A method and system for estimating the envelope size of an airborne AE type flat panel antenna

Abstract

The present invention discloses a method and system for estimating the envelope size of an airborne A-E type flat panel antenna, and relates to the field of antenna measurement technology, including obtaining a sample data set of an A-E type flat panel antenna; constructing and training an end-to-end deep learning model based on the sample data set of an A-E type flat panel antenna; placing the A-E type flat panel antenna to be estimated on a multi-axis measurement platform, and collecting the geometric parameters of the A-E type flat panel antenna through a three-dimensional scanner; applying a working frequency band signal to the A-E type flat panel antenna using a spectrum analyzer, and measuring and recording the corresponding electromagnetic wave parameters under the working frequency band; inputting the geometric parameters and electromagnetic wave parameters of the A-E type flat panel antenna to be estimated into the trained end-to-end deep learning model, and outputting an envelope size estimation value. The present invention adopts an integrated learning strategy to fuse the prediction results of multiple heterogeneous models, thereby improving the accuracy and robustness of the prediction, and realizes adaptive adjustment of model parameters through laser ranging and error analysis, thereby ensuring that the prediction accuracy always meets the requirements.

Description

Technical Field

[0001] The present invention relates to the field of antenna measurement technology, and in particular to a method and system for estimating the envelope size of an airborne AE type flat panel antenna. Background Art

[0002] The traditional airborne AE type flat antenna envelope size measurement mainly relies on manual measurement or simple mechanical measurement equipment. These methods have problems such as low measurement accuracy, poor efficiency, and poor repeatability. Especially in the mass production and testing of antennas, traditional measurement methods are difficult to meet the needs of quickly and accurately obtaining the antenna envelope size. At the same time, due to the complex structure of airborne flat antennas, there is a complex nonlinear relationship between its envelope size and the geometric parameters and electromagnetic characteristics of the antenna, and conventional mathematical models are difficult to accurately describe this relationship.

[0003] Existing antenna size estimation methods often only consider a single geometric feature or electromagnetic characteristic, and lack the ability to comprehensively analyze multi-dimensional parameters. In addition, these methods generally lack adaptive learning and optimization mechanisms, and cannot dynamically adjust the estimation model according to actual measurement results, making it difficult to ensure the accuracy and reliability of the estimation. Summary of the invention

[0004] In view of the shortcomings of the existing airborne AE type flat antenna envelope size measurement method in terms of accuracy, efficiency and adaptability, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to construct an intelligent prediction method based on deep learning to achieve fast and accurate prediction of the envelope size of the airborne AE type flat antenna and have adaptive optimization capabilities.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a method for estimating the envelope size of an airborne AE-type flat-panel antenna, which includes obtaining a sample data set of the AE-type flat-panel antenna and performing standardization processing on the sample data set of the AE-type flat-panel antenna, wherein the sample data set of the AE-type flat-panel antenna includes geometric parameters, electromagnetic wave parameters, and envelope size; constructing and training an end-to-end deep learning model based on the sample data set of the AE-type flat-panel antenna, and the end-to-end deep learning model adopts a multi-level structure design; placing the AE-type flat-panel antenna to be estimated on a multi-axis measurement platform, and collecting the geometric parameters of the AE-type flat-panel antenna through a three-dimensional scanner; using a spectrum analyzer to apply an operating frequency band signal to the AE-type flat-panel antenna, and measuring and recording the corresponding electromagnetic wave parameters under the operating frequency band; inputting the geometric parameters and electromagnetic wave parameters of the AE-type flat-panel antenna to be estimated into the trained end-to-end deep learning model, and outputting an envelope size estimation value.

[0008] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna described in the present invention, wherein: constructing and training an end-to-end deep learning model includes the following steps: designing a dual-channel parallel neural network structure as the basic architecture of the end-to-end deep learning model according to the geometric features and electromagnetic characteristics of the AE type flat antenna; dividing the AE type flat antenna sample data set into a training data set and a verification data set according to a preset ratio; defining a joint loss function, and using the AdamW optimization algorithm to train the end-to-end deep learning model; evaluating the performance of the end-to-end deep learning model based on the verification data set, and automatically optimizing the structural parameters of the end-to-end deep learning model in combination with the Bayesian optimization algorithm; monitoring the model prediction error, and when the prediction error exceeds an acceptable threshold, starting the optimization adjustment process; using an integrated learning strategy to train at least two end-to-end deep learning models with different neural network structures, and fusing the prediction results of each model through a voting mechanism.

[0009] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna of the present invention, the specific formula of the joint loss function is as follows:

[0010]

[0011] in, is the weight coefficient, n is the number of samples in the batch, is the true envelope size value of the i-th sample, is the predicted envelope size value of the i-th sample, is the regularization coefficient, is the jth weight parameter of the model, is the true importance distribution of the kth feature channel, is the predicted importance distribution of the kth feature channel, is the weight of the geometric parameter channel, is the weight of the electromagnetic wave parameter channel.

[0012] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna described in the present invention, the end-to-end deep learning model is trained using the AdamW optimization algorithm, including the following steps: using the AdamW optimizer to initialize the model training parameters, and setting the initial learning rate and weight decay coefficient; introducing a cosine annealing learning rate adjustment strategy to gradually reduce the learning rate; in each training batch, respectively calculating the feature losses of the geometric parameter channel and the electromagnetic wave parameter channel, and adaptively adjusting the weights of the two channel losses through the attention mechanism; applying the gradient clipping technology to limit the gradient values ​​generated during the training process to a preset interval; using the early stopping mechanism, when the loss of the verification data set has not decreased for consecutive preset rounds, saving the current optimal model parameters, and continuing the training after reducing the learning rate; dynamically adjusting the training batch size according to the loss value fluctuation range during the training process.

[0013] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna described in the present invention, the following steps are included: automatically optimizing the structural parameters of the end-to-end deep learning model in combination with the Bayesian optimization algorithm: defining the search space of network structure parameters and hyperparameters, and defining the objective function; using the Bayesian optimization algorithm to explore the search space based on the Gaussian process regression model to find the optimal network structure and hyperparameter configuration; after each training round, updating the posterior probability distribution of the Bayesian optimization algorithm, and recommending the next set of network structure and hyperparameter combinations based on the posterior probability distribution; performing iterative optimization until the performance indicators of the verification data set converge, and selecting the network structure parameters and hyperparameters corresponding to the optimal objective function value.

[0014] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna described in the present invention, starting the optimization adjustment process includes the following steps: setting an acceptable threshold value of the prediction error, and comparing the prediction error with the acceptable threshold value; when the prediction error is greater than the acceptable threshold value, entering the parameter adjustment process; when the prediction error is less than or equal to the acceptable threshold value, it means that the model prediction accuracy temporarily meets the requirements and no adjustment is required; using the adjusted model parameters, re-executing the envelope size estimation process and recalculating the prediction error; comparing the new round of prediction error with the acceptable threshold value until the error is reduced below the acceptable threshold value or the maximum number of iterations is reached.

[0015] As a preferred solution of the method for estimating the envelope size of the airborne AE type flat antenna described in the present invention, the geometric parameters include antenna length, antenna width, antenna height and antenna tilt angle; the electromagnetic wave parameters include wavelength, gain parameter and radiation efficiency.

[0016] In a second aspect, an embodiment of the present invention provides an estimation system for the envelope size of an airborne AE-type flat-panel antenna, which includes a sample processing module for acquiring a sample data set of an AE-type flat-panel antenna and performing standardization on the sample data set of an AE-type flat-panel antenna; a modeling module for constructing and training an end-to-end deep learning model based on the sample data set of an AE-type flat-panel antenna; a real-time data acquisition module for acquiring geometric parameters and electromagnetic wave parameters of the AE-type flat-panel antenna to be estimated; and a prediction module for inputting the geometric parameters and electromagnetic wave parameters of the AE-type flat-panel antenna to be estimated into the trained end-to-end deep learning model, and outputting an envelope size estimation value.

[0017] The beneficial effects of the present invention are as follows: the present invention adopts a multi-axis measurement platform and a high-precision three-dimensional scanner to collect geometric parameters, and cooperates with shock-proof pads and digital filtering processing to effectively improve the measurement accuracy and stability. The electromagnetic wave parameters are measured using a spectrum analyzer and a standard gain antenna in an electromagnetic wave darkroom, eliminating external interference and ensuring the accuracy of the measurement data. A dual-channel parallel neural network structure is designed, and the relevant features are adaptively enhanced through the attention mechanism, thereby improving the feature extraction ability of the model. The joint loss function is used for training to enhance the generalization performance of the model. The Bayesian optimization algorithm is introduced to automatically search the network structure and hyperparameters, reducing the cost of manual parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 The framework flow chart of the method for estimating the envelope size of airborne AE type flat panel antenna.

[0020] Figure 2 Flowchart for establishing and training a deep learning model for estimating the envelope size of airborne AE-type flat-panel antennas.

[0021] Figure 3 Flow chart for calculating the prediction error of the method for estimating the envelope size of the airborne AE type flat panel antenna. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1, reference Figure 1~Figure 3 , which is the first embodiment of the present invention, and provides a method for estimating the envelope size of an airborne AE type flat antenna. The framework flow chart is as follows Figure 1 As shown, including,

[0026] S1: Obtain an AE type flat panel antenna sample dataset and perform standardization on the AE type flat panel antenna sample dataset.

[0027] Specifically, the AE type flat antenna sample data set includes geometric parameters, electromagnetic wave parameters and their corresponding antenna envelope size. Among them, the antenna geometric parameters and antenna electromagnetic wave parameters are used as input features, and the antenna envelope size is used as label data. The AE type flat antenna sample data is hierarchically organized according to antenna type and application scenario to ensure the integrity and diversity of the training data. When dividing the data, in order to prevent data leakage, it is necessary to ensure that the same type of antenna samples do not appear in both the training data set and the verification data set. In addition, the data set is standardized to eliminate the scale differences between different features.

[0028] In addition, the data sources of geometric parameters, electromagnetic wave parameters and their corresponding antenna envelope dimensions mainly include: design documents of existing AE type flat antennas, experimental test data, simulation analysis results, etc.

[0029] S2: Build and train an end-to-end deep learning model based on the AE type flat antenna sample dataset. The end-to-end deep learning model adopts a multi-level structure design.

[0030] Among them, the deep learning model establishment and training flow chart is as follows Figure 2 As shown, the details are as follows:

[0031] S2.1: Based on the geometric features and electromagnetic characteristics of the AE type flat antenna, a dual-channel parallel neural network structure is designed as the basic architecture of the end-to-end deep learning model.

[0032] The basic architecture includes: a dual-channel input layer, which is used to receive geometric parameters and electromagnetic wave parameters respectively; the ResNet architecture is used, and 32, 64, and 128 3×3 convolution kernels are used for feature extraction in turn. The step size of each convolution layer is 1, and the same padding is used to keep the feature map size unchanged, and each layer is followed by BatchNormalization and ReLU activation functions; a residual connection module is used to prevent gradient disappearance; CBAM is used as a channel attention mechanism module to adaptively enhance the spatial and channel dimension information of related features; and finally a dynamic fully connected layer and an output layer for feature fusion and size prediction. The number of neurons in the dynamic fully connected layer is adaptively adjusted according to the input feature dimension, specifically: the number of neurons in the first layer is twice the input feature dimension, the second layer is 1 times the input feature dimension, and the last layer corresponds to the predicted target dimension.

[0033] S2.2: Divide the standardized AE type flat antenna sample data set into a training data set and a validation data set according to a preset ratio.

[0034] In this embodiment, the preset ratio is 8:2.

[0035] S2.3: Define the joint loss function and use the AdamW optimization algorithm to train the end-to-end deep learning model.

[0036] Preferably, the joint loss function comprehensively considers prediction accuracy (mean square error loss term), model generalization (regularization loss term), feature selection (attention mechanism loss term) and dual-channel balance (inter-channel balance loss term). The specific formula is as follows:

[0037]

[0038] in, is the weight coefficient, n is the number of samples in the batch, is the true envelope size value of the i-th sample, is the predicted envelope size value of the i-th sample, is the regularization coefficient, is the jth weight parameter of the model, is the true importance distribution of the kth feature channel, is the predicted importance distribution of the kth feature channel, is the weight of the geometric parameter channel, is the weight of the electromagnetic wave parameter channel.

[0039] In this embodiment, (dominant, ensuring basic prediction accuracy), (to prevent overfitting), , (Optimize feature extraction). is the mean square error loss, is the regularization loss, is the attention mechanism loss, For channel balancing loss, the BatchNorm layer parameters use standard configuration.

[0040] Preferably, the design of the joint loss function fully considers the particularity of the AE type flat antenna prediction task, ensures the basic prediction accuracy through the MSE term and assigns a maximum weight of 0.5, uses the L2 regularization term weight of 0.2 to suppress the overfitting of the model on highly correlated geometric and electromagnetic features, and combines the attention mechanism loss term weight of 0.15 to guide the model to automatically learn the importance distribution of features under different working frequency bands, and cooperates with the channel balance loss term weight of 0.15 to ensure that the dual-channel parallel network makes full use of the two types of input features, and finally achieves high-precision estimation of the antenna envelope size.

[0041] Furthermore, the training process of the deep learning model is as follows: the AdamW optimizer is used to initialize the model training parameters, the initial learning rate is set to 0.0003, and it is linearly increased to 0.001 within the first 5 epochs, and the maximum training rounds are set to 200; the cosine annealing learning rate adjustment strategy is introduced to gradually reduce the learning rate with the training rounds, as follows: a large learning rate is maintained in the early stage of training to accelerate model convergence. When the training rounds reach 1 / 3 of the total rounds, the learning rate is reduced to 0.77 times the initial value; when the training rounds reach 2 / 3 of the total rounds, the learning rate is further reduced to 0.33 times the initial value; in the later stage of training, the learning rate is reduced to a smaller value to achieve fine tuning of the model parameters and ensure prediction accuracy; in addition, when the loss of the verification data set has not decreased significantly for 5 consecutive rounds, the learning rate decay mechanism is additionally triggered to multiply the current learning rate by 0.1, and it is triggered at most 2 times to prevent the model from falling into the local optimum.

[0042] In each training batch, the feature losses of the geometric parameter channel and the electromagnetic wave parameter channel are calculated respectively, and the weights of the two channel losses are adaptively adjusted through the attention mechanism; the gradient clipping technology is applied to limit the gradient values ​​generated during the training process to the preset interval [-1,1]. This gradient clipping mechanism stabilizes the training process by limiting the range of the gradient. When it is detected that the gradient value exceeds the preset range, the modulus of the gradient vector is scaled proportionally so that it does not exceed the threshold, while keeping the gradient direction unchanged, effectively preventing drastic fluctuations in model parameters due to excessive gradients.

[0043] Furthermore, an early stopping mechanism is adopted. When the loss of the validation dataset does not decrease for 10 consecutive rounds, the current optimal model parameters are saved, and the training is continued after the learning rate is appropriately reduced. According to the loss fluctuation during the training process, the training batch size is dynamically adjusted as follows: the initial batch size is set to 32. When the standard deviation of the loss fluctuation for 3 consecutive rounds of training is greater than 0.1, the batch size is increased by 50% to improve the training stability; when the standard deviation of the loss fluctuation for 3 consecutive rounds of training is less than 0.01, the batch size is reduced by 25% to improve the training efficiency; the adjustment range of the batch size is limited to [16, 128], and each adjustment interval is not less than 5 training rounds to ensure that the model has enough time to adapt to the new batch size. To ensure the stability of dynamic batch adjustment, after each adjustment, the learning rate is adjusted accordingly according to the ratio of the new and old batch sizes.

[0044] The optimal dynamic batch size adjustment strategy can better adapt to the needs of different training stages than the fixed batch size, effectively balance the training speed and model stability, and has faster convergence speed and better training effect than the traditional fixed batch training method. By combining the early stopping mechanism, gradient clipping and dynamic batch adjustment, it can significantly improve the training efficiency while ensuring the stability of model training, reduce computing resource consumption, and obtain higher prediction accuracy and generalization performance.

[0045] S2.4: Evaluate the performance of the end-to-end deep learning model based on the validation dataset, and automatically optimize the structural parameters of the end-to-end deep learning model using the Bayesian optimization algorithm.

[0046] Specifically, define the search space of network structure parameters and hyperparameters, including the number of convolution kernels, convolution kernel size, pooling layer parameters, number of fully connected layer units, etc.; define the objective function , the specific formula is as follows:

[0047]

[0048] in, is the weight coefficient, is the mean square error of the validation dataset, is the mean absolute error of the validation dataset, is the normalized value for the number of model parameters.

[0049] In addition, the objective function evaluates the model performance through the dual indicators of MSE and MAE to ensure the prediction accuracy; the number of parameters C is introduced params Constraints are used to avoid overly complex models; weight coefficients are used to balance the importance of various indicators.

[0050] Furthermore, a Bayesian optimization algorithm was used to explore the search space based on the Gaussian process regression model to find the optimal network structure and hyperparameter configuration. Specifically, the Gaussian process regression model uses a radial basis function (RBF) kernel, the initial value of the length scale is set to 1.0, and the signal variance is 1.0; the exploration and utilization are balanced by maximizing the acquisition function EI, where the acquisition function takes into account the predicted mean and variance, so that the algorithm can weigh between known good areas and unknown areas, and each time the hyperparameter combination with the largest acquisition function value is selected for evaluation, and the evaluation results are used to update the Gaussian process model; at the same time, a random exploration strategy based on Thompson sampling is introduced to sample from the Gaussian process posterior distribution with a probability of 0.1 to increase the diversity of the search; for discrete hyperparameters such as the number of convolution kernels, the continuous optimization results are converted into discrete values ​​by rounding.

[0051] After each training round, the posterior probability distribution of the Bayesian optimization algorithm is updated, and the next set of network structure and hyperparameter combinations are recommended based on the posterior probability distribution; iterative optimization is performed until the performance indicators of the verification data set converge, and the network structure parameters and hyperparameters corresponding to the optimal objective function value are selected.

[0052] The validation data set performance indicators meet the following convergence criteria: the relative improvement of the validation data set loss in 5 consecutive iterations is less than 0.1%, and the fluctuation range of the best performance indicators in the last 10 iterations does not exceed 1%. In addition, when the number of iterations reaches the preset maximum number of iterations or the optimization time exceeds the preset limit, the convergence judgment will also be triggered.

[0053] Preferably, the Bayesian optimization algorithm can efficiently locate the global optimal solution in the search space, which has significant advantages over random search or grid search. Through this automated network structure and hyperparameter optimization process, the model performance can be greatly improved without the need for repeated manual adjustments, greatly reducing the development cost and cycle. The optimized deep learning model not only has better prediction performance, but also shows better generalization ability and stability, and can maintain high prediction accuracy in different working frequency bands and application scenarios. At the same time, the efficient search strategy of Bayesian optimization reduces the required computing resources and optimization time, providing reliable guarantees for the actual deployment and application of the model.

[0054] S2.5: Monitor the model prediction error and start the optimization and adjustment process when the prediction error exceeds the acceptable threshold.

[0055] Specifically, the calculation flow chart of the prediction error is as follows: Figure 3As shown, the actual antenna envelope size of each sample in the AE type flat antenna sample data set is obtained; the actual envelope size is compared with the envelope size estimate output by the model, and the relative error and absolute error between the two are calculated. The relative error is the percentage difference between the predicted value and the actual value, and the absolute error is the numerical difference between the predicted value and the actual value. It should be noted that the prediction error includes relative error and absolute error.

[0056] Furthermore, starting the optimization adjustment process includes the following steps: setting an acceptable threshold of the prediction error (for example, the relative error threshold is set to 3%, and the absolute error threshold is set to 0.5 mm), and comparing the prediction error with the acceptable threshold, wherein the prediction error is determined by the relative error and the absolute error; when the prediction error is greater than the acceptable threshold, entering the parameter adjustment process; when the prediction error is less than or equal to the acceptable threshold, it means that the model prediction accuracy temporarily meets the requirements and no adjustment is required; using the adjusted model parameters, re-execute the envelope size estimation process (ie, S4), and recalculate the prediction error; compare the new round of prediction error with the acceptable threshold until the prediction error is reduced below the acceptable threshold or the maximum number of iterations is reached.

[0057] The parameter adjustment process is as follows: first, analyze the error distribution characteristics, identify the main sources of errors, and determine the network level that needs to be optimized; use a combination of grid search and Bayesian optimization to automatically adjust network parameters, including adjusting the number of convolution kernels, updating the convolution kernel size, and optimizing the residual connection structure.

[0058] It should be noted that the acceptable threshold is determined by comprehensively considering the installation requirements of the AE type flat panel antenna in actual engineering applications and the accuracy limitations of the measurement system.

[0059] S2.6: Adopt an integrated learning strategy to train at least two end-to-end deep learning models with different neural network structures, and fuse the prediction results of each model through a voting mechanism to obtain an estimated value of the envelope size.

[0060] Furthermore, at least two deep learning models with different network structures are designed. These models differ in the number of layers, convolution kernel parameters, fully connected layer settings, etc. The integrated learning adopts at least two heterogeneous models with the best performance.

[0061] Furthermore, these deep learning models with heterogeneous network structures are trained separately, using the same training data set and joint loss function. When a new sample to be predicted is input, the new sample is simultaneously input into multiple trained deep learning models. For continuous value prediction, a weighted average voting mechanism is adopted, and the weight is proportional to the performance of each model on the validation set. A simple voting mechanism is used to count the envelope size estimates given by each model, and the value with the highest frequency of occurrence is selected as the final prediction result.

[0062] Preferably, compared with a single model, the ensemble learning method can give full play to the advantages of multiple models and improve the accuracy and robustness of the overall prediction.

[0063] S3: Place the AE-type flat-panel antenna to be estimated on a multi-axis measurement platform, and collect geometric parameters of the AE-type flat-panel antenna using a three-dimensional scanner.

[0064] The geometric parameters include antenna length, antenna width, antenna height and antenna tilt angle.

[0065] Specifically, the AE type flat panel antenna to be tested is placed on a multi-axis measurement platform, which includes a precision displacement table in three directions of XYZ, and high-precision positioning is achieved through a precision displacement mechanism; the AE type flat panel antenna to be tested is adjusted to a horizontal reference position through the precision displacement table; the multi-axis measurement platform is provided with shock-absorbing pads to reduce the influence of external vibration on the measurement results.

[0066] Furthermore, a high-precision three-dimensional scanner is arranged above the multi-axis measurement platform, and the three-dimensional scanner uses a multi-point scanning method to scan and measure the AE type flat-panel antenna to be measured; the scanning accuracy of the three-dimensional scanner is 0.1mm, and the scanning range covers the entire surface of the AE type flat-panel antenna to be measured; the geometric parameters of the AE type flat-panel antenna are collected by the three-dimensional scanner; the three-dimensional scanner collects data points at a preset sampling interval every 1mm, and performs an all-round scan of the AE type flat-panel antenna to be measured to obtain the complete geometric outline of the AE type flat-panel antenna to be measured; the original data collected by the three-dimensional scanner is processed by digital filtering to eliminate measurement noise and improve the measurement accuracy of the geometric parameters.

[0067] S4: Use a spectrum analyzer to apply a working frequency band signal to the estimated AE type flat antenna, and measure and record the corresponding electromagnetic wave parameters under the working frequency band.

[0068] Among them, electromagnetic wave parameters include wavelength, gain parameter, and radiation efficiency.

[0069] Specifically, an AE-type flat antenna is placed in an electromagnetic wave darkroom, and the walls of the electromagnetic wave darkroom are covered with electromagnetic wave absorbing materials to eliminate electromagnetic wave reflection and interference; a high-performance spectrum analyzer is connected to the feeding port of the AE-type flat antenna through a coaxial cable, and the frequency range of the spectrum analyzer is 0.1-20GHz, and the frequency resolution is 1MHz; the spectrum analyzer generates a sweep signal, and the frequency range of the sweep signal covers the entire working frequency band of the AE-type flat antenna, and scans according to a preset step frequency.

[0070] During the frequency scanning process, standard measuring equipment is used to obtain electromagnetic wave characteristic parameters, including wavelength parameters, gain parameters and radiation efficiency parameters, through a variety of parameter measurement methods under standardized environmental conditions; the measurement process strictly controls environmental conditions to ensure the accuracy and repeatability of the measurement results.

[0071] S5: The geometric parameters and electromagnetic wave parameters of the AE type flat antenna to be estimated are input into the trained end-to-end deep learning model, and an estimated envelope size is output.

[0072] Specifically, the collected antenna geometric parameters and electromagnetic wave parameters are normalized and preprocessed; the normalized geometric parameters and electromagnetic wave parameters are received respectively through a multi-channel input architecture, and feature extraction and fusion are performed; a deep neural network architecture is used, combined with residual connection and attention mechanism, to extract and integrate key features; the envelope size prediction value is obtained through a fully connected layer and nonlinear transformation, and denormalization is performed to obtain the final estimation result.

[0073] Furthermore, the present embodiment also provides an airborne AE type flat-panel antenna envelope size estimation system, including a sample processing module, used to obtain an AE type flat-panel antenna sample data set and perform standardization on the AE type flat-panel antenna sample data set; a modeling module, used to build and train an end-to-end deep learning model based on the AE type flat-panel antenna sample data set; a real-time data acquisition module, used to collect geometric parameters and electromagnetic wave parameters of the AE type flat-panel antenna to be estimated; a prediction module, used to input the geometric parameters and electromagnetic wave parameters of the AE type flat-panel antenna to be estimated into the trained end-to-end deep learning model, and output an envelope size estimation value.

[0074] In summary, the present invention adopts a multi-axis measurement platform and a high-precision three-dimensional scanner to collect geometric parameters, and cooperates with shock-proof pads and digital filtering processing to effectively improve the measurement accuracy and stability. A spectrum analyzer and a standard gain antenna are used to measure electromagnetic wave parameters in an electromagnetic wave darkroom to eliminate external interference and ensure the accuracy of the measurement data. A dual-channel parallel neural network structure is designed, and the relevant features are adaptively enhanced through the attention mechanism to improve the feature extraction ability of the model. The joint loss function is used for training to enhance the generalization performance of the model. The Bayesian optimization algorithm is introduced to automatically search the network structure and hyperparameters, reducing the cost of manual parameter adjustment.

[0075] Example 2, reference Figure 1~Figure 3 , which is the second embodiment of the present invention, and this embodiment provides a method for estimating the envelope size of an airborne AE type flat-panel antenna. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0076] In order to verify the effectiveness of the proposed airborne AE flat panel antenna envelope size estimation method, an aviation research institute conducted a simulation experiment. Thirty different types of AE flat panel antennas were selected as test samples. The operating frequency range of these antennas was 2~18GHz, and the antenna size varied in the range of 100~500mm.

[0077] During the measurement platform construction phase, a three-axis precision displacement measuring table was used to construct a multi-axis measurement platform, and a 5mm thick rubber anti-vibration pad was laid on the platform surface. A high-precision laser scanner was used to measure geometric parameters. The scanner was installed 2 meters above the measurement platform, and the scanning range was 600mm×600mm×300mm. In the experiment, the scanner sampled the antenna surface at intervals of 1mm, and each antenna sample obtained about 20,000 data points. A 4th-order Butterworth low-pass filter was used to reduce the noise of the raw data, and the cutoff frequency was set to 100Hz.

[0078] The electromagnetic parameter measurement was carried out in a 5m×5m×3m anechoic chamber, with the walls of the chamber covered with 50mm thick silicon carbide absorbing material. A high-precision spectrum analyzer was used to generate a sweep frequency signal of 0.1~20GHz, with a step setting of 10MHz. The standard gain antenna used a double-ridged horn antenna with a measurement distance of 3m. Under environmental conditions of 25±2℃ and relative humidity of 65%±5%, each antenna sample was measured three times, and the average value was taken as the final result.

[0079] The training dataset of the deep learning model contains 2000 AE type flat antenna samples, and the validation dataset and test set each contain 250 samples. During the model training process, the initial learning rate is set to 0.001, the batch size is 32, and the number of training rounds is 200. The balance factor in the joint loss function , The Bayesian optimization process was iterated 100 times, and the final network structure parameters were: 32 3×3 convolution kernels in the first layer, 64 3×3 convolution kernels in the second layer, and 128 3×3 convolution kernels in the third layer.

[0080] Experimental results show that among the 30 test samples, the average relative error of envelope size prediction by this method is 2.8%, the maximum relative error is 3.9%, and the standard deviation is 0.6%. Among them, the prediction error of 25 samples is lower than the acceptable threshold of 3%, and only 5 samples need parameter adjustment and re-prediction. After parameter optimization, the prediction errors of these 5 samples are reduced to less than 2.5%.

[0081] In order to deeply evaluate the performance advantages of this method, the performance of the traditional manual measurement method, the machine vision-based method and the method of the present invention in various technical indicators are compared. The specific results are shown in Table 1.

[0082] Table 1 Performance comparison between the present invention and the prior art

[0083]

[0084] As can be seen from Table 1, the measurement accuracy of this method is improved to ±0.1mm, which is 5 times that of the traditional manual measurement method; the measurement time efficiency is improved by 4~6 times, and the prediction error rate is controlled within the range of 2%-4%, which is more than 50% lower than the existing method. At the same time, this method has strong environmental adaptability and anti-interference ability, good system stability, and can meet the rapid detection needs in the mass production of airborne AE type flat antennas.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for estimating the envelope size of an airborne AE type flat panel antenna, characterized in that: include, Acquire a sample data set of an AE type flat panel antenna, and perform standardization processing on the sample data set of the AE type flat panel antenna, wherein the sample data set of the AE type flat panel antenna includes geometric parameters, electromagnetic wave parameters, and antenna envelope size; Constructing and training an end-to-end deep learning model based on the AE type flat antenna sample data set, wherein the end-to-end deep learning model adopts a multi-level structure design; Placing the AE-type flat-panel antenna to be estimated on a multi-axis measurement platform, and collecting geometric parameters of the AE-type flat-panel antenna by a three-dimensional scanner; A spectrum analyzer is used to apply a working frequency band signal to the AE type flat antenna, and the corresponding electromagnetic wave parameters under the working frequency band are measured and recorded; The geometric parameters and electromagnetic wave parameters of the AE type flat antenna to be estimated are input into the trained end-to-end deep learning model, and the estimated envelope size is output.

2. The method for estimating the envelope size of an airborne AE type flat panel antenna according to claim 1, characterized in that: The construction and training of the end-to-end deep learning model includes the following steps: According to the geometric features and electromagnetic characteristics of the AE type flat antenna, a dual-channel parallel neural network structure is designed as the basic architecture of the end-to-end deep learning model; Dividing the AE type flat antenna sample data set into a training data set and a verification data set according to a preset ratio; Define a joint loss function and use the AdamW optimization algorithm to train the end-to-end deep learning model; Evaluating the performance of the end-to-end deep learning model based on the validation data set, and automatically optimizing the structural parameters of the end-to-end deep learning model in combination with a Bayesian optimization algorithm; Monitor the model prediction error and start the optimization and adjustment process when the prediction error exceeds the acceptable threshold; An integrated learning strategy is adopted to train at least two end-to-end deep learning models with different neural network structures, and the prediction results of each model are fused through a voting mechanism.

3. The method for estimating the envelope size of an airborne AE type flat antenna according to claim 2, characterized in that: The specific formula of the joint loss function is as follows: in, is the weight coefficient, n is the number of samples in the batch, is the true envelope size value of the i-th sample, is the predicted envelope size value of the i-th sample, is the regularization coefficient, is the jth weight parameter of the model, is the true importance distribution of the kth feature channel, is the predicted importance distribution of the kth feature channel, is the weight of the geometric parameter channel, is the weight of the electromagnetic wave parameter channel.

4. The method for estimating the envelope size of an airborne AE type flat panel antenna according to claim 2, characterized in that: The use of the AdamW optimization algorithm to train the end-to-end deep learning model includes the following steps: Use AdamW optimizer to initialize model training parameters; Introduce the cosine annealing learning rate adjustment strategy to gradually reduce the learning rate; In each training batch, the feature losses of the geometric parameter channel and the electromagnetic wave parameter channel are calculated respectively, and the weights of the two channel losses are adaptively adjusted through the attention mechanism; Apply gradient clipping technology to limit the gradient values ​​generated during training to a preset range; An early stopping mechanism is used. When the loss of the validation dataset does not decrease for consecutive preset rounds, the current optimal model parameters are saved and the training continues after the learning rate is reduced. The training batch size is dynamically adjusted according to the fluctuation range of the loss value during training.

5. The method for estimating the envelope size of an airborne AE type flat antenna according to claim 2, characterized in that: The automatic optimization of the structural parameters of the end-to-end deep learning model in combination with the Bayesian optimization algorithm comprises the following steps: Define the search space of network structure parameters and hyperparameters, and define the objective function; Use the Bayesian optimization algorithm to explore the search space based on the Gaussian process regression model to find the optimal network structure and hyperparameter configuration; After each training round, the posterior probability distribution of the Bayesian optimization algorithm is updated, and the next set of network structure and hyperparameter combinations are recommended based on the posterior probability distribution; Perform iterative optimization until the performance indicators of the validation data set converge, and select the network structure parameters and hyperparameters corresponding to the optimal objective function value.

6. The method for estimating the envelope size of an airborne AE type flat panel antenna according to claim 2, characterized in that: The startup optimization adjustment process includes the following steps: Set an acceptable threshold for prediction error and compare the prediction error with the acceptable threshold; When the prediction error is greater than the acceptable threshold, the parameter adjustment process begins; When the prediction error is less than or equal to the acceptable threshold, it means that the prediction accuracy of the model temporarily meets the requirements and no adjustment is required; Using the adjusted model parameters, the envelope size estimation process is re-executed and the prediction error is recalculated; The prediction error of the new round is compared with the acceptable threshold until the error decreases below the acceptable threshold or the maximum number of iterations is reached.

7. The method for estimating the envelope size of an airborne AE type flat antenna according to claim 1, characterized in that: The geometric parameters include antenna length, antenna width, antenna height and antenna tilt angle; the electromagnetic wave parameters include wavelength, gain parameter and radiation efficiency.

8. A system for estimating the envelope size of an airborne AE type flat panel antenna, based on the method for estimating the envelope size of an airborne AE type flat panel antenna according to any one of claims 1 to 7, characterized in that: Also includes, A sample processing module is used to obtain a sample data set of an AE type flat panel antenna and perform standardization processing on the sample data set of an AE type flat panel antenna; A modeling module, used to build and train an end-to-end deep learning model based on the AE type flat antenna sample data set; A real-time data acquisition module is used to collect geometric parameters and electromagnetic wave parameters of the AE type flat antenna to be estimated; The prediction module is used to input the geometric parameters and electromagnetic wave parameters of the AE type flat antenna to be estimated into the trained end-to-end deep learning model, and output the estimated envelope size.

Images