High-precision design and optimization method for EMI filter

Abstract

The invention aims to provide a high-precision design and optimization method for an EMI (Electro-Magnetic Interference) filter aiming at the corresponding defects in the prior art, a mixed neural network architecture model is constructed by combining a feedforward neural network and a convolutional neural network, and the filter is subjected to reverse modeling by utilizing the mixed neural network architecture model, so that the filtering precision is improved. According to the method, the predicted value, the topological structure and the element parameters of the filter are obtained, then the element parameters are subjected to multi-objective optimization by using a multi-objective genetic algorithm to obtain the optimized element parameters, and the optimized element parameters and the topological structure of the filter jointly form an optimization scheme of the filter, so that the design efficiency and the optimization precision of the EMI filter can be remarkably improved; and the design period of the EMI filter is shortened, and the market demand of rapid iterative updating of current high-frequency electronic equipment can be met.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic compatibility (EMC) technology, specifically to a high-precision design and optimization method for EMI filters. Background Technology

[0002] An EMI filter is an electronic device that uses the impedance of inductors to high-frequency currents and the bypassing effect of capacitors to high-frequency voltages to limit electromagnetic interference within a certain range, thereby ensuring the normal operation of electronic equipment or systems. Traditional EMI filters are primarily developed by engineers through trial and error based on equivalent circuit and electromagnetic models, gradually determining the parameters of the LC components and then verifying the design through extensive experimentation. However, this approach is highly dependent on the engineer's experience and has several shortcomings. ① Insufficient accuracy: The equivalent circuit model adopts the lumped parameter assumption, which often ignores the parasitic parameters of components (such as the distributed capacitance of inductors) and the changes of parasitic parameters with frequency, temperature, current and other factors under complex operating conditions, making it difficult to accurately describe the dynamic characteristics of EMI filters; the electromagnetic model only analyzes the electromagnetic field distribution theoretically, and when dealing with large-scale complex systems, due to the limitations of computing resources and time, simplification assumptions are often required, which leads to a decrease in the accuracy of the model; ② Inefficient: In the traditional EMI filter design process, every adjustment to the filter parameters requires the re-fabrication of a physical prototype and comprehensive performance testing in an electromagnetic compatibility testing laboratory. The entire design process requires a large number of experimental verifications, which is not only time-consuming and labor-intensive, but also the design cycle of a single solution usually takes 2 to 4 weeks, which is far from meeting the market demand for rapid iteration and updates of current high-frequency electronic equipment. ③ Difficulty in handling nonlinear characteristics: Both the equivalent circuit model and the electromagnetic model are based on the linearization assumption. For power electronic systems with strong nonlinear characteristics (such as switches, power supplies, etc.), the equivalent circuit model and the electromagnetic model are difficult to accurately capture and describe these complex nonlinear behaviors, and cannot achieve effective modeling of EMI filters in such systems. ④ Poor adaptability: Traditional EMI filters usually use a fixed topology (such as the common L-type, etc.), which leads to limitations in the frequency response characteristics of the filter. Faced with complex and ever-changing EMI spectrum, the filter's suppression effect is poor in certain frequency bands. ⑤ High cost: In the traditional design process, in order to test the performance of filters with different parameters and topologies, physical prototypes need to be made multiple times. This not only consumes a lot of raw materials (such as inductors, capacitors and other components), but also increases labor costs and the cost of using testing equipment, resulting in a great waste of resources and a significant increase in the overall design cost of the filter.

[0003] Therefore, how to reduce the development cost and time of EMI filters, and improve the design efficiency and accuracy of EMI filters, has always been a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a high-precision design and optimization method for EMI filters. This method combines feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model. The hybrid neural network architecture model is then used to perform reverse modeling of the filter to obtain the filter's predicted values, topology, and component parameters. A multi-objective genetic algorithm is then used to perform multi-objective optimization of the component parameters to obtain the optimized component parameters. These optimized parameters, together with the filter's predicted values ​​and topology, constitute the filter's optimization scheme.

[0005] The objective of this invention is achieved through the following approach: A high-precision design and optimization method for EMI filters includes the following steps: 1) Construct a multi-dimensional filter insertion loss testing system to collect the insertion loss of EMI filters; 2) Combine feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model; 3) The insertion loss of the EMI filter is collected using a multi-dimensional filter insertion loss test system, and the collected insertion loss is standardized to obtain the standardized insertion loss; 4) Input the standardized insertion loss into the hybrid neural network architecture model, and use the hybrid neural network architecture model to perform reverse modeling of the filter to obtain the topology of the EMI filter and several component parameters; 5) The component parameters of the EMI filter are optimized using a multi-objective genetic algorithm to obtain several optimized component parameters; 6) Combine the topology of the EMI filter and several optimized component parameters to form an optimized EMI filter scheme.

[0006] Preferably, the multi-dimensional filter insertion loss test system includes a multi-port test module, a multi-impedance load switching matrix module, and a dynamic test condition synchronization module, wherein... The multi-port test module uses a four-port vector network analyzer to construct a synchronous measurement link with forward transmission and reverse isolation. The test ports integrate N-type / APC-7 type calibration interfaces. The multi-impedance load switching matrix module integrates a programmable impedance synthesis module, which realizes nanosecond-level switching of several load combinations through a PIN switch array, and the VSWR compensation algorithm eliminates impedance mismatch error in real time. The dynamic test condition synchronization module is a temperature control-electric control collaborative system based on the PXIe bus, which realizes the time alignment of temperature cycle and impedance switching events, and records the filter package deformation compensation parameters.

[0007] Preferably, the hybrid neural network architecture model is obtained through the following methods: 2-1) Collect input parameters and performance data of the EMI filter under different operating conditions; 2-2) Combine feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model, and set the optimizer of the hybrid neural network architecture model to adaptive moment estimation and the loss function to weighted cross-entropy; 2-3) Preprocess the input parameters and performance index data collected in step 2-1), and use the processed input parameters and performance index data to train the hybrid neural network architecture model.

[0008] Preferably, the hybrid neural network architecture model includes an input layer, a feature extraction layer, a conditional correlation layer, and an output layer, wherein... The input layer receives a standardized three-dimensional insertion loss tensor, which includes the insertion loss amplitude-frequency curve, phase-frequency response, and filter body parameters. The feature extraction layer employs a 4-layer residual network to extract the joint amplitude and phase frequency features of the insertion loss. Each layer contains a 3×3 convolutional kernel, batch normalization, and ReLU activation function. The size of the output feature map of the convolutional layer is dynamically adjusted by the kernel size, stride, and input dimension. An adaptive max pooling layer is cascaded after the convolutional layer. The pooling window size is optimized according to the frequency band segmentation strategy, supporting multi-scale feature extraction from wideband coarse-grained to narrowband fine-grained. The condition association layer: uses a bidirectional gated cyclic unit to model the dynamic coupling relationship between different test conditions and insertion loss characteristics, and realizes context association across load conditions through hidden state propagation; The output layer is designed as a partially fully connected network guided by a frequency domain attention mechanism, which outputs the insertion loss prediction values ​​for several frequency points. Adjacent frequency points share 50% of the hidden layer neuron weights, and the generalization ability of 6GHz wideband modeling is improved by constraining the spectrum continuity.

[0009] Preferably, it also involves the definition of a multi-objective optimization problem, specifically including: A) Determine the optimization variables, including inductance, capacitance, resistance, and layout spacing; B) The objective functions are determined to be maximizing insertion loss, minimizing volume, and minimizing cost; C) The constraints include resonant frequency offset rate < 5% and temperature rise ≤ 40℃.

[0010] Preferably, the process also involves verifying an optimized scheme for a pair of EMI filters, specifically including: S1) Establish a finite element simulation model of the EMI filter based on the optimization scheme of the EMI filter, and conduct simulation experiments using electromagnetic simulation software to verify the parameters of each optimized component in the EMI filter optimization scheme and obtain the predicted value of the insertion loss of the EMI filter. S2) Based on the finite element simulation model of the EMI filter, a filter prototype is made using 3D printing technology; S3) Based on the predicted insertion loss value of the EMI filter, the filter prototype is tested and verified using a multi-physics coupling test chamber to verify the filter insertion loss characteristics.

[0011] Preferably, in S3), the specific method for verifying the insertion loss characteristics of the filter prototype using a multiphysics coupling test chamber based on the predicted insertion loss value of the EMI filter includes: S3-1) Measure the insertion loss of the filter prototype using the experimental verification method; S3-3) Calculate the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter; S3-4) Based on the deviation calculated in step S3-3), verify the insertion loss characteristics of the filter in the following manner: If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is ≤5%, then the filter prototype passes the measured verification of the filter insertion loss characteristics. If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is greater than 5%, the adaptive calibration feedback mechanism is triggered, and the error gradient is backpropagated to step 5) and the optimization weights are adjusted.

[0012] Preferably, the experimental verification method specifically includes: S3-1-1) Connect a programmable impedance matrix to the input / output terminals of the filter prototype, and use a four-port vector network analyzer to simultaneously measure the insertion loss amplitude-frequency characteristics of the filter prototype, while recording the stability coefficients of the S21 / S12 parameters. S3-1-2) Temperature cycling test is performed in the multiphysics coupling test chamber. The insertion loss drift characteristics of the filter dielectric substrate under a fixed temperature change rate are captured by thermocouple array and dielectric constant monitoring probe. At the same time, infrared thermal imager is used to quantify the group delay fluctuation value caused by packaging deformation.

[0013] Preferably, each neuron in the output layer is connected to 4 groups of hidden layer neurons, and adjacent neurons in the output layer share 3 groups of hidden layer connections. The hidden layer contains 150 groups of neurons, with 2 neurons in each group, and some of them are connected to ANN.

[0014] The beneficial effects of this invention are as follows: Preferably, the process also involves verifying an optimized scheme for a pair of EMI filters, specifically including: S1) Establish a finite element simulation model of the EMI filter based on the optimization scheme of the EMI filter, and conduct simulation experiments using electromagnetic simulation software to verify the parameters of each optimized component in the EMI filter optimization scheme and obtain the predicted value of the insertion loss of the EMI filter. S2) Based on the finite element simulation model of the EMI filter, a filter prototype is made using 3D printing technology; S3) Based on the predicted insertion loss value of the EMI filter, the filter prototype is tested and verified using a multi-physics coupling test chamber to verify the filter insertion loss characteristics.

[0015] Preferably, in S3), the specific method for verifying the insertion loss characteristics of the filter prototype using a multiphysics coupling test chamber based on the predicted insertion loss value of the EMI filter includes: S3-1) Measure the insertion loss of the filter prototype using the experimental verification method; S3-3) Calculate the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter; S3-4) Based on the deviation calculated in step S3-3), verify the insertion loss characteristics of the filter in the following manner: If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is ≤5%, then the filter prototype passes the measured verification of the filter insertion loss characteristics. If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is greater than 5%, the filter prototype fails the measured verification of the filter insertion loss characteristics, triggering the adaptive calibration feedback mechanism to backpropagate the error gradient to step 5) and adjust the optimization weights.

[0016] This invention uses electromagnetic simulation software to construct a finite element simulation model of an EMI filter, which can quickly and efficiently verify the parameters of each component in the optimization scheme and obtain the predicted value of the insertion loss of the EMI filter, providing comparative data for subsequent experimental verification. Compared with traditional actual testing methods, it does not require the production of a large number of physical samples and can simulate and analyze different parameter combinations in a short time, saving time and costs.

[0017] Preferably, each neuron in the output layer is connected to 4 groups of hidden layer neurons, and adjacent neurons in the output layer share 3 groups of hidden layer connections. The hidden layer contains 150 groups of neurons, with 2 neurons in each group, and some of them are connected to ANN.

[0018] This invention sets each neuron in the output layer to connect to four groups of hidden layer neurons, enabling the extensive collection of hidden layer features and comprehensive analysis of input information. Setting adjacent output neurons in the output layer to share three sets of connections enhances the relevance and smoothness of the output, improves the model's stability and generalization ability, and allows the model to accurately predict even new data. Setting the hidden layer to contain 150 groups, each with two neurons, and partially connecting to ANNs effectively avoids overfitting, reasonably controls complexity, accelerates training and inference, and improves efficiency.

[0019] The advantages of this invention are as follows: ① This invention, through the efficient modeling capability of neural networks and the rapid optimization capability of optimization algorithms, can significantly improve the design efficiency and optimization accuracy of EMI filters, reduce the design cycle of EMI filters, and meet the market demand for rapid iteration and updates of current high-frequency electronic equipment. ② This invention reduces the number of times a physical prototype of a filter needs to be manufactured by combining simulation experiments and actual tests, thereby reducing the consumption of raw materials (such as inductors, capacitors, etc.), lowering labor costs and the cost of using testing equipment, and significantly reducing the overall design cost of the filter. ③ This invention achieves fully automated design of EMI filters from topology prediction to parameter optimization through the deep integration of deep learning and multi-objective optimization algorithms. It is applicable to electromagnetic interference suppression in new energy vehicle motor controllers, photovoltaic inverters and high-speed communication equipment.

[0020] Glossary Feedforward Neural Network (FNN) is a unidirectional neural network. Information starts from the input layer, passes through a series of intermediate layers (hidden layers), and finally flows to the output layer. In this process, information is transmitted in only one direction, without feedback connections. That is, there is no backward connection from the output layer or hidden layer to the input layer or other layers.

[0021] Convolutional Neural Network (CNN) is a deep learning model specifically designed for processing data with a grid structure (such as images and audio). It automatically extracts features from data through components such as convolutional layers, pooling layers, and fully connected layers, and has efficient feature extraction capabilities and good generalization performance.

[0022] Multiphysics coupling test chamber: This is an experimental device used to simulate and study the interaction of multiple physical fields. It is usually made of high-strength, well-sealed materials to ensure the stability and safety of the internal environment. It can withstand certain pressure and temperature changes and can simultaneously simulate multiple physical fields, such as temperature field, humidity field, electromagnetic field, mechanical field, etc., to reproduce the complex multiphysics coupling environment that may occur in actual applications.

[0023] The IEC 61000 standard series is an international standard for electromagnetic compatibility (EMC), designed to ensure that electrical and electronic equipment can function properly and coordinate with each other in an electromagnetic environment without causing unacceptable interference. In this invention, IEC 61000-4-21 specifically refers to Electromagnetic Compatibility (EMC) - Part 4-21, namely, Test and Measurement Techniques - Reverberation Chamber Test Methods.

[0024] MIL-STD-220C: Military standard, insertion loss measurement method. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the method in this embodiment. Detailed Implementation

[0026] like Figures 1 to 2 As shown, a high-precision design and optimization method for EMI filters is characterized by the following steps: 1) Construct a multi-dimensional filter insertion loss testing system to collect the insertion loss of EMI filters; 2) Combine feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model; 3) The insertion loss of the EMI filter is collected using a multi-dimensional filter insertion loss test system, and the collected insertion loss is standardized to obtain the standardized insertion loss; 4) Input the standardized insertion loss into the hybrid neural network architecture model, and use the hybrid neural network architecture model to perform reverse modeling of the filter to obtain the topology of the EMI filter and several component parameters; 5) The component parameters of the EMI filter are optimized using a multi-objective genetic algorithm to obtain several optimized component parameters; 6) Combine the topology of the EMI filter and several optimized component parameters to form an optimized EMI filter scheme.

[0027] Based on the above method, the following is an example: 1) Construct a multi-dimensional filter insertion loss test system to collect the insertion loss of EMI filters. The multi-dimensional filter insertion loss test system uses a vector network analyzer, a programmable signal source and a power probe array to accurately measure the forward transmission characteristics and reverse isolation characteristics of the filter under wide-band sweep frequency signal excitation and multi-load conditions. The frequency coverage range extends to 10MHz-6GHz.

[0028] In this embodiment, the multi-dimensional filter insertion loss test system includes a multi-port test module, a multi-impedance load switching matrix module, and a dynamic test condition synchronization module, wherein... The multi-port test module uses a four-port vector network analyzer to construct a synchronous measurement link for forward transmission (S21) and reverse isolation (S12). The test port integrates an N-type / APC-7 type calibration interface, the sweep frequency step accuracy is set to 50kHz, and the dynamic range reaches 130dB. The multi-impedance load switching matrix module integrates a programmable impedance synthesis module (50Ω / 75Ω / custom Z0), which realizes nanosecond-level switching of 2^8 load combinations through a PIN switch array, and the standing wave ratio (VSWR) compensation algorithm eliminates impedance mismatch error in real time. The dynamic test condition synchronization module is a temperature control-electric control collaborative system based on the PXIe bus. It realizes the time alignment of temperature cycling (-40℃→+85℃ temperature change rate 5℃ / min) and impedance switching events, with a synchronization accuracy better than 10μs, and records the filter package deformation compensation parameters.

[0029] In this embodiment, a high-precision calibration database also needs to be constructed to store data such as insertion loss amplitude-frequency curves, phase-frequency response, group delay characteristics, temperature drift characteristics, and filter structure parameters.

[0030] It is worth noting that the test object in this embodiment, namely the filter, is a multilayer dielectric integrated waveguide (SIW) bandstop filter with a center frequency of 3.5GHz (5Gn78 band), a stopband bandwidth of 500MHz, in-band rejection ≥40dB, power capacity of 40dBm, support for 50Ω / 75Ω dual impedance mode, and an operating temperature range of -40℃ to +85℃. The specific method for optimizing the multilayer dielectric integrated waveguide (SIW) bandstop filter is as follows: 2) Specific methods for constructing a hybrid neural network architecture model by combining feedforward neural networks (FNN) and convolutional neural networks (CNN) include: 2-1) Collect input parameters and performance data of multilayer dielectric integrated waveguide (SIW) bandstop filters under different operating conditions; 2-2) Combining the feedforward neural network (FNN) and the convolutional neural network (CNN), a hybrid neural network architecture model is constructed, and the optimizer of the hybrid neural network architecture model is set to the adaptive moment estimation optimizer (AdamW), and the loss function is set to the weighted cross-entropy function; In this embodiment, the structure of the hybrid neural network architecture model includes an input layer, a feature extraction layer, a conditional association layer, and an output layer, wherein... The input layer receives a standardized three-dimensional insertion loss tensor (dimensions are the number of frequency points × the number of test conditions × the number of load combinations). The tensor data includes the insertion loss amplitude-frequency curve, phase-frequency response, and filter body parameters (order / topology / material dielectric constant, etc.). The feature extraction layer employs a 4-layer ResNet residual network to extract joint amplitude and phase frequency features of the insertion loss. Each layer contains a 3×3 convolutional kernel, batch normalization, and a ReLU activation function. The size of the output feature map of the convolutional layer is dynamically adjusted by the kernel size (K), stride (S), and input dimension (D). An adaptive max-pooling layer is cascaded after the convolutional layer. The pooling window size (P) is optimized according to the frequency band segmentation strategy, supporting multi-scale feature extraction from wideband coarse-grained to narrowband fine-grained. The condition association layer employs a bidirectional gated cyclic unit (BiGRU) to model the dynamic coupling relationship between different test conditions (temperature / impedance / excitation power) and insertion loss characteristics, and achieves context association across load conditions through hidden state propagation. The output layer is designed as a partially fully connected network guided by a frequency domain attention mechanism, which outputs the insertion loss prediction values ​​for 151 frequency points. Adjacent frequency points share 50% of the hidden layer neuron weights, and the generalization ability of 6GHz wideband modeling is improved by constraining the spectrum continuity.

[0031] In this embodiment, each neuron in the output layer is connected to 4 groups of hidden layer neurons, and adjacent neurons in the output layer share 3 groups of hidden layer connections. The hidden layer contains 150 groups of neurons, with 2 neurons in each group, and some of them are connected to ANN.

[0032] 2-3) Preprocess the input parameters and performance index data collected in step 2-1), and use the processed input parameters and performance index data to train the hybrid neural network architecture model until the root mean square error (RMSE) of the validation set is ≤1 dB.

[0033] 3) The insertion loss of the multi-dimensional filter insertion loss test system was collected to obtain the insertion loss of the multi-layer dielectric integrated waveguide (SIW) band-stop filter, and the collected insertion loss was standardized to obtain the standardized insertion loss. Data acquisition conditions: cold start, load combination: random switching of 50Ω / 75Ω / 100Ω (switching rate 50 ns, duty cycle 1:1:1), impedance mismatch: a mismatch disturbance with VSWR=2.0 is injected at 6 GHz frequency, lasting for 10 μs pulse; Data collected: 9,600 sets of S-parameter matrices and 2,048 sets of harmonic suppression data.

[0034] 4) Input the standardized insertion loss into the hybrid neural network architecture model, and use the hybrid neural network architecture model to perform reverse modeling of the filter to obtain the topology of the EMI filter and several component parameters; Input data: a three-dimensional tensor of 256 (frequency points) × 50 (number of test conditions) × 4 (number of load combinations); Training results: Topological classification accuracy on the test set was 96.7%, and parameter prediction error was <2.5%.

[0035] 5) The component parameters of the EMI filter are optimized using a multi-objective genetic algorithm to obtain several optimized component parameters; This embodiment also involves the definition of a multi-objective optimization problem, specifically including: A) The optimization variables include inductance (0.1μH–10mH), capacitance (1pF–100μF), resistance (0.1Ω–10kΩ), and layout spacing (0.5mm–5mm). B) The objective functions are determined to be maximizing insertion loss, minimizing volume, and minimizing cost; C) The constraints include resonant frequency offset rate < 5% and temperature rise ≤ 40℃.

[0036] The optimization strategies of the improved NSGA-II algorithm include: ① Dynamically adjust the reference point weights and automatically balance the priority of the objective function based on the population distribution; ② Introduce a simulated annealing mechanism to control the intensity of mutation and avoid premature convergence.

[0037] The initial scheme and optimization results in this embodiment are as follows: Initial scheme: π-type filter, ; Optimization results: After 300 iterations, the insertion loss is 35 dB@30 MHz, the volume is 28 cm³, and the cost is 42 yuan.

[0038] 6) Combine the topology of the EMI filter and several optimized component parameters to form an optimized EMI filter scheme.

[0039] This embodiment also involves the process of verifying an optimization scheme for a pair of EMI filters, specifically including: S1) Establish a finite element simulation model of the EMI filter based on the optimization scheme of the EMI filter, and conduct simulation experiments using electromagnetic simulation software (ANSYS HFSS) to verify the parameters of each optimized component in the EMI filter optimization scheme and obtain the predicted value of the insertion loss of the EMI filter. S2) Based on the finite element simulation model of the EMI filter, a filter prototype is made using 3D printing technology; S3) Combining the joint test specifications of IEC 61000-4-21 and MIL-STD-220C, the specific methods for verifying the insertion loss characteristics of the filter prototype using a multi-physics coupling test chamber based on the predicted insertion loss value of the EMI filter include: S3-1) Using a field verification method, the insertion loss of the filter prototype is measured. The field verification method specifically includes: S3-1-1) Connect a programmable impedance matrix to the input / output terminals of the filter prototype, and use a four-port vector network analyzer to simultaneously measure the insertion loss amplitude-frequency characteristics of the filter prototype, while recording the stability coefficients of the S21 / S12 parameters. S3-1-2) Temperature cycling test is performed in the multiphysics coupling test chamber. The insertion loss drift characteristics of the filter dielectric substrate under a fixed temperature change rate are captured by thermocouple array and dielectric constant monitoring probe. At the same time, infrared thermal imager is used to quantify the group delay fluctuation value caused by packaging deformation.

[0040] S3-2) Calculate the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter; S3-3) Based on the deviation calculated in step S3-2), verify the insertion loss characteristics of the filter in the following manner: If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is ≤5% under wide temperature range (-40℃~+85℃) and multiple impedance (50Ω / 75Ω) conditions, then the filter prototype passes the measured verification of the filter insertion loss characteristics. If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is greater than 5% under wide temperature range (-40℃ to +85℃) and multiple impedance (50Ω / 75Ω) conditions, the filter prototype fails the measured verification of the filter insertion loss characteristics, triggering the adaptive calibration feedback mechanism to backpropagate the error gradient to step 5) and adjust the optimization weights.

[0041] The simulation and measured results of this embodiment are as follows: Simulation results: ANSYS HFSS shows an insertion loss of 33.5 dB at 30 MHz; Actual measurement results: The measured value in the anechoic chamber was 32.8 dB, with a deviation of 2.1%, which meets the CISPR 25 Class 4 standard.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A high-precision design and optimization method for EMI filters, characterized in that, Includes the following steps: 1) Construct a multi-dimensional filter insertion loss testing system to collect the insertion loss of EMI filters; 2) Combine feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model; 3) The insertion loss of the EMI filter is collected using a multi-dimensional filter insertion loss test system, and the collected insertion loss is standardized to obtain the standardized insertion loss; 4) Input the standardized insertion loss into the hybrid neural network architecture model, and use the hybrid neural network architecture model to perform reverse modeling of the filter to obtain the topology of the EMI filter and several component parameters; 5) The component parameters of the EMI filter are optimized using a multi-objective genetic algorithm to obtain several optimized component parameters; 6) Combine the topology of the EMI filter and several optimized component parameters to form an optimized EMI filter scheme.

2. The high-precision design and optimization method according to claim 1, characterized in that, The multi-dimensional filter insertion loss testing system includes a multi-port testing module, a multi-impedance load switching matrix module, and a dynamic test condition synchronization module. The multi-port test module uses a four-port vector network analyzer to construct a synchronous measurement link with forward transmission and reverse isolation. The test ports integrate N-type / APC-7 type calibration interfaces. The multi-impedance load switching matrix module integrates a programmable impedance synthesis module, which realizes nanosecond-level switching of several load combinations through a PIN switch array, and the VSWR compensation algorithm eliminates impedance mismatch error in real time. The dynamic test condition synchronization module is a temperature control-electric control collaborative system based on the PXIe bus, which realizes the time alignment of temperature cycle and impedance switching events, and records the filter package deformation compensation parameters.

3. The high-precision design and optimization method according to claim 1, characterized in that, The hybrid neural network architecture model was obtained through the following methods: 2-1) Collect input parameters and performance data of the EMI filter under different operating conditions; 2-2) Combine feedforward neural networks and convolutional neural networks to construct a hybrid neural network architecture model, and set the optimizer of the hybrid neural network architecture model to adaptive moment estimation and the loss function to weighted cross-entropy; 2-3) Preprocess the input parameters and performance index data collected in step 2-1), and use the processed input parameters and performance index data to train the hybrid neural network architecture model.

4. The high-precision design and optimization method according to claim 1 or 3, characterized in that, The hybrid neural network architecture model comprises an input layer, a feature extraction layer, a conditional association layer, and an output layer. The input layer receives a standardized three-dimensional insertion loss tensor, which includes the insertion loss amplitude-frequency curve, phase-frequency response, and filter body parameters. The feature extraction layer employs a 4-layer residual network to extract the joint amplitude and phase frequency features of the insertion loss. Each layer contains a 3×3 convolutional kernel, batch normalization, and ReLU activation function. The size of the output feature map of the convolutional layer is dynamically adjusted by the kernel size, stride, and input dimension. An adaptive max pooling layer is cascaded after the convolutional layer. The pooling window size is optimized according to the frequency band segmentation strategy, supporting multi-scale feature extraction from wideband coarse-grained to narrowband fine-grained. The condition association layer: uses a bidirectional gated cyclic unit to model the dynamic coupling relationship between different test conditions and insertion loss characteristics, and realizes context association across load conditions through hidden state propagation; The output layer is designed as a partially fully connected network guided by a frequency domain attention mechanism, which outputs the insertion loss prediction values ​​for several frequency points. Adjacent frequency points share 50% of the hidden layer neuron weights, and the generalization ability of 6GHz wideband modeling is improved by constraining the spectrum continuity.

5. The high-precision design and optimization method according to claim 1, characterized in that, It also involves the definition of a multi-objective optimization problem, specifically including: A) Determine the optimization variables, including inductance, capacitance, resistance, and layout spacing; B) The objective functions are determined to be maximizing insertion loss, minimizing volume, and minimizing cost; C) The constraints include resonant frequency offset rate < 5% and temperature rise ≤ 40℃.

6. The high-precision design and optimization method according to claim 1, characterized in that, It also involves the process of verifying an optimized EMI filter scheme, specifically including: S1) Establish a finite element simulation model of the EMI filter based on the optimization scheme of the EMI filter, and conduct simulation experiments using electromagnetic simulation software to verify the parameters of each optimized component in the EMI filter optimization scheme and obtain the predicted value of the insertion loss of the EMI filter. S2) Based on the finite element simulation model of the EMI filter, a filter prototype is made using 3D printing technology; S3) Based on the predicted insertion loss value of the EMI filter, the filter prototype is tested and verified using a multi-physics coupling test chamber to verify the filter insertion loss characteristics.

7. The high-precision design and optimization method according to claim 6, characterized in that, In S3), the specific methods for verifying the insertion loss characteristics of the filter prototype using a multiphysics coupling test chamber based on the predicted insertion loss value of the EMI filter include: S3-1) Measure the insertion loss of the filter prototype using the experimental verification method; S3-3) Calculate the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter; S3-4) Based on the deviation calculated in step S3-3), verify the insertion loss characteristics of the filter in the following manner: If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is ≤5%, then the filter prototype passes the measured verification of the filter insertion loss characteristics. If the deviation between the measured insertion loss of the filter prototype and the predicted insertion loss of the EMI filter is greater than 5%, the filter prototype fails the measured verification of the filter insertion loss characteristics, triggering the adaptive calibration feedback mechanism to backpropagate the error gradient to step 5) and adjust the optimization weights.

8. The high-precision design and optimization method according to claim 7, characterized in that, The experimental verification method specifically includes: S3-1-1) Connect a programmable impedance matrix to the input / output terminals of the filter prototype, and use a four-port vector network analyzer to simultaneously measure the insertion loss amplitude-frequency characteristics of the filter prototype, while recording the stability coefficients of the S21 / S12 parameters. S3-1-2) Temperature cycling test is performed in the multiphysics coupling test chamber. The insertion loss drift characteristics of the filter dielectric substrate under a fixed temperature change rate are captured by thermocouple array and dielectric constant monitoring probe. At the same time, infrared thermal imager is used to quantify the group delay fluctuation value caused by packaging deformation.

9. The high-precision design and optimization method according to claim 4, characterized in that, Each neuron in the output layer is connected to 4 groups of hidden layer neurons, and adjacent neurons in the output layer share 3 groups of hidden layer connections. The hidden layer contains 150 groups of neurons, with 2 neurons in each group, and some of them are connected to ANN.

Images