A machine learning optimization method for antenna decoupling structure based on masked autoencoder

Abstract

The present invention discloses a machine learning optimization method for an antenna decoupling structure based on a masked autoencoder, which belongs to the technical field of microwave antenna design. The method of the present invention comprises the following steps: generation of sample data, classification of sample data, pre-training of a neural network, judging whether the generated decoupling structure meets the optimization index, and fine-tuning the neural network according to design requirements. The method of the present invention combines machine learning technology, analyzes and learns the shape features of the decoupling structure in the sample set, extracts the shape features, and hands them over to the autoencoder neural network for learning, so that the trained neural network model has the ability to design a high-performance antenna decoupling structure; can be used for the optimization and design of various planar electromagnetic structures; can optimize the corresponding decoupling structure according to the isolation requirements, and use it for the intelligent design of antennas, which can significantly improve the efficiency of antenna design.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave antenna design, and in particular relates to a machine learning optimization method for an antenna decoupling structure based on a masked autoencoder. Background Art

[0002] Masked Autoencoders (MAE) are a type of autoencoder neural network with image processing capabilities. Like general autoencoders, masked autoencoders have important applications in tasks such as feature extraction and noise elimination. The special feature of masked autoencoders is that they can receive incomplete input information and then complete the input information according to actual requirements. Therefore, masked autoencoders can also be used for computer vision tasks such as image recognition and image restoration. Using masked autoencoders for the intelligent design of planar electromagnetic structures is an important means to achieve intelligent electromagnetic design by "masking" the electromagnetic structure to be designed and then using the autoencoder to specify the generation of an electromagnetic structure with target performance.

[0003] The prior art "Equivalent circuit theory-assisted deep learning for accelerated generative design of metasurfaces" discloses a method for generating metasurfaces based on equivalent circuit theory. Metasurface image samples with specific performance are generated in a targeted manner through equivalent circuit theory, and then handed over to a general autoencoder for learning. During the use phase of the model, the intermediate variables of the autoencoder are optimized with the help of a genetic algorithm to obtain a metasurface unit with target performance. However, this method requires the designer to have considerable electromagnetic theoretical knowledge to specifically design training samples, and the metasurfaces designed by this method have poor diversity and cannot complete all optimization tasks well.

[0004] The prior art "Prior-Knowledge-Guided Deep-Learning-Enabled Synthesis for Broadband and Large Phase Shift Range Metacells in Metalens Antenna" discloses a method for designing metasurfaces using a generative adversarial network. By training the metasurface, a metasurface unit that meets the phase requirements in a specific frequency band is generated. However, this method can only complete the design of a single optimization target (such as phase), and the designer needs to have corresponding prior knowledge during the optimization process. Summary of the invention

[0005] The object of the present invention is to overcome the defects of the above-mentioned prior art and provide a machine learning optimization method for an antenna decoupling structure based on a masked autoencoder.

[0006] The technical problem proposed by the present invention is solved in this way:

[0007] A machine learning optimization method for an antenna decoupling structure based on a masked autoencoder comprises the following steps:

[0008] Step 1. Generate sample data

[0009] Batch generate N random shapes as the shape of the decoupling structure, N is a positive integer, and use the top view of the two-unit microstrip antenna with different decoupling structure shapes and the same other size parameters as the input data of the sample set;

[0010] Electromagnetic simulation software is used to perform batch simulations on N two-unit microstrip antennas with different decoupling structure shapes and the same other size parameters, and N groups of reflection coefficients and transmission coefficients within a set frequency band are obtained as sample set data;

[0011] The sample data in the sample set is divided into training data and test data;

[0012] Step 2. Classification of sample data

[0013] Step 2-1. Determine the number of categories to be classified according to the range of the set frequency band and the numerical distribution of the reflection coefficient and the transmission coefficient in the sample set;

[0014] Step 2-2. Classify all samples in the sample set based on K-means;

[0015] Step 2-3. Label each sample with its category number;

[0016] Step 3. Use the image of the category number and the randomly masked isolated structure in the training sample as the input of the autoencoder neural network, and the image of the shape of the complete isolated structure as the output of the neural network to pre-train the neural network.

[0017] Step 4. Determine whether the output structure of the trained model meets the optimization index;

[0018] Step 4-1. Determine the initial structure, run the model once, and obtain the complete structure of the model output;

[0019] Step 4-2. Use simulation software to verify whether the output structure meets the optimization index. If it does, the optimization process ends. If not, go to step 5.

[0020] Step 5. Fine-tune the pre-trained neural network according to design requirements;

[0021] Step 5-1. Collect the newly generated samples and classify them based on the existing samples based on K-nearestneighbor;

[0022] Step 5-2. Use all samples of the category where the new sample belongs and the category where the optimization target belongs to to continue training the decoding network;

[0023] Step 5-3. Go to step 4;

[0024] Furthermore, when running the model in step 4-1, the current category label of the initial structure after passing through the encoding network is replaced with the target category label.

[0025] Furthermore, when fine-tuning the model in step 5-2, only the free parameters of the decoding network are updated, and the free parameters of the encoding network are not updated.

[0026] The beneficial effects of the present invention are:

[0027] The method of the present invention uses a masked autoencoder neural network to directly perform this on the part of the electromagnetic structure that needs to be designed and optimized, and then redesigns and restores the masked part according to the target requirements by calling the trained neural network model. The present invention can be used to design any planar electromagnetic structure; the present invention can generate a specific electromagnetic structure according to the target requirements, and use it for the intelligent design of antennas, which can significantly improve the efficiency of antenna design. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0029] Figure 2 It is a schematic diagram of the structure of the masked autoencoder neural network;

[0030] Figure 3 Schematic diagram of some training data;

[0031] Figure 4 Flowchart of shape changes during optimization for a set of test specimens;

[0032] Figure 5 This is a comparison chart of the isolation between the antenna units with and without isolation structures after optimization. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0034] This embodiment provides a neural network modeling method for antenna shape based on computer vision, and its flow chart is as follows: Figure 1 As shown, the following steps are included:

[0035] Step 1. Generate sample data

[0036] The two-unit microstrip patch antenna includes a radiation patch, a dielectric substrate and a metal floor; the radiation patch is located on the upper surface of the dielectric substrate, and the metal floor is located on the lower surface of the dielectric substrate;

[0037] Batch generate 2000 random shapes as the shapes of the decoupling structure, use the S parameters of the corresponding decoupling shapes as the input data of the sample set, and the decoupling structure shape diagram as the output data of the sample set;

[0038] The commercial electromagnetic simulation software CST was used to batch simulate 2,000 microstrip antennas with different decoupling structure shapes and the same other size parameters, and 2,000 sets of reflection coefficients and transmission absorption curves within the set frequency band of 1.5 GHz to 3.5 GHz were obtained as sample set data;

[0039] 1800 sample data in the sample set are used as training data. The schematic diagram of some training data is as follows Figure 3 As shown, other sample data are used as test data;

[0040] Step 2. Classification of sample data

[0041] Step 2-1. Take the reflection coefficient in the sample set and the transmission coefficient of the corresponding frequency band as characteristic parameters;

[0042] Step 2-2. Use the K-means algorithm to divide all samples in the sample set into 20 categories according to their characteristic parameters;

[0043] Step 2-3. Number each sample from 1 to 20 according to its classification result;

[0044] Step 3. Use the image of the category number and the randomly masked isolated structure in the training sample as the input of the autoencoder neural network, and the image of the shape of the complete isolated structure as the output of the neural network to pre-train the neural network.

[0045] Step 4. Determine whether the output structure of the trained model meets the optimization index;

[0046] Step 4-1. Determine the initial structure, run the model once, and obtain the complete structure of the model output;

[0047] Step 4-2. Use simulation software to verify whether the output structure meets the optimization index. If it does, the optimization process ends. If not, go to step 5.

[0048] Step 5. Fine-tune the pre-trained neural network according to design requirements;

[0049] Step 5-1. Collect the newly generated samples and classify them based on the existing samples based on K-nearestneighbor;

[0050] Step 5-2. Use all samples of the category where the new sample belongs and the category where the optimization target belongs to to continue training the decoding network;

[0051] Step 5-3. Go to step 4.

[0052] When running the model in step 4-1, the current category label of the initial structure after passing through the encoding network is replaced with the target category label.

[0053] When fine-tuning the model in step 5-2, only the free parameters of the decoding network are updated, and the free parameters of the encoding network are not updated.

[0054] The above descriptions are only preferred embodiments of the present invention, and all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

Claims

1. A machine learning optimization method for antenna decoupling structure based on masked autoencoder, It is characterized in that The following steps are involved: Step 1. Generate sample data Batch generate N random shapes as the shape of the decoupling structure, N is a positive integer, and use the top view of a two-unit microstrip antenna with different decoupling structure shapes and the same other size parameters as the input data of the sample set. The color of the radiation patch in the antenna top view is different from the color of the dielectric substrate; Electromagnetic simulation software is used to perform batch simulations on N two-unit microstrip antennas with different decoupling structure shapes and the same other size parameters, and N groups of reflection coefficients and transmission coefficients within a set frequency band are obtained as sample set data; The sample data in the sample set is divided into training data and test data; Step 2. Classification of sample data Step 2-1. Determine the number of categories to be classified according to the range of the set frequency band and the numerical distribution of the reflection coefficient and the transmission coefficient in the sample set; Step 2-2. Classify all samples in the sample set based on K-means; Step 2-3. Label each sample with its category number; Step 3. Use the image of the category number and the randomly masked isolated structure in the training sample as the input of the autoencoder neural network, and the image of the shape of the complete isolated structure as the output of the neural network to pre-train the neural network; Step 4. Determine whether the output structure of the trained model meets the optimization index; Step 4-1. Determine the initial structure, run the model once, and obtain the complete structure of the model output; Step 4-2. Use simulation software to verify whether the output structure meets the optimization index. If it does, the optimization process ends. If not, go to step 5. Step 5. Fine-tune the pre-trained neural network according to design requirements; Step 5-1. Collect the newly generated samples and classify them based on the existing samples based on K-nearest neighbor; Step 5-2. Use all samples of the category where the new sample belongs and the category where the optimization target belongs to to continue training the decoding network; Step 5-3. Go to step 4.

2. The machine learning optimization method for the antenna decoupling structure based on the masked autoencoder according to claim 1, It is characterized in that 90% of the sample data in the sample set is used as training data, and the rest of the sample data is used as test data.

3. The machine learning optimization method for the antenna decoupling structure based on the masked autoencoder according to claim 1, It is characterized in that N=2000, set the frequency band to 1.5GHz-3.5GHz.

4. The machine learning optimization method for the antenna decoupling structure based on the masked autoencoder according to claim 1, It is characterized in that In step 2-1, the number of types of antenna decoupling structures is set to 20.

5. The machine learning optimization method for the antenna decoupling structure based on the masked autoencoder according to claim 1, It is characterized in that When running the model in step 4-1, the current category label of the initial structure after passing through the encoding network is replaced with the target category label.

6. The machine learning optimization method for the antenna decoupling structure based on the masked autoencoder according to claim 1, It is characterized in that When fine-tuning the model in step 5-2, only the free parameters of the decoding network are updated, and the free parameters of the encoding network are not updated.

Images