Microwave power divider performance optimization method based on pattern recognition

Abstract

The invention discloses a microwave power divider performance optimization method based on pattern recognition, and the method comprises the steps: collecting the structure parameters and performance data of a microwave power divider, and generating a standardized data set; establishing an improved Hopfield neural network model, and training and outputting a comprehensive evaluation quantity; initializing a lion group algorithm, sampling to generate an initial solution, and screening a feasible set; executing optimization iteration updating of the lion group optimization algorithm; judging whether the optimization result meets a performance target and a physical constraint or not; inputting the candidate optimization structure parameters into electromagnetic simulation to verify errors, and updating model parameters if the errors exceed the limit; and outputting the final optimization structure parameters and the corresponding performance indexes, and returning to the model. According to the method, the improved Hopfield neural network model and the lion group optimization algorithm are fused, an intelligent optimization mechanism driven by physical constraints is established, and self-adaptive optimization and high-precision performance prediction of structural parameters of the microwave power divider are achieved.

Description

Technical Field

[0001] This invention relates to the field of microwave and radio frequency circuit design technology, and in particular to a method for optimizing the performance of microwave power dividers based on pattern recognition. Background Technology

[0002] Currently, microwave power dividers, as key passive components in radio frequency and microwave systems, are widely used in communications, radar, and navigation. Their performance stability and parameter accuracy directly affect the amplitude, phase, and bandwidth characteristics of system signals. Traditional power divider design methods mainly rely on a combination of theoretical calculations, equivalent circuit analysis, and electromagnetic simulation optimization, using parameter scanning and manual tuning to approximate target performance values. However, with the rise of high-frequency broadband and multi-port structures, the design variables of power dividers have increased significantly, and the efficiency and accuracy of traditional design methods in high-dimensional parameter spaces have gradually decreased. These methods typically rely on design experience, lacking adaptive and global search capabilities in the optimization process, resulting in massive computational loads, slow optimization speeds, and performance results that are sensitive to initial conditions, making it difficult to meet the miniaturization, broadband, and intelligent requirements of modern radio frequency systems.

[0003] With the development of artificial intelligence and data-driven technologies, pattern recognition and machine learning methods are increasingly being introduced into microwave circuit performance prediction and structural optimization design. Existing research has attempted to establish mapping relationships between structural parameters and performance indicators using models such as neural networks and support vector machines, and to combine these with swarm intelligence algorithms for automatic optimization. However, most existing methods are purely data-driven models, neglecting physical laws such as energy conservation, impedance matching, and signal symmetry in microwave device design, resulting in prediction results lacking physical consistency and interpretability. Furthermore, traditional optimization algorithms are prone to getting trapped in local optima in complex multimodal spaces, exhibiting insufficient model generalization and transferability.

[0004] Therefore, how to provide a method for optimizing the performance of microwave power dividers based on pattern recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a microwave power divider performance optimization method based on pattern recognition. This invention utilizes artificial intelligence methods such as pattern recognition, physical constraint learning, and the lion pack optimization algorithm to detail the entire process from sample data construction, model training, physical constraint learning to global optimization of structural parameters and simulation verification. By introducing an input layer, correlation layer, output layer, and physical constraint control layer into the Hopfield neural network model, an improved Hopfield neural network model is constructed, achieving pattern recognition and physical consistency learning between structural parameters and performance indicators. Simultaneously, combining the lion pack optimization algorithm and introducing a dynamic hunting mechanism and adaptive step size adjustment strategy on the basis of the traditional lion pack optimization algorithm, global optimization and rapid convergence in the high-dimensional parameter space are achieved. In the performance prediction stage, this invention considers physical laws such as energy conservation, impedance matching, and signal symmetry. In the optimization stage, it introduces a multi-layer feedback and adaptive control mechanism, realizing a synergistic closed loop of performance evaluation, parameter optimization, and model self-learning. Compared with existing technologies, this invention has advantages such as high optimization accuracy, fast convergence efficiency, strong physical consistency, and excellent interpretability.

[0006] A microwave power divider performance optimization method based on pattern recognition according to an embodiment of the present invention includes: Sample data of the structural parameters and corresponding performance indicators of the microwave power divider were collected, and the structural parameters of the microwave power divider were preprocessed to generate a standardized dataset. An improved Hopfield neural network model was established, which was trained using the structural parameters of the microwave power divider as input vectors, and output a comprehensive evaluation value. The search population for the lion pack optimization algorithm is initialized. Based on the range of values ​​of the structural parameters of the microwave power divider, an initial solution set is generated using the Latin hypercube sampling method, and the first generation population is obtained by physical feasibility screening. The optimization iteration is performed by inputting the structural parameters of each individual in the population into the improved Hopfield neural network model, obtaining the optimized comprehensive evaluation value, calculating the comprehensive fitness value by combining the standardized target performance index, and performing iterative updates of the lion flock optimization algorithm based on the comprehensive fitness value. Determine whether the optimization results of the population meet the set performance target and the physical consistency constraint threshold. If they do, record the structural parameters of the best individual as candidate optimization structural parameters; otherwise, return to the optimization iteration process. The candidate optimized structural parameters are input into an electromagnetic simulation tool for simulation verification. The simulation performance indicators are obtained and compared with the prediction results of the improved Hopfield neural network model. If the error exceeds the error tolerance threshold, the parameters of the improved Hopfield neural network model are updated and the training process is returned. Under the condition that the simulation performance indicators meet the set performance target and the error does not exceed the error tolerance threshold, the final optimized structural parameters and corresponding performance indicators are output, and the final optimization results are returned to the improved Hopfield neural network model for updating.

[0007] Optionally, the structural parameters and corresponding performance indicators of the microwave power divider include: The transmission line width, transmission line length, transmission line spacing, and isolation resistor value of a microwave power divider correspond to the performance indicators, which are the upper limit of return loss, upper limit of insertion loss, lower limit of isolation, upper limit of amplitude imbalance, upper limit of phase imbalance, and bandwidth requirements of each sample after combining multiple samples within the range of values ​​of each structural parameter. Optionally, the preprocessing of the structural parameters of the microwave power divider includes: The structural parameters of the microwave power divider are standardized in terms of dimensions and outliers are removed. The zero-mean unit variance method is used for normalization to generate a standardized dataset.

[0008] Optionally, the output comprehensive evaluation quantity includes: Based on the Hopfield neural network model, the underlying connection structure and energy evolution mechanism are modified to form an improved Hopfield neural network model with an input layer, a correlation layer, an output layer, and a physical constraint control layer. The layers are connected by asymmetric weighted connections, and dynamic constraints and hierarchical feedback mechanisms are introduced in the energy evolution process. The input layer receives the structural parameters of the microwave power divider, introduces a normalized mapping function and an input weight matrix, and establishes a mapping relationship between the structural parameter vector and the internal nodes of the improved Hopfield neural network model. The correlation layer computation improves the energy coupling relationship between nodes inside the Hopfield neural network model and forms a multi-channel asymmetric connection structure inside the improved Hopfield neural network model; The output layer adds a calculation unit for energy change rate and physical deviation, summarizes the energy state of each node in the improved Hopfield neural network model, and generates a comprehensive energy value and energy change trend. The physical constraint control layer introduces energy conservation, impedance matching, and signal symmetry into the energy function of the improved Hopfield neural network model to regulate the energy evolution process in real time. Establish inter-layer feedback channels to interconnect and transmit energy state information between the input layer, correlation layer, output layer and physical constraint control layer; Under the energy reduction criterion, the improved Hopfield neural network model is iteratively trained based on the structural parameters of the microwave power divider, and the weight coefficients, threshold bias and constraint factors are continuously adjusted. When the energy function converges and the physical constraint error is reduced to the set range, the model parameters are fixed and the energy mapping learning of the training phase is completed. When the improved Hopfield neural network model training converges and the physical constraints meet the threshold conditions, it outputs energy state information, physical consistency evaluation value, and optimized comprehensive evaluation value formed by weighted fusion of energy state, physical constraint deviation and performance prediction results during the training phase.

[0009] Optionally, the process of performing physical feasibility screening to obtain the first-generation population includes: The range of values ​​for the structural parameters of each microwave power divider is determined based on the standardized dataset, and normalization is performed according to the maximum and minimum values ​​to generate a normalized parameter space. In the normalized parameter space, the combination of structural parameters of the microwave power divider is generated by Latin hypercube sampling, forming the initial solution set of structural parameters of the microwave power divider. Based on the physical consistency conditions including energy conservation, impedance matching and signal symmetry, the initial solution set is screened, and the feasible set that meets the physical constraint threshold is retained. Based on population size and sex ratio, the feasible set is divided into male lions, female lions, and wandering individuals, and energy feedback lions, constraint evaluation lions, and leader lions are added. The energy feedback lion group adaptively adjusts according to the rate of energy change, the constraint evaluation lion group determines physical consistency and constraint satisfaction, and the leader lion is the optimal individual selected from male lions based on the comprehensive fitness value, forming a multi-level lion group structure. The initial hunting step size is set by combining the structural parameter sensitivity of each individual in the population. Boundary correction is performed on individuals that exceed the range to obtain a stable first-generation population. Standardized target performance indicators are then determined based on the performance indicator boundaries.

[0010] Optionally, the step of calculating the comprehensive fitness value by combining the standardized target performance index and performing iterative updates of the lion pride optimization algorithm based on the comprehensive fitness value includes: The structural parameter vectors of each individual in the population are input into the improved Hopfield neural network model to obtain the optimized comprehensive evaluation and energy state values. Based on the difference between the standardized target performance index and the prediction performance of the improved Hopfield neural network model, the target deviation is calculated; and the comprehensive fitness value is formed by combining the optimized comprehensive evaluation value and the physical consistency penalty value. The hunting direction is determined based on the comprehensive fitness value. At the same time, the energy feedback lion pride adaptively adjusts the step size according to the energy change rate. The step size decreases when the energy change rate is faster. An inertial adjustment factor and a convergence adjustment factor are introduced to smooth the iteration process. The coordinated hunt is carried out based on the position of the lead lion and the center position of the lionesses, while the energy feedback lions also participate in correcting the hunting direction and optimizing the hunting direction. The free lions perform random diffusion within the structural parameter space, and the constrained evaluation lions perform physical consistency screening on the diffusion results, retaining individuals that satisfy energy conservation and impedance matching. According to the set pairing probability, parent individuals are selected from the leader lion and the liones, and structural parameters are recombined and minor variations are performed to generate new individuals. After physical consistency is checked, they are added to the population. The elite individuals with the highest overall fitness value are retained. Individuals that hunt, migrate, and reproduce are merged and sorted. The best individuals are selected to form the next generation of the population. The optimal structural parameters and the optimized overall evaluation value are recorded.

[0011] Optionally, determining whether the optimization result of the population satisfies the set performance target and the physical consistency constraint threshold includes: The structural parameter vector of the individual to be evaluated is determined, and standardized target performance indicators and physical consistency constraint thresholds are set. The physical consistency constraints include impedance matching, energy conservation, structural symmetry, and geometric boundaries. The overall performance deviation value is obtained by weighting and group standardizing the target deviation. Calculate the physical consistency deviation of the structural parameters, which consists of impedance matching deviation, energy conservation deviation, structural symmetry deviation, and geometric overrun penalty. The judgment is based on the comprehensive performance deviation value and the physical consistency deviation value. When neither of them exceeds the threshold, it is determined that the performance target and physical consistency constraints are met; otherwise, it is determined that they are not met. At the same time, the joint judgment value is calculated. Record the structural parameters of individuals that meet the conditions as candidate optimized structural parameters, and save the candidate optimized structural parameters, the optimized comprehensive evaluation value, and the joint judgment value.

[0012] Optionally, updating and improving the parameters of the Hopfield neural network model and returning to perform the training process includes: Based on the candidate optimized structural parameters, the port excitation, reference impedance, and frequency sampling points of the electromagnetic simulation environment are set. In the electromagnetic simulation environment, the performance index of the candidate optimized structural parameters is calculated, and the results at each frequency point are interpolated and statistically analyzed to obtain the simulation performance index sequence and the representative value of the frequency band. The performance prediction results corresponding to the optimized comprehensive evaluation quantity are compared with the simulation performance indicators one by one. The absolute error and relative error of each indicator are calculated, and the total error measure is obtained by weighted summation of the errors of each indicator according to the weight. When the total error metric exceeds the error tolerance threshold, the improved Hopfield neural network model parameters are updated, the inter-layer weights, node threshold biases and physical constraint factors are adjusted, and a new total error metric is calculated; multiple iterative corrections are performed while maintaining the physical consistency constraint; if the error still does not converge after multiple corrections, the improved Hopfield neural network model is retrained. When the retraining mechanism is triggered, the updated and improved Hopfield neural network model is returned to the training phase and relearned using a standardized dataset to correct the global parameter distribution. When the error metric converges after retraining and meets the physical consistency constraint, the candidate optimized structural parameters are confirmed as valid results and the corresponding performance indicators and the optimized comprehensive evaluation are recorded.

[0013] Optionally, the output finally optimizes the structural parameters and corresponding performance indicators, and returns the final optimization result to the improved Hopfield neural network model, including: Input the candidate optimized structural parameters into the electromagnetic simulation tool to obtain the simulation performance index vector and set the target performance vector and index weight vector; Compare the simulation performance index vector with the target performance vector, calculate the comprehensive satisfaction metric, and determine the pass when the set threshold is met and all key indicators reach the target threshold; Obtain the prediction performance vector corresponding to the optimized comprehensive evaluation quantity, calculate the total error metric and physical consistency deviation, and pass the verification when both are within the target threshold. The candidate optimized structural parameters that have passed the verification are solidified as the final optimized structural parameters. The corresponding simulation performance index, comprehensive satisfaction metric, total error metric and physical consistency evaluation value are output, and the corresponding comprehensive evaluation quantity after optimization is recorded. The final optimized structural parameters, corresponding simulation performance indicators, and error verification results are returned to the model for incremental correction and updating of the standardized dataset to determine the tolerance range of structural parameters and the test acceptance threshold.

[0014] The beneficial effects of this invention are: This invention constructs a microwave power divider performance optimization method based on pattern recognition, establishing a performance optimization system that combines data-driven and physical constraints, significantly improving the design efficiency and performance accuracy of microwave power dividers. The method introduces an input layer, correlation layer, output layer, and physical constraint control layer into the Hopfield neural network model, forming an improved Hopfield neural network model. This improved model automatically follows electromagnetic physical laws such as energy conservation, impedance matching, and signal symmetry while learning the mapping relationship between structural parameters and performance indicators, thus ensuring the physical consistency and interpretability of performance prediction results. By using a pattern recognition algorithm to perform deep learning on the nonlinear relationship between structural parameters and performance indicators, high-precision modeling and comprehensive evaluation of power divider performance are achieved.

[0015] During the optimization process, this invention combines the output results of an improved lion pack optimization algorithm and an improved Hopfield neural network model to construct a comprehensive fitness value. It proposes a dynamic hunting mechanism based on energy state feedback and an adaptive step size adjustment strategy, enabling the optimization process to achieve an adaptive balance between global search and local convergence. In each iteration, the algorithm simultaneously considers performance objectives, energy state, and physical constraints, achieving collaborative optimization under multiple objectives and constraints, effectively avoiding the problem of traditional optimization algorithms getting trapped in local optima. Through a closed-loop training mechanism with the electromagnetic simulation verification module, this invention achieves self-correction of consistency between model predictions and simulation results, making the optimization results more stable and reliable. This invention possesses advantages such as high optimization accuracy, fast convergence speed, strong physical consistency, excellent interpretability, and high engineering feasibility, providing an efficient and scalable solution for the intelligent design and performance optimization of microwave power dividers and other RF devices. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a microwave power divider performance optimization method based on pattern recognition proposed in this invention; Figure 2 This is a schematic diagram illustrating the structure of the improved Hopfield neural network model and the lion pack optimization algorithm working together in a microwave power divider performance optimization method based on pattern recognition proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 A method for optimizing the performance of a microwave power divider based on pattern recognition, comprising: Sample data of the structural parameters and corresponding performance indicators of the microwave power divider were collected, and the structural parameters of the microwave power divider were preprocessed to generate a standardized dataset. An improved Hopfield neural network model was established, which was trained using the structural parameters of the microwave power divider as input vectors, and output a comprehensive evaluation value. The search population for the lion pack optimization algorithm is initialized. Based on the range of values ​​of the structural parameters of the microwave power divider, an initial solution set is generated using the Latin hypercube sampling method, and the first generation population is obtained by physical feasibility screening. The optimization iteration is performed by inputting the structural parameters of each individual in the population into the improved Hopfield neural network model, obtaining the optimized comprehensive evaluation value, calculating the comprehensive fitness value by combining the standardized target performance index, and performing iterative updates of the lion flock optimization algorithm based on the comprehensive fitness value. Determine whether the optimization results of the population meet the set performance target and the physical consistency constraint threshold. If they do, record the structural parameters of the best individual as candidate optimization structural parameters; otherwise, return to the optimization iteration process. The candidate optimized structural parameters are input into an electromagnetic simulation tool for simulation verification. The simulation performance indicators are obtained and compared with the prediction results of the improved Hopfield neural network model. If the error exceeds the error tolerance threshold, the parameters of the improved Hopfield neural network model are updated and the training process is returned. Under the condition that the simulation performance indicators meet the set performance target and the error does not exceed the error tolerance threshold, the final optimized structural parameters and corresponding performance indicators are output, and the final optimization results are returned to the improved Hopfield neural network model for updating.

[0020] In this embodiment, the structural parameters and corresponding performance indicators of the microwave power divider include: The transmission line width, transmission line length, transmission line spacing, and isolation resistor value of a microwave power divider correspond to the performance indicators, which are the upper limit of return loss, upper limit of insertion loss, lower limit of isolation, upper limit of amplitude imbalance, upper limit of phase imbalance, and bandwidth requirements of each sample after combining multiple samples within the range of values ​​of each structural parameter. In this embodiment, the preprocessing of the structural parameters of the microwave power divider includes: The structural parameters of the microwave power divider are standardized in terms of dimensions and outliers are removed. The zero-mean unit variance method is used for normalization to generate a standardized dataset.

[0021] In this embodiment, the output comprehensive evaluation quantity includes: Based on the Hopfield neural network model, the underlying connection structure and energy evolution mechanism are modified to form an improved Hopfield neural network model with an input layer, a correlation layer, an output layer, and a physical constraint control layer. The layers are connected by asymmetric weighted connections, and dynamic constraints and hierarchical feedback mechanisms are introduced in the energy evolution process. The input layer receives the structural parameters of the microwave power divider, introduces a normalized mapping function and an input weight matrix, and establishes a mapping relationship between the structural parameter vector and the internal nodes of the improved Hopfield neural network model. The association layer calculates and improves the energy coupling relationship between nodes within the Hopfield neural network model, and forms a multi-channel asymmetric connection structure within the improved Hopfield neural network model. Specifically, the calculation improves the energy coupling relationship between nodes within the Hopfield neural network model as follows: Based on the differences in the amplitude and direction of energy changes of internal nodes in the improved Hopfield neural network model, the coupling strength is determined by energy difference. The output layer adds a calculation unit for energy change rate and physical deviation, summarizes the energy states of each node within the improved Hopfield neural network model, and generates a comprehensive energy value and energy change trend. Specifically, the generation of the comprehensive energy value involves: The energy states of each node within the improved Hopfield neural network model are weighted and summed. The weights are determined based on the importance and energy coupling strength of the nodes within the improved Hopfield neural network model. The energy change rate and physical deviation correction results are also incorporated. The energy change rate calculation unit corrects the dynamic response of the comprehensive energy value by real-time monitoring of the rate of change of the energy function of the improved Hopfield neural network model. The physical deviation calculation unit calculates the deviation value based on three physical constraints: energy conservation, impedance matching, and signal symmetry, and applies physical consistency correction to the comprehensive energy value. The energy conservation deviation is obtained by calculating the power transfer ratio of each input and output port separately and taking the normalized difference between the sum of the input port power and the output port power. The impedance matching error is obtained by comparing the real part difference between the input port impedance and the standard characteristic impedance and calculating the normalized deviation. The signal symmetry deviation is obtained by calculating the normalized value of the amplitude difference and phase difference between the output ports. The physical constraint control layer introduces energy conservation, impedance matching, and signal symmetry into the energy function of the improved Hopfield neural network model to regulate the energy evolution process in real time. Specifically, this real-time regulation involves continuously monitoring the energy changes and physical deviations between nodes of the improved Hopfield neural network model. Energy conservation ensures the balance between input and output energy, impedance matching adjusts the transmission relationship between nodes, and signal symmetry verifies the amplitude-phase consistency of the output port. When any constraint deviates from the threshold, a penalty feedback signal is generated to dynamically correct the inter-layer weights and node thresholds of the improved Hopfield neural network model. Establish inter-layer feedback channels to interconnect and transmit energy state information between the input layer, correlation layer, output layer and physical constraint control layer; Under the energy reduction criterion, the improved Hopfield neural network model is iteratively trained based on the structural parameters of the microwave power divider, and the weight coefficients, threshold bias and constraint factors are continuously adjusted. When the energy function converges and the physical constraint error is reduced to the set range, the model parameters are fixed and the energy mapping learning of the training phase is completed. When the improved Hopfield neural network model training converges and the physical constraints meet the threshold condition, the output includes energy state information, physical consistency evaluation value, and an optimized comprehensive evaluation value formed by weighted fusion of energy state, physical constraint deviation, and performance prediction results during the training phase. Among these: The physical consistency evaluation value is obtained by calculating the deviation value of each of the three types of physical constraints, namely energy conservation, impedance matching and signal symmetry, and normalizing them according to the set threshold. The weights are set according to the importance of each constraint, and the three satisfaction values ​​are weighted and summed. The value ranges from 0 to 1. The closer the value is to 1, the higher the physical consistency of the output result of the improved Hopfield neural network model.

[0022] In this embodiment, the step of performing physical feasibility screening to obtain the first-generation population includes: The range of values ​​for the structural parameters of each microwave power divider is determined based on the standardized dataset, and normalization is performed according to the maximum and minimum values ​​to generate a normalized parameter space. The structural parameter combination of the microwave power divider is generated using Latin hypercube sampling within the normalized parameter space, forming the initial solution set of the microwave power divider's structural parameters. Specifically, the generation of the structural parameter combination of the microwave power divider using Latin hypercube sampling is as follows: The range of values ​​for each structural parameter is divided into several equally probable segments, and a sampling point is randomly selected in each equally probable segment. The sampling points are then combined in a random matching manner to generate multiple sets of multidimensional structural parameter samples, forming a comprehensive and evenly distributed combination of structural parameters for the microwave power divider. Based on the physical consistency conditions including energy conservation, impedance matching, and signal symmetry, the initial solution set is screened, retaining feasible sets that meet the physical constraint thresholds. Specifically, the screening process involves calculating the input-output port power difference for each set of structural parameters in the initial solution set of the microwave power divider to verify the energy conservation condition, comparing the deviation between the input port impedance and the standard characteristic impedance to determine the impedance matching degree, and detecting the amplitude and phase difference between the output ports to evaluate the signal symmetry. Only when all three physical consistency indicators are within the set threshold range is the structural parameter set determined to be a physically feasible solution; otherwise, it is discarded. Based on population size and sex ratio, the feasible set is divided into male lions, female lions, and wandering individuals, and energy feedback lions, constraint evaluation lions, and leader lions are added. The energy feedback lion group adaptively adjusts according to the rate of energy change, the constraint evaluation lion group determines physical consistency and constraint satisfaction, and the leader lion is the optimal individual selected from male lions based on the comprehensive fitness value, forming a multi-level lion group structure. The initial hunting step size is set by combining the structural parameter sensitivity of each individual in the population. Boundary corrections are performed on individuals exceeding the range to obtain a stable first-generation population. Standardized target performance indicators are then determined based on the performance indicator boundaries, including: Standardized target performance metrics are generated by normalizing and aligning the performance metrics in the standardized dataset according to a uniform scale.

[0023] In this embodiment, the step of calculating the comprehensive fitness value by combining standardized target performance indicators and performing iterative updates of the lion pride optimization algorithm based on the comprehensive fitness value includes: The structural parameter vectors of each individual in the population are input into the improved Hopfield neural network model to obtain the optimized comprehensive evaluation and energy state values. Based on the difference between the standardized target performance index and the predicted performance of the improved Hopfield neural network model, the target bias is calculated; and combined with the optimized comprehensive evaluation value and the physical consistency penalty value, a comprehensive fitness value is constructed, wherein: The target deviation is calculated by calculating the absolute difference between each performance index predicted by the improved Hopfield neural network model and the standardized target performance index, and assigning different weights to each performance index according to its importance, and then summing the weighted deviations of each performance index. The physical consistency penalty is obtained by linearly weighting the energy conservation deviation, impedance matching error, and signal symmetry deviation according to a set weight. The overall fitness value is obtained by weighted fusion of the inverse term of the target deviation, the positive term of the optimized overall evaluation, and the negative term of the physical consistency penalty. The hunting direction is determined based on the comprehensive fitness value, while the energy feedback lion pride adaptively adjusts its step size according to the rate of energy change; the step size decreases as the rate of energy change increases. An inertia adjustment factor and a convergence adjustment factor are introduced to smooth the iterative process, wherein: The hunting direction is determined based on the comprehensive vector relationship between the individual's current position, the leader lion's position, and the group center's position. The initial hunting direction is determined by calculating the distance vector of the current individual relative to the leader lion and the group center in the parameter space, and combining the changing trend of the comprehensive fitness value gradient. At the same time, inertial weights and random perturbations are applied to the initial hunting direction vector. The hunting is carried out in a coordinated manner based on the position of the lead lion and the center position of the lionesses. At the same time, the energy feedback lions cooperate to correct the hunting direction and optimize the hunting direction. The optimization of the hunting direction is achieved by the energy feedback lions applying a decay correction to the initial hunting direction vector based on the energy change rate of the improved Hopfield neural network model. The free lions perform random diffusion within the structural parameter space, and the constrained evaluation lions perform physical consistency screening on the diffusion results, retaining individuals that satisfy energy conservation and impedance matching. According to the set pairing probability, parent individuals are selected from the leader lion and the liones, and structural parameters are recombined and minor variations are performed to generate new individuals. After physical consistency is checked, they are added to the population. The elite individuals with the highest overall fitness value are retained. Individuals that hunt, migrate, and reproduce are merged and sorted. The best individuals are selected to form the next generation of the population. The optimal structural parameters and the optimized overall evaluation value are recorded.

[0024] In this embodiment, determining whether the optimization result of the population satisfies the set performance target and the physical consistency constraint threshold includes: The structural parameter vector of the individual to be evaluated is determined, and standardized target performance indicators and physical consistency constraint thresholds are set. The physical consistency constraints include impedance matching, energy conservation, structural symmetry, and geometric boundaries. The overall performance deviation value is obtained by weighting and group standardizing the target deviation. Calculate the physical consistency deviation of the structural parameters, which consists of impedance matching deviation, energy conservation deviation, structural symmetry deviation, and geometric overrun penalty. The judgment is based on the comprehensive performance deviation value and the physical consistency deviation value. When neither exceeds the threshold, it is determined that the performance target and physical consistency constraints are met; otherwise, it is determined that they are not met. A joint judgment value is also calculated, where: The joint judgment value is calculated by weighted fusion of the comprehensive performance deviation value and the physical consistency deviation value. Record the structural parameters of individuals that meet the conditions as candidate optimized structural parameters, and save the candidate optimized structural parameters, the optimized comprehensive evaluation value, and the joint judgment value.

[0025] In this embodiment, updating and improving the parameters of the Hopfield neural network model and returning to execute the training process includes: Based on the candidate optimized structural parameters, the port excitation, reference impedance, and frequency sampling points of the electromagnetic simulation environment are set. In the electromagnetic simulation environment, the performance index of the candidate optimized structural parameters is calculated, and the results at each frequency point are interpolated and statistically analyzed to obtain the simulation performance index sequence and the representative value of the frequency band. The performance prediction results corresponding to the optimized comprehensive evaluation quantity are compared with the simulation performance indicators one by one. The absolute error and relative error of each indicator are calculated, and the total error measure is obtained by weighted summation of the errors of each indicator according to the weight. When the total error metric exceeds the error tolerance threshold, the improved Hopfield neural network model parameters are updated, the inter-layer weights, node threshold biases and physical constraint factors are adjusted, and a new total error metric is calculated; multiple iterative corrections are performed while maintaining the physical consistency constraint; if the error still does not converge after multiple corrections, the improved Hopfield neural network model is retrained. When the retraining mechanism is triggered, the updated and improved Hopfield neural network model is returned to the training phase and relearned using a standardized dataset to correct the global parameter distribution. When the error metric converges after retraining and meets the physical consistency constraint, the candidate optimized structural parameters are confirmed as valid results and the corresponding performance indicators and the optimized comprehensive evaluation are recorded.

[0026] In this embodiment, the step of outputting the final optimized structural parameters and corresponding performance indicators, and returning the final optimization result to the improved Hopfield neural network model, includes: Input the candidate optimized structural parameters into the electromagnetic simulation tool to obtain the simulation performance index vector and set the target performance vector and index weight vector; Compare the simulation performance index vector with the target performance vector, calculate the comprehensive satisfaction metric, and determine the pass when the set threshold is met and all key indicators reach the target threshold, where: The comprehensive satisfaction metric is calculated based on the deviation ratio between each simulation performance index and the target performance index. The absolute value of the difference between the target value and the predicted value of each performance index is used as the satisfaction level. The satisfaction results of each index are weighted according to their importance and then summed by weight to obtain the comprehensive satisfaction metric. The value ranges from 0 to 1. The closer the value is to 1, the closer the simulation result is to the target performance. Obtain the prediction performance vector corresponding to the optimized comprehensive evaluation quantity, calculate the total error metric and physical consistency deviation, and pass the verification when both are within the target threshold. The candidate optimized structural parameters that have passed the verification are solidified as the final optimized structural parameters. The corresponding simulation performance index, comprehensive satisfaction metric, total error metric and physical consistency evaluation value are output, and the corresponding comprehensive evaluation quantity after optimization is recorded. The final optimized structural parameters, corresponding simulation performance indicators, and error verification results are returned to the model for incremental correction and updating of the standardized dataset to determine the tolerance range of structural parameters and the test acceptance threshold.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the performance optimization design of a 2.4GHz Wilkinson power divider used in the RF front-end of a 5G small base station in an experimental area. This power divider requires low insertion loss, high isolation, and good amplitude-phase consistency while maintaining a 50Ω impedance match. Traditional methods, such as manual empirical methods and manual parameter scanning, often require repeated adjustments to structural dimensions and resistance values, resulting in long optimization cycles, numerous simulations, and results heavily influenced by the designer's experience. This makes it difficult to simultaneously meet multiple performance indicators within a limited timeframe, leading to prominent issues such as high return loss and insufficient isolation performance.

[0028] In practical applications, a Taconic RF-35 dielectric substrate (dielectric constant 3.5, thickness 1.524 mm) was selected as the experimental prototype, and optimization experiments were conducted with a center frequency of 2.4 GHz. The initial design employed traditional formulas and empirical fine-tuning, resulting in a sample exhibiting a return loss of −19.8 dB, insertion loss of 0.28 dB, and isolation of 24.2 dB in CST simulations. Subsequently, the optimization method proposed in this invention was incorporated into the same design flow, allowing for multi-dimensional adjustments to the transmission line width, length, spacing, and isolation resistance. The system automatically derives parameter combinations and generates candidate structures through pattern recognition and intelligent optimization. After optimization, the obtained design parameters are: transmission line width 1.05 mm, length 30.2 mm, spacing 0.45 mm, and isolation resistance 100.4 Ω.

[0029] The results show that the optimized power divider achieves a return loss of −33.2 dB at the center frequency, a insertion loss of 0.14 dB, an isolation of 37.0 dB, an amplitude imbalance of only 0.05 dB, a phase imbalance of 0.8°, and an effective bandwidth expansion to 26.8%. These results were verified in actual prototype testing using a vector network analyzer, with the measured curves showing good agreement with the simulation results. Compared to traditional methods, the design cycle was shortened from 4 hours to less than 1 hour, and the number of optimization iterations was reduced by approximately 60%, while maintaining high physical consistency and stability. This invention not only improves design efficiency and prediction accuracy in microwave power divider performance optimization but also significantly enhances robustness to dielectric fluctuations and structural errors, demonstrating good engineering feasibility and promotional value.

[0030] Table 1: Performance Verification Statistics of the Method of the Invention in Wilkinson Power Divider Optimization

[0031] As shown in Table 1, the microwave power divider performance optimization method based on pattern recognition proposed in this invention outperforms existing methods in both performance improvement and optimization efficiency. Using a 2.4GHz Wilkinson power divider as the test object, a comparison was made between traditional manual design, particle swarm optimization (PSO), Bayesian optimization (BO), Hopfield neural network model combined with lion swarm optimization algorithm, and the method of this invention, all within the same electromagnetic simulation environment. The results show that traditional manual experience and manual parameter scanning methods rely on the designer's experience and judgment, resulting in a return loss of −19.8dB, an isolation of 24.2dB, and an insertion loss maintained at around 0.28dB. Furthermore, the design time exceeds 4 hours, indicating low optimization efficiency and significant fluctuations in results.

[0032] The automation level was improved by employing particle swarm optimization and Bayesian optimization methods, but the optimization stability was insufficient due to the influence of random initial samples and model kernel functions. Particle swarm optimization could improve the return loss to around -25.6dB in most cases, but it was still prone to getting trapped in local optima. Although the Bayesian optimization method converged faster, it was unstable under complex multi-parameter conditions, especially in impedance matching and phase balance, where the errors were more obvious. The basic Hopfield neural network combined with the basic lion swarm optimization algorithm outperformed the previous two methods in global search capability, improving the return loss to -29.1dB and the isolation to 33.0dB, but it still had problems with high-frequency phase deviation and local energy convergence instability.

[0033] This invention proposes a multi-layered energy constraint architecture based on the Hopfield neural network model. It innovatively introduces a four-layer structure: input, correlation, output, and physical constraints. Combined with the adaptive step size adjustment and hunting strategy of the lion pack optimization algorithm, it achieves a dynamic balance between energy convergence and physical consistency constraints. Experimental results show that the optimized power divider at 2.4 GHz achieves a return loss of −33.2 dB, an isolation of 37.0 dB, a insertion loss of 0.14 dB, an amplitude imbalance of only 0.05 dB, and a phase imbalance controlled within 0.8°. Compared with traditional methods, the design time is reduced by more than 60%, optimization convergence is faster, and stability is significantly improved. These results fully verify the superiority of the proposed method in terms of design accuracy, energy constraint preservation, and cross-frequency generalization performance, providing a feasible technical path for the high-performance intelligent design of microwave power dividers.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for performance optimization of a microwave power divider based on pattern recognition, characterized in that, The method comprises the following steps: Collecting sample data of structural parameters of the microwave power divider and corresponding performance indicators, preprocessing the structural parameters of the microwave power divider, and generating a standardized data set; An improved Hopfield neural network model is established, the structural parameters of the microwave power divider are taken as an input vector for training, and a comprehensive evaluation quantity is output; The search population of the lion swarm optimization algorithm is initialized, the initial solution set is generated by using the Latin hypercube sampling method according to the value range of the structural parameters of the microwave power divider, and the first generation population is obtained after physical feasibility screening; In the optimization iteration process, the structural parameters of each individual in the population are input into the improved Hopfield neural network model to obtain the optimized comprehensive evaluation quantity, the comprehensive fitness value is calculated in combination with the standardized target performance indicators, and the iteration update of the lion swarm optimization algorithm is performed according to the comprehensive fitness value; It is judged whether the optimization result of the population meets the set performance target and satisfies the physical consistency constraint threshold, if yes, the structural parameters of the optimal individual are recorded as the candidate optimization structural parameters, otherwise the optimization iteration process is returned; The candidate optimization structural parameters are input into an electromagnetic simulation tool for simulation verification, the simulation performance indicators are obtained, and error comparison is performed with the prediction results of the improved Hopfield neural network model, if the error exceeds the error tolerance threshold, the parameters of the improved Hopfield neural network model are updated and the training process is returned; Under the condition that the simulation performance indicators meet the set performance target and the error does not exceed the error tolerance threshold, the final optimization structural parameters and corresponding performance indicators are output, and the final optimization result is returned to the improved Hopfield neural network model for updating.

2. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The structural parameters of the microwave power divider and the corresponding performance indicators comprise: The transmission line width, transmission line length, transmission line spacing and isolation resistance value of the microwave power divider, and the corresponding performance indicators are the upper limit of return loss, the upper limit of insertion loss, the lower limit of isolation, the upper limit of amplitude imbalance, the upper limit of phase imbalance and bandwidth requirement of each sample after a plurality of samples are generated by combining the value range of each structural parameter; According to the microwave power divider performance optimization method based on pattern recognition in claim 1, the preprocessing of the structural parameters of the microwave power divider comprises: The dimensional uniformity and abnormal value elimination of the structural parameters of the microwave power divider are performed, the zero mean unit variance method is used for normalization processing, and a standardized data set is generated.

3. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The output comprehensive evaluation quantity comprises: Based on the Hopfield neural network model, the bottom connection structure and energy evolution mechanism are modified to form an improved Hopfield neural network model with an input layer, an associated layer, an output layer and a physical constraint control layer; the layers are connected through asymmetric weighting, and a dynamic constraint and hierarchical feedback mechanism is introduced in the energy evolution process; The input layer receives the structural parameters of the microwave power divider, introduces a normalization mapping function and an input weight matrix, and establishes a mapping relationship from the structural parameter vector to the internal nodes of the improved Hopfield neural network model; The correlation layer computation improves the energy coupling relationship between nodes inside the Hopfield neural network model and forms a multi-channel asymmetric connection structure inside the improved Hopfield neural network model; The output layer adds a calculation unit for energy change rate and physical deviation, summarizes the energy state of each node in the improved Hopfield neural network model, and generates a comprehensive energy value and energy change trend. The physical constraint control layer introduces energy conservation, impedance matching, and signal symmetry into the energy function of the improved Hopfield neural network model to regulate the energy evolution process in real time. Establish inter-layer feedback channels to interconnect and transmit energy state information between the input layer, correlation layer, output layer and physical constraint control layer; Under the energy reduction criterion, the improved Hopfield neural network model is iteratively trained based on the structural parameters of the microwave power divider, and the weight coefficients, threshold bias and constraint factors are continuously adjusted. When the energy function converges and the physical constraint error is reduced to the set range, the model parameters are fixed and the energy mapping learning of the training phase is completed. When the improved Hopfield neural network model training converges and the physical constraints meet the threshold conditions, it outputs energy state information, physical consistency evaluation value, and optimized comprehensive evaluation value formed by weighted fusion of energy state, physical constraint deviation and performance prediction results during the training phase.

4. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The process of conducting physical feasibility screening to obtain the first-generation population includes: The range of values ​​for the structural parameters of each microwave power divider is determined based on the standardized dataset, and normalization is performed according to the maximum and minimum values ​​to generate a normalized parameter space. In the normalized parameter space, the combination of structural parameters of the microwave power divider is generated by Latin hypercube sampling, forming the initial solution set of structural parameters of the microwave power divider. Based on the physical consistency conditions including energy conservation, impedance matching and signal symmetry, the initial solution set is screened, and the feasible set that meets the physical constraint threshold is retained. Based on population size and sex ratio, the feasible set is divided into male lions, female lions, and wandering individuals, and energy feedback lions, constraint evaluation lions, and leader lions are added. The energy feedback lion group adaptively adjusts according to the rate of energy change, the constraint evaluation lion group determines physical consistency and constraint satisfaction, and the leader lion is the optimal individual selected from male lions based on the comprehensive fitness value, forming a multi-level lion group structure. The initial hunting step size is set by combining the structural parameter sensitivity of each individual in the population. Boundary correction is performed on individuals that exceed the range to obtain a stable first-generation population. Standardized target performance indicators are then determined based on the performance indicator boundaries.

5. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The step of calculating the comprehensive fitness value by combining standardized target performance indicators and performing iterative updates of the lion pride optimization algorithm based on the comprehensive fitness value includes: The structural parameter vectors of each individual in the population are input into the improved Hopfield neural network model to obtain the optimized comprehensive evaluation and energy state values. Based on the difference between the standardized target performance index and the prediction performance of the improved Hopfield neural network model, the target deviation is calculated; and the comprehensive fitness value is formed by combining the optimized comprehensive evaluation value and the physical consistency penalty value. The hunting direction is determined based on the comprehensive fitness value. At the same time, the energy feedback lion pride adaptively adjusts the step size according to the energy change rate. The step size decreases when the energy change rate is faster. An inertial adjustment factor and a convergence adjustment factor are introduced to smooth the iteration process. The coordinated hunt is carried out based on the position of the lead lion and the center position of the lionesses, while the energy feedback lions also participate in correcting the hunting direction and optimizing the hunting direction. The free lions perform random diffusion within the structural parameter space, and the constrained evaluation lions perform physical consistency screening on the diffusion results, retaining individuals that satisfy energy conservation and impedance matching. According to the set pairing probability, parent individuals are selected from the leader lion and the liones, and structural parameters are recombined and minor variations are performed to generate new individuals. After physical consistency is checked, they are added to the population. The elite individuals with the highest overall fitness value are retained. Individuals that hunt, migrate, and reproduce are merged and sorted. The best individuals are selected to form the next generation of the population. The optimal structural parameters and the optimized overall evaluation value are recorded.

6. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The determination of whether the optimization result of the population meets the set performance target and the physical consistency constraint threshold includes: The structural parameter vector of the individual to be evaluated is determined, and standardized target performance indicators and physical consistency constraint thresholds are set. The physical consistency constraints include impedance matching, energy conservation, structural symmetry, and geometric boundaries. The overall performance deviation value is obtained by weighting and group standardizing the target deviation. Calculate the physical consistency deviation of the structural parameters, which consists of impedance matching deviation, energy conservation deviation, structural symmetry deviation, and geometric overrun penalty. The judgment is based on the comprehensive performance deviation value and the physical consistency deviation value. When neither of them exceeds the threshold, it is determined that the performance target and physical consistency constraints are met; otherwise, it is determined that they are not met. At the same time, the joint judgment value is calculated. Record the structural parameters of individuals that meet the conditions as candidate optimized structural parameters, and save the candidate optimized structural parameters, the optimized comprehensive evaluation value, and the joint judgment value.

7. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The process of updating and improving the parameters of the Hopfield neural network model and returning to the training process includes: Based on the candidate optimized structural parameters, the port excitation, reference impedance, and frequency sampling points of the electromagnetic simulation environment are set. In the electromagnetic simulation environment, the performance index of the candidate optimized structural parameters is calculated, and the results at each frequency point are interpolated and statistically analyzed to obtain the simulation performance index sequence and the representative value of the frequency band. The performance prediction results corresponding to the optimized comprehensive evaluation quantity are compared with the simulation performance indicators one by one. The absolute error and relative error of each indicator are calculated, and the total error measure is obtained by weighted summation of the errors of each indicator according to the weight. When the total error metric exceeds the error tolerance threshold, the improved Hopfield neural network model parameters are updated, the inter-layer weights, node threshold biases and physical constraint factors are adjusted, and a new total error metric is calculated; multiple iterative corrections are performed while maintaining the physical consistency constraint; if the error still does not converge after multiple corrections, the improved Hopfield neural network model is retrained. When the retraining mechanism is triggered, the updated and improved Hopfield neural network model is returned to the training phase and relearned using a standardized dataset to correct the global parameter distribution. When the error metric converges after retraining and meets the physical consistency constraint, the candidate optimized structural parameters are confirmed as valid results and the corresponding performance indicators and the optimized comprehensive evaluation are recorded.

8. The method for performance optimization of a microwave power divider based on pattern recognition according to claim 1, characterized in that, The output finally optimizes the structural parameters and corresponding performance indicators, and returns the final optimization result to the improved Hopfield neural network model, including: Input the candidate optimized structural parameters into the electromagnetic simulation tool to obtain the simulation performance index vector and set the target performance vector and index weight vector; Compare the simulation performance index vector with the target performance vector, calculate the comprehensive satisfaction metric, and determine the pass when the set threshold is met and all key indicators reach the target threshold; Obtain the prediction performance vector corresponding to the optimized comprehensive evaluation quantity, calculate the total error metric and physical consistency deviation, and pass the verification when both are within the target threshold. The candidate optimized structural parameters that have passed the verification are solidified as the final optimized structural parameters. The corresponding simulation performance index, comprehensive satisfaction metric, total error metric and physical consistency evaluation value are output, and the corresponding comprehensive evaluation quantity after optimization is recorded. The final optimized structural parameters, corresponding simulation performance indicators, and error verification results are returned to the model for incremental correction and updating of the standardized dataset to determine the tolerance range of structural parameters and the test acceptance threshold.

Images