assuming that the system is linear and consists of a forward operator f and a feedback operator g, the output relationship is as follows:

y _i (n)＝f(x’ _i (n))；

x’ _i (n)＝x _i (n)+g(y _i (n))；

the output information is thus:

step three: establishing a neural network model training process: firstly, randomly selecting K groups of random inputs I (N) for a preset array model, and calculating an array radiation directional diagram output value corresponding to each group of inputs by adopting a full-wave simulation mode; then, substituting known K groups of input and output values into the neural network model by adopting a conjugate gradient algorithm, setting a preset training rate, predicting precision and cycle times, and finishing the establishment of the neural network model;

step four: establishing a matrix similarity function:

for newly generated random input I (N), the output values of the array radiation directional diagrams are respectively calculated by adopting a full-wave model simulation method and a neural network model method, the similarity between the output values and the array radiation directional diagrams is calculated by using a similarity function in a matrix form with the same scale, a precision cut-off value is set, and the calculation correctness of the neural network model is verified.