General multi-stage power divider performance prediction method based on deep neural network

Abstract

The invention belongs to the technical field of microwave and millimeter wave device design, and provides a universal multi-stage power divider performance prediction method based on a deep neural network. The method comprises the following steps: firstly, carrying out S parameter acquisition based on single-stage power dividers, cascading the single-stage power dividers to obtain a multi-stage power divider, and then carrying out S parameter acquisition on the multi-stage power divider; then, data cleaning is carried out on the collected S parameters, normalization is carried out to form a training set, a deep neural network is trained according to the training set, and a multi-stage power divider performance prediction model is obtained; and finally, inputting the S parameter sequence of the single-stage power divider of the multi-stage power divider to be predicted into the multi-stage power divider performance prediction model, and outputting an S parameter prediction value of the multi-stage power divider to be predicted by the multi-stage power divider performance prediction model. In conclusion, the multi-stage cascade performance is quickly predicted through the S parameter of the single-stage power divider, and the problems that an existing method is time-consuming in calculation, large in data size, poor in model reusability and the like are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave and millimeter wave device design, and specifically provides a general multi-stage power splitter performance prediction method based on deep neural network. Background Art

[0002] In the design of microwave passive circuit components, electromagnetic simulation software is often used to analyze the performance of multi-stage power divider cascades. However, this method consumes large amounts of computational resources, and optimizing the structural parameters of multi-stage power dividers requires repeated simulations, with the time cost increasing exponentially with the number of stages. To address this problem, an artificial intelligence-based performance prediction method for multi-stage power dividers has been proposed. Currently, power divider performance prediction methods based on artificial neural networks (ANNs) predict performance based on input structural parameters. These methods are only applicable to single-stage structures and cannot handle the multi-pole nonlinear response of multi-stage cascades. Convolutional neural network (CNN)-based methods design devices by segmenting structural images, but are limited by mesh accuracy and data volume explosion, and processing errors can lead to reduced prediction accuracy. Therefore, a data-efficient performance prediction method for multi-stage power dividers that is independent of specific structures is urgently needed. Summary of the Invention

[0003] The purpose of the present invention is to provide a general multi-stage power splitter performance prediction method based on deep neural network, which quickly predicts the multi-stage cascade performance through the S parameters of the single-stage power splitter, solving the problems of time-consuming calculation, large data volume, and poor model reusability of the existing methods.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] A general multi-stage power splitter performance prediction method based on deep neural network, characterized by comprising the following steps:

[0006] Step S1: S-parameter acquisition is performed based on a single-stage power divider structure, and the single-stage power divider is cascaded to obtain a multi-stage power divider, and then S-parameter acquisition is performed on the multi-stage power divider;

[0007] Step S2: Clean the collected S parameters to remove noise and outliers, and normalize the data to form a training set;

[0008] Step S3: construct a deep neural network, train the deep neural network according to the training set, and obtain a multi-stage power splitter performance prediction model after the training is completed;

[0009] Step S4: inputting the S parameter sequence of a single-stage power splitter of the multi-stage power splitter to be predicted into the multi-stage power splitter performance prediction model, and the multi-stage power splitter performance prediction model outputs the S parameter prediction value of the multi-stage power splitter to be predicted.

[0010] Further, in step S1, the multi-stage power divider adopts N-stage cascade, N is the cascade number of the target multi-stage power divider; the structure parameters of each single-stage unit in the multi-stage power divider are the same, and each stage output port is connected in accordance with the port impedance matching rule of Wilkinson power divider.

[0011] Further, in step S1, the S parameter acquisition process is:

[0012] For each structure parameter of the single-stage power divider, set its value range and sampling interval, and perform parameter scanning on all structure parameters to form a plurality of structure parameter arrays;

[0013] For each structure parameter array, model the single-stage power divider and the multi-stage power divider through an electromagnetic simulation software, obtain the S parameters of the single-stage power divider and the multi-stage power divider under the group of structure parameters through simulation, and the data acquisition range covers the target frequency band.

[0014] Further, the normalization interval of step S2 is [-1, 1].

[0015] Further, in the training set of step S2, the S parameter sequence of the single-stage power divider is taken as the input feature vector X, and the S parameter sequence of the multi-stage power divider is taken as the output label Y to form a training sample.

[0016] Further, in step S3, the deep neural network adopts a multi-layer fully connected neural network structure, the dimensions of the input layer and the output layer match the length of the S parameter sequence, and the hidden layer is 4 layers.

[0017] Further, in step S3, the training process is: taking mean square error (MSE) as the loss function, using the back propagation algorithm to optimize the network parameters, and completing the training.

[0018] Based on the above technical scheme, the present application has the advantages that a general multi-stage power divider performance prediction method based on deep neural network is provided, which can quickly predict the multi-stage cascade performance through the S parameters of the single-stage power divider; and the specific advantages are as follows:

[0019] 1) Significantly improve the calculation efficiency and shorten the design cycle of the multi-stage power divider;

[0020] The traditional multi-stage power divider full-wave simulation takes more than 3 hours and 30 minutes, the present application quickly predicts through single-stage simulation combined with deep neural network (DNN), and the overall time consumption is compressed to within 1 minute, the speed is improved by more than 3 times; avoids the exponential computational complexity of multi-stage structure parameter optimization, supports real-time iterative adjustment of design parameters, and meets the rapid design requirements of high-frequency complex systems;

[0021] 2) The data demand is greatly reduced, and the engineering practicability is enhanced;

[0022] The traditional artificial intelligence method needs thousands to tens of thousands of structure parameters - performance mapping samples, and the method only needs 450 groups of single-stage power divider S parameter data to realize effective training, and the data amount is reduced by more than 80 %; the data preprocessing is simplified to S parameter normalization, without complex structure parameter cleaning, so that the data preparation cost is significantly reduced, and the method is especially suitable for engineering scenes of Wilkinson power divider types with different topologies;

[0023] 3) Reduce the consumption of computing resources, and realize green design;

[0024] Compared with the traditional method, the method can save more than 60 % of computing resources by reducing the number of full-wave simulation times and optimizing the model complexity, reduce the hardware cost and energy consumption, and is suitable for scenes sensitive to lightweight design such as satellite communication load and 5G base station antenna feed network, and is conducive to promoting the development of microwave technology in the direction of high efficiency and low consumption. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 It is a flowchart of the multi-stage power divider performance prediction method based on artificial intelligence in the application.

[0026] Figure 2 It is a structure diagram of a single-stage power divider in the embodiment of the application.

[0027] Figure 3 It is a structure diagram of a multi-stage power divider in the embodiment of the application.

[0028] Figure 4 It is a structure diagram of a deep neural network in the embodiment of the application.

[0029] Figure 5 It is a comparison diagram of return loss (S11) prediction results and full-wave simulation results of a 1-to-4 power divider in the embodiment of the application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and beneficial effects of the application more clear and apparent, the application will be further described in detail below with reference to the drawings and embodiments.

[0031] The embodiment provides a general multi-stage power divider performance prediction method based on a deep neural network, and the flowchart is as shown in Figure 1 The method specifically comprises the following steps:

[0032] Step S1: S parameter acquisition based on a single-stage power divider structure, and obtaining a multi-stage power divider by using a single-stage power divider for cascade, and then performing S parameter acquisition on the multi-stage power divider;

[0033] The multi-stage power splitter adopts an N-stage cascade, where N is the cascade number of the target multi-stage power splitter, for example, a 2-stage cascade power splitter corresponds to a 1-to-4 power splitter. The structural parameters of each single-stage unit in the multi-stage power splitter are the same, and the port impedance matching rule of the Wilkinson power splitter is followed, and the output ports of each stage are connected through a 50Ω transmission line.

[0034] like Figure 2 The figure shows a schematic diagram of the structure of a single-stage power divider in this embodiment, wherein the structural parameters include the length and width of each microstrip line; for each structural parameter, its value range and sampling interval are set (e.g., the value range of the structural parameter L3 is 1 mm to 4.4 mm, and the sampling interval is 0.2 mm), and all structural parameters are scanned (e.g., the structural parameter L3: starting from 1 mm, the sampling points are set at intervals of 0.2 mm, 1 mm, 1.2 mm, 1.4 mm, ..., until 4.4 mm), and several structural parameter arrays are formed; for each structural parameter array, single-stage power dividers and multi-stage power dividers are modeled using electromagnetic simulation software, as shown in FIG. Figure 3 The figure shows the structure of a two-stage cascade power splitter. The S parameters (such as input port return loss (S11) and transmission coefficient (S21) of each output port) of the single-stage power splitter and the multi-stage power splitter under this set of structural parameters are obtained through simulation. The data acquisition range covers the target frequency band, which is specifically 6 GHz to 14 GHz in this embodiment.

[0035] Based on this, by changing the structural parameters such as the length and width of the microstrip line, the corresponding S parameters are obtained. The microstrip line width (such as W1 and S2) affects the characteristic impedance, and the microstrip line length (such as L1 and L3) affects the phase delay. The mapping relationship between the structural parameters (geometric dimensions) and the S parameters is established through electromagnetic simulation. Among them, the S parameters contain amplitude and phase information, reflecting the electromagnetic response characteristics of the single-stage power divider and its corresponding multi-stage power divider. In this embodiment, a total of no less than 450 sets of data are obtained;

[0036] Step S2: Clean the collected S parameters to remove noise and outliers, and normalize the data to the [-1, 1] range;

[0037] The S-parameter data collected by the simulation software has a large span, ranging from -3dB to -50dB, which is not conducive to neural network training. Therefore, it is necessary to normalize the data to the range of [-1,1] and remove noise and abnormal data.

[0038] Normalize the S parameter data collected in step S1, map the value range to [-1, 1], use the S parameter sequence of the single-stage power splitter as the input feature vector X, and the S parameter sequence of the multi-stage power splitter as the output label Y to form a training set;

[0039] Step S3: construct a deep neural network, train the deep neural network according to the training set, and obtain a multi-stage power splitter performance prediction model after the training is completed;

[0040] A multi-layer fully connected neural network structure is used. The input layer dimension matches the length of the S-parameter sequence of a single-stage power splitter. The hidden layer has four layers, and the output layer dimension matches the S-parameter prediction value of a multi-stage power splitter. The model parameters are optimized using a backpropagation algorithm, with mean square error (MSE) as the loss function. The model is trained until the validation set meets the preset conditions.

[0041] In this embodiment, a multi-layer fully connected neural network is Figure 4 As shown in the figure, the dimensions of the input and output layers match the 81-dimensional data of the power splitter S parameters (81 points sampled at 0.1 GHz intervals in the range of 6 to 14 GHz). Each hidden layer contains 1000 neurons to ensure that the S parameters with strong nonlinearity can be fitted. The activation function is ReLu. During the training process, the optimizer is Adam (learning rate 0.001), the training epochs are 500, and the batch size is 45.

[0042] Step S4: inputting the S parameter sequence of a single-stage power splitter of the multi-stage power splitter to be predicted into the multi-stage power splitter performance prediction model, and the multi-stage power splitter performance prediction model outputs the S parameter prediction value of the multi-stage power splitter to be predicted;

[0043] In this embodiment, a test set is used to verify the consistency between the S parameter prediction values ​​output by the multi-stage power splitter performance prediction model and the simulation results;

[0044] After deep neural network training, the training set MSE = 0.01198, MAE = 0.0003; the validation set MSE = 0.0207, MAE = 0.0015. The return loss (S11) prediction results of the 1-to-4 power splitter in the embodiment of the present invention are compared with the full-wave simulation results. Figure 4 As shown, a comparison between the predicted curve and the full-wave simulation shows that the key frequency points and pole positions have a degree of consistency of >95%. In addition, based on the return loss (S11) prediction results of the 1-to-4 power splitter, the unit structure parameters of the 1-to-4 power splitter selected in this embodiment are shown in Table 1, thereby realizing the structural design of the 1-to-4 power splitter.

[0045] Table 1

[0046]

[0047]

[0048] The above description is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A general multi-stage power splitter performance prediction method based on deep neural network, characterized in that: The following steps are involved: Step S1: S-parameter acquisition is performed based on a single-stage power divider structure, and the single-stage power divider is cascaded to obtain a multi-stage power divider, and then S-parameter acquisition is performed on the multi-stage power divider; Step S2: Clean the collected S parameters to remove noise and outliers, and normalize the data to form a training set; Step S3: construct a deep neural network, train the deep neural network according to the training set, and obtain a multi-stage power splitter performance prediction model after the training is completed; Step S4: inputting the S parameter sequence of a single-stage power splitter of the multi-stage power splitter to be predicted into the multi-stage power splitter performance prediction model, and the multi-stage power splitter performance prediction model outputs the S parameter prediction value of the multi-stage power splitter to be predicted.

2. The general multi-stage power splitter performance prediction method based on deep neural network according to claim 1 is characterized in that: In step S1, the multi-stage power splitter adopts N-stage cascade, where N is the cascade number of the target multi-stage power splitter; the structural parameters of each single-stage unit in the multi-stage power splitter are the same, and the output ports of each stage are connected according to the port impedance matching rule of the Wilkinson power splitter.

3. The general multi-stage power splitter performance prediction method based on deep neural network according to claim 1 is characterized in that: In step S1, the S parameter acquisition process is as follows: For each structural parameter of the single-stage power divider, a value range and a sampling interval are set, all structural parameters are scanned, and a plurality of structural parameter arrays are formed by combining them; For each structural parameter array, single-stage power dividers and multi-stage power dividers are modeled using electromagnetic simulation software. The S parameters of the single-stage power dividers and multi-stage power dividers under this set of structural parameters are obtained through simulation, and the data acquisition range covers the target frequency band.

4. The general multi-stage power splitter performance prediction method based on deep neural network according to claim 1 is characterized in that: The normalization interval of step S2 is [-1, 1].

5. The general multi-stage power splitter performance prediction method based on deep neural network according to claim 1 is characterized in that: In the training set of step S2, the S parameter sequence of the single-stage power splitter is used as the input feature vector X, and the S parameter sequence of the multi-stage power splitter is used as the output label Y to constitute the training samples.

6. The method for predicting performance of a general multi-stage power splitter based on a deep neural network according to claim 1, characterized in that: In step S3, the deep neural network adopts a multi-layer fully connected neural network structure, the dimensions of the input layer and the output layer match the length of the S parameter sequence, and the hidden layer is 4 layers.

7. The method for predicting performance of a general multi-stage power splitter based on a deep neural network according to claim 1, characterized in that: The training process in step S3 is: using mean square error (MSE) as the loss function, optimizing network parameters using the back propagation algorithm to complete the training.

Images