A filter physical awareness optimization method based on residual attention and transposed convolution

Abstract

This invention discloses a filter physical perception optimization method based on residual attention and transposed convolution, comprising the following steps: constructing a filter characteristic prediction surrogate model based on residual self-attention mechanism and transposed convolution decoding to establish a nonlinear mapping relationship between filter geometric parameters and electromagnetic response parameters; training the filter characteristic prediction surrogate model using an S-parameter weighted mean square error loss function to obtain a trained filter characteristic prediction surrogate model; and using a physical perception-improved differential evolution algorithm, employing the trained filter characteristic prediction surrogate model as a predictor to find the optimal combination of geometric parameters that satisfies the filter design specifications in the solution space. This method is a filter optimization design method that balances high-precision prediction and high-efficiency optimization. This method can predict the S-parameters of microwave filters more accurately than traditional neural networks, effectively replacing time-consuming full-wave electromagnetic simulation.

Description

Technical Field

[0001] This invention belongs to the field of microwave filter technology and relates to a filter physical perception optimization method based on residual attention and transposed convolution. Background Technology

[0002] Microwave filters are core frequency-selective devices in modern wireless communication and radar detection. With the development of 5G / 6G communication technologies, the performance requirements for filters are becoming increasingly stringent. Currently, mainstream design methods heavily rely on full-wave electromagnetic simulation software (such as HFSS and CST) for iterative optimization. Due to the highly sensitive and strong nonlinear mapping between the filter's geometry and its S-parameter response, and the lengthy time required for a single high-precision full-wave simulation, the optimization process based on traditional algorithms, involving thousands of iterations, often takes weeks, severely restricting product development cycles. Although existing technologies attempt to introduce general neural networks to construct surrogate models to replace time-consuming electromagnetic simulations, the S-parameter curves of microwave filters have a large dynamic range, and existing general network structures and conventional mean square error loss functions tend to fit the average trend, making it difficult to accurately capture the S-parameter curves, leading to distorted predictions by the surrogate model.

[0003] Furthermore, in the global iterative optimization phase, traditional global optimization algorithms typically evaluate fitness solely based on the numerical error between the predicted curve and the target curve, lacking the ability to perceive the physical waveform characteristics of the S-parameters. This purely numerical comparison often leads to two typical defects in the optimization results: first, transmission zero-point frequency drift, resulting in the failure of out-of-band suppression performance; second, a skewed passband response, failing to meet the stringent requirements for passband flatness.

[0004] In summary, existing design methods have significant shortcomings in terms of design efficiency, the accuracy of surrogate models in predicting S-parameters, and the engineering usability of optimization results. There is an urgent need for a highly efficient automated filter optimization design method. Summary of the Invention

[0005] To address the issues of distorted S-parameter prediction and low efficiency in optimization algorithms in surrogate models, and the fact that existing optimization algorithms often rely on statistical indicators such as global mean square error as evaluation functions, lacking effective constraints on the local S-parameter curves, which leads to physical performance defects such as transmission zero-point frequency offset and passband amplitude response tilt in the optimization results, the technical solution adopted in this invention is: A filter physical awareness optimization method based on residual attention and transposed convolution includes the following steps: A filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding is constructed to establish the nonlinear mapping relationship between filter geometric parameters and electromagnetic response parameters. The filter characteristic prediction surrogate model is trained using the S-parameter weighted mean square error loss function to obtain the trained filter characteristic prediction surrogate model. The differential evolution algorithm based on physical perception is used as a predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

[0006] Furthermore: the filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding includes: an encoder and a decoder; The encoder includes: Mapping module: used to map the geometric parameters of the filter as input to a high-dimensional latent feature space; Feature extraction module: used to extract features from the mapped feature vector transmitted by the mapping module, and capture the nonlinear fluctuations of S-parameters caused by small changes in geometric parameters; Fusion module: used to fuse the features extracted by the feature extraction module with the original input, retaining the original information while adaptively focusing on the key geometric features that have the greatest impact on the filter response; The decoder includes: Multiple cascaded transposed convolutional blocks map the extracted deep semantic features back to the physical space of the target frequency points, and finally output the S-parameter response curve containing 150 frequency points, realizing end-to-end fast prediction from geometric structure to electromagnetic properties.

[0007] Furthermore, the mapping module includes: initially increasing the dimensionality of the geometric structure parameters of the input filter through a linear layer, and then, in conjunction with the LeakyReLU activation function, mapping the low-dimensional geometric physical parameters to a high-dimensional latent feature space.

[0008] Furthermore: the feature extraction module includes a residual block, which adopts a two-layer perceptron structure, including a main path and skip connections; The main path sequentially includes: a first linear layer, a first batch of normalized layers, a first LeakyReLU activation layer, a second linear layer, and a second batch of normalized layers; The skip connection sums the input signal of the residual block directly with the output of the main path, and the sum is then activated and output by LeakyReLU.

[0009] Furthermore, the fusion module introduces a self-attention mechanism. The feature vectors extracted by the feature extraction module are transformed by linear transformation to generate a query matrix Q, a key matrix K, and a value matrix V. By calculating the dot product of Q and K and performing scaling and Softmax normalization, an attention weight map is obtained. Then, it is multiplied by V with weights, and then fused with the original input through an additive skip connection. The attention-enhanced feature vector is transformed by a linear layer and reshaped from a 1D vector into a feature map format suitable for convolution operations.

[0010] Furthermore: the transposed convolution block sequentially performs transposed convolution, batch normalization, and LeakyReLU activation operations for upsampling, gradually restoring the spatial dimension of the features and reconstructing frequency domain details; The decoded high-dimensional feature map is flattened by the Flatten layer and fed into the output head, which consists of the third linear layer, the second LeakyReLU layer, and the fourth linear layer.

[0011] Furthermore, the S-parameter weighted mean square error loss function employs a dynamic weight allocation strategy; when the true amplitude of the S-parameter is detected to be lower than a preset threshold, the weight coefficient (weight) of the corresponding frequency point in the loss function is calculated using the following formula:

[0012] weight: represents the dynamic weight coefficient assigned to the current frequency point when calculating the loss function; y true : Represents the true amplitude of the S-parameters obtained from full-wave electromagnetic simulation; T: Represents the preset amplitude threshold, used to distinguish whether it is a deep resonance / transmission zero region. α: Represents the dynamic adjustment coefficient or variation scaling factor, used to control the growth rate of weights as depth increases.

[0013] Furthermore, the differential evolution algorithm based on physical perception constructs a fitness evaluation function that integrates physical feature constraints, comprising two core mechanisms: One is the S11 passband edge anchoring strategy. When calculating fitness, in addition to considering the overall return loss of the passband, several frequency points within a preset range inside the passband boundary are extracted. When the S11 amplitude of the edge frequency point is detected to be worse than the design index, a penalty weight much higher than that of the ordinary frequency points in the passband is applied, so that the optimization direction prioritizes to reduce the reflection coefficient of the frequency band edge. The second is the S21 transmission zero-point frequency traction mechanism, which automatically searches for local minimum points of the predicted S21 curve within the preset stopband frequency range, obtains its actual frequency position, calculates the deviation distance between this frequency and the design target zero-point frequency, and accordingly transforms the lateral error in the frequency dimension into a fitness penalty term to guide the parameters to evolve towards the target zero-point frequency.

[0014] A filter physical awareness optimization device based on residual attention and transposed convolution, comprising: Module: Used to build a filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding, and to establish the nonlinear mapping relationship between the filter's geometric structure parameters and electromagnetic response parameters; Training module: Used to train the filter characteristic prediction surrogate model using the S-parameter weighted mean square error loss function, and obtain the trained filter characteristic prediction surrogate model; Optimization module: Used for the differential evolution algorithm based on physical perception improvement, it uses the trained filter characteristic prediction surrogate model as a predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

[0015] A readable storage medium storing a program module, characterized in that the program module, when run in a processor, can implement any of the methods described above.

[0016] This invention provides a filter physical perception optimization method based on residual attention and transposed convolution, which is a filter optimization design method that balances high-precision prediction and high-efficiency optimization. This method can predict the S-parameters of microwave filters more accurately than traditional neural networks, effectively replacing time-consuming full-wave electromagnetic simulation.

[0017] By using the surrogate model constructed in this invention to replace traditional full-wave electromagnetic simulation, the prediction time for a single S-parameter is significantly reduced from minutes to milliseconds, improving computational efficiency by four to five orders of magnitude. This makes it possible to complete tens of thousands of global iterative optimizations in an extremely short time, thereby significantly reducing the development cycle of microwave filters.

[0018] This invention achieves the prediction of filter S-parameters through the synergistic effect of a "residual-attention-transposed convolution" hybrid architecture and a weighted loss function. The hybrid architecture utilizes a self-attention mechanism to capture the global coupling of geometric parameters and employs transposed convolution to ensure the smooth continuity of the spectral curve. Simultaneously, the weighted loss function, through dynamic weight allocation, forces the model to focus on high-attenuation regions. This combination effectively overcomes the smoothing distortion problem commonly found in conventional predictions and improves the model's prediction accuracy in the high-attenuation extreme regions of S-parameters.

[0019] This invention constructs a differential evolutionary optimization algorithm system that integrates physical feature constraints. Utilizing passband edge anchoring and frequency pulling strategies, it effectively eliminates passband amplitude response tilt, ensuring excellent passband flatness and deep return loss performance. Simultaneously, it can strictly limit the transmission zero point within a preset frequency band, precisely controlling out-of-band suppression characteristics. Ultimately, without manual intervention, it directly outputs filter structure parameters that meet engineering requirements. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 : Overall flowchart of the present invention; Figure 2 : A schematic diagram of the filter characteristic prediction network architecture based on residual self-attention mechanism and transposed convolution decoding in this invention; Figure 3 This invention provides a schematic diagram of a cavity filter embodiment used to verify the effectiveness of the algorithm, along with the parameters to be optimized. Figure 4 Comparison of the surrogate model prediction value and the full-wave simulation value of the S11 parameter in this embodiment of the invention; Figure 5 Comparison of the surrogate model prediction value and the full-wave simulation value of parameter S21 in this embodiment of the invention; Figure 6 The S-parameter response curve corresponding to the optimal filter geometry output after optimization in this embodiment of the invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Figure 1 : Overall flowchart of the present invention; A filter physical awareness optimization method based on residual attention and transposed convolution includes the following steps: S1: Construct a surrogate model for predicting filter characteristics based on residual self-attention mechanism and transposed convolution decoding, which is used to establish the nonlinear mapping relationship between filter geometric parameters and electromagnetic response parameters; S2: The filter characteristic prediction surrogate model is trained using the S-parameter weighted mean square error loss function to obtain the trained filter characteristic prediction surrogate model. S3: Based on the physical perception-based improved differential evolution algorithm, the trained filter characteristic prediction surrogate model is used as the predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

[0025] Steps S1 / S2 / S3 are executed sequentially; Figure 2 : A schematic diagram of the filter characteristic prediction network architecture based on residual self-attention mechanism and transposed convolution decoding in this invention; Furthermore: the filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding includes: an encoder and a decoder; The encoder includes: The mapping module is used to map the filter's geometric structure parameters (which can be a normalized 7-dimensional geometric parameter vector, the specific number of dimensions of the geometric parameter vector is determined by the filter's structure) to a high-dimensional latent feature space as input. Feature extraction module: used to extract features from the mapped feature vector transmitted by the mapping module, and capture the nonlinear fluctuations of S-parameters caused by small changes in geometric parameters; Fusion module: used to fuse the features extracted by the feature extraction module with the original input, retaining the original information while adaptively focusing on the key geometric features that have the greatest impact on the filter response; The decoder includes multiple cascaded transposed convolutional blocks that map the extracted deep semantic features back to the physical space of the target frequency points, ultimately outputting an S-parameter response curve containing 150 frequency points, achieving end-to-end fast prediction from geometric structure to electromagnetic properties. The preferred number of transposed convolutional blocks is 3. Furthermore, the mapping module includes: initially increasing the dimensionality of the input filter geometric structure parameters through a linear layer (Linear, L), and then, in conjunction with the LeakyReLU activation function, mapping the low-dimensional geometric physical parameters to a high-dimensional latent feature space, laying the foundation for subsequent deep feature extraction; Furthermore, the feature extraction module includes residual blocks, which contain main paths and skip connections; forming the core feature extraction module. The residual block employs a two-layer perceptron structure, specifically consisting of a first linear layer (L), a first batch normalization layer (BN, B), a LeakyReLU activation layer (~), a second linear layer (L), and a second batch normalization layer (BN, B). To address the vanishing gradient problem in deep networks and enhance feature transfer efficiency, the skip connections directly sum the input signal of the residual block with the output of the main path. The summed result is then activated by LeakyReLU before being output. This design effectively ensures that the model can capture the nonlinear fluctuations in S-parameters caused by minute changes in geometric parameters.

[0026] Furthermore, the fusion module introduces a self-attention mechanism. The feature vectors extracted by the feature extraction module are transformed by linear transformation to generate a query matrix Q, a key matrix K, and a value matrix V. By calculating the dot product of Q and K and performing scaling and Softmax normalization, an attention weight map is obtained. This map is then multiplied by V with a weight, and then fused with the original input through an additive skip connection. The attention-enhanced feature vector is transformed by a linear layer and reshaped from a 1D vector into a feature map format suitable for convolution operations.

[0027] Furthermore, the transposed convolutional block sequentially performs transposed convolution (ConvTrans), batch normalization (BN, B), and LeakyReLU (~) activation operations to upsample, gradually restoring the spatial dimension of the features and reconstructing the frequency domain details; The decoded high-dimensional feature map is flattened by the Flatten layer and fed into the output header, which consists of a third linear layer (L), a second LeakyReLU layer (~), and a fourth linear layer (L). This layer maps the extracted deep semantic features back to the physical space of the target frequency points, and finally outputs the S-parameter (S11 and S21) response curves containing 150 frequency points, realizing end-to-end fast prediction from geometric structure to electromagnetic properties.

[0028] Compared with the commonly used proxy models based on stacked simple fully connected layers (MLPs) in existing technologies, the architecture proposed in this invention has significant substantial features and advancements: First, existing technologies are prone to gradient vanishing and training difficulties when the network is deepened. The residual structure introduced in this invention ensures the effective propagation of gradients in deep networks, significantly improving the model's ability to fit the highly nonlinear mapping relationship between filter geometric parameters and electromagnetic response. Second, existing technologies typically process input geometric variables in isolation, while this invention utilizes a self-attention mechanism to automatically capture the potential physical dependencies between different geometric parameters, improving the physical consistency of prediction results. Third, addressing the shortcomings of existing technologies that directly regress high-dimensional frequency points, resulting in jagged curves and poor local smoothness, this invention innovatively adopts a transposed convolution decoding strategy. It utilizes the local perception characteristics of convolution operations for feature upsampling, ensuring that the generated S-parameter curves have continuous and smooth physical characteristics in the frequency domain, thereby more accurately reconstructing the sharp resonant points and transmission zeros of the filter.

[0029] Furthermore, the S-parameter weighted mean square error loss function employs a dynamic weight allocation strategy; when the true amplitude of the S-parameter is detected to be lower than a preset threshold, the weight coefficient (weight) of the corresponding frequency point in the loss function is calculated using the following formula:

[0030] weight: represents the dynamic weight coefficient assigned to the current frequency point when calculating the loss function.

[0031] ytrue: Represents the true amplitude of the S-parameters obtained from the full-wave electromagnetic simulation (usually in dB).

[0032] T: represents the preset amplitude threshold (used to distinguish whether it is a deep resonance / transmission zero region; in the example, the value is 10, corresponding to -10dB).

[0033] α: Represents the dynamic adjustment coefficient (or variation scaling factor, used to control the growth rate of weights as depth increases).

[0034] When calculating the error between the predicted and true values, the amplitude of the true S-parameters is dynamically detected: This method introduces a dynamic weight allocation mechanism to detect the amplitude of the true S-parameters in real time when calculating the fitting error between the predicted and true values. Using a preset passband matching threshold (e.g., -10 dB to -15 dB) as the dividing point, when the amplitude of the true S-parameters is detected to be lower than the threshold, the algorithm dynamically generates weight coefficients based on the absolute magnitude of the amplitude, establishing a nonlinear mapping relationship of "the lower the amplitude, the greater the penalty weight," thereby significantly amplifying the gradient signal in the backpropagation of this region. When the amplitude is higher than the threshold, the baseline weight remains unchanged. Through this differentiated weighting strategy, the model is forced to focus its attention on the resonant poles that determine the filter performance during training, thereby greatly improving the prediction accuracy of key features of the S-parameter curve.

[0035] Existing technologies typically employ standard mean square error (MSE), causing models to tend to fit a flat stopband while ignoring deep resonant points within extremely narrow bandwidths, resulting in prediction distortion. This invention, through dynamic weighting, forces the model to focus on and accurately fit the resonant characteristics of the filter.

[0036] Furthermore, the differential evolution algorithm based on physical perception constructs a fitness evaluation function that integrates physical feature constraints, comprising two core mechanisms: One is the S11 passband edge anchoring strategy. When calculating fitness, in addition to considering the overall return loss of the passband, several frequency points within a preset range inside the passband boundary are extracted. When the S11 amplitude of the detected edge frequency point is less than the design index, a penalty weight much higher than that of the ordinary frequency points in the passband (such as 2000 times the weight) is applied, so that the optimization direction prioritizes to reduce the reflection coefficient of the frequency band edge. The second is the S21 transmission zero-point frequency traction mechanism, which automatically searches for local minimum points of the predicted S21 curve within the preset stopband frequency range, obtains its actual frequency position, calculates the deviation distance between this frequency and the design target zero-point frequency, and accordingly transforms the lateral error in the frequency dimension into a fitness penalty term to guide the parameters to evolve towards the target zero-point frequency.

[0037] Compared with existing technologies, the differential evolution algorithm based on physical perception proposed in this invention has significant advantages: its unique edge anchoring strategy effectively eliminates the common tilting defects in the passband response and ensures passband flatness; the introduced frequency pulling mechanism compensates for the blind spot of single amplitude evaluation, and precisely locks the transmission zero point to the target frequency through lateral constraints, ensuring the high selectivity of the filter; at the same time, the application of dynamic mutation factors breaks the fixed parameter limitation, significantly enhances population diversity, effectively avoids the algorithm from getting trapped in local optima in complex parameter space, and achieves efficient global optimization.

[0038] A filter physical awareness optimization device based on residual attention and transposed convolution, comprising: Building modules: These are used to construct a filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding, and to establish the nonlinear mapping relationship between filter geometry and electromagnetic response parameters. Training module: Used to train the filter characteristic prediction surrogate model using the S-parameter weighted mean square error loss function, and obtain the trained filter characteristic prediction surrogate model; Optimization module: Used for the differential evolution algorithm based on physical perception improvement, it uses the trained filter characteristic prediction surrogate model as a predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

[0039] A readable storage medium storing a program module, characterized in that the program module, when run in a processor, can implement the method as described in any one of the above.

[0040] Example 1 A filter physical awareness optimization method based on residual attention and transposed convolution includes the following steps: S1: Construct a filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding to establish the nonlinear mapping relationship between filter geometry and electromagnetic response parameters; S2: The filter characteristic prediction surrogate model is trained using the S-parameter weighted mean square error loss function to obtain the trained filter characteristic prediction surrogate model. S3: Based on the physical perception-based improved differential evolution algorithm, the trained filter characteristic prediction surrogate model is used as the predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

[0041] This embodiment focuses on a fourth-order microwave filter ( Figure 3 The filter was optimized using a method that incorporates seven key geometric design parameters (Table 1). Simulations were performed at frequencies ranging from 25.0 GHz to 31.0 GHz, with a frequency scan step size of 150 frequency samples.

[0042] Table 1

[0043] Construction and training of the proxy model: 1. Dataset Construction A full-wave simulation was performed using Python-HFSS co-simulation, randomly selecting 6000 combinations of geometric parameters in the parameter space via Latin hypercube sampling to obtain the corresponding S11 and S21 frequency response data. Z-score normalization was applied to the input geometric parameters X and the output S-parameter Y to accelerate neural network convergence.

[0044] 2. Construction and Weighted Training of the Proxy Model A filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding, as proposed in this invention, is constructed, and the AdamW optimizer is used for iterative parameter updates with an initial learning rate of 0.001. In the loss calculation stage of model training, a weighted mean square error loss function strategy is applied: by real-time detection of the true amplitude of the full-wave simulation S-parameters, for deep resonance regions with amplitudes less than -10dB, the loss is calculated according to the formula...

[0045] weight: represents the dynamic weight coefficient assigned to the current frequency point when calculating the loss function.

[0046] y true : Represents the true amplitude of the S-parameters obtained from the full-wave electromagnetic simulation (the unit is usually dB).

[0047] T: Represents the preset amplitude threshold (used to distinguish whether it is a deep resonance / transmission zero region).

[0048] α: Represents the dynamic adjustment coefficient (or variation scaling factor, used to control the growth rate of weights as depth increases).

[0049] In this embodiment, a preset amplitude threshold is used. T The value is 10 (corresponding to -10dB), which is the dynamic adjustment coefficient. α The value is 0.1; Dynamically increasing the loss weight at this frequency point forces the network to focus on fitting the trough features. Final validation results show that the trained model achieves a global R-squared score exceeding 0.82 on the test set for S11 and exceeding 0.97 for S21, and accurately matches the S-parameters corresponding to the physical parameter combinations (such as...). Figure 4 , Figure 5 This provides a high-fidelity prediction benchmark for subsequent optimization.

[0050] Automatic Filter Optimization Design Based on Improved Differential Evolution 1. Set design goals Passband frequency: 27.2 GHz - 29.0 GHz.

[0051] Passband specifications: S11 < -20 dB, S21 > -0.1 dB.

[0052] Stopband specifications: Two deep transmission zeros exist near 26 GHz–27 GHz and 29 GHz–30 GHz to achieve high out-of-band rejection.

[0053] 2. Execute the improved differential evolution algorithm The population size is set to 200, and the maximum number of iterations is set to 100. During the algorithm's execution, the fitness function performs the following physical perception logic: Passband edge anchoring: The algorithm focuses on checking the S11 predictions at 27.2 GHz (left edge) and 29.0 GHz (right edge). If edge warping is found (e.g., greater than -15 dB), a penalty weight of 2000 is applied.

[0054] Zero-point frequency pulling: The algorithm automatically searches for the minimum values ​​of the S21 curve in the ranges of 26 GHz–27 GHz and 29 GHz–30 GHz. If the minimum frequency deviates from this band, a lateral pulling penalty is applied based on the frequency deviation distance.

[0055] 3. Optimize Result Analysis The algorithm stops running when it reaches the maximum number of iterations. Each iteration takes only about 0.3-0.5 seconds, and the total optimization time is 84 seconds.

[0056] The final output of the optimal geometric parameters is: Table 2

[0057] 4. Verification and Demonstration of Beneficial Effects The final optimized values ​​obtained after the above optimization were input into HFSS for full-wave electromagnetic simulation verification, and the results are as follows: Figure 6 As shown: Dashed line (S11): In the passband range of 27.2 GHz to 29.0 GHz, the return loss is less than -20dB, and the curve is generally flat without significant tilting, proving that the "edge anchoring strategy" of the present invention effectively suppresses passband tilting.

[0058] Solid line (S21): The transmission zero on the left is located at 26.42 GHz with a depth of -52.19 dB; the transmission zero on the right is located at 29.8 GHz with a depth of -41.95 dB. These two zero positions are highly consistent with the design target, proving that the "frequency pulling mechanism" of this invention successfully controls the zero at the specified frequency, significantly enhancing the out-of-band rejection capability of the filter.

[0059] This embodiment demonstrates that, by leveraging a trained surrogate model and optimization algorithm, the present invention can design microwave filters with excellent electrical performance within a relatively short computation time. Compared to traditional methods that require several hours of simulation and debugging, this significantly improves efficiency, and the designed structure directly meets engineering application requirements without the need for manual fine-tuning.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A filter physical awareness optimization method based on residual attention and transposed convolution, characterized in that: Includes the following steps: A filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding is constructed to establish the nonlinear mapping relationship between filter geometric parameters and electromagnetic response parameters. The filter characteristic prediction surrogate model is trained using the S-parameter weighted mean square error loss function to obtain the trained filter characteristic prediction surrogate model. The differential evolution algorithm based on physical perception is used as a predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

2. The filter physical awareness optimization method based on residual attention and transposed convolution according to claim 1, characterized in that: The filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding includes: an encoder and a decoder; The encoder includes: Mapping module: used to map the geometric parameters of the filter as input to a high-dimensional latent feature space; Feature extraction module: used to extract features from the mapped feature vector transmitted by the mapping module, and capture the nonlinear fluctuations of S-parameters caused by small changes in geometric parameters; Fusion module: used to fuse the features extracted by the feature extraction module with the original input, retaining the original information while adaptively focusing on the key geometric features that have the greatest impact on the filter response; The decoder includes: Multiple cascaded transposed convolutional blocks map the extracted deep semantic features back to the physical space of the target frequency points, and finally output the S-parameter response curve containing 150 frequency points, realizing end-to-end fast prediction from geometric structure to electromagnetic properties.

3. The filter physical perception optimization method based on residual attention and transposed convolution according to claim 2, characterized in that: The mapping module includes: initially increasing the dimensionality of the geometric structure parameters of the input filter through a linear layer, and then mapping the low-dimensional geometric physical parameters to a high-dimensional latent feature space in conjunction with the LeakyReLU activation function.

4. The filter physical perception optimization method based on residual attention and transposed convolution according to claim 2, characterized in that: The feature extraction module includes a residual block, which employs a two-layer perceptron structure containing a main path and skip connections. The main path sequentially includes: a first linear layer, a first batch of normalized layers, a first LeakyReLU activation layer, a second linear layer, and a second batch of normalized layers; The skip connection sums the input signal of the residual block directly with the output of the main path, and the sum is then activated and output by LeakyReLU.

5. The filter physical perception optimization method based on residual attention and transposed convolution according to claim 2, characterized in that: The fusion module introduces a self-attention mechanism. The feature vectors extracted by the feature extraction module are transformed into a query matrix Q, a key matrix K, and a value matrix V through linear transformation. By calculating the dot product of Q and K and performing scaling and Softmax normalization, an attention weight map is obtained. Then, it is multiplied by V with a weight, and then the attention-weighted features are fused with the original input through an additive skip connection. The attention-enhanced feature vector is transformed through a linear layer and reshaped from a 1D vector into a feature map format suitable for convolution operations.

6. The filter physical awareness optimization method based on residual attention and transposed convolution according to claim 2, characterized in that: The transposed convolutional block sequentially performs transposed convolution, batch normalization, and LeakyReLU activation operations to upsample, gradually restoring the spatial dimension of the features and reconstructing frequency domain details; The decoded high-dimensional feature map is flattened by the Flatten layer and fed into the output head, which consists of the third linear layer, the second LeakyReLU layer, and the fourth linear layer.

7. The filter physical awareness optimization method based on residual attention and transposed convolution according to claim 1, characterized in that: The S-parameter weighted mean square error loss function employs a dynamic weight allocation strategy. When the true amplitude of the S-parameter is detected to be lower than a preset threshold, the weight coefficient (weight) of the corresponding frequency point in the loss function is calculated using the following formula: weight: represents the dynamic weight coefficient assigned to the current frequency point when calculating the loss function; y true : Represents the true amplitude of the S-parameters obtained from full-wave electromagnetic simulation; T: Represents the preset amplitude threshold, used to distinguish whether it is a deep resonance / transmission zero region. α: Represents the dynamic adjustment coefficient or variation scaling factor, used to control the growth rate of weights as depth increases.

8. The filter physical awareness optimization method based on residual attention and transposed convolution according to claim 1, characterized in that: The differential evolution algorithm based on physical perception constructs a fitness evaluation function that incorporates physical feature constraints, and includes two core mechanisms: One is the S11 passband edge anchoring strategy. When calculating fitness, in addition to considering the overall return loss of the passband, several frequency points within a preset range inside the passband boundary are extracted. When the S11 amplitude of the edge frequency point is detected to be worse than the design index, a penalty weight much higher than that of the ordinary frequency points in the passband is applied, so that the optimization direction prioritizes to reduce the reflection coefficient of the frequency band edge. The second is the S21 transmission zero-point frequency traction mechanism, which automatically searches for local minimum points of the predicted S21 curve within the preset stopband frequency range, obtains its actual frequency position, calculates the deviation distance between this frequency and the design target zero-point frequency, and accordingly transforms the lateral error in the frequency dimension into a fitness penalty term to guide the parameters to evolve towards the target zero-point frequency.

9. A filter physical awareness optimization device based on residual attention and transposed convolution, characterized in that: include: Module: Used to build a filter characteristic prediction proxy model based on residual self-attention mechanism and transposed convolution decoding, and to establish the nonlinear mapping relationship between the filter's geometric structure parameters and electromagnetic response parameters; Training module: Used to train the filter characteristic prediction surrogate model using the S-parameter weighted mean square error loss function, and obtain the trained filter characteristic prediction surrogate model; Optimization module: Used for the differential evolution algorithm based on physical perception improvement, it uses the trained filter characteristic prediction surrogate model as a predictor to find the optimal combination of geometric parameters that meets the filter design specifications in the solution space.

10. A readable storage medium storing a program module, characterized in that, The program module, when run in a processor, can implement the method as described in any one of claims 1-8.

Images