Deep learning-based piezoelectric ceramic filter size prediction model construction method, construction system and prediction method
By constructing a deep learning-based CNN-LSTM model and utilizing the impedance spectrum data of piezoelectric ceramic samples, direct prediction from the main frequency of interference noise to the size of piezoelectric ceramics was achieved. This solves the problems of long design cycles and high costs in existing technologies and improves the efficiency and accuracy of filter design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING NORMAL UNIVERSITY
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for designing piezoelectric ceramic filters suffer from long design cycles, high costs, and a lack of direct mapping from interference frequencies to piezoelectric ceramic dimensions, making it difficult to achieve frequency matching quickly and accurately.
A deep learning-based approach was adopted, which constructs a CNN-LSTM model and trains the model using impedance spectrum data of piezoelectric ceramic samples to achieve direct prediction from the dominant frequency of interference noise to the size of piezoelectric ceramics. The model parameters are then optimized by combining the frequency-impedance curve and the regression loss function.
It enables fast and accurate prediction of piezoelectric ceramic filter dimensions, reduces design costs and cycle time, improves frequency matching accuracy and engineering adaptability, and is suitable for filter design under complex operating conditions.
Smart Images
Figure CN122133457A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electromagnetic compatibility and filtering technology, specifically to a method, system, and method for constructing a piezoelectric ceramic filter size prediction model based on deep learning. Background Technology
[0002] With the rapid development of power electronics and switching power supply technologies, various DC / DC and AC / DC converters are widely used in power systems, industrial control, communication equipment, and household appliances. Because switching devices periodically turn on and off at high frequencies, these devices generate conducted and radiated noise during operation. Interference noise in the 10kHz–200kHz frequency band is typically closely related to the switching frequency and its lower harmonics, easily coupling into the power grid and sensitive signal links. This not only affects the stable operation of the equipment itself but may also lead to non-compliance with relevant electromagnetic compatibility standards, necessitating effective filtering and suppression measures.
[0003] In existing technologies, the suppression of interference noise in switching power supplies mainly employs LC filters and active filters based on lumped components such as inductors and capacitors, as well as filter structures that utilize piezoelectric ceramics, surface acoustic wave devices, and other components with resonant characteristics to form "notch filters" in specific frequency bands. Among these, piezoelectric ceramics have advantages such as small size, low loss, and ease of integration, and their mechanical resonance and electrical impedance characteristics have significant application potential in suppressing interference in the mid-to-low frequency band. However, in the design process of using piezoelectric ceramics to achieve interference suppression at specific frequencies, existing technologies suffer from drawbacks such as reliance on complex analytical models and equivalent circuits, reliance on trial-and-error multiple-round prototypes, long design cycles, high costs, and a disconnect between interference spectrum analysis and filter size design. Existing machine learning applications in this area are mostly limited to parameter fitting, lacking a direct "frequency → size" mapping; that is, existing research rarely achieves direct prediction of the piezoelectric ceramic size matching the target interference frequency or spectrum.
[0004] Therefore, research on how to quickly and accurately complete the reverse design of "given frequency → corresponding piezoelectric ceramic size" under complex working conditions is of great significance to the technological development of this field. Summary of the Invention
[0005] The technical objective of this invention is to provide a method, system, and method for constructing a piezoelectric ceramic filter size prediction model based on deep learning. The aim is to quickly and accurately predict the appropriate size of the piezoelectric ceramic, the core component of the piezoelectric ceramic filter, by using a deep learning-based prediction model to suppress specific interference noise in the working environment.
[0006] To achieve the above-mentioned technical objectives, the present invention provides technical solutions to the problems from the following aspects.
[0007] In a first aspect, the present invention discloses a method for constructing a piezoelectric ceramic filter size prediction model based on deep learning, characterized by comprising the following steps: Step S1: Select piezoelectric ceramic samples of various sizes and obtain impedance spectrum data of each sample by frequency sweep test; Step S2: Preprocess the obtained impedance spectrum data to construct the frequency-impedance curves of each sample; sample based on the frequency-impedance curves to obtain the impedance spectrum two-dimensional matrix corresponding to each sample. The impedance spectrum two-dimensional matrix consists of the frequency parameters of each sampling point and the impedance parameters corresponding to them; use the impedance spectrum two-dimensional matrix as data samples and the size parameters of the corresponding samples as sample labels to summarize all labeled data samples to obtain the sample set for training the model. Step S3: Construct a deep learning model; Step S4: Use the sample set obtained in step S2 to perform supervised training on the deep learning model constructed in step S3. During the training process, the impedance spectrum two-dimensional matrix is used as input and the size parameters of the corresponding piezoelectric ceramic sample are used as output. The model parameters are iteratively updated based on the regression loss function. Step S5: Select any interference noise frequency from the preset range of interference noise frequencies for the test model; Step S6: Using the dominant frequency of the interference noise selected in step S5 as the center frequency of the piezoelectric ceramic of the target filter, and combining it with the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. The curve is sampled to obtain a two-dimensional matrix of the ideal impedance spectrum. The two-dimensional matrix of the ideal impedance spectrum is input into the deep learning model that has completed the stage training to predict the size parameters of the piezoelectric ceramic that suppresses the target interference noise. Step S7: Based on the size parameters predicted in Step S6, select the corresponding piezoelectric ceramic and measure its resonant frequency. Compare the measured main resonant frequency with the interference noise main frequency selected in Step S5 to verify the accuracy of the model prediction result. If the deviation between the main resonant frequency and the interference noise main frequency is not greater than the preset tolerance, the prediction result is determined to be accurate, the model training ends, and the target prediction model is obtained. Otherwise, the deep learning model is retrained and optimized until the termination condition is met.
[0008] Based on the above solutions, further improvements or preferred solutions include: Furthermore, in step S1, when selecting piezoelectric ceramic samples, the main resonant frequencies of each sample are statistically analyzed so that the main resonant frequency distribution of all samples covers the frequency range of 10kHz to 200kHz.
[0009] Furthermore, the regression loss function used to train the model in step S4 is as follows: ; In the formula, L is the loss value, and N is the number of samples. This is the actual size vector of the piezoelectric ceramic sample corresponding to the nth sample. The size vector predicted by the model. It is a 2-norm.
[0010] Furthermore, the deep learning model is a CNN-LSTM model.
[0011] Further: The CNN-LSTM model includes at least two levels of convolution-pooling structure, with the kernel size set along the frequency direction to extract resonance peaks and anti-resonance valleys at different frequency scales. The LSTM model of the CNN-LSTM model adopts a multi-layer stacked structure to simultaneously capture the correlation of the impedance spectrum in both the positive and negative directions of the frequency axis.
[0012] Furthermore, in the CNN-LSTM model: The CNN model uses a one-dimensional convolutional structure, with the kernel size K set to an odd number between 3 and 9 along the frequency axis. The convolution is calculated using the following formula: ; In the formula, For the first Layer input features, For the first The layer outputs features, where k is the index of the feature map on the frequency axis. For the first The weight of the r-th element in the convolutional kernel, where r is the index within the kernel. For bias terms, It is a non-linear activation function; Each layer of an LSTM model contains several memory units, and its state updates include the input gate. Forgotten Gate Output gate With unit state ,satisfy: ; ; ; ; ; ; In the formula, For a moment The input vector, For a moment The hidden state, For element-wise multiplication, where are the parameters to be trained. Candidate memory units, To calculate the hidden state, The hyperbolic tangent activation function is used. This is the input weight matrix corresponding to the input gate. This is the input weight matrix corresponding to the forget gate. The input weight matrix corresponds to the output gate. This is the input weight matrix corresponding to the candidate memory units. This is the cyclic weight matrix corresponding to the input gate. This is the cyclic weight matrix corresponding to the forget gate. This is the cyclic weight matrix corresponding to the output gate. This is the cyclic weight matrix corresponding to the candidate memory units. This is the bias vector corresponding to the input gate. This is the bias vector corresponding to the forget gate. This is the bias vector corresponding to the output gate. This is the bias vector corresponding to the candidate memory unit.
[0013] Furthermore, when the deviation between the piezoelectric ceramic main resonant frequency selected in step S7 based on the prediction result and the interference noise main frequency used for verification is greater than the preset tolerance, new data samples and labels are generated based on the impedance spectrum data of the piezoelectric ceramic and its actual size, and fed back into the sample set to retrain and optimize the deep learning model.
[0014] In a second aspect, the present invention discloses a piezoelectric ceramic filter size prediction model construction system for implementing the method described in any of the preceding claims, characterized in that it comprises: 1) Impedance spectrum acquisition module, used to perform frequency sweep tests on piezoelectric ceramic samples of various sizes to obtain impedance spectrum data covering the frequency band from 10kHz to 200kHz; 2) Data preprocessing module, used to preprocess the impedance spectrum data and construct a sample set for training the model; 3) Model building and training module: Construct a CNN-LSTM model and train the constructed CNN-LSTM model using the sample set to obtain a piezoelectric ceramic size prediction model; 4) Interference and noise analysis module, used to perform spectrum analysis on interference and noise in various working scenarios of the filter, and identify the main frequency of interference and noise in the 10kHz to 200kHz frequency band; 5) Size prediction module, which combines the selected interference noise main frequency with the ideal bandwidth and quality factor designed for the filter to generate input data with the same pattern as the sample data, and imports it into the trained prediction model to predict the size parameters of the piezoelectric ceramic that can suppress the target interference noise. 6) Verification and optimization module, used to perform impedance spectrum measurement on the piezoelectric ceramic selected based on the prediction results, compare its actual main resonant frequency with the main frequency of the interference noise used for testing, and when the deviation between the two exceeds the preset tolerance, generate new samples based on the impedance spectrum data and actual size parameters of the currently selected piezoelectric ceramic, and feed them back into the sample set to retrain and optimize the model.
[0015] Thirdly, this invention discloses a method for predicting the size of a piezoelectric ceramic filter based on deep learning, implemented using a target prediction model constructed by any of the above methods, characterized by including the following steps: Collect target interference noise in the actual working scenario of the filter and determine the current dominant frequency of the target interference noise; Using the dominant frequency of the current target interference noise as the center frequency of the filter piezoelectric ceramic, and combining the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. The curve is then sampled to obtain a two-dimensional matrix of the ideal impedance spectrum. This two-dimensional matrix of the ideal impedance spectrum is then input into the target prediction model that has been trained to predict the size parameters of the piezoelectric ceramic that suppress the current target interference noise.
[0016] Furthermore, the prediction method uses spectral analysis to determine the dominant frequency of the target interference noise; In the spectrum analysis method, the interference noise spectrum It has the following expression: ; In the formula For time-domain signals, Let m be the number of sampling points, e be the base of the natural logarithm, and j be the imaginary unit. Pi amplitude Used to detect the main frequency of target interference noise. .
[0017] The present invention has the following technical effects: 1) This invention collects impedance spectrum data of a large number of piezoelectric ceramic samples, uses the impedance data and geometric dimensions after specific processing as sample data to train a deep learning model (such as CNN-LSTM model), forming a nonlinear mapping from impedance spectrum features to size parameters. Combined with the identification results of the main frequency of interference noise in the working scene, it realizes the direct reverse design of "target interference main frequency → predicted ceramic parameters", avoiding the complex formula derivation and multiple trial and error in traditional methods, and can effectively reduce design costs and design cycle. 2) This invention adopts a data-driven modeling approach, starting directly from the real impedance spectrum measured by devices such as vector network analyzers, and automatically learns the influence of different sample size factors on its resonance characteristics. This overcomes the limitations of existing technologies that use analytical models and equivalent circuits to fully reflect the actual situation of the filter, and is suitable for engineering application scenarios with complex process conditions and environmental factors. 3) The prediction model constructed in this invention has high prediction accuracy and generalization ability, which can improve the suppression effect of piezoelectric ceramic filters for specific interference frequencies and has good engineering adaptive optimization characteristics. Attached Figure Description
[0018] Figure 1 This is a flowchart of the construction method of the present invention; Figure 2 The structural block diagram of the system for this invention; Figure 3 This is a schematic diagram of a two-dimensional impedance spectrum matrix in one embodiment of the present invention; Figure 4 This is a schematic diagram of the interference noise spectrum and main frequency extraction according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the verification of the model training and prediction results of the present invention. Detailed Implementation
[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are intended to limit the scope of protection of the present invention.
[0020] Example 1: like Figure 1 This embodiment provides a method for constructing a piezoelectric ceramic filter size prediction model based on deep learning. The specific implementation process of the construction method is as follows: Step S1: Select a large number of piezoelectric ceramic samples of various sizes, and use a vector analyzer to perform frequency sweep tests on all samples to obtain impedance spectrum data of samples of different sizes.
[0021] Regarding the selection of samples, this embodiment takes a disc-shaped piezoelectric ceramic as an example. Its geometric parameters include thickness and diameter. Taking a piezoelectric ceramic filter used to suppress interference noise in the working environment of switching devices as an example, when selecting samples, it is necessary to make the main resonant frequencies of all samples as evenly distributed as possible, and to be able to cover the noise range of 10kHz~200kHz commonly found in the working environment of switching power supplies, or in other words, each sample should have at least one resonant peak falling within this frequency range.
[0022] Step S2: Preprocess the obtained impedance spectrum data, including normalizing the impedance amplitude and / or phase, using smoothing filters or frequency domain filters to suppress measurement noise, and constructing frequency-impedance curves for each sample; the frequency-impedance curve spectrum is composed of the frequency and the normalized impedance amplitude and / or phase values, such as... Figure 3 As shown.
[0023] Next, for the frequency-impedance curves of all samples, the following operations were performed: Multiple sampling points are selected at equal intervals on the frequency axis, and the impedance parameters (impedance amplitude and / or phase value) corresponding to each sampling point on the curve are determined. The obtained frequency-impedance parameter pairs are arranged in the sampling order to construct a two-dimensional impedance spectrum matrix of the current sample. Here, the planning of the number of sampling points and the step size should be able to completely include the main spectral characteristics of the corresponding piezoelectric ceramic sample, such as the positions of the peaks and valleys of the curve.
[0024] Finally, using the impedance spectrum two-dimensional matrix as data samples and the size parameters of the corresponding samples as sample labels, all labeled data samples are summarized to obtain the sample set for training the model.
[0025] Step S3: Construct a deep learning model.
[0026] This embodiment preferably uses a CNN-LSTM model. By setting the convolutional and pooling layers of the CNN model, multi-scale local features of the impedance spectrum are extracted to reduce the feature dimensionality. The LSTM model is set as a multi-layer structure to identify long-term dependencies and resonant mode features between impedance spectra.
[0027] The specific settings are as follows: 1) The CNN model includes at least two levels of convolution-pooling structure, with the kernel size set along the frequency direction to extract resonance peaks and anti-resonance valleys at different frequency scales; The convolution structure uses a one-dimensional convolution structure, and the kernel size K is set to an odd number between 3 and 9 in the frequency axis direction. The convolution calculation method is as follows: ; In the formula, For the first Layer input features, For the first The layer outputs features, where k is the index of the feature map on the frequency axis. For the first The weight of the r-th element in the convolutional kernel, where r is the index within the kernel. For bias terms, It is a non-linear activation function; Each layer of an LSTM model contains several memory units, and its state updates include the input gate. Forgotten Gate Output gate With unit state ,satisfy: ; ; ; ; ; ; In the formula, For a moment The input vector, For a moment The hidden state, For element-wise multiplication, where are the parameters to be trained. Candidate memory units, To calculate the hidden state, The hyperbolic tangent activation function is used. This is the input weight matrix corresponding to the input gate. This is the input weight matrix corresponding to the forget gate. The input weight matrix corresponds to the output gate. This is the input weight matrix corresponding to the candidate memory units. This is the cyclic weight matrix corresponding to the input gate. This is the cyclic weight matrix corresponding to the forget gate. This is the cyclic weight matrix corresponding to the output gate. This is the cyclic weight matrix corresponding to the candidate memory units. This is the bias vector corresponding to the input gate. This is the bias vector corresponding to the forget gate. This is the bias vector corresponding to the output gate. This is the bias vector corresponding to the candidate memory unit.
[0028] 2) The LSTM model employs a stacked structure of two or more layers to simultaneously capture the correlation of the impedance spectrum in both the positive and negative directions of the frequency axis.
[0029] Step S4: Use the sample set obtained in step S2 to perform supervised training on the CNN-LSTM model constructed in step S3. During the training process, the impedance spectrum two-dimensional matrix is used as input, and the corresponding geometric dimensions (thickness, diameter) of the piezoelectric ceramic sample is used as output. Through supervised training, the model learns the nonlinear mapping relationship between the spectral morphology features and the geometric structure. The model parameters are iteratively updated based on the regression loss function, thereby obtaining a data-driven size prediction model that does not require explicit physical modeling.
[0030] During the training process, the mean squared error function is used as the regression loss function. The loss is calculated separately for the training set and the validation set in the sample set. Overfitting is avoided by using an early stopping strategy or a learning rate decay strategy, thereby improving the generalization ability of the size prediction model.
[0031] The specific expression for the mean square error function is as follows: ; In the formula, L is the loss value, and N is the number of samples. This is the actual size vector of the piezoelectric ceramic sample corresponding to the nth sample. The size vector predicted by the model. It is a 2-norm.
[0032] Step S5: Determine the range of possible interference noise frequencies in the pre-determined filter operating scenario, and arbitrarily select one interference noise frequency as the model test data.
[0033] The dominant frequency of the interference noise can be determined by spectral analysis methods such as Fourier transform and power spectral density estimation. Frequency peaks in the spectrum that exceed a preset power threshold are detected, and one or more frequency peaks with the largest amplitude are taken as the dominant frequency of the target interference noise.
[0034] Step S6: Using the dominant frequency of the interference noise selected in step S5 as the center frequency of the piezoelectric ceramic of the filter to be designed, and combining it with the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. Then, sampling points are planned in the same way as in step S2, and the curve is sampled to obtain an ideal impedance spectrum two-dimensional matrix composed of multiple frequency-impedance parameter pairs arranged in the sampling order. That is, the ideal impedance spectrum two-dimensional matrix is composed of the frequency of each sampling point and the ideal impedance parameter that corresponds to it. The two-dimensional matrix of the ideal impedance spectrum is input into the CNN-LSTM model that has completed phased training. By optimizing the search or inverse reasoning, the input space of the CNN-LSTM model is traversed or iterated, and the size parameter that minimizes the deviation between the main resonant frequency and the main frequency of the interference noise used for testing is selected as the target size parameter output for the disk-shaped piezoelectric ceramic.
[0035] Step S7: Based on the size parameters predicted in step S6, select the corresponding piezoelectric ceramic and measure its resonant frequency. Compare the measured main resonant frequency with the interference noise main frequency selected in step S5 to test the accuracy of the model prediction results. If the deviation between the main resonant frequency and the main frequency of the interference noise is not greater than the preset tolerance, the prediction result is determined to be accurate, the model training ends, and the target prediction model is obtained. If the deviation between the main resonant frequency and the main frequency of the interference noise is greater than the preset tolerance, the CNN-LSTM model needs to be incrementally trained or retrained to iteratively improve the model's prediction accuracy until the termination condition is met.
[0036] During incremental training or retraining, new data samples and labels can be generated based on the actual impedance spectrum data and size of the piezoelectric ceramics used during testing. After being fed back into the sample set, the model can be optimized and trained.
[0037] In this embodiment, the above step numbers are not used to strictly constrain the implementation order, and some steps may be executed earlier or later depending on the actual situation.
[0038] Example 2: Based on the same design concept as in Embodiment 1, this embodiment provides a construction system for performing the construction method of Embodiment 1.
[0039] like Figure 2 As shown, the construction system includes the following components: 1) Impedance spectrum acquisition module, used to perform frequency sweep tests on piezoelectric ceramic samples of various sizes to obtain impedance spectrum data covering the frequency band from 10kHz to 200kHz. Each sample has at least one resonant peak falling within this frequency band. 2) Data preprocessing module, used to preprocess the impedance spectrum data and construct a sample set for training the model, including a training set and a validation set, etc. 3) Model building and training module: Construct a CNN-LSTM model and train the constructed CNN-LSTM model using the sample set to obtain a piezoelectric ceramic size prediction model; 4) Interference and noise analysis module, used to perform spectrum analysis on interference and noise in various working scenarios of the filter, and identify the main frequency of interference and noise in the 10kHz to 200kHz frequency band; 5) Size prediction module, which combines the main frequency of interference noise with the ideal bandwidth and quality factor designed for the filter to generate two-dimensional impedance spectrum matrix input data with the same pattern as the sample data. This data is then imported into the trained prediction model to predict the size parameters of the filter piezoelectric ceramic that can suppress the target interference noise. 6) The verification and optimization module is used to perform impedance spectrum measurements on the piezoelectric ceramics selected based on the prediction results, compare their actual main resonant frequency with the main frequency of the interference noise used for verification, and when the deviation between the two exceeds the preset tolerance, generate new samples based on the impedance spectrum data and actual size parameters of the current piezoelectric ceramics, and feed them back into the sample set to retrain and optimize the model.
[0040] This embodiment and Embodiment 1 belong to the same general inventive concept. The process of building and training the model, as well as the method of building the two-dimensional impedance spectrum matrix, are the same as in Embodiment 1, and will not be repeated here.
[0041] Example 3: This embodiment provides a piezoelectric ceramic filter size prediction method based on deep learning, which is implemented using the prediction model constructed as described in Embodiment 1.
[0042] In this embodiment, when predicting the piezoelectric ceramic size parameters of the filter using the target prediction model that has completed final training, the target interference noise in the working scenario where the filter will be applied is first measured, and the dominant frequency of the current target interference noise is determined. Then, using the dominant frequency of the current target interference noise as the center frequency of the piezoelectric ceramic to be designed, and combining it with the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. This curve is sampled to obtain a two-dimensional matrix of the ideal impedance spectrum, which is then input into the target prediction model that has completed training to predict the piezoelectric ceramic size parameters that suppress the current target interference noise.
[0043] In this embodiment, the spectral analysis method for determining the dominant frequency of the target interference noise uses the Discrete Fourier Transform (DFT). Calculate according to the following formula: ; In the formula For time-domain signals, Let m be the number of sampling points, e be the base of the natural logarithm, and j be the imaginary unit. Pi amplitude Used to detect the main frequency of target interference noise. .
[0044] like Figure 5 The right figure shows the predicted dimensions obtained using the prediction model constructed in this invention in several experiments, and the comparison results with the actual dimensions. The actual dimensions refer to the dimensions of existing piezoelectric ceramics that can suppress a certain interference noise signal, and the predicted dimensions refer to the predicted dimensions output by the model for that interference noise signal. Figure 5 As can be seen, the predicted results output by the model are basically consistent with the appropriate real size, which verifies the accuracy and reliability of the prediction model constructed in this invention.
[0045] The above description is merely an implementation example of the present invention. It should be noted that those skilled in the art can make other adjustments besides the embodiments, and such adjustments should also be considered within the scope of protection of the present invention without departing from the principle of the present invention.
Claims
1. A method for constructing a piezoelectric ceramic filter size prediction model based on deep learning, characterized in that, Includes the following steps: Step S1: Select piezoelectric ceramic samples of various sizes and obtain impedance spectrum data of each sample by frequency sweep test; Step S2: Preprocess the obtained impedance spectrum data to construct the frequency-impedance curves for each sample; Based on the frequency-impedance curve, sampling is performed to obtain a two-dimensional impedance spectrum matrix corresponding to each sample. The two-dimensional impedance spectrum matrix consists of the frequency parameters of each sampling point and the impedance parameters that correspond to them one by one. Using the two-dimensional impedance spectrum matrix as data samples and the size parameters of the corresponding samples as sample labels, all labeled data samples are summarized to obtain the sample set for training the model. Step S3: Construct a deep learning model; Step S4: Use the sample set obtained in step S2 to perform supervised training on the deep learning model constructed in step S3. During the training process, the impedance spectrum two-dimensional matrix is used as input and the size parameters of the corresponding piezoelectric ceramic sample are used as output. The model parameters are iteratively updated based on the regression loss function. Step S5: Select any interference noise frequency from the preset range of interference noise frequencies for the test model; Step S6: Using the dominant frequency of the interference noise selected in step S5 as the center frequency of the piezoelectric ceramic of the target filter, and combining it with the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. The curve is sampled to obtain a two-dimensional matrix of the ideal impedance spectrum. The two-dimensional matrix of the ideal impedance spectrum is input into the deep learning model that has completed the stage training to predict the size parameters of the piezoelectric ceramic that suppresses the target interference noise. Step S7: Based on the size parameters predicted in step S6, select the corresponding piezoelectric ceramic and measure its resonant frequency. Compare the measured main resonant frequency with the interference noise main frequency selected in step S5 to verify the accuracy of the model prediction results. If the deviation between the main resonant frequency and the main frequency of the interference noise is not greater than the preset tolerance, the prediction result is determined to be accurate, the model training ends, and the target prediction model is obtained; otherwise, the deep learning model is retrained and optimized until the termination condition is met.
2. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 1, characterized in that, In step S1, when selecting piezoelectric ceramic samples, the main resonant frequencies of each sample are statistically analyzed so that the main resonant frequency distribution of all samples covers the frequency range of 10kHz to 200kHz.
3. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 1, characterized in that, The regression loss function used to train the model in step S4 is as follows: ； In the formula, L is the loss value, and N is the number of samples. This is the actual size vector of the piezoelectric ceramic sample corresponding to the nth sample. The size vector predicted by the model. It is a 2-norm.
4. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 1, characterized in that, The deep learning model constructed in step S3 is a CNN-LSTM model.
5. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 4, characterized in that: The CNN model in the CNN-LSTM model includes at least two levels of convolution-pooling structure, with the convolution kernel size set along the frequency direction to extract resonance peaks and anti-resonance valleys at different frequency scales; The LSTM model in the CNN-LSTM model uses a multi-layer stacked structure to simultaneously capture the correlation of the impedance spectrum in both the positive and negative directions of the frequency axis.
6. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 4, characterized in that, In the CNN-LSTM model: The CNN model uses a one-dimensional convolutional structure, with the kernel size K set to an odd number between 3 and 9 along the frequency axis. The convolution is calculated using the following formula: ； In the formula, For the first Layer input features, For the first The layer outputs features, where k is the index of the feature map on the frequency axis. For the first The weight of the r-th element in the convolutional kernel, where r is the index within the kernel. For bias terms, It is a non-linear activation function; Each layer of an LSTM model contains several memory units, and its state updates include the input gate. Forgotten Gate Output gate With unit state ,satisfy: ； ； ； ； ； ； In the formula, For a moment The input vector, For a moment The hidden state, For element-wise multiplication, where are the parameters to be trained. Candidate memory units, To calculate the hidden state, The hyperbolic tangent activation function is used. This is the input weight matrix corresponding to the input gate. This is the input weight matrix corresponding to the forget gate. The input weight matrix corresponds to the output gate. This is the input weight matrix corresponding to the candidate memory units. This is the cyclic weight matrix corresponding to the input gate. This is the cyclic weight matrix corresponding to the forget gate. This is the cyclic weight matrix corresponding to the output gate. This is the cyclic weight matrix corresponding to the candidate memory units. This is the bias vector corresponding to the input gate. This is the bias vector corresponding to the forget gate. This is the bias vector corresponding to the output gate. This is the bias vector corresponding to the candidate memory unit.
7. The method for constructing a piezoelectric ceramic filter size prediction model based on deep learning as described in claim 1, characterized in that, When the deviation between the piezoelectric ceramic main resonant frequency selected in step S7 based on the prediction result and the interference noise main frequency used for verification is greater than the preset tolerance, new data samples and labels are generated based on the impedance spectrum data and actual size of the piezoelectric ceramic, and fed back into the sample set to retrain and optimize the deep learning model.
8. A system for constructing a piezoelectric ceramic filter size prediction model for implementing the method as described in any one of claims 1-7, characterized in that, include: 1) Impedance spectrum acquisition module, used to perform frequency sweep tests on piezoelectric ceramic samples of various sizes to obtain impedance spectrum data covering the frequency band from 10kHz to 200kHz; 2) Data preprocessing module, used to preprocess the impedance spectrum data and construct a sample set for training the model; 3) Model building and training module: Construct a CNN-LSTM model and train the constructed CNN-LSTM model using the sample set to obtain a piezoelectric ceramic size prediction model; 4) Interference and noise analysis module, used to perform spectrum analysis on interference and noise in various working scenarios of the filter, and identify the main frequency of interference and noise in the 10kHz to 200kHz frequency band; 5) Size prediction module, which combines the selected interference noise main frequency with the ideal bandwidth and quality factor designed for the filter to generate input data with the same pattern as the sample data, and imports it into the trained prediction model to predict the size parameters of the piezoelectric ceramic that can suppress the target interference noise. 6) Verification and optimization module, used to perform impedance spectrum measurement on the piezoelectric ceramic selected based on the prediction results, compare its actual main resonant frequency with the main frequency of the interference noise used for testing, and when the deviation between the two exceeds the preset tolerance, generate new samples based on the impedance spectrum data and actual size parameters of the currently selected piezoelectric ceramic, and feed them back into the sample set to retrain and optimize the model.
9. A deep learning-based method for predicting the size of piezoelectric ceramic filters, implemented using a target prediction model constructed according to any one of claims 1-7, characterized in that, Includes the following steps: Collect target interference noise in the actual working scenario of the filter and determine the current dominant frequency of the target interference noise; Using the dominant frequency of the current target interference noise as the center frequency of the filter piezoelectric ceramic, and combining the ideal bandwidth and quality factor designed for the filter, an ideal frequency-impedance curve is generated. The curve is then sampled to obtain a two-dimensional matrix of the ideal impedance spectrum. This two-dimensional matrix of the ideal impedance spectrum is then input into the target prediction model that has been trained to predict the size parameters of the piezoelectric ceramic that suppress the current target interference noise.
10. The piezoelectric ceramic filter size prediction method as described in claim 9, characterized in that, The dominant frequency of the interference noise was determined using spectral analysis. In the spectrum analysis method, the interference noise spectrum It has the following expression: ； In the formula For time-domain signals, Let m be the number of sampling points, e be the base of the natural logarithm, and j be the imaginary unit. Pi amplitude Used to detect the main frequency of target interference noise. .