Design Method and Device of N-Frequency Unequal Power Divider Based on BP Neural Network

Abstract

The present invention discloses a design method for an N-frequency unequal power divider based on a BP neural network: obtaining relevant parameters of the N-frequency unequal power divider as training samples of the BP neural network; based on the training samples, using a genetic algorithm to train the number of hidden layer nodes, weights, and bias values of the BP neural network to construct a corresponding improved BP neural network model; using the genetic algorithm to perform a global search for the optimal solution of the improved BP neural network model, so as to obtain the optimal design parameters of the N-frequency unequal power divider. The present invention only needs to model the problem to be solved using an improved BP neural network and use a genetic algorithm for optimization to obtain the optimal power divider design parameters; avoiding the problem that the error function of the traditional BP neural network is easy to fail to find the global optimal point and is easy to fall into a local minimum point during the training process. The present invention also provides a corresponding design device for an N-frequency unequal power divider based on a BP neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency and microwave, and more specifically, relates to a design method and device for an N-frequency unequal power divider based on a BP neural network. Background Technique

[0002] With the rapid development of wireless communication systems, the need to support multi-band communication capabilities in the same system has emerged. For example, systems such as GSM, WCDMA, LTE, 2.4G / 5G / 6G, and Wi-Fi need to operate in a single system. Among them, the power divider is an important part of microwave circuits, which has the functions of signal distribution and combination. However, the traditional Wilkinson power divider can only operate at a single frequency and its odd harmonics, far from meeting the requirements of multi-band, high integration, and low cost. This requires components to operate at three or more frequencies.

[0003] The traditional design methods for multi-frequency power dividers mainly improve the power divider in terms of structure. When calculating the structure parameters based on the transmission line theory, it is necessary to solve transcendental equations, which not only involve complex calculations but also have strict constraints on the characteristic impedance and electrical length of each transmission line. The finally obtained solution is only an approximate solution, so it is impossible to ensure that the structure parameters of the power divider reach the optimal values simultaneously. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement requirements of the prior art, the present invention provides a design method for an N-frequency unequal power divider based on a BP neural network, which can obtain the optimal design parameters of the power divider only according to the relevant parameters of the power divider without being restricted by the initial value and without the need to understand the internal properties of the power divider.

[0005] To achieve the above object, according to one aspect of the present invention, there is provided a design method for an N-frequency unequal power divider based on a BP neural network, including the following steps:

[0006] Obtain the relevant parameters of the N-frequency unequal power divider as the training samples of the BP neural network;

[0007] Based on the training samples, use the genetic algorithm to train the number of hidden layer nodes, weights, and bias values of the BP neural network, and construct a corresponding improved BP neural network model;

[0008] Use the genetic algorithm to perform a global search for the optimal solution of the improved BP neural network model, so as to obtain the optimal design parameters of the N-frequency unequal power divider.

[0009] In an embodiment of the present invention, the N-frequency unequal power divider is composed of N transmission lines and N isolation resistors. Z s, Z2, and Z3 are the impedances of input port ① and two output ports ② and ③ respectively. In a traditional power divider, the quarter-wavelength impedance transformer is replaced by N transmission lines with different characteristic impedances. The characteristic impedances of the N transmission lines are Z 1a , Z 2a , …, Z Na from port ① to port ②, and Z 1b , Z 2b , …, Z Nb from port ① to port ③. Their electrical lengths are θ1, θ2, …, θ N . To achieve good isolation, resistors R1, R2, …, R N are placed between the two output ports. There are the following relationships:

[0010]

[0011] where P3 and P2 are the output powers of port ③ and port ② respectively, and k is a proportionality coefficient used to define the ratio of the output powers of P3 and P2.

[0012] In an embodiment of the present invention, the BP neural network consists of three layers: an input layer, a hidden layer, and an output layer. Based on training samples, the characteristic impedances Z 1a , Z 2a , …, Z Na , electrical lengths θ1, θ2, …, θ N , and isolation resistors R1, R2, …, R N of the input layer are obtained, and an input matrix x: [Z 1a , Z 2a , …, Z Na , θ1, θ2, …, θ N , R1, R2, …, R N is obtained. The sum of the squares F of the S parameters of the training samples obtained by the simulation software is used as the output result y of the output layer.

[0013] In an embodiment of the present invention, the sum of the squares F of the S parameters is specifically:

[0014]

[0015] where f1, f2, …, f N are N frequency points in the network, where S11 is the return loss of port ① and S23 is the isolation between port ② and port ③.

[0016] In an embodiment of the present invention, when S11 and S23 reach their minimum values simultaneously, that is, when F reaches its minimum value, the design parameters of the optimal N-frequency unequal power divider are obtained, namely the characteristic impedances Z 1a , Z 2a, …, Z Na , electrical lengths θ1, θ2, …, θ N , isolation resistances R1, R2, …, R N .

[0017] In an embodiment of the present invention, the number of hidden layer neurons is defined as q, and the set BP neural network structure is m-q-1, that is, there are m nodes in the input layer, q neuron nodes in the hidden layer, and 1 node in the output layer. The number of weights from the input layer to the hidden layer is defined as w1num = m*q, the number of thresholds is q, and the number of weights from the hidden layer to the output layer is defined as w2mum = q*1 = q; the transfer function of the neuron is defined as

[0018] In an embodiment of the present invention, the training samples are specifically as follows:

[0019] First, preprocessing operations are performed on n input matrices x and output results y, where the input matrix is an x*n matrix and the output is a y*n matrix, and n is the number of samples. The processing process is to perform a normalization operation on each item of the input matrix, standardize the data row by row, and map the minimum and maximum values of each row to [-1 1]. The formula is:

[0020]

[0021] Then, the original sample data is split into a training set with a first ratio, a test set with a second ratio, and a validation set with a third ratio. The sum of the first ratio, the second ratio, and the third ratio is 100%.

[0022] In an embodiment of the present invention, in the training stage of the BP neural network, a genetic algorithm is used to update the number of hidden layer nodes, weights, and thresholds of the BP neural network. When the iteration reaches the maximum number of times or the mean square error function error reaches a preset value, the update stops, and the trained number of hidden layer nodes, weights, and thresholds are obtained.

[0023] In an embodiment of the present invention, the obtained trained number of hidden layer nodes, weights, and thresholds are substituted into the BP neural network, and Levenberg-Marquardt is used to train the BP neural network to avoid falling into a local optimum, and the final BP neural network model is obtained.

[0024] According to another aspect of the present invention, there is also provided a design device for an N-frequency unequal power divider based on a BP neural network, including at least one processor and a memory. The at least one processor and the memory are connected through a data bus. The memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to complete the design method of the N-frequency unequal power divider based on the BP neural network.

[0025] Generally speaking, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:

[0026] (1) Compared with the traditional design method, the method of the present invention is not limited by the initial value and does not need to understand the internal properties of the power divider. Only by using the improved backpropagation (BP) neural network of the present invention to model the problem to be solved (such as the impedance and electrical length of each section of the power divider) and using the genetic algorithm for optimization, the optimal design parameters of the power divider can be obtained.

[0027] (2) The method of the present invention can effectively reduce the complexity of the BP neural network structure, avoid the problem that the error function of the traditional BP neural network is easy to fail to find the global optimal point and is easy to fall into the local minimum point during the training process. Thus, the design efficiency is effectively improved, fully demonstrating its superiority. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic flow chart of the design method of the N-frequency unequal power divider based on the BP neural network in the embodiment of the present invention;

[0029] Figure 2 It is the structure of the three-frequency unequal power divider provided by the embodiment of the present invention;

[0030] Figure 3 It is the BP neural network structure provided by the embodiment of the present invention;

[0031] Figure 4 It is the schematic flow chart of the algorithm provided by the embodiment of the present invention;

[0032] Figure 5 It is the S-parameters under the parameter configuration of the three-frequency unequal power divider in Document 1;

[0033] Figure 6 It is the S-parameters under the parameter configuration of the three-frequency unequal power divider obtained by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0035] The present invention combines an improved BP neural network with a genetic algorithm (GA) and applies it to the optimal design of an N-band unequal power divider. Based on previous research, for an N-band unequal Wilkinson power divider operating at N arbitrary frequencies with an arbitrary power distribution ratio, the present invention proposes an optimal design method for a power divider based on the combination of a BP neural network and a genetic algorithm (Improved Back Propagation Genetic Algorithm, IBPGA).

[0036] As Figure 1 shown, the present invention provides a design method for an N-band unequal power divider based on a BP neural network, including the following steps:

[0037] (1) Obtain the relevant parameters of the N-band unequal power divider as the training samples of the BP neural network;

[0038] (2) Based on the training samples, use the genetic algorithm to train the number of hidden layer nodes, weights, and bias values of the BP neural network, and construct a corresponding improved BP neural network model;

[0039] (3) Use the genetic algorithm to perform a global search for the optimal solution of the improved BP neural network model, so as to obtain the optimal design parameters of the N-band unequal power divider.

[0040] Specifically, the present invention takes a three-band unequal Wilkinson power divider as an example to illustrate the method of the present invention. The structure of the three-band unequal Wilkinson power divider is as Figure 2 shown and consists of three transmission lines and three isolation resistors.

[0041] Use the genetic algorithm to train the number of hidden layer nodes, weights, and bias values of the BP neural network, construct a corresponding improved BP neural network model, and finally use the genetic algorithm to perform a global search for the optimal solution of the improved BP neural network model, so as to obtain the optimal design parameters of the three-band unequal power divider.

[0042] The technical solution adopted to solve this technical problem is:

[0043] To achieve the above object, the embodiments of the present invention include the following steps:

[0044] 1. Acquisition of training samples

[0045] As Figure 2 shown, the three-band unequal power divider consists of three transmission lines and three isolation resistors. Z s, Z2, and Z3 are the impedances of input port ① and two output ports ② and ③ respectively. In the traditional power divider, the quarter-wavelength impedance transformer is replaced by three transmission lines with different characteristic impedances. The characteristic impedances of the three transmission lines are Z 1a 、Z 2a 、Z 3a 、from port ① to port ②, and Z 1b 、Z 2b 、Z 3b from port ① to port ③. Their electrical lengths are θ1, θ2, and θ3. To achieve good isolation, resistors R1, R2, and R3 are placed between the two output ports. There are the following relationships:

[0046]

[0047] where P3 and P2 are the output powers of port ③ and port ② respectively, and k is a proportionality coefficient used to define the ratio of the output powers of P3 and P2.

[0048] The circuit of the three-frequency unequal power divider is simulated using a simulation software (such as the electromagnetic simulation software Advanced Design System (ADS)) to obtain the S-parameters of the circuit. Among them, S11 is the return loss of port ①, S22 is the return loss of port ②, and S23 is the isolation between port ② and port ③. The obtained simulation data is used as the input and output of the BP neural network. Among them, the input data of the BP neural network are the characteristic impedances Z 1a 、Z 2a 、Z 3a 、electrical lengths θ1, θ2, and θ3, isolation resistors R1, R2, and R3, a total of 9 parameters; the output is the sum of the squares of the S-parameters at 3 frequency points f1, f2, and f3 in the network, denoted as F:

[0049]

[0050] where N = 3 is the required 3 frequency points. When S11 and S23 reach the minimum values simultaneously, that is, when F reaches the minimum, the design parameters of the optimal three-frequency unequal power divider can be obtained, namely the characteristic impedances Z 1a 、Z 2a 、Z 3a 、electrical lengths θ1, θ2, and θ3, isolation resistors R1, R2, and R3.

[0051] 2. Construction of the improved BP neural network.

[0052] As Figure 3 shown, the BP neural network consists of three layers: the input layer, the hidden layer, and the output layer. Based on the training samples, the characteristic impedances Z 1a 、Z2a , Z 3a , electrical lengths θ1, θ2, θ3, isolation resistances R1, R2, and R3, to obtain the input matrix x:

[0053] [Z 1a , Z 2a , Z 3a , θ1, θ2, θ3, R1, R2, R3]

[0054] Use the sum of squares F (formula (1)) of the training sample S-parameters (the S-parameters are obtained by simulation with ADS simulation software) obtained by ADS simulation as the output result y of the output layer.

[0055] Define the number of neurons in the hidden layer as q, and the set BP neural network structure as m-q-1, that is, there are m nodes in the input layer (at this time m = 9 parameters), q neuron nodes in the hidden layer, and 1 node in the output layer. (This value of q is the optimization objective and is obtained after optimization by the GA algorithm, so an unknown variable needs to be defined currently)

[0056] Define the number of weights from the input layer to the hidden layer w1num = m*q, and the number of thresholds is q (i.e., equal to the number of neurons in the hidden layer);

[0057] Define the number of weights from the hidden layer to the output layer w2mum = q*1 = q, and the number of thresholds is 1;

[0058] Define the transfer function of the neuron as the Sigmoid function

[0059]

[0060] First, perform preprocessing operations on n (number of samples) input matrices x and output results y (the input matrix is an x*n matrix, and the output is a y*n matrix). The processing process is to normalize each item of the input matrix, perform standardization processing on the data row by row, and map the minimum and maximum values of each row to [-1 1]. The formula is:

[0061]

[0062] Then, divide the original sample data into a 70% training set, a 15% test set, and a 15% validation set.

[0063] 3. Training, validation, and optimization of the improved BP neural network

[0064] To find the minimum value of F, the design parameters of the optimal three-frequency unequal power divider, that is, Z 1a , Z 2a , Z 3a, electrical lengths θ1, θ2, θ3, isolation resistances R1, R2, and R3. In this subsection, the constructed improved BP neural network is trained, optimized, and optimized. As Figure 4 shown, it is a schematic flowchart of the design method of a three-frequency unequal power divider based on an improved BP neural network in an embodiment of the present invention:

[0065] (1) Training of the improved BP neural network

[0066] The genetic algorithm is used to update the number of hidden layer nodes, weights, and thresholds of the BP neural network. When the iteration reaches the maximum number of times or the error of the Mean Squared Error (MSE) function reaches a preset value, the update stops, and the trained number of hidden layer nodes, weights, and thresholds are obtained. The specific implementation steps are as follows:

[0067] Step 1: Input the initialized number of hidden layer nodes, weights, and thresholds of the BP neural network to generate an initial population containing individuals with the number of hidden layer nodes, weights, and thresholds;

[0068] Step 2: According to the ADS simulation software, obtain n input matrices x and output results y, and construct an individual fitness MSE function;

[0069] The individual fitness value selects the Mean Squared Error (MSE) function of the performance function of the BP neural network.

[0070]

[0071] where y i and represent the predicted value and the target value of the model, y i is the predicted result value output by the BP neural network, is the target value obtained through the ADS simulation software.

[0072] Step 3: Combine the genetic algorithm to initialize the number of hidden layer nodes, weights, and thresholds of the BP neural network, encode the number of hidden layer nodes, weights, and thresholds, and use the BP neural network to train to obtain the error function MSE.

[0073] In the genetic algorithm, the population size is preset to 50, the maximum number of genetic generations is 100 selection operations, and the preset selection operation is the roulette wheel method;

[0074] Step 4: Perform a crossover operation with a preset crossover probability of 0.9;

[0075] Step 5: Perform a mutation operation with a preset mutation probability of 0.08;

[0076] Step 6 When the set error condition or the maximum number of iterations is met, the genetic algorithm stops updating, and the globally optimal number of network nodes, weights, and thresholds are obtained.

[0077] It should be noted that the settings of the above parameters are only for examples, and can be flexibly set according to needs and actual situations in practice.

[0078] (2) Prediction of the final BP neural network model

[0079] Substitute the globally optimal number of hidden layer nodes, weights, and thresholds obtained in section (1) into the BP neural network to obtain the final BP neural network model.

[0080] To avoid the problem of falling into local optimum, the present invention uses Levenberg-Marquardt to train the BP neural network, and the specific implementation process is as follows.

[0081] Step 1 Set the maximum number of iterations to 500, the performance target to 0, the maximum number of verification failures to 10, and the maximum value of the Levenberg-Marquardt algorithm error mu to 10 10 .

[0082] The constraint conditions are set such that the weights and thresholds of the globally optimal hidden layer are between -3 and +3, and the number of hidden layer nodes is between 5 and 100.

[0083] Step 2 Optimize based on the mean square error MSE function in Levenberg-Marquardt

[0084] Step 3 Obtain the final number of hidden layer nodes, weights, and thresholds, and then substitute them into the BP neural network to obtain the final BP neural network model.

[0085] It should be noted that using the Levenberg-Marquardt algorithm to train the BP neural network model proposed by the present invention can achieve the optimal training results and convergence, and is suitable for solving the problems proposed by the present invention.

[0086] It should be noted that the settings of the above parameters are only for examples, and can be flexibly set according to needs and actual situations in practice.

[0087] (3) Optimize the BP neural network model using the genetic algorithm

[0088] This part mainly uses the genetic algorithm to perform a global search for the optimal solution of the output F (Formula 1) of the improved BP neural network model, and takes F as the individual fitness value. When F is globally minimized, the optimal design parameters of the triple-frequency unequal power divider can be obtained.

[0089] Step 1 Initialize the genetic algorithm with a preset population size of 50 and a maximum number of genetic generations of 100;

[0090] Step 2 Construct a fitness function based on the output F of the BP neural network model;

[0091] Step 3 Selection operation, with the preset selection operation being the roulette wheel method

[0092] Step 4 Perform the crossover operation with a preset crossover probability of 0.9;

[0093] Step 5 Perform the mutation operation with a preset mutation probability of 0.08;

[0094] Step 6 When F converges to the minimum value or reaches the maximum number of iterations, the genetic algorithm stops updating, and the structural parameters [Z 1a 、Z 2a 、Z 3a 、θ1, θ2, θ3, R1, R2, R3] of the globally optimal three - frequency unequal - power divider are obtained.

[0095] Obviously, the present invention can also be used for the design of a four - frequency unequal - power divider.

[0096] It should be noted that the settings of the above - mentioned various parameters are only for examples, and can be flexibly set according to needs and actual situations in practice.

[0097] The four - frequency unequal - power divider can be composed of four transmission lines and 4 isolation resistors. Z s , Z2, Z3 are the impedances of the input port ① and the two output ports ②, ③ respectively. In the traditional power divider, the quarter - wavelength impedance transformer is replaced by four transmission lines with different characteristic impedances. The characteristic impedance of the four - section transmission line from port ① to port ② is Z 1a 、Z 2a 、Z 3a 、Z 4a The characteristic impedance from port ① to port ③ is Z 1b 、Z 2b 、Z 3b 、Z 4b , and its electrical lengths are θ1, θ2, θ3, and θ4. In order to achieve good isolation, resistors R1, R2, R3, and R4 are placed between the two output ports. There are the following relationships:

[0098]

[0099] Among them, P3 and P2 are the output powers of port ③ and port ②, and k is a proportionality coefficient used to define the ratio of the output powers of P3 and P2.

[0100] The four - frequency unequal power divider circuit is simulated using the electromagnetic simulation software Advanced Design System (ADS) to obtain the S - parameters of the circuit. Among them, the input data of the BP neural network are the characteristic impedances Z 1a 、Z 2a 、Z 3a 、Z 4a 、electrical lengths θ1, θ2, θ3, and θ4, isolation resistors R1, R2, R3, and R4. There are a total of 12 parameters; the output is the sum of the squares of the S - parameters at 4 frequency points f1, f2, f3, and f4 in the network, denoted as F:

[0101]

[0102] where N = 4 is the required 4 frequency points. Based on the above architecture, using the method of the present invention, when S11 and S23 reach the minimum values simultaneously, that is, when F reaches the minimum, the design parameters of the optimal four - frequency unequal power divider can be obtained, namely the characteristic impedances Z 1a 、Z 2a 、Z 3a 、Z 4a 、electrical lengths θ1, θ2, θ3, and θ4, isolation resistors R1, R2, R3, and R4.

[0103] It should be noted that: Since in the current design applications of unequal power dividers, applications above 4 frequencies are very few. When actually applying an unequal power divider above 4 frequencies, the frequency interval between multiple frequency points needs to be considered. If the interval is too small, it is extremely easy to become a broadband filtering unequal power divider, and the ability to suppress useless frequency points is greatly weakened, and the application scenario range is reduced. Therefore, the present invention focuses on elaborating the three - frequency and four - frequency unequal power divider schemes, and this idea can be extended to the design of the N - frequency unequal power divider scheme and is still applicable.

[0104] Furthermore, the present invention also provides an N - frequency unequal power divider design device based on a BP neural network, including at least one processor and a memory. The at least one processor and the memory are connected through a data bus. The memory stores instructions executable by the at least one processor. After being executed by the processor, the instructions are used to complete the above - mentioned N - frequency unequal power divider design method based on a BP neural network.

[0105] 4. Simulation and Verification

[0106] An embodiment of the present invention provides a design method for a three - band unequal power divider based on an improved BP neural network, which can be applied to the design of a three - band unequal power divider. In this part, a three - band unequal power divider with a power ratio of 2:1 (k = 2) is designed using this method. The three frequencies are required to be f1 = 1 GHz, f2 = 2 GHz, and f3 = 3 GHz. Among them, Zs = 50 Ω, so that Z2 = 70.7106 Ω and Z3 = 35.3553 Ω can be obtained to verify the effectiveness of the proposed design method. (K is a known value, which can be 2, 3, 4, etc. according to the power ratio requirement, and Zs is a known value; thus, Z2 and Z3 are derived from K and Zs; Z1a, Z2a, Z3a, Z1b, Z2b, and Z3b are all unknown, but they are related through the known K value). The specific simulation steps are as follows:

[0107] (1) Construct an improved BP neural network

[0108] Construct the input layer, hidden layer, and output layer of the BP neural network. The input feature impedances Z 1a 、Z 2a 、Z 3a , electrical lengths θ1, θ2, θ3, isolation resistances R1, R2, and R3 are input to the input layer, a total of 9 parameters. The set BP neural network structure is 9 - q - 1, that is, 9 nodes in the input layer, q neuron nodes in the hidden layer, and 1 node in the output layer. The output of the output layer is the sum of the squares of the S - parameters F (Formula 1).

[0109] (2) Acquisition of training samples

[0110] Use the electromagnetic simulation software ADS to simulate the circuit to obtain simulation data as the input and output of the BP neural network. The input data are the feature impedances Z 1a 、Z 2a 、Z 3a , electrical lengths θ1, θ2, θ3, isolation resistances R1, R2, and R3; the output data is the sum of the squares of the S - parameters F (Formula 1).

[0111] (3) Training, optimization, and optimization of the improved BP neural network

[0112] 1) Based on the training samples obtained from the ADS simulation software in step (2), that is, n input matrices x and output results y, construct an individual fitness MSE function. Use the genetic algorithm to update the number of hidden layer nodes, weights, and thresholds of the BP neural network. Use the BP neural network to train to obtain the MSE function. Stop updating when the number of iterations reaches the maximum value or the MSE error reaches the preset value, and obtain the trained number of hidden layer nodes, weights, and thresholds.

[0113] 2) Substitute the number of global optimal hidden layer nodes, weights, and thresholds into the BP neural network, and use the Levenberg-Marquardt algorithm to train the BP neural network to obtain the final BP neural network model.

[0114] Among them, the mean square error MSE function of the individual fitness value is used to calculate the error, the maximum number of iterations is set to 500, the performance target is set to 0, the maximum number of validation failures is set to 10, and the maximum value of mu is set to 10. 10 . The weights and thresholds are between -3 and +3, and the number of hidden layer nodes is between 5 and 100. Substitute the number of hidden layer nodes, weights, and thresholds into the BP neural network, and finally obtain the final BP neural network model.

[0115] The simulation results are shown in Table 1, indicating that the regression R values obtained by the proposed algorithm IGABP of the present invention are all above 90%, representing a strong correlation between the predicted output and the target output, and proving that the established BP neural network model is very accurate.

[0116] Table 1 Regression R values of the BP neural network

[0117]

[0118] 3) Take the output F (Formula 1) of the improved BP neural network model as the fitness function, and use the genetic algorithm to globally search for the optimal solution of the fitness function F (Formula 1). Take the characteristic impedance Z 1a , Z 2a , Z 3a , electrical lengths θ1, θ2, θ3, isolation resistances R1, R2, and R3, a total of 9 parameters as the initial population, where the following conditions are met:

[0119] Z 1a > Z 2a > Z 3a > 0;

[0120] 0 < qi| i＝1,2,3 < 90;

[0121] In the genetic algorithm part, the preset population size is 50, the maximum number of genetic generations is 100, the selection operation is the roulette wheel method, the crossover probability is 0.9, and the mutation probability is 0.08. Continuously iterate. When F reaches the global minimum, the optimal design parameters of the three-frequency unequal power divider can be obtained.

[0122] (4) Simulation results

[0123] Substitute the optimal design parameters of the three-frequency unequal power divider obtained in step (3) into the ADS simulation software to obtain the corresponding S parameters, Figure 5 and Figure 6For the S-parameter comparison results, the simulation results show that S11, S22, S33, and S23 obtained from the ideal transmission line model are superior to the design method in Reference [1] at the three required frequency points f1, f2, and f3.

[0124] The following table shows the comparison between the optimal solution obtained by the method of the present invention and the value obtained by the method in Reference [1].

[0125] Table 2 Comparison of Optimal Solutions

[0126]

[0127]

[0128] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

[0129] References

[0130] [1] A.Qaroot and N.Dib and A.Gheethan, “DESIGN METHODOLOGY OF MULTI-FREQUENCY UN-EQUALSPLIT WILKINSON POWER DIVIDERS USING TRANSMISSION LINETRANSFORMERS,” Progress In Electromagnetics Research B, Vol.22, 1–21, 2010.

Claims

1. A design method of an N-frequency unequal power divider based on a BP neural network, characterized in that, It includes the following steps: Obtain the relevant parameters of the N-frequency unequal power divider as the training samples of the BP neural network; among them, the N-frequency unequal power divider includes an input port ① and two output ports ② and ③; use the sum of the squares of the S-parameters F obtained by the simulation software as the output result y of the output layer. The sum of the squares of the S-parameters F is specifically: , where f 1, f 2,..., f N are the N frequency points in the network, where S11 is the return loss of port ① and S23 is the isolation between port ② and port ③; Based on the training samples, use the genetic algorithm to train the number of hidden layer nodes, weights, and bias values of the BP neural network, and construct a corresponding improved BP neural network model; Use the genetic algorithm to perform a global search for the optimal solution of the improved BP neural network model, so as to obtain the optimal design parameters of the N-frequency unequal power divider.

2. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 1, characterized in that, The N-way unequal power divider is composed of N transmission lines and N isolation resistors. Z s , Z 2, Z 3 are the impedances of the input port ① and the two output ports ② and ③ respectively. In the traditional power divider, the quarter-wavelength impedance transformer is replaced by N transmission lines with different characteristic impedances. Among them, the characteristic impedances of the N transmission lines from port ① to port ② are Z 1a , Z 2a , …, Z Na , and the characteristic impedances from port ① to port ③ are Z 1b , Z 2b , …, Z Nb , and their electrical lengths are θ 1, θ 2, …, θ N . In order to achieve good isolation, resistors R 1, R 2, …, R N are placed between the two output ports, and there are the following relationships: Among them, P 3 and P 2 are the output powers of port ③ and port ②, k is the proportionality coefficient, used to define P 3 and P the ratio of the output powers of 2.

3. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 2, characterized in that The BP neural network consists of three layers: an input layer, a hidden layer, and an output layer. Based on the training samples, the characteristic impedance Z of the input layer is obtained 1a 、Z 2a 、…、Z Na , the electrical length θ 1、 θ 2、…、 θ N , the isolation resistance R 1、 R 2、…、 R N , and the input matrix x is obtained: [Z 1a 、Z 2a 、…、Z Na 、 θ 1、 θ 2、…、 θ N 、 R 1、 R 2、…、 R N .

4. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 1, characterized in that, When both S11 and S23 reach their minimum values, that is, when F reaches its minimum, the design parameters of the optimal N-way unequal power divider are obtained, namely the characteristic impedances Z 1a , Z 2a , …, Z Na , the electrical lengths θ 1, θ 2, …, θ N , and the isolation resistances R 1, R 2, …, R N .

5. The design method of an N-frequency unequal power divider based on a BP neural network according to claim 3, wherein Define the number of neurons in the hidden layer as q, and the set BP neural network structure as m-q-1, that is, there are m nodes in the input layer, q neuron nodes in the hidden layer, and 1 node in the output layer. Define the number of weights from the input layer to the hidden layer w1num = m * q, the number of thresholds as q, and define the number of weights from the hidden layer to the output layer w2mum = q * 1 = q; Define the transfer function of the neuron as the Sigmoid function .

6. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 1 or 2, characterized in that, The specific training samples are as follows: First, perform preprocessing operations on n input matrices x, where n is the number of samples. The processing process is to normalize each item of the input matrix, perform standardization processing on the data row by row, and map the minimum and maximum values of each row to [-1, 1]. The formula is: y is the output result. Then, the original sample data is split into a training set with a first ratio, a test set with a second ratio, and a validation set with a third ratio. The sum of the first ratio, the second ratio, and the third ratio is 100%.

7. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 1 or 2, characterized in that, In the training stage of the BP neural network, use the genetic algorithm to update the number of hidden layer nodes, weights, and thresholds of the BP neural network. Stop updating when the iteration reaches the maximum number of times or the mean square error function error reaches the preset value, and obtain the trained number of hidden layer nodes, weights, and thresholds.

8. The design method of the N-frequency unequal power divider based on the BP neural network according to claim 7, characterized in that, Substitute the obtained trained number of hidden layer nodes, weights, and thresholds into the BP neural network, and use Levenberg-Marquardt to train the BP neural network to avoid falling into local optima, and obtain the final BP neural network model.

9. An N-frequency unequal power divider design device based on a BP neural network, characterized in that: It includes at least one processor and a memory. The at least one processor and the memory are connected through a data bus. The memory stores instructions that can be executed by the at least one processor. After the instructions are executed by the processor, they are used to complete the N-frequency unequal power divider design method according to any one of claims 1-8.

Images