A piezoelectric ceramic filter reverse design method, system and storage medium

Abstract

The application discloses a piezoelectric ceramic filter reverse design method and system and a storage medium, obtains a discrete conduction interference peak frequency of a target power supply, collects admittance frequency response data of rectangular piezoelectric ceramics of different sizes and calibrates a finite element model, and constructs a data set containing size parameters, main modal frequency, peak admittance and B modal splitting state; generates physical baseline frequency and physical enhancement features according to the inverse ratio relationship between piezoelectric resonance frequency and characteristic size; constructs a Transformer model, performs target frequency matching optimization under the constraints of size range, thickness and aspect ratio based on the trained prediction model, and outputs the length, width and thickness of the piezoelectric ceramic. The application can quickly deduce the piezoelectric ceramic size with corresponding multi-modal resonance characteristics from the target interference frequency, improve the accuracy of split peak identification and reverse design, and reduce the common mode inductance, weight and volume required by the electromagnetic interference filter.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted piezoelectric device design technology, specifically relating to a piezoelectric ceramic filter reverse design method, system, and storage medium. Background Technology

[0002] Isolated flyback converters are widely used in consumer electronics adapters, chargers, and low-power offline power supplies. Power switching devices generate high rates of voltage and current change during turn-on and turn-off, which interact with transformer leakage inductance, inter-winding capacitance, and distributed capacitance in the circuit, forming locally concentrated common-mode interference spikes in the conducted interference spectrum.

[0003] Existing technologies typically use Y capacitors to provide a return path for common-mode current, and combine them with common-mode inductors to achieve broadband attenuation. However, the capacitance of Y capacitors is limited by leakage current safety, and simply increasing the common-mode inductor will increase the size, weight, and cost of the filter. In scenarios where interference mainly manifests as several discrete peaks, broadband filtering is prone to over-design.

[0004] Piezoelectric ceramics exhibit low impedance near their mechanical resonant frequency, making them suitable as narrowband bypass elements for selectively suppressing interference at target frequencies. The resonant frequency and peak admittance of piezoelectric ceramics are influenced by their length, width, thickness, material parameters, and boundary conditions, and may exhibit multimodal coupling and split peaks. Traditional finite element method point-by-point scanning and manual trial-and-error design methods are computationally intensive; ordinary neural networks lack physical constraints, easily producing results that do not conform to the resonance law, and struggle to simultaneously identify whether the B mode splits and which branch is the dominant peak.

[0005] Therefore, a design method is needed that can integrate physical priors, accurately characterize multimodal and split peaks, and automatically back-calculate the piezoelectric ceramic size from the actual interference frequency. Summary of the Invention

[0006] Purpose of the invention: To address the technical problems and defects existing in the prior art, this invention provides a reverse design method, system, and storage medium for piezoelectric ceramic filters, which can solve the problems of high computational cost, insufficient split peak identification capability, weak physical rationality of prediction results, and difficulty in directly determining device size from target interference frequency in existing design methods.

[0007] Technical Solution: To achieve the above-mentioned objectives, this invention adopts the following technical solution: a reverse design method for piezoelectric ceramic filters based on a physically guided peak query network, comprising the following steps:

[0008] S1, Interference Peak Location: Obtain the conducted interference spectrum of the power supply to be suppressed, and determine the target interference peak frequency from the conducted interference spectrum;

[0009] S2, Training dataset construction: Obtain the admittance frequency response data of piezoelectric ceramic samples with different geometric dimensions, extract the resonant frequency, peak admittance and split peak state of multiple main modes from the admittance frequency response data, and construct the training dataset;

[0010] S3, constructs the physical baseline frequency based on the piezoelectric ceramic geometry, and then performs physical enhancement features on the physical baseline frequency;

[0011] S4, Physical guidance prediction model construction: The Transformer model is used as the physical guidance peak query prediction model. The physical enhancement features are used as the model input, and the admittance frequency response data of each main mode of piezoelectric ceramics are used as the prediction output. The physical guidance peak query prediction model is trained and optimized.

[0012] S5. Based on the trained and optimized prediction model, the length, width and thickness of the piezoelectric ceramic are used as optimization variables. Under the constraints of engineering feasibility, the target interference peak frequency matching optimization is performed to obtain the piezoelectric ceramic size that matches the target interference peak frequency.

[0013] Furthermore, the admittance frequency response data of the piezoelectric ceramic samples with different geometric dimensions mentioned in step S2 are obtained by the following method:

[0014] S21, the measured admittance frequency response of multiple sets of piezoelectric ceramic samples with different geometric dimensions was measured using a vector network analyzer;

[0015] S22, using the measured admittance frequency response as the calibration benchmark, the piezoelectric coupling parameters, dielectric parameters and mechanical damping parameters in the piezoelectric ceramic finite element model are calibrated and simulated.

[0016] S23. In the finite element model after calibration and simulation, the length A, width B and thickness C of the piezoelectric ceramic are parametrically scanned to obtain the simulation results of the full-band admittance frequency response corresponding to each set of dimensions.

[0017] S24, use both simulation results and experimental results as the training dataset.

[0018] Furthermore, the main mode includes a length radial mode A, a width radial mode B, and a thickness mode C; the width radial mode B includes two sub-segments, namely a low-frequency branch B1 and a high-frequency branch B2.

[0019] Furthermore, the physical baseline frequency mentioned in step S3 is determined by the following formula:

[0020] , In the formula, f A0 f B0 and f C0These are the physical baseline frequencies of the principal modes in the length, width, and thickness directions, respectively. The equivalent wave velocity in the in-plane direction. The equivalent wave velocity in the thickness direction; the physical enhancement features include the piezoelectric ceramic length A, piezoelectric ceramic width B, piezoelectric ceramic thickness C, logA, logB, logC, size ratio, and the reciprocals of the geometric dimensions 1 / A, 1 / B, and 1 / C, as well as the physical baseline frequency.

[0021] The physical baseline frequency of this invention is a first-order approximation obtained under the condition of neglecting complex modal coupling, geometric scaling effects, and actual boundary perturbations. It is used to characterize the basic inverse relationship between the main modal frequency and the corresponding feature size, and is not directly used as the final resonant frequency. The physical guided peak query prediction model uses the physical baseline frequency as the reference starting point for frequency prediction, and further learns the frequency correction caused by factors such as size scaling, piezoelectric coupling, modal splitting, and boundary conditions, thereby obtaining the final predicted frequency of each main mode.

[0022] 5. Furthermore, the width radial mode B includes the center frequencies f of two splits. Bc and peak spacing Δf B Determined by the following formula:

[0023] , In the formula, This indicates the frequency of the low-frequency branch B1 in the split circuit. This indicates the frequency of the high-frequency branch B2 in the split.

[0024] Furthermore, the prediction frequencies of each principal mode are confirmed by the following formula:

[0025] , In the formula, This represents the frequency prediction output (i.e., admittance frequency response data) for each main mode of the piezoelectric ceramic. Indicates the physical baseline frequency. This indicates the amount of correction to the physical baseline frequency.

[0026] This invention also provides a piezoelectric ceramic filter reverse design system based on a physically guided peak query network, comprising:

[0027] The spectrum acquisition module is used to acquire the conducted interference spectrum of the power source to be suppressed and determine the target interference peak frequency;

[0028] The dataset construction module is used to acquire admittance frequency response data of piezoelectric ceramics of different sizes and generate modal frequencies, peak admittance and split peak labels;

[0029] The physical feature building module is used to generate physical baseline frequencies and physical enhancement features;

[0030] The model training and prediction module is used to train the physical-guided peak query prediction model and output the multimodal resonance characteristics;

[0031] The inverse optimization module is used to solve for the piezoelectric ceramic size from the target interference peak frequency under engineering feasibility constraints.

[0032] The results output module is used to output the piezoelectric ceramic dimensions, predicted resonant frequency, peak admittance, and splitting state.

[0033] The present invention also provides a computer-readable storage medium storing a computer program capable of executing the above-described method.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0035] (1) By using analytical physical baseline frequency constraints to learn the network, the learnable space is reduced and the physical rationality of the prediction results is improved;

[0036] (2) The B-mode bimodal peaks are represented as center frequency and peak spacing, and the splitting state and dominant branch are jointly predicted, which can handle unsplit, weakly split and strongly split spectra;

[0037] (3) Based on the frozen proxy model, reverse optimization can be performed, and the device size can be quickly obtained from the actual EMI peak frequency, reducing repeated finite element scanning and manual trial and error;

[0038] (4) Piezoelectric ceramics provide a low-impedance return path at the target resonant frequency, which can achieve interference suppression with a smaller common-mode inductor, thereby reducing the weight and volume of the filter. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the reverse design process of the piezoelectric ceramic filter for suppressing discrete conducted interference peaks as described in this invention.

[0040] Figure 2 This is a structural diagram of the physical-guided peak query multi-task prediction model in an embodiment of the present invention;

[0041] Figure 3 This is a comparison chart of the predicted piezoelectric ceramic and the measured admittance spectrum in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram showing the connection position of the piezoelectric ceramic filter in the flyback converter in an embodiment of the present invention;

[0043] Figure 5This is a comparison diagram of the conducted interference spectrum before and after the piezoelectric ceramic is connected in an embodiment of the present invention;

[0044] Figure 6 This is a comparison chart of the weight and volume of the traditional filtering scheme and the piezoelectric ceramic filtering scheme in the embodiments of the present invention. Detailed Implementation

[0045] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0046] like Figure 1 As shown, the technical solution of the piezoelectric ceramic filter reverse design method of the present invention is as follows:

[0047] Obtain the conducted interference spectrum of the power source to be suppressed and determine the target interference peak frequency; collect admittance frequency response data of piezoelectric ceramics of different sizes, calibrate the finite element model through measured data and generate a parameterized dataset; extract the frequency and peak admittance of mode A, branch B1, branch B2 and mode C from the admittance frequency response, and establish the existence label of mode B split and the dominant branch label.

[0048] A physical baseline frequency is constructed based on the approximate inverse relationship between the piezoelectric resonant frequency and the feature size. The original size, logarithmic size, size ratio, reciprocal scale, and physical baseline frequency are used together as physical enhancement features. The global coupling relationship is learned through a Transformer encoder. The peak structure is decoded using learnable queries corresponding to different modes. The frequency, peak admittance, and split state are jointly predicted through a multi-task output head.

[0049] The trained prediction model is frozen, and target frequency matching optimization with engineering feasibility constraints is performed using the length, width, and thickness of the piezoelectric ceramic as optimization variables. The output is the piezoelectric ceramic size that matches the target interference peak. A sample is prepared and connected to the power supply common-mode interference return path, and the model is verified through admittance testing and conducted EMI testing. The specific steps are described in detail below:

[0050] S1: Training Dataset Construction and Finite Element Calibration

[0051] PZT-5A rectangular piezoelectric ceramic was selected as the sample. A vector network analyzer was used to connect the sample electrodes, and the complex admittance was measured within a preset frequency range to obtain the complete admittance frequency response of samples of different sizes. The sample geometry is represented by length A, width B, and thickness C.

[0052] A finite element model of the piezoelectric ceramic with the same dimensions as the actual object was established in COMSOL Multiphysics. The constitutive relation of the piezoelectric material, electrode boundaries, and frequency domain solution were set. Using the main resonant frequency, anti-resonant frequency, and peak admittance of the measured admittance curve as calibration targets, the piezoelectric coupling parameters, free permittivity, and mechanical damping parameters of PZT-5A were adjusted until the simulated curve and the measured curve reached the predetermined error requirement within the target modal region.

[0053] After calibration, a parameter scan is performed within the permissible ranges of A, B, and C to calculate the full-band admittance curve for each size group. Local maxima are detected for each curve, and the A mode, B1 branch, B2 branch, and C mode are selected based on peak value, salience, peak width, frequency interval, and amplitude ratio, with the corresponding frequencies and peak admittance recorded. For the B mode, the existence of a valid B2 branch is determined based on a multidimensional weighted score, and the dominant branch between B1 and B2 is identified when bimodalities exist. This forms the training dataset for the samples.

[0054] The geometric dimensions are normalized; the peak admittance of different modes is normalized after logarithmic transformation. For samples without effective B2 branches, a B2 mask is set so that the sample is not included in the calculation of B2 frequency and admittance regression loss.

[0055] S2: Physically Guided Peak Query Model

[0056] The resonance of piezoelectric ceramics is essentially the result of elastic waves propagating within the ceramic and forming standing waves. According to the wave equation f=v / λ, the resonant frequency is determined by both the elastic wave propagation velocity v and the standing wave wavelength. For the first-order principal mode formed along the length, width, or thickness of a rectangular piezoelectric ceramic, the corresponding characteristic dimension can be approximated as equal to half the standing wave wavelength, i.e., λ. A ≈2A、λ B ≈2B and λ C ≈2C. Reference Figure 2 First, the physical baseline frequency is calculated based on the approximate inverse relationship between the size of the rectangular piezoelectric resonator and the dominant mode frequency:

[0057]

[0058] The physical augmentation input is composed of A, B, C, logA, logB, logC, A / B, A / C, B / C, 1 / A, 1 / B, 1 / C, and the physical baseline frequency, and is linearly projected to a unified dimension. In one embodiment of this invention, the network employs a 4-layer Transformer encoder, each layer including a 6-head self-attention network and a feedforward network, with an embedding dimension of 192. The decoding end uses a 2-layer Transformer decoder, and learnable queries are set corresponding to the center of mode A, the center of mode B, the distance between modes B, and mode C, respectively. The number of layers, heads, and dimensions can be adjusted according to the data scale.

[0059] The two split peaks of the B mode are represented by the center frequency f. Bc and peak spacing Δf B The expression is as follows:

[0060]

[0061] In the formula, This indicates the frequency of the low-frequency branch B1 in the split circuit. This represents the frequency of the high-frequency branch B2 in the split network. The network outputs the relative correction of each mode with respect to the physical baseline, and according to... Obtain the predicted frequency. This represents the frequency prediction output (i.e., admittance frequency response data) for each main mode of the piezoelectric ceramic. Indicates the physical baseline frequency. This represents the correction amount for the physical baseline frequency. For the B mode, the B1 and B2 frequencies are reconstructed based on the predicted center frequency and peak spacing. The multi-task output header also outputs the peak admittance of each mode, the probability of B2 presence, and the dominant branch category.

[0062] The model training loss consists of frequency loss, admittance loss, and split / classification loss.

[0063]

[0064] In the formula, This represents the total training loss of the model and the modal frequency loss. Peak admittance loss and split state classification loss .in:

[0065] The frequency loss calculates the error for branches A, B1, C, and the effective B2 branch; the admittance loss calculates the error for each effective peak admittance; the split state classification loss includes B2 presence classification and dominant branch classification. For samples without an effective B2 branch, the frequency and admittance errors of the B2 branch are masked. Additionally, a sequence consistency loss may be included to penalize predictions that do not meet the predetermined modal frequency order. Mini-batch training is performed using the AdamW optimizer and cosine annealing learning rate, and the model with the minimum total loss in the validation set is saved.

[0066] Optionally, nearest neighbor residual correction is performed after positive prediction. K nearest training samples are selected in the normalized size space, weighted according to distance, and the weighted residuals between the true values ​​of the nearest samples and the model predictions are summed to the current prediction. The value of K, the distance metric, and the weighting function can be determined from the validation set.

[0067] S3: Reverse Dimensioning Design under Target Frequency Constraints

[0068] Obtain the conducted interference spectrum of the flyback converter to be suppressed, and select the discrete peak frequencies to be suppressed as the target vector f*. Freeze the trained forward prediction model, set the size x=[A,B,C]T as an optimizable variable, and construct the objective function. P(x) is used to limit the allowable range, minimum thickness, aspect ratio, manufacturing resolution, and other engineering conditions of A, B, and C.

[0069] The first stage uses multiple initial values ​​or preset candidate grids for coarse search, quickly evaluating the objective function on the prediction model and retaining several low-loss candidates. The second stage uses each candidate as an initial value, leverages the differentiable property of the input size in the prediction model for gradient refinement, and performs boundary projection or penalty constraints after each update, ultimately selecting the size that minimizes the objective function and satisfies engineering conditions.

[0070] In a series of experiments, the main conducted interference peaks of the flyback converter were located at 163 kHz and 326 kHz. Considering static capacitance, admittance peak, and bonding and assembly stability, the thickness C was fixed at 1 mm. The model output was 8.5 mm × 4.56 mm × 1 mm; considering manufacturing precision, an 8.5 mm × 4.5 mm × 1 mm sample was prepared. The measured main peaks were approximately 165 kHz and 328 kHz, with a B-mode splitting peak appearing near 356 kHz, such as... Figure 3 As shown, the predicted peak position is consistent with the measured curve, indicating that this reverse design process can obtain the manufacturable dimensions from the target frequency.

[0071] Circuit applications of piezoelectric ceramic filter elements: Refer to Figure 4The prepared rectangular piezoelectric ceramic filter element is connected between the secondary winding of the transformer and the drain of the MOSFET. Common-mode noise generated during MOSFET switching transients returns to the noise source via this low-impedance resonant branch, thereby reducing interference propagating through parasitic capacitance and external lines. The piezoelectric ceramic filter element can be combined with a common-mode inductor to form a complete filter network.

[0072] For an 8.5 mm × 4.5 mm × 1 mm PZT-5A piezoelectric ceramic, its two principal in-plane modes near 165 kHz and 328 kHz can be represented by an extended BVD circuit consisting of a static capacitance C0 connected in parallel with two series-connected RLC moving branches. Based on the geometry and dielectric constant, C0 is approximately 0.72 nF; estimated from the resonant and anti-resonant frequencies, each of the two moving capacitances is approximately 63 pF, and the low-frequency equivalent capacitance is approximately 0.85 nF.

[0073] The conducted interference spectrum under unfiltered, piezoelectric ceramic filtered, and Y-capacitor filtered conditions was measured using a linear impedance stabilization network. Figure 5 A set of comparative results is presented. In the complete filter network, the piezoelectric ceramic solution uses a 2.2 mH common-mode inductor, while the traditional Y-capacitor solution uses a 20 mH common-mode inductor, reducing the required common-mode inductance by approximately 89%. This corresponds to a reduction of approximately 75.6% in weight and approximately 60.5% in volume for key passive components. Figure 6 As shown. The above values ​​are used to illustrate the effect of one embodiment and do not limit the scope of protection of the present invention.

[0074] S5: Systems, electronic devices, and storage media

[0075] This embodiment provides a piezoelectric ceramic filter reverse design system, including a spectrum acquisition module, a dataset construction module, a physical feature construction module, a model training and prediction module, a reverse optimization module, and a result output module. Each module can be implemented by software, hardware, or a combination of both. The result output module can display the target frequency, recommended size, predicted A / B1 / B2 / C mode frequencies and peak admittances, B2 existence probability, dominant branch, and constraint satisfaction status, and export a design report.

[0076] This embodiment also provides an electronic device, including a processor, a memory, a communication interface, and a computer program stored in the memory. When the processor executes the computer program, it implements the method of any of the foregoing embodiments. The electronic device may be a personal computer, a workstation, a server, or an embedded computing platform.

[0077] This embodiment also provides a computer-readable storage medium storing a computer program for implementing the aforementioned method. The storage medium includes, but is not limited to, a hard disk, an optical disk, flash memory, a solid-state drive, and non-volatile memory.

[0078] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A reverse design method for piezoelectric ceramic filters, characterized in that... Includes the following steps: S1, Interference Peak Location: Obtain the conducted interference spectrum of the power supply to be suppressed, and determine the target interference peak frequency from the conducted interference spectrum; S2, Training dataset construction: Obtain the admittance frequency response data of piezoelectric ceramic samples with different geometric dimensions, extract the resonant frequency, peak admittance and split peak state of multiple main modes from the admittance frequency response data, and construct the training dataset; S3, constructs the physical baseline frequency based on the piezoelectric ceramic geometry, and then performs physical enhancement features on the physical baseline frequency; S4, Physical guidance prediction model construction: The Transformer model is used as the physical guidance peak query prediction model. The physical enhancement features are used as the model input, and the admittance frequency response data of each main mode of piezoelectric ceramics are used as the prediction output. The physical guidance peak query prediction model is trained and optimized. S5. Based on the trained and optimized prediction model, the length, width and thickness of the piezoelectric ceramic are used as optimization variables. Under the constraints of engineering feasibility, the target interference peak frequency matching optimization is performed to obtain the piezoelectric ceramic size that matches the target interference peak frequency.

2. The reverse design method for piezoelectric ceramic filters according to claim 1, characterized in that: The admittance frequency response data of the piezoelectric ceramic samples with different geometric dimensions mentioned in step S2 were obtained by the following method: S21, the measured admittance frequency response of multiple sets of piezoelectric ceramic samples with different geometric dimensions was measured using a vector network analyzer; S22, using the measured admittance frequency response as the calibration benchmark, the piezoelectric coupling parameters, dielectric parameters and mechanical damping parameters in the piezoelectric ceramic finite element model are calibrated and simulated. S23. In the finite element model after calibration and simulation, the length A, width B and thickness C of the piezoelectric ceramic are parametrically scanned to obtain the simulation results of the full-band admittance frequency response corresponding to each set of dimensions. S24, use both simulation results and experimental results as the training dataset.

3. The reverse design method for piezoelectric ceramic filters according to claim 1, characterized in that: The main modes include a length radial mode A, a width radial mode B, and a thickness mode C; the width radial mode B includes two sub-modes, namely a low-frequency branch B1 and a high-frequency branch B2.

4. The reverse design method for piezoelectric ceramic filters according to claim 1, characterized in that: The physical baseline frequency mentioned in step S3 is determined by the following formula. ， In the formula, f A0 f B0 and f C0 These are the physical baseline frequencies of the principal modes in the length, width, and thickness directions, respectively. The equivalent wave velocity in the in-plane direction. Equivalent wave velocity in the thickness direction; The physical enhancement features include the piezoelectric ceramic length A, piezoelectric ceramic width B, piezoelectric ceramic thickness C, logA, logB, logC, size ratio, and the reciprocals of the geometric dimensions 1 / A, 1 / B, and 1 / C, as well as the physical baseline frequency.

5. The reverse design method for piezoelectric ceramic filters according to claim 3, characterized in that: The width-radial mode B includes the center frequencies f of the two splits. Bc and peak spacing Δf B Determined by the following formula: ， In the formula, This indicates the frequency of the low-frequency branch B1 in the split circuit. This indicates the frequency of the high-frequency branch B2 in the split.

6. The reverse design method for piezoelectric ceramic filters according to claim 5, characterized in that: The prediction frequencies of each principal mode are determined by the following formula: ， In the formula, This represents the frequency prediction output (i.e., admittance frequency response data) for each main mode of the piezoelectric ceramic. Indicates the physical baseline frequency. This indicates the amount of correction to the physical baseline frequency.

7. A piezoelectric ceramic filter reverse design system based on a physically guided peak query network, characterized in that, include: The spectrum acquisition module is used to acquire the conducted interference spectrum of the power source to be suppressed and determine the target interference peak frequency; The dataset construction module is used to acquire admittance frequency response data of piezoelectric ceramics of different sizes and generate modal frequencies, peak admittance and split peak labels; The physical feature building module is used to generate physical baseline frequencies and physical enhancement features; The model training and prediction module is used to train the physical-guided peak query prediction model and output the multimodal resonance characteristics; The inverse optimization module is used to solve for the piezoelectric ceramic size from the target interference peak frequency under engineering feasibility constraints. The results output module is used to output the piezoelectric ceramic dimensions, predicted resonant frequency, peak admittance, and splitting state.

8. A computer-readable storage medium, characterized in that: The device contains a computer program that can execute the method according to any one of claims 1-6.

Images