Intelligent tuning method and system for surface acoustic wave filter

Abstract

The invention relates to the technical field of machine learning, in particular to an intelligent tuning method and system for a surface acoustic wave filter, and the method comprises the following steps: obtaining the parameter data of the filter, generating a hidden space vector and a basic parameter manifold, monitoring the temperature and excitation duration, generating a state vector, and calculating a reconstruction tensor; and adding and updating weight parameters to form a temporary weight, mapping a target index, calculating a geodesic distance, screening an optimal point, and decoding to obtain an optimal tuning parameter combination. According to the method, the characteristic data of the surface acoustic wave filter is mapped into the potential parameter space, the weight is dynamically updated, adaptive tuning of performance offset in operation is achieved, monitoring and normalization processing are conducted in combination with the temperature, the power excitation duration and other states, the tuning parameter offset trend is corrected in time, and the tuning precision is improved. Depending on external element compensation or frequency loop adjustment is avoided, parameter selection accuracy is improved through geodesic distance calculation in performance index mapping and optimal point screening, and it is guaranteed that the filter keeps stable frequency characteristics under complex conditions.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a smart tuning method and system for surface acoustic wave filters. Background Technology

[0002] Machine learning technology involves constructing mathematical models to extract features, recognize patterns, and optimize parameters from data. Its core aspects include using neural networks for nonlinear mapping of input data, achieving multi-level feature representation through deep learning structures, classifying and regressing data samples based on support vector machines, and employing reinforcement learning strategies to complete decision-making processes in dynamic environments. This technology encompasses a complete system from data acquisition and preprocessing to model training, validation, and inference, and is widely used in speech recognition, image analysis, predictive control, and signal processing. Among these, traditional intelligent tuning methods for surface acoustic wave (SAW) filters refer to methods that compensate for changes in filter operating frequency or performance parameters by adding capacitors or inductors to external circuits, or by using voltage-controlled oscillators (VCOs) to change the input signal frequency. These methods primarily rely on changes in the parameters of electronic components or adjustments to the frequency control loop to complete the filter tuning operation.

[0003] Traditional surface acoustic wave (SAW) filter tuning methods rely on external circuitry to load capacitors and inductors or on voltage-controlled oscillators to change the input signal frequency to compensate for offset. When the frequency or performance changes, corrections can only be made based on changes in component parameters. This makes it difficult to respond in real time to performance degradation caused by temperature fluctuations or prolonged high-power excitation. In complex environments, this method results in tuning lag, often leading to untimely filter performance recovery, output signal distortion, or unstable communication links. In practical applications, this manifests as reduced system reliability and insufficient adaptability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a smart tuning method and system for surface acoustic wave filters.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a smart tuning method for a surface acoustic wave filter, comprising the following steps: S1: Obtain the S-parameter dataset of the surface acoustic wave filter, input it into the autoencoder network to generate the latent space coordinate vector and the reconstructed SAW tuning parameter combination, and obtain the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vector based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination. S2: Monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector into the supernet to calculate the manifold reconstruction tensor; S3: Perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters; S4: Map the expected SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point. Calculate the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by the temporary weight parameters. Select the point with the smallest geodesic distance as the optimal coordinate point. Input the optimal coordinate point to the updated decoder to decode and obtain the optimal SAW tuning parameter combination.

[0006] As a further aspect of the present invention, the decoder's original weight parameters and the basic SAW parameter manifold composed of the latent space coordinate vectors include an initial weight set, parameter point distribution, and performance mapping relationship. The manifold reconstruction tensor specifically includes temperature drift compensation and power aging compensation. The temporary weight parameters include an instantaneous weight matrix and an instantaneous bias vector. The optimal SAW tuning parameter combination specifically refers to the tuning capacitor setting value, the tuning inductor setting value, and the frequency offset correction value.

[0007] As a further aspect of the present invention, step S1 includes: obtaining the S-parameter dataset and splitting the S-parameter dataset into a training set and a validation set; The training set is input into the encoder submodule of the autoencoder network to obtain encoder output features. The encoder submodule adopts a multilayer perceptron structure, and the activation function of each layer is a linear rectified unit. The encoder output features are input to the decoder submodule, the structure of which is symmetrical to that of the encoder submodule, to generate the reconstructed SAW tuning parameter combination. The mean square error between the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination is calculated, and the weights of the autoencoder network are adjusted using the Adam optimizer. Based on this, the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vectors are obtained.

[0008] As a further aspect of the present invention, step S2 includes: monitoring the real-time temperature reading sequence obtained by the SAW filter within a specific time interval, and smoothing the real-time temperature reading sequence by weighted averaging to generate a smoothed temperature reading; Calculate the total high-power excitation duration of the SAW filter within the cumulative operating time. The high-power excitation duration is defined as the duration during which the operating power of the SAW filter exceeds a preset threshold. Generate the cumulative high-power excitation duration. The smoothed temperature reading and the cumulative high-power excitation duration are linearly normalized to generate the SAW filter state vector; The SAW filter state vector is input into the supernetwork, which is a multilayer perceptron structure whose output layer dimension matches the shape of the original weight parameters of the decoder, to generate the manifold reconstruction tensor.

[0009] As a further aspect of the present invention, step S3 includes: calling the initial weight set, the parameter point distribution, and the performance mapping relationship in the original weight parameters of the decoder; The temperature drift compensation amount and the power aging compensation amount in the manifold reconstruction tensor are called, and the temperature drift compensation amount is added element-wise to the weight matrix in the initial weight set, and the power aging compensation amount is added element-wise to the bias vector in the initial weight set to generate the instantaneous weight matrix and the instantaneous bias vector. The instantaneous weight matrix and the instantaneous bias vector are combined to form the temporary weight parameters. The temporary weight parameters are then used to replace the original weight parameters of the decoder to generate an updated decoder.

[0010] As a further aspect of the present invention, step S4 includes: mapping the desired SAW performance index onto the basic SAW parameter manifold through a pre-trained multilayer perceptron to generate the target coordinate point; The transient manifold reconstructed by the temporary weight parameters is constructed, and the transient manifold is obtained by adjusting the original weight parameters of the decoder in real time; The geodesic distance from the target coordinate point within the instantaneous manifold is calculated using the following formula: ; in, Represents the interior point of the instantaneous manifold With point Geodesic distance between them and Representative point With point In the Coordinate values ​​in each dimension Representing the Weight coefficients for each dimension are used to generate geodesic distances; The point with the smallest distance from the geodesic line is selected as the optimal coordinate point.

[0011] As a further embodiment of the present invention, step S4 further includes: inputting the optimal coordinate point to the updated decoder for decoding, and mapping the optimal coordinate point from the latent space back to the SAW tuning parameter space through the multilayer perceptron backpropagation algorithm to generate the optimal SAW tuning parameter combination.

[0012] As a further aspect of the present invention, the temperature drift compensation amount The calculation formula is: in, This represents the temperature drift compensation amount. This represents the temperature compensation coefficient, and its value range is... to This coefficient was obtained through regression analysis based on the correlation between historical SAW filter performance data and temperature changes. This represents the real-time temperature reading obtained by monitoring the SAW filter. The reference temperature value representing the SAW filter under standard test conditions is set as follows: Celsius.

[0013] As a further aspect of the present invention, the power aging compensation amount is jointly determined by the cumulative high-power excitation duration and the aging characteristics of the SAW filter material, and is compensated in a non-linear incremental manner. The power aging compensation amount The calculation formula is: ; in, This represents the power aging compensation amount. This represents the aging coefficient of the SAW filter material. This represents the cumulative high-power excitation duration. This represents the aging time reference constant used to normalize the cumulative high-power excitation duration.

[0014] A surface acoustic wave (SAW) filter intelligent tuning system, the system being used to implement the above-mentioned SAW filter intelligent tuning method, the system comprising: The S-parameter acquisition and encoding module is used to acquire the S-parameter dataset of the surface acoustic wave filter and input it into the autoencoder network to generate the latent space coordinate vector and the reconstructed SAW tuning parameter combination. Based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination, the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vector are obtained. The state monitoring and tensor calculation module is used to monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector to the supernet to calculate the manifold reconstruction tensor. The weight update module is used to perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters. The tuning parameter output module is used to map the desired SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point, calculate the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by the temporary weight parameters, select the point with the minimum geodesic distance as the optimal coordinate point, and input the optimal coordinate point to the updated decoder to decode and obtain the optimal SAW tuning parameter combination.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by mapping the characteristic data of the surface acoustic wave filter under different operating conditions into a potential parameter space and dynamically updating the parameter weights, the adaptive tuning of the filter to performance deviations in a real-time environment is realized. Combined with the monitoring and normalization of state factors such as temperature and power excitation duration, the deviation trend of the tuning parameters can be corrected in a timely manner, so that parameter optimization no longer depends on the compensation of external components or the adjustment of the frequency loop. In the process of mapping the target performance index and selecting the optimal point, the accuracy of parameter selection is improved by geodesic distance calculation, thereby ensuring that the filter can maintain stable frequency characteristics and performance consistency under complex conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent tuning method for surface acoustic wave filters of the present invention. Figure 2 This is a flowchart illustrating the basic SAW parameter manifold generation process of this invention. Figure 3 This is a flowchart of the manifold reconstruction tensor generation process of the present invention; Figure 4 This is a flowchart of the decoder weight parameter update process of the present invention; Figure 5 The flowchart for generating the optimal SAW tuning parameter combination of this invention is shown below. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a smart tuning method for a surface acoustic wave filter, comprising the following steps: S1: Obtain the S-parameter dataset of the surface acoustic wave filter, input it into the autoencoder network to generate latent space coordinate vectors and reconstruct the SAW tuning parameter combination, and obtain the original weight parameters of the decoder and the basic SAW parameter manifold composed of latent space coordinate vectors based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination. Step S1 includes: obtaining the S-parameter dataset and splitting the S-parameter dataset into a training set and a validation set; The training set is input into the encoder submodule of the autoencoder network to obtain the encoder output features. The encoder submodule adopts a multilayer perceptron structure, and the activation function of each layer is a linear rectified unit. The encoder output features are input to the decoder submodule. The structure of the decoder submodule is symmetrical with that of the encoder submodule to generate a combination of reconstructed SAW tuning parameters. The mean square error between the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination is calculated, and the Adam optimizer is used to adjust the weights of the autoencoder network. Based on this, the original weight parameters of the decoder and the basic SAW parameter manifold composed of latent space coordinate vectors are obtained. The decoder's original weight parameters and the underlying SAW parameter manifold, composed of latent space coordinate vectors, include the initial weight set, parameter point distribution, and performance mapping relationship.

[0020] The process of obtaining the S-parameter dataset for surface acoustic wave (SAW) filters began with a frequency sweep test on 10,000 identical SAW filter samples using a vector network analyzer (VNA), specifically a Keysight N5227B. The test environment was maintained at 25°C and 50% relative humidity. Each filter was equipped with a tunable matching network, and its tuning parameter combination included three tuning capacitors (…). , , ) and two tuned inductors ( , Before testing, a set of tuning parameters was randomly set for each filter sample, with capacitance values ​​ranging from 0.5pF to 5.0pF and inductance values ​​ranging from 1.0nH to 10.0nH. The VNA's test frequency range was set to cover the passband and transition band of the SAW filter; for example, for a filter with a center frequency of 800MHz, the test range was set to 700MHz to 900MHz, with 401 scan points. For each scan point, the S-parameter matrix was recorded. (Input reflection coefficient) and The amplitude and phase of the (forward transmission coefficient) thus constitute a... 10,000 sets of S-parameter eigenvectors with real values. These 10,000 sets of S-parameter eigenvectors and their corresponding 5 SAW tuning parameters are combined ( , , , , These two sets of data together constitute the original S-parameter dataset. Subsequently, this dataset is divided into two parts: 80% and 20%. Specifically, 8000 sets are randomly selected from the 10000 data sets as the training set, and the remaining 2000 sets are used as the validation set.

[0021] The S-parameter feature vectors from the training set are used as input to the autoencoder network. The encoder submodule of this network employs a five-layer multilayer perceptron structure. The input layer has 1604 neurons, corresponding to the dimension of the S-parameter feature vectors. This is followed by three hidden layers with 1024, 512, and 256 neurons respectively. The output layer is the hidden space layer, with 3 neurons, generating a three-dimensional hidden space coordinate vector. Neurons in each layer (except the input layer) use a Rectified Linear Unit (ReLU) as the activation function. For example, inputting the first set of S-parameter data vectors from the training set into the encoder, after layer-by-layer weighted summation and ReLU activation, yields a three-dimensional encoder output feature, i.e., a hidden space coordinate vector, denoted as . Next, we will use this latent space coordinate vector. The input is fed into the decoder submodule. The structure of the decoder submodule is strictly symmetrical to that of the encoder submodule. Its input layer has 3 neurons, the number of neurons in the three hidden layers are 256, 512, and 1024 respectively, and the number of neurons in the output layer is 5, corresponding to the five SAW tuning parameters. The decoder also uses ReLU as the activation function for the hidden layers, and a linear activation function for the output layer. After the latent space coordinate vector is processed layer by layer by the decoder submodule, a set of reconstructed SAW tuning parameter combinations is generated. For example, for the input vector... The decoder output reconstructed SAW tuning parameter combination is as follows: .

[0022] Subsequently, the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination corresponding to this set of S-parameters are calculated. The mean squared error (MSE) between parameters is calculated as follows: first, the parameters are normalized; then, the squared error of each parameter is calculated; finally, the average value is taken. For example, if the error value is... This error value is backpropagated, and the Adam optimizer is used to adjust all weight matrices and bias vectors in the autoencoder network. The parameters of the Adam optimizer are set as follows: learning rate... The exponential decay rate estimated by the first moment The exponential decay rate estimated by the second moment This training process is repeated across the entire training set until the mean squared error on the validation set no longer decreases for 10 consecutive epochs or reaches the preset maximum number of epochs (e.g., 500 epochs). After training, the weights and biases of the decoder submodule are fixed and used as the decoder's original weight parameters. Simultaneously, the 8000 sets of training data are processed by the trained encoder to generate a set of 8000 three-dimensional latent space coordinate vectors, forming the basic SAW parameter manifold. The decoder's original weight parameters and the basic SAW parameter manifold composed of latent space coordinate vectors together constitute an initial model. This model includes the initial weight set, the distribution of parameter points in the three-dimensional latent space, and the performance mapping relationship from any point in this space back to a specific SAW tuning parameter combination.

[0023] Please see Figure 1 and Figure 3 S2: Monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector into the supernet to calculate the manifold reconstruction tensor. Step S2 includes: monitoring the real-time temperature reading sequence obtained by the SAW filter within a specific time interval, and smoothing the real-time temperature reading sequence by weighted averaging to generate smoothed temperature readings; Calculate the total high-power excitation duration of the SAW filter within the cumulative operating time. The high-power excitation duration is defined as the duration during which the operating power of the SAW filter exceeds a preset threshold. Generate the cumulative high-power excitation duration. The smoothed temperature readings and the cumulative high-power excitation duration are linearly normalized to generate the SAW filter state vector. The SAW filter state vector is input into the supernetwork, which is a multilayer perceptron structure whose output layer dimension matches the shape of the original weight parameters of the decoder, generating a manifold reconstruction tensor. The manifold reconstruction tensor is specifically the temperature drift compensation and power aging compensation. Temperature drift compensation The calculation formula is: ; in, This represents the amount of temperature drift compensation. This represents the temperature compensation coefficient, and its value range is... to This coefficient was obtained through regression analysis based on the correlation between historical SAW filter performance data and temperature changes. This represents the real-time temperature reading obtained from monitoring by the SAW filter. The reference temperature value representing the SAW filter under standard test conditions is set to... Celsius; The power aging compensation amount is determined by the cumulative high-power excitation duration and the aging characteristics of the SAW filter material, and is compensated in a non-linear incremental manner. Power aging compensation The calculation formula is: ; in, This represents the amount of power aging compensation. This represents the aging coefficient of the SAW filter material. Represents the cumulative high-power excitation duration. This represents the aging time reference constant used to normalize the cumulative high-power excitation duration.

[0024] To monitor the condition of the SAW filter, a digital temperature sensor (model: MaximIntegrated DS18B20) is first installed at a specific location on its package, such as the center of the metal cover. This sensor samples at a frequency of 1Hz, continuously acquiring real-time temperature readings. For example, a sequence of 10 readings is obtained over a 10-second time interval. The unit is degrees Celsius (°C). To eliminate short-term fluctuations, a weighted average method is used to smooth the five most recent data points. The weighting coefficients are set to... ,in The corresponding latest data point.

[0025] The calculation process for smoothing temperature readings is as follows: .

[0026] This smooth temperature reading This is 45.805℃. Simultaneously, the input power is monitored via a directional coupler connected in series with the SAW filter input and an RF power detector (model: Analog Devices ADL5902). A high-power excitation threshold is set to +24dBm. When the detected input power exceeds this threshold, a timer begins to accumulate duration. This timer continues to accumulate throughout the entire operating life of the filter. Assuming the filter has accumulated 5000 hours of operation, the total time during which the detected input power exceeded +24dBm is 120.5 hours; this is the accumulated high-power excitation duration. .

[0027] Next, the smoothed temperature readings and the cumulative high-power excitation duration are linearly normalized. The SAW filter is designed to operate from -40℃ to +85℃, therefore the temperature normalization calculation is as follows: For the cumulative high-power excitation duration, an aging reference upper limit of 10,000 hours is set, and its normalized calculation is as follows: These two normalized values ​​constitute the SAW filter state vector. This state vector The input is fed into a pre-trained supernetwork. This supernetwork is a four-layer multilayer perceptron, with 2 neurons in the input layer, 16 neurons in the two hidden layers, and 32 neurons in the two hidden layers. The dimension of the output layer is exactly the same as the total number of weight parameters in the decoder in S1. After the state vector undergoes a nonlinear transformation by the supernetwork, a manifold reconstruction tensor is generated. This tensor is numerically represented by two parts: one part is the temperature drift compensation amount used to adjust the decoder weight matrix. The other part is the power aging compensation amount used to adjust the decoder bias vector. Temperature drift compensation The calculation formula is: In this formula, This represents the temperature drift compensation amount, which is a scalar and will be subsequently expanded into a tensor that matches the shape of the decoder weight matrix through the output layer of the supernetwork. The value is the smoothed temperature reading obtained from the aforementioned calculation, which is 45.805℃. This is the reference temperature value under standard testing conditions, set at 25℃. It is the temperature compensation coefficient. The value of was determined through the following experiment: Fifty SAW filter samples were selected and tested at 10℃ intervals within a temperature range of -40℃ to +85℃. The drift of the center frequency at each temperature point relative to 25℃ was recorded. Linear regression analysis was performed on the frequency drift data and temperature change data of all samples. The experimental data are summarized in Table 1.

[0028] Table 1 shows the experimental data on temperature-frequency drift. As shown in Table 1, by performing linear fitting on these data points, the relationship between frequency drift and temperature change is obtained, and the normalized result of the slope is the temperature compensation coefficient. Calculations show that... The set value is 0.085. The physical meaning of this value is that for every 1°C change in temperature, the relevant weighting parameters need to be compensated by 0.085 units. Substitute the parameter values ​​into the formula for calculation: Power aging compensation amount The calculation formula is: .

[0029] In this formula, The power aging compensation amount is a scalar that will be expanded into a tensor that matches the shape of the decoder bias vector through the output layer of the supernetwork. This refers to the cumulative high-power excitation duration obtained earlier, which is 120.5 hours. It is the aging coefficient of the SAW filter material. These are aging time reference constants used to normalize the cumulative high-power excitation duration. These two parameters were determined through accelerated aging experiments: 30 filter samples were selected and continuously subjected to +30 dBm RF power at +85°C. The degradation of insertion loss was measured every 200 hours, for a total of 2000 hours. The relationship between loss degradation and cumulative excitation duration was fitted to a logarithmic model. Thus determine and Based on the fitting results of the experimental data, the following settings were made: It is 0.15. It lasts for 600 hours.

[0030] Substitute the parameter values ​​into the formula to calculate: .

[0031] This formula introduces a base-10 logarithmic function to simulate the nonlinear aging characteristics of SAW filters, where performance initially degrades significantly with increasing high-power operating time, followed by a gradual plateauing. Based on the above calculations, the current temperature conditions necessitate parameter adjustments ( The aging effect is greater than that caused by the cumulative high power duration. The hypernetwork combines these two scalar values. and As a foundation, a manifold reconstruction tensor is generated that perfectly matches the weights and biases of each layer of the decoder.

[0032] Please see Figure 1 and Figure 4S3: Perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters; Step S3 includes: calling the initial weight set, parameter point distribution, and performance mapping relationship from the decoder's original weight parameters; The temperature drift compensation and power aging compensation in the manifold reconstruction tensor are called. The temperature drift compensation is added element-wise to the weight matrix in the initial weight set, and the power aging compensation is added element-wise to the bias vector in the initial weight set to generate the instantaneous weight matrix and instantaneous bias vector. The instantaneous weight matrix and instantaneous bias vector are combined to form temporary weight parameters. The temporary weight parameters are then used to replace the original weight parameters of the decoder to generate an updated decoder. Temporary weight parameters include the instantaneous weight matrix and the instantaneous bias vector.

[0033] This step begins by calling the raw weight parameters of the decoder, trained and stored in S1, and the manifold reconstruction tensor generated by the hypernetwork in S2. The raw weight parameters of the decoder are stored in a hierarchical manner; for example, for the first hidden layer of the decoder (256 neurons), its initial weight set includes a dimension of... The weight matrix and a dimension of The bias vector. The manifold reconstruction tensor also has a matching structure, containing a Temperature drift compensation matrix and a The power aging compensation vector. The specific parameter update operation is element-wise matrix addition. The temperature drift compensation amount calculated in S2 is used. With power aging compensation (These two values ​​are the fundamental scalars for the compensation tensor generated by the supernetwork.) The last layer of the supernetwork, based on its learned mappings, expands and generates a compensation matrix and compensation vector with the exact same shape as the weights and biases of each layer in the decoder. For example, for the first hidden layer of the decoder, the supernetwork outputs a... Temperature drift compensation matrix and Power aging compensation vector Assume the original weight matrix of this layer... a certain element The value is -0.5432, corresponding to the element in the temperature drift compensation matrix. The value is ,in This is the scaling factor learned by the hypernetwork. Then the new instantaneous weight elements... The value is Similarly, assume the original bias vector of this layer a certain element The value is 0.1234, corresponding to the element in the power aging compensation vector. The value is ,in This is the scaling factor learned by the hypernetwork. Then the new instantaneous bias element... The value is .

[0034] This element-wise addition operation is performed on all weight matrices and bias vectors of all layers in the decoder. Taking the second hidden layer (512 neurons) in the decoder submodule as an example, its original weight matrix dimension is... The bias vector dimension is The hypernetwork correspondingly generates a Temperature drift compensation matrix and a The power aging compensation vector is generated. The original weight matrix is ​​added to the temperature drift compensation matrix to generate the instantaneous weight matrix for this layer; the original bias vector is added to the power aging compensation vector to generate the instantaneous bias vector for this layer. This process iterates through all hidden and output layers of the decoder until all original weight parameters have been added to their corresponding compensation tensors. After all layers have been calculated, all newly generated instantaneous weight matrices and instantaneous bias vectors are combined to form a complete set of temporary weight parameters. This set of temporary weight parameters is identical in structure and dimension to the original weight parameters of the decoder. Subsequently, the weight parameters of the decoder model in memory are replaced with this set of temporary weight parameters. This replacement operation is instantaneous, generating an updated decoder with parameters adjusted in real time.

[0035] The scaling factor learned by the hypernetwork specifically refers to the fact that the output layer of the hypernetwork is designed to have the same structure and number of parameters as the decoder, but instead of directly outputting compensation values, it outputs an independent scaling factor for each weight and bias. For example, for a decoder with dimension 1... The weight matrix, and the corresponding output layer part of the supernetwork also has There are 10 neurons, and the output value of each neuron is the scaling factor of the corresponding weight element. This scaling factor, learned through backpropagation during training, determines the base temperature drift compensation amount. The degree of influence on each specific weight. Similarly, for the bias vector, the hypernetwork also outputs a set of one-to-one scaling factors. Used to adjust the base power aging compensation amount The effect on each bias element.

[0036] Please see Figure 1 and Figure 5S4: Map the expected SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point. Calculate the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by the temporary weight parameters. Select the point with the smallest geodesic distance as the optimal coordinate point. Input the optimal coordinate point to the updated decoder to decode and obtain the optimal combination of SAW tuning parameters. Step S4 includes: mapping the desired SAW performance metrics onto the basic SAW parameter manifold through a pre-trained multilayer perceptron to generate target coordinate points; A transient manifold reconstructed from temporary weight parameters is constructed, which is obtained by adjusting the original weight parameters of the decoder in real time. The geodesic distance to the target coordinate point within the instantaneous manifold is calculated using the following formula: in, Represents the interior points of the instantaneous manifold With point Geodesic distance between them and Representative point With point In the Coordinate values ​​in each dimension Representing the Weight coefficients for each dimension are used to generate geodesic distances; Select the point with the smallest geodesic distance as the optimal coordinate point; Step S4 also includes: inputting the optimal coordinate point to the updated decoder for decoding, and using the multilayer perceptron backpropagation algorithm to map the optimal coordinate point from the latent space back to the SAW tuning parameter space to generate the optimal SAW tuning parameter combination. The optimal combination of SAW tuning parameters specifically refers to the tuning capacitor setting, tuning inductor setting, and frequency offset correction value.

[0037] This step begins with a desired SAW performance specification, for example, a target center frequency of 799.8 MHz, insertion loss of less than 2.5 dB, and in-band ripple of less than 0.5 dB. This set of performance specifications... As input, the vector is fed into a pre-trained multilayer perceptron specifically designed for mapping performance metrics to latent space. This network maps the performance metric vector onto the underlying SAW parametric manifold constructed in S1, generating a 3D target coordinate point. For example, the output target coordinate point... for Next, on the instantaneous manifold reconstructed by the temporary weight parameters of the decoder updated in S3, the distance from each point on the manifold to the target coordinate point is calculated. The geodesic distance. This calculation does not traverse all points on the manifold, but instead selects a representative set of discrete points (e.g., 1000 sample points) from the underlying SAW parametric manifold as candidates.

[0038] For each candidate point Use the following formula to calculate its distance to the target point. Geodesic distance: ; In this formula, This represents the candidate points within the instantaneous manifold. With the target point The geodesic distance between them. It is the dimension of the hidden space, here. . and They are points With point In the Coordinates in the dimension. It is the first The weighting coefficients for each dimension reflect the sensitivity of movement along that dimension to the final SAW performance parameters. The determination process is as follows: On the basic SAW parameter manifold, with the origin as the center, make small perturbations along the three coordinate axes (e.g., The step size is determined, and the corresponding SAW tuning parameter combination is obtained through decoding. Then, its S-parameter performance is calculated using circuit simulation software (such as ADS). The changes in key performance indicators such as center frequency and insertion loss caused by perturbations in each dimension are statistically analyzed. The magnitude of the changes is normalized and used as weighting coefficients.

[0039] For example, experiments show that the first dimension has the greatest impact on the center frequency, the second dimension has the greatest impact on the insertion loss, and the third dimension has a significant impact on the bandwidth. The weighting coefficients were calculated and set as follows: This formula introduces dimensional weight coefficients. This allows the calculation of geodesic distance to take into account the differences in the impact of different latent space dimensions on actual performance. The calculation is performed using two candidate points on the manifold as an example. Candidate point 1, Candidate point 2, Target point .

[0040] Calculate the geodesic distance to candidate point 1: ; Calculate the geodesic distance to candidate point 2: ; The result shows that the geodesic distance from candidate point 2 to the target point (0.090) is less than the geodesic distance from candidate point 1 (0.153). After traversing all 1000 candidate points, the point with the smallest geodesic distance is selected. For example, after comparison, the final target point is determined. This is the optimal coordinate point. Finally, this optimal coordinate point... The input is fed into the updated decoder generated in S3. This decoder performs a forward propagation calculation, transforming the 3D latent space coordinate vector through layer-by-layer nonlinear transformations using time-varying compensated weights and biases, ultimately generating a 5D vector at the output layer. This vector represents the optimal combination of SAW tuning parameters. For example, the decoder output might be... This combination corresponds to a specific tuning element setting and a frequency offset correction value, specifically referring to the tuning capacitor setting value. The tuning inductor setting value is And the frequency offset correction value is -150kHz.

[0041] The aforementioned pre-trained multilayer perceptron specifically designed for mapping performance metrics to latent space refers to a four-layer multilayer perceptron. Its input layer contains three neurons, corresponding to the target center frequency, insertion loss, and in-band ripple, respectively. Following this are two hidden layers with 128 and 64 neurons, respectively, each using a Rectified Linear Unit (ReLU) as its activation function. The output layer contains three neurons, employing a linear activation function, directly outputting the 3D target coordinates mapped to the underlying SAW parametric manifold. This network is trained using the coordinates in the underlying SAW parametric manifold and their corresponding performance metrics (obtained through decoder and circuit simulation) as the training dataset for supervised learning.

[0042] A surface acoustic wave (SAW) filter intelligent tuning system is provided. This system is used to execute the aforementioned SAW filter intelligent tuning method. The system includes: The S-parameter acquisition and encoding module is used to acquire the S-parameter dataset of the surface acoustic wave filter and input it into the autoencoder network to generate latent space coordinate vectors and reconstruct the SAW tuning parameter combination. Based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination, the original weight parameters of the decoder and the basic SAW parameter manifold composed of latent space coordinate vectors are obtained. The state monitoring and tensor calculation module is used to monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector to the supernet to calculate the manifold reconstruction tensor. The weight update module is used to perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters. The tuning parameter output module is used to map the desired SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point. It calculates the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by temporary weight parameters, selects the point with the minimum geodesic distance as the optimal coordinate point, and inputs the optimal coordinate point to the updated decoder to obtain the optimal combination of SAW tuning parameters.

[0043] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. A smart tuning method for a surface acoustic wave filter, characterized in that, Includes the following steps: S1: Obtain the S-parameter dataset of the surface acoustic wave filter, input it into the autoencoder network to generate the latent space coordinate vector and the reconstructed SAW tuning parameter combination, and obtain the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vector based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination. S2: Monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector into the supernet to calculate the manifold reconstruction tensor; S3: Perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters; S4: Map the expected SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point. Calculate the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by the temporary weight parameters. Select the point with the smallest geodesic distance as the optimal coordinate point. Input the optimal coordinate point to the updated decoder to decode and obtain the optimal SAW tuning parameter combination.

2. The intelligent tuning method for surface acoustic wave filters according to claim 1, characterized in that, The decoder's original weight parameters and the basic SAW parameter manifold formed by the latent space coordinate vectors include the initial weight set, parameter point distribution, and performance mapping relationship. The manifold reconstruction tensor specifically includes temperature drift compensation and power aging compensation. The temporary weight parameters include the instantaneous weight matrix and the instantaneous bias vector. The optimal SAW tuning parameter combination specifically refers to the tuning capacitor setting value, the tuning inductor setting value, and the frequency offset correction value.

3. The intelligent tuning method for surface acoustic wave filters according to claim 2, characterized in that, The S1 step includes: obtaining the S-parameter dataset and splitting the S-parameter dataset into a training set and a validation set; The training set is input into the encoder submodule of the autoencoder network to obtain encoder output features. The encoder submodule adopts a multilayer perceptron structure, and the activation function of each layer is a linear rectified unit. The encoder output features are input to the decoder submodule, the structure of which is symmetrical to that of the encoder submodule, to generate the reconstructed SAW tuning parameter combination. The mean square error between the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination is calculated, and the weights of the autoencoder network are adjusted using the Adam optimizer. Based on this, the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vectors are obtained.

4. The intelligent tuning method for a surface acoustic wave filter according to claim 3, characterized in that, The S2 step includes: monitoring the real-time temperature reading sequence obtained by the SAW filter within a specific time interval, and smoothing the real-time temperature reading sequence by weighted averaging to generate smooth temperature readings; Calculate the total high-power excitation duration of the SAW filter within the cumulative operating time. The high-power excitation duration is defined as the duration during which the operating power of the SAW filter exceeds a preset threshold. Generate the cumulative high-power excitation duration. The smoothed temperature reading and the cumulative high-power excitation duration are linearly normalized to generate the SAW filter state vector; The SAW filter state vector is input into the supernetwork, which is a multilayer perceptron structure whose output layer dimension matches the shape of the original weight parameters of the decoder, to generate the manifold reconstruction tensor.

5. The intelligent tuning method for a surface acoustic wave filter according to claim 4, characterized in that, The S3 step includes: calling the initial weight set, the parameter point distribution and the performance mapping relationship in the original weight parameters of the decoder; The temperature drift compensation amount and the power aging compensation amount in the manifold reconstruction tensor are called, and the temperature drift compensation amount is added element-wise to the weight matrix in the initial weight set, and the power aging compensation amount is added element-wise to the bias vector in the initial weight set to generate the instantaneous weight matrix and the instantaneous bias vector. The instantaneous weight matrix and the instantaneous bias vector are combined to form the temporary weight parameters. The temporary weight parameters are then used to replace the original weight parameters of the decoder to generate an updated decoder.

6. The intelligent tuning method for a surface acoustic wave filter according to claim 5, characterized in that, The S4 step includes: mapping the desired SAW performance metric onto the basic SAW parameter manifold through a pre-trained multilayer perceptron to generate the target coordinate point; The transient manifold reconstructed by the temporary weight parameters is constructed, and the transient manifold is obtained by adjusting the original weight parameters of the decoder in real time; The geodesic distance from the target coordinate point within the instantaneous manifold is calculated using the following formula: ； in, Represents the interior point of the instantaneous manifold With point Geodesic distance between them and Representative point With point In the Coordinate values ​​in each dimension Representing the Weight coefficients for each dimension are used to generate geodesic distances; The point with the smallest distance from the geodesic line is selected as the optimal coordinate point.

7. The intelligent tuning method for a surface acoustic wave filter according to claim 6, characterized in that, The S4 step further includes: inputting the optimal coordinate point to the updated decoder for decoding, and mapping the optimal coordinate point from the latent space back to the SAW tuning parameter space through the multilayer perceptron backpropagation algorithm to generate the optimal SAW tuning parameter combination.

8. The intelligent tuning method for a surface acoustic wave filter according to claim 7, characterized in that, The temperature drift compensation amount The calculation formula is: in, This represents the temperature drift compensation amount. This represents the temperature compensation coefficient, and its value range is... to This coefficient was obtained through regression analysis based on the correlation between historical SAW filter performance data and temperature changes. This represents the real-time temperature reading obtained by monitoring the SAW filter. The reference temperature value representing the SAW filter under standard test conditions is set as follows: Celsius.

9. The intelligent tuning method for a surface acoustic wave filter according to claim 8, characterized in that, The power aging compensation amount is jointly determined by the cumulative high-power excitation duration and the aging characteristics of the SAW filter material, and is compensated in a non-linear incremental manner. The power aging compensation amount The calculation formula is: ； in, This represents the power aging compensation amount. This represents the aging coefficient of the SAW filter material. This represents the cumulative high-power excitation duration. This represents the aging time reference constant used to normalize the cumulative high-power excitation duration.

10. A smart tuning system for a surface acoustic wave filter, characterized in that, The system is used to implement the intelligent tuning method for surface acoustic wave filters according to any one of claims 1-9, and the system comprises: The S-parameter acquisition and encoding module is used to acquire the S-parameter dataset of the surface acoustic wave filter and input it into the autoencoder network to generate the latent space coordinate vector and the reconstructed SAW tuning parameter combination. Based on the error adjustment network of the reconstructed SAW tuning parameter combination and the original SAW tuning parameter combination, the original weight parameters of the decoder and the basic SAW parameter manifold composed of the latent space coordinate vector are obtained. The state monitoring and tensor calculation module is used to monitor the SAW filter to obtain real-time temperature readings and cumulative high-power excitation duration, generate and normalize the SAW filter state vector, and input the SAW filter state vector to the supernet to calculate the manifold reconstruction tensor. The weight update module is used to perform element-wise matrix addition on the original weight parameters of the decoder and the manifold reconstruction tensor to generate temporary weight parameters, and replace the original weight parameters of the decoder with the temporary weight parameters. The tuning parameter output module is used to map the desired SAW performance index to the basic SAW parameter manifold to obtain the target coordinate point, calculate the geodesic distance to the target coordinate point within the instantaneous manifold reconstructed by the temporary weight parameters, select the point with the minimum geodesic distance as the optimal coordinate point, and input the optimal coordinate point to the updated decoder to decode and obtain the optimal SAW tuning parameter combination.

Images